def test_warnings():
a = Input(shape=(3,), name='input_a')
b = Input(shape=(3,), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
model = Model([a, b], [a_2, b_2])
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
model.compile(optimizer, loss, metrics=[], loss_weights=loss_weights,
sample_weight_mode=None)
def gen_data(batch_sz):
while True:
yield ([np.random.random((batch_sz, 3)), np.random.random((batch_sz, 3))],
[np.random.random((batch_sz, 4)), np.random.random((batch_sz, 3))])
with pytest.warns(Warning) as w:
out = model.fit_generator(gen_data(4), steps_per_epoch=10, use_multiprocessing=True, workers=2)
warning_raised = any(['Sequence' in str(w_.message) for w_ in w])
assert warning_raised, 'No warning raised when using generator with processes.'
with pytest.warns(None) as w:
out = model.fit_generator(RandomSequence(3), steps_per_epoch=4, use_multiprocessing=True, workers=2)
assert all(['Sequence' not in str(w_.message) for w_ in w]), 'A warning was raised for Sequence.'
def test_model_multiple_calls():
x1 = Input(shape=(20,))
y1 = sequential([
Dense(10),
Dense(1),
])(x1)
m1 = Model(x1, y1)
x2 = Input(shape=(25,))
y2 = sequential([
Dense(20),
m1
])(x2)
m2 = Model(x2, y2)
m2.compile('adam', 'mse')
x3 = Input(shape=(20,))
y3 = sequential([
Dense(25),
m2
])(x3)
m3 = Model(x3, y3)
m3.compile('adam', 'mse')
m3.train_on_batch(np.zeros((32, 20)), np.zeros((32, 1)))
def test_model_custom_target_tensors():
a = Input(shape=(3,), name='input_a')
b = Input(shape=(3,), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
y = K.placeholder([10, 4], name='y')
y1 = K.placeholder([10, 3], name='y1')
y2 = K.placeholder([7, 5], name='y2')
model = Model([a, b], [a_2, b_2])
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
# test list of target tensors
with pytest.raises(ValueError):
model.compile(optimizer, loss, metrics=[], loss_weights=loss_weights,
sample_weight_mode=None, target_tensors=[y, y1, y2])
model.compile(optimizer, loss, metrics=[], loss_weights=loss_weights,
sample_weight_mode=None, target_tensors=[y, y1])
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
{y: np.random.random((10, 4)),
y1: np.random.random((10, 3))})
# test dictionary of target_tensors
with pytest.raises(ValueError):
model.compile(optimizer, loss,
metrics=[],
loss_weights=loss_weights,
sample_weight_mode=None,
target_tensors={'does_not_exist': y2})
# test dictionary of target_tensors
model.compile(optimizer, loss,
metrics=[],
loss_weights=loss_weights,
sample_weight_mode=None,
target_tensors={'dense_1': y, 'dropout': y1})
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
{y: np.random.random((10, 4)),
y1: np.random.random((10, 3))})
if K.backend() == 'tensorflow':
import tensorflow as tf
# test with custom TF placeholder as target
pl_target_a = tf.placeholder('float32', shape=(None, 4))
model.compile(optimizer='rmsprop', loss='mse',
target_tensors={'dense_1': pl_target_a})
model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
def test_sparse_input_validation_split():
test_input = sparse.random(6, 3, density=0.25).tocsr()
in1 = Input(shape=(3,), sparse=True)
out1 = Dense(4)(in1)
test_output = np.random.random((6, 4))
model = Model(in1, out1)
model.compile('rmsprop', 'mse')
model.fit(test_input, test_output, epochs=1, batch_size=2, validation_split=0.2)
def m():
x = Input(shape=(input_size + output_size, nb_chars))
m_realness = sequential([
LSTM(14),
Dense(1, activation='sigmoid'),
])(x)
m = Model([x], [m_realness])
m.compile(Adam(), 'mse')
return m
def decoder_dummy(label_sizes, nb_filter=16, data_shape=(1, 64, 64), nb_bits=12,
optimizer='adam'):
input = Input(shape=data_shape)
x = input
outputs, losses = decoder_end_block(x, label_sizes, nb_bits,
activation=lambda: ELU())
model = Model(input, list(outputs.values()))
model.compile(optimizer, loss=list(losses.values()),
loss_weights={k: decoder_loss_weights(k) for k in losses.keys()})
return model
def test_sparse_placeholder_fit():
test_inputs = [sparse.random(6, 3, density=0.25).tocsr() for _ in range(2)]
test_outputs = [sparse.random(6, i, density=0.25).tocsr() for i in range(3, 5)]
in1 = Input(shape=(3,))
in2 = Input(shape=(3,), sparse=True)
out1 = Dropout(0.5, name='dropout')(in1)
out2 = Dense(4, name='dense_1')(in2)
model = Model([in1, in2], [out1, out2])
model.predict(test_inputs, batch_size=2)
model.compile('rmsprop', 'mse')
model.fit(test_inputs, test_outputs, epochs=1, batch_size=2, validation_split=0.5)
model.evaluate(test_inputs, test_outputs, batch_size=2)
def decoder_baseline(label_sizes, nb_bits=12, data_shape=(1, 64, 64),
depth=1, nb_filter=16, optimizer='adam'):
n = nb_filter
input = Input(shape=data_shape)
x = sequential([
conv2d_block(n, depth=depth, pooling='max'), # 32x32
conv2d_block(2*n, depth=depth, pooling='max'), # 16x16
conv2d_block(4*n, depth=depth, pooling='max'), # 8x8
conv2d_block(8*n, depth=depth, pooling='max'), # 4x4
])(input)
outputs, losses = decoder_end_block(x, label_sizes, nb_bits,
activation=lambda: ELU())
model = Model(input, list(outputs.values()))
model.compile(optimizer, loss=list(losses.values()),)
return model
def test_render_gan_builder_generator_extended():
labels_shape = (27,)
z_dim_offset = 50
builder = RenderGAN(lambda x: tag3d_network_dense(x, nb_units=4),
generator_units=4, discriminator_units=4,
z_dim_offset=z_dim_offset,
labels_shape=(27,))
bs = 19
z, z_offset, labels = data(builder, bs)
real = np.zeros((bs,) + builder.data_shape)
labels_input = Input(shape=labels_shape)
z = Input(shape=(z_dim_offset,))
fake = builder.generator_given_z_and_labels([z, labels_input])
m = Model([z, labels_input], [fake])
m.compile('adam', 'mse')
m.train_on_batch([z_offset, labels], real)
def test_model_with_partial_loss():
a = Input(shape=(3, ), name='input_a')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
a_3 = dp(a_2)
model = Model(a, [a_2, a_3])
optimizer = 'rmsprop'
loss = {'dropout': 'mse'}
model.compile(optimizer, loss, metrics=['mae'])
input_a_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
# test train_on_batch
out = model.train_on_batch(input_a_np, output_a_np)
out = model.test_on_batch(input_a_np, output_a_np)
# fit
out = model.fit(input_a_np, [output_a_np])
# evaluate
out = model.evaluate(input_a_np, [output_a_np])
# Same without dropout.
a = Input(shape=(3, ), name='input_a')
a_2 = Dense(4, name='dense_1')(a)
a_3 = Dense(4, name='dense_2')(a_2)
model = Model(a, [a_2, a_3])
optimizer = 'rmsprop'
loss = {'dense_2': 'mse'}
model.compile(optimizer, loss, metrics={'dense_1': 'mae'})
# test train_on_batch
out = model.train_on_batch(input_a_np, output_a_np)
out = model.test_on_batch(input_a_np, output_a_np)
# fit
out = model.fit(input_a_np, [output_a_np])
# evaluate
out = model.evaluate(input_a_np, [output_a_np])
def simple_gan():
z = Input(batch_shape=simple_gan_z_shape, name='z')
generator = sequential([
Dense(4*simple_gan_nb_z, activation='relu', name='g1'),
Dense(4*simple_gan_nb_z, activation='relu', name='g2'),
Dense(simple_gan_nb_out, name='g_loss'),
])(z)
d_input = Input(batch_shape=simple_gan_real_shape, name='data')
discriminator = sequential([
Dense(400, input_dim=2, name='d1'),
LeakyReLU(0.3),
Dense(400, name='d2'),
LeakyReLU(0.3),
Dense(1, activation='sigmoid', name='d_loss')
])(d_input)
g = Model(z, generator)
g.compile(Adam(lr=0.0002, beta_1=0.5), {'g_loss': 'binary_crossentropy'})
d = Model(d_input, discriminator)
d.compile(Adam(lr=0.0002, beta_1=0.5), {'d_loss': 'binary_crossentropy'})
return GAN(g, d)
def test_trainable_weights_count_consistency():
"""Tests the trainable weights consistency check of Model.
This verifies that a warning is shown if model.trainable is modified
and the model is summarized/run without a new call to .compile()
Reproduce issue #8121
"""
a = Input(shape=(3, ), name='input_a')
model1 = Model(inputs=a, outputs=Dense(1)(a))
model1.trainable = False
b = Input(shape=(3, ), name='input_b')
y = model1(b)
model2 = Model(inputs=b, outputs=Dense(1)(y))
model2.compile(optimizer='adam', loss='mse')
model1.trainable = True
# Should warn on .summary()
with pytest.warns(UserWarning) as w:
model2.summary()
warning_raised = any(['Discrepancy' in str(w_.message) for w_ in w])
assert warning_raised, 'No warning raised when trainable is modified without .compile.'
# And on .fit()
with pytest.warns(UserWarning) as w:
model2.fit(x=np.zeros((5, 3)), y=np.zeros((5, 1)))
warning_raised = any(['Discrepancy' in str(w_.message) for w_ in w])
assert warning_raised, 'No warning raised when trainable is modified without .compile.'
# And shouldn't warn if we recompile
model2.compile(optimizer='adam', loss='mse')
with pytest.warns(None) as w:
model2.summary()
assert len(
w
) == 0, "Warning raised even when .compile() is called after modifying .trainable"
def test_model_with_partial_loss():
a = Input(shape=(3,), name='input_a')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
a_3 = dp(a_2)
model = Model(a, [a_2, a_3])
optimizer = 'rmsprop'
loss = {'dropout': 'mse'}
model.compile(optimizer, loss, metrics=['mae'])
input_a_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
# test train_on_batch
out = model.train_on_batch(input_a_np, output_a_np)
out = model.test_on_batch(input_a_np, output_a_np)
# fit
out = model.fit(input_a_np, [output_a_np])
# evaluate
out = model.evaluate(input_a_np, [output_a_np])
# Same without dropout.
a = Input(shape=(3,), name='input_a')
a_2 = Dense(4, name='dense_1')(a)
a_3 = Dense(4, name='dense_2')(a_2)
model = Model(a, [a_2, a_3])
optimizer = 'rmsprop'
loss = {'dense_2': 'mse'}
model.compile(optimizer, loss, metrics={'dense_1': 'mae'})
# test train_on_batch
out = model.train_on_batch(input_a_np, output_a_np)
out = model.test_on_batch(input_a_np, output_a_np)
# fit
out = model.fit(input_a_np, [output_a_np])
# evaluate
out = model.evaluate(input_a_np, [output_a_np])
def train_f_enc(self, steps_list, epoch=50):
print("training f_enc")
f_swap0 = Sequential(name='f_swap0')
f_swap0.add(self.f_enc)
f_swap0.add(Dense(FIELD_DEPTH))
f_swap0.add(Activation('softmax', name='softmax_swap0'))
f_swap1 = Sequential(name='f_swap1')
f_swap1.add(self.f_enc)
f_swap1.add(Dense(FIELD_DEPTH))
f_swap1.add(Activation('softmax', name='softmax_swap1'))
env_model = Model(self.f_enc.inputs, [f_swap0.output, f_swap1.output],
name="env_model")
env_model.compile(optimizer='adam',
loss=['categorical_crossentropy'] * 2)
for ep in range(epoch):
losses = []
for idx, steps_dict in enumerate(steps_list):
prev = None
for step in steps_dict['steps']:
x = self.convert_input(step.input)[:2]
env_values = step.input.env.reshape((3, -1))
p1 = np.clip(env_values[0].argmax() - 1, 0, 9)
p2 = np.clip(env_values[1].argmax() - 1, 0, 9)
p3 = np.clip(env_values[2].argmax() - 1, 0, 9)
now = (p1, p2, p3)
if prev == now:
continue
prev = now
y0 = to_one_hot_array(min(p1, p2), FIELD_DEPTH)
y1 = to_one_hot_array(max(p1, p2), FIELD_DEPTH)
y = [yy.reshape((self.batch_size, -1)) for yy in [y0, y1]]
loss = env_model.train_on_batch(x, y)
losses.append(loss)
print("ep %3d: loss=%s" % (ep, np.average(losses)))
if np.average(losses) < 1e-06:
break
def decoder_resnet(label_sizes, nb_filter=16, data_shape=(1, 64, 64), nb_bits=12,
resnet_depth=(3, 4, 6, 3),
optimizer='adam'):
def _bn_relu_conv(nb_filter, nb_row=3, nb_col=3, subsample=1):
return sequential([
BatchNormalization(mode=0, axis=1),
ELU(),
Convolution2D(nb_filter=nb_filter, nb_row=nb_row, nb_col=nb_col,
subsample=(subsample, subsample), init="he_normal", border_mode="same")
])
def f(nb_filter, subsample=1):
return sequential([
_bn_relu_conv(nb_filter, subsample=subsample),
_bn_relu_conv(nb_filter),
])
input = Input(shape=data_shape)
fitlers_by_depth = [nb_filter * 2**i for i in range(len(resnet_depth))]
print("fitlers_by_depth", fitlers_by_depth)
x = _bn_relu_conv(nb_filter, 3, 3, subsample=2)(input)
for i, (n, d) in enumerate(zip(fitlers_by_depth, resnet_depth)):
for di in range(d):
if di == 0 and i != 0:
shortcut = _bn_relu_conv(n, 1, 1, subsample=2)
subsample = 2
else:
shortcut = lambda x: x
subsample = 1
x = merge([shortcut(x), f(n, subsample)(x)], mode='sum')
outputs, losses = decoder_end_block(x, label_sizes, nb_bits,
activation=lambda: ELU())
model = Model(input, list(outputs.values()))
model.compile(optimizer, loss=list(losses.values()),
loss_weights={k: decoder_loss_weights(k) for k in losses.keys()})
return model
def train_f_enc(self, steps_list, epoch=50):
print("training f_enc")
f_add0 = Sequential(name='f_add0')
f_add0.add(self.f_enc)
f_add0.add(Dense(FIELD_DEPTH))
f_add0.add(Activation('softmax', name='softmax_add0'))
f_add1 = Sequential(name='f_add1')
f_add1.add(self.f_enc)
f_add1.add(Dense(FIELD_DEPTH))
f_add1.add(Activation('softmax', name='softmax_add1'))
env_model = Model(self.f_enc.inputs, [f_add0.output, f_add1.output], name="env_model")
env_model.compile(optimizer='adam', loss=['categorical_crossentropy']*2)
for ep in range(epoch):
losses = []
for idx, steps_dict in enumerate(steps_list):
prev = None
for step in steps_dict['steps']:
x = self.convert_input(step.input)[:2]
env_values = step.input.env.reshape((4, -1))
in1 = np.clip(env_values[0].argmax() - 1, 0, 9)
in2 = np.clip(env_values[1].argmax() - 1, 0, 9)
carry = np.clip(env_values[2].argmax() - 1, 0, 9)
y_num = in1 + in2 + carry
now = (in1, in2, carry)
if prev == now:
continue
prev = now
y0 = to_one_hot_array((y_num % 10)+1, FIELD_DEPTH)
y1 = to_one_hot_array((y_num // 10)+1, FIELD_DEPTH)
y = [yy.reshape((self.batch_size, -1)) for yy in [y0, y1]]
loss = env_model.train_on_batch(x, y)
losses.append(loss)
print("ep %3d: loss=%s" % (ep, np.average(losses)))
if np.average(losses) < 1e-06:
break
def create_squeeze_net():
inp = Input(shape=getInputDim())
x = Conv2D(64, (3, 3), padding='same', activation='relu')(inp)
x = MaxPooling2D((3, 3), strides=2)(x)
x = create_fire_mod(x, 64, 128)
x = MaxPooling2D((3, 3), strides=2)(x)
x = create_fire_mod(x, 64, 128)
x = MaxPooling2D((3, 3), strides=2)(x)
x = Conv2D(32, (1, 1), activation='relu')(x)
x = Flatten()(x)
x = Dense(64, activation='relu')(x)
x = Dense(1, activation='sigmoid')(x)
model = Model(inp, x)
model.compile(loss='binary_crossentropy',
optimizer='adagrad',
metrics=[
'binary_accuracy',
])
return model
def create_model():
tokens = get_tokens()
num_tokens = len(tokens) + 1
input_data = Input(name='speech_data_input', shape=(500, 13))
layer_dense_1 = Dense(256, activation="relu", use_bias=True, kernel_initializer='he_normal')(input_data)
layer_dropout_1 = Dropout(0.4)(layer_dense_1)
layer_dense_2 = Dense(512, activation="relu", use_bias=True, kernel_initializer='he_normal')(layer_dropout_1)
layer_gru1 = GRU(512, return_sequences=True, kernel_initializer='he_normal', dropout=0.4)(layer_dense_2)
layer_gru2 = GRU(512, return_sequences=True, go_backwards=True, kernel_initializer='he_normal', dropout=0.4)(layer_gru1)
layer_dense_3 = Dense(256, activation="relu", use_bias=True, kernel_initializer='he_normal')(layer_gru2)
layer_dropout_2 = Dropout(0.4)(layer_dense_3)
layer_dense_4 = Dense(num_tokens, activation="relu", use_bias=True, kernel_initializer='he_normal')(layer_dropout_2)
output = Activation('softmax', name='Activation0')(layer_dense_4)
#ctc
labels = Input(name='speech_labels', shape=[70], dtype='int64')
input_length = Input(name='input_length', shape=[1], dtype='int64')
label_length = Input(name='label_length', shape=[1], dtype='int64')
loss_out = Lambda(ctc_lambda, output_shape=(1,), name='ctc')([labels, output, input_length, label_length])
model = Model(inputs=[input_data, labels, input_length, label_length], outputs=loss_out)
adad = Adadelta(lr=0.01, rho=0.95, epsilon=K.epsilon())
model.compile(loss={'ctc': lambda y_true, output: output}, optimizer=adad)
print("model compiled successful!")
return model
def test_trainable_weights_count_consistency():
"""Tests the trainable weights consistency check of Model.
This verifies that a warning is shown if model.trainable is modified
and the model is summarized/run without a new call to .compile()
Reproduce issue #8121
"""
a = Input(shape=(3,), name='input_a')
model1 = Model(inputs=a, outputs=Dense(1)(a))
model1.trainable = False
b = Input(shape=(3,), name='input_b')
y = model1(b)
model2 = Model(inputs=b, outputs=Dense(1)(y))
model2.compile(optimizer='adam', loss='mse')
model1.trainable = True
# Should warn on .summary()
with pytest.warns(UserWarning) as w:
model2.summary()
warning_raised = any(['Discrepancy' in str(w_.message) for w_ in w])
assert warning_raised, 'No warning raised when trainable is modified without .compile.'
# And on .fit()
with pytest.warns(UserWarning) as w:
model2.fit(x=np.zeros((5, 3)), y=np.zeros((5, 1)))
warning_raised = any(['Discrepancy' in str(w_.message) for w_ in w])
assert warning_raised, 'No warning raised when trainable is modified without .compile.'
# And shouldn't warn if we recompile
model2.compile(optimizer='adam', loss='mse')
with pytest.warns(None) as w:
model2.summary()
assert len(w) == 0, "Warning raised even when .compile() is called after modifying .trainable"
def test_warnings():
a = Input(shape=(3,), name='input_a')
b = Input(shape=(3,), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
model = Model([a, b], [a_2, b_2])
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
model.compile(optimizer, loss, metrics=[], loss_weights=loss_weights,
sample_weight_mode=None)
def gen_data(batch_sz):
while True:
yield ([np.random.random((batch_sz, 3)),
np.random.random((batch_sz, 3))],
[np.random.random((batch_sz, 4)),
np.random.random((batch_sz, 3))])
with pytest.warns(Warning) as w:
out = model.fit_generator(gen_data(4),
steps_per_epoch=10,
use_multiprocessing=True,
workers=2)
warning_raised = any(['Sequence' in str(w_.message) for w_ in w])
assert warning_raised, 'No warning raised when using generator with processes.'
with pytest.warns(None) as w:
out = model.fit_generator(RandomSequence(3),
steps_per_epoch=4,
use_multiprocessing=True,
workers=2)
assert all(['Sequence' not in str(w_.message) for w_ in w]), 'A warning was raised for Sequence.'
def decoder_baseline(label_sizes,
nb_bits=12,
data_shape=(1, 64, 64),
depth=1,
nb_filter=16,
optimizer='adam'):
n = nb_filter
input = Input(shape=data_shape)
x = sequential([
conv2d_block(n, depth=depth, pooling='max'), # 32x32
conv2d_block(2 * n, depth=depth, pooling='max'), # 16x16
conv2d_block(4 * n, depth=depth, pooling='max'), # 8x8
conv2d_block(8 * n, depth=depth, pooling='max'), # 4x4
])(input)
outputs, losses = decoder_end_block(x,
label_sizes,
nb_bits,
activation=lambda: ELU())
model = Model(input, list(outputs.values()))
model.compile(
optimizer,
loss=list(losses.values()),
)
return model
def get_model_with_classification_head(self, base_model):
base_model.summary()
x = base_model.output
x = GlobalAveragePooling2D(name='flatten_1')(x)
#x = Flatten(name='flatten_1')(x)
x = Dense(1024, activation='relu', name='fc6_1')(x)
x = Dense(1024, activation='relu', name='fc7_2')(x)
x = Dropout(0.5)(x)
predictions = Dense(self.nb_classes,
activation='softmax',
name='fc8_3')(x)
# this is the model we will train
model = Model(inputs=base_model.input, outputs=predictions)
for layer in base_model.layers:
layer.trainable = False
model.compile(optimizer="rmsprop", loss=LOSS, metrics=['accuracy'])
model.summary()
return model
class DefogGAN():
def __init__(self):
self.project_dir = (os.path.dirname(__file__))
self.data_path = os.path.join(self.project_dir, "./data/starCraft")
self.result_work_path = os.path.join(self.project_dir,
'./result/DefogGAN')
self.making_ing_path = os.path.join(self.project_dir,
'./result/DefogGAN')
self.createFolder(self.result_work_path)
is_multi_gpu = True
self.is_validation_check = True
self.save_weights = True
#self.num_of_replay = 3500
self.n_epochs = 1000
self.batch_size = 512
self.save_interval = 1000
self.num_of_making_img = 5
optimizer = Adam(0.0001, beta_1=0.5, beta_2=0.9)
self.discriminator = self.build_discriminator()
self.discriminator.compile(loss='binary_crossentropy',
optimizer=optimizer,
metrics=['accuracy'])
self.generator = self.build_generator()
fog_img = Input(shape=(82, 32, 32))
gen_missing = self.generator(fog_img)
self.discriminator.trainable = False
valid = self.discriminator(gen_missing)
self.combined = Model(fog_img, [gen_missing, valid])
if is_multi_gpu:
self.combined = multi_gpu_model(self.combined,
gpus=self.get_count_of_gpu())
weight = K.variable(
np.array([0.75, 0.1875, 0.0468, 0.012, 0.003, 0.0007]))
self.combined.compile(loss=[
self.weighted_pyramidal_loss(weights=weight), 'binary_crossentropy'
],
loss_weights=[0.999, 0.001],
optimizer=optimizer)
def accumulate_resolution(self, x, threshhold=0.5):
new = np.zeros((x.shape[0], 32, 32))
new_enemy = np.zeros((x.shape[0], 32, 32))
new_self = np.zeros((x.shape[0], 32, 32))
for n in range(x.shape[0]):
for w in range(x.shape[2]):
for h in range(x.shape[3]):
is_enemy = False
is_self = False
for u in range(x.shape[1]):
if u < 32:
# enemy
if not int(x[n][u][w][h] + threshhold) == 0:
is_enemy = True
new_enemy[n][w][h] -= int(x[n][u][w][h] +
threshhold)
else:
# self
if not int(x[n][u][w][h] + threshhold) == 0:
is_self = True
new_self[n][w][h] += int(x[n][u][w][h] +
threshhold)
if is_enemy and is_self:
new[n][w][h] = 0
elif not is_enemy and not is_self:
new[n][w][h] = -30
elif is_enemy and not is_self:
new[n][w][h] = new_enemy[n][w][h]
elif not is_enemy and is_self:
new[n][w][h] = new_self[n][w][h]
return new
def make_pickle(self):
for i in ['train', 'validation', 'test']:
for j in ['x', 'y']:
f = open('{}/{}_{}_data_set.csv'.format(self.data_path, j, i),
'r',
encoding='utf-8')
read_csv_file = csv.reader(f)
is_first = True
for line in read_csv_file:
if is_first:
tensor = np.zeros((int(line[0]), int(line[1]),
int(line[2]), int(line[3])))
is_first = False
else:
tensor[int(line[0])][int(line[1])][int(line[2])][int(
line[3])] = float(line[4])
#print(int(line[0]), int(line[1]), int(line[2]), int(line[3]), tensor[int(line[0])][int(line[1])][int(line[2])][int(line[3])])
f.close()
with open('{}/{}_{}_dataset.pkl'.format(self.data_path, j, i),
'wb') as f:
pickle.dump(tensor, f, pickle.HIGHEST_PROTOCOL)
def get_pickle_data(self):
with open('{}/x_train_dataset.pkl'.format(self.data_path), 'rb') as f:
x_train = pickle.load(f)
with open('{}/x_validation_dataset.pkl'.format(self.data_path),
'rb') as f:
x_validation = pickle.load(f)
with open('{}/x_test_dataset.pkl'.format(self.data_path), 'rb') as f:
x_test = pickle.load(f)
with open('{}/y_train_dataset.pkl'.format(self.data_path), 'rb') as f:
y_train = pickle.load(f)
with open('{}/y_validation_dataset.pkl'.format(self.data_path),
'rb') as f:
y_validation = pickle.load(f)
with open('{}/y_test_dataset.pkl'.format(self.data_path), 'rb') as f:
y_test = pickle.load(f)
return x_train, x_validation, x_test, y_train, y_validation, y_test
def get_sample_data(self):
f = open('{}/sample_data.csv'.format(self.data_path),
'r',
encoding='utf-8')
read_csv_file = csv.reader(f)
tensor_fog = np.zeros((self.num_of_replay, 82, 32, 32))
tensor_real = np.zeros((self.num_of_replay, 66, 32, 32))
for line in read_csv_file:
if len(line) == 1:
num_of_replay = int(line[0])
if len(line) == 5:
tensor_fog[num_of_replay][int(line[0])][int(line[1])][int(
line[2])] = int(line[3])
if int(line[0]) < 66:
tensor_real[num_of_replay][int(line[0])][int(line[1])][int(
line[2])] = int(line[4])
f.close()
# shuffle tensor index
train_set_index = random.sample(
range(0, self.num_of_replay),
int(self.num_of_replay * self.train_rate))
temp_set = [
i for i in range(0, self.num_of_replay) if not i in train_set_index
]
validation_set_index = temp_set[:int(self.num_of_replay *
self.validation_rate)]
test_set_index = temp_set[int(self.num_of_replay * self.test_rate):]
return tensor_fog[train_set_index], tensor_fog[
validation_set_index], tensor_fog[test_set_index], tensor_real[
train_set_index], tensor_real[
validation_set_index], tensor_real[test_set_index]
def check_pickle(self):
for i in ['train', 'validation', 'test']:
for j in ['x', 'y']:
fname = '{}/{}_{}_dataset.pkl'.format(self.data_path, j, i)
if not os.path.isfile(fname):
return False
return True
def train_defogGAN(self):
min_loss = 99999999999
valid = np.ones((self.batch_size, 1))
fake = np.zeros((self.batch_size, 1))
#x_train, x_validation, x_test, y_train, y_validation, y_test = self.get_sample_data()
is_pickle = self.check_pickle()
if not is_pickle:
self.make_pickle()
x_train, x_validation, x_test, y_train, y_validation, y_test = self.get_pickle_data(
)
best_img = None
last_epoch_img = None
for epoch in range(self.n_epochs + 1):
n_batches = int(x_train.shape[0] / self.batch_size)
for i in range(math.ceil(n_batches)):
start_batch = i * self.batch_size
end_batch = (1 + i) * self.batch_size
end_batch = x_train.shape[0] if end_batch > x_train.shape[
0] else (1 + i) * self.batch_size
images_train_x = x_train[start_batch:end_batch, :, :, :]
images_train_y = y_train[start_batch:end_batch, :, :, :]
gen_missing = self.generator.predict(images_train_x)
d_loss_real = self.discriminator.train_on_batch(
images_train_y, valid)
d_loss_fake = self.discriminator.train_on_batch(
gen_missing, fake)
d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)
# Train Generator
g_loss = self.combined.train_on_batch(images_train_x,
[images_train_y, valid])
if self.is_validation_check:
validation_recon_img = self.generator.predict(x_validation)
validation_loss = self.get_MSE_Value(validation_recon_img,
y_validation)
else:
validation_loss = 0
print(
"%d [D loss: %f, acc: %.2f%%] [G loss: %f, mse: %f, validation_mse : %f]"
% (epoch, d_loss[0], 100 * d_loss[1], g_loss[0], g_loss[1],
validation_loss))
if epoch % self.save_interval == 0:
last_epoch_img = self.generator.predict(x_test)
if self.save_weights:
self.createFolder('{}/interval_model/'.format(
self.result_work_path))
self.generator.save_weights(
'{}/interval_model/model_weight_{}.h5'.format(
self.result_work_path, epoch))
if min_loss > validation_loss and self.is_validation_check:
min_loss = validation_loss
best_img = self.generator.predict(x_test)
if self.save_weights:
if os.path.exists('{}/best_model'.format(
self.result_work_path)):
shutil.rmtree('{}/best_model'.format(
self.result_work_path))
self.createFolder('{}/best_model/'.format(
self.result_work_path))
self.generator.save_weights(
'{}/best_model/model_weight_{}.h5'.format(
self.result_work_path, epoch))
self.make_test(x_test, last_epoch_img, best_img, y_test)
def make_test(self, x_test, last_epoch_img, best_img, y_test):
n = self.num_of_making_img
x_test = x_test[:n]
x_test = x_test[:, :66, :, :]
last_epoch_img = last_epoch_img[:n]
best_img = best_img[:n]
y_test = y_test[:n]
result = np.zeros((0, n, 32, 32)) # model , index of replay , x, y
result = np.concatenate(
(result, self.accumulate_resolution(x_test).reshape(1, n, 32, 32)),
axis=0)
result = np.concatenate(
(result, self.accumulate_resolution(last_epoch_img).reshape(
1, n, 32, 32)),
axis=0)
result = np.concatenate(
(result, self.accumulate_resolution(best_img).reshape(
1, n, 32, 32)),
axis=0)
result = np.concatenate(
(result, self.accumulate_resolution(y_test).reshape(1, n, 32, 32)),
axis=0)
print(result.shape)
self.make_img(self.making_ing_path, result)
print('Success make image')
def make_img(self, path, map):
print(map.shape) # (7,30,32,32)
plt_threshhold = -20
model_names = [
'fog_exposed', 'last_epoch', 'best_epoch', 'Ground_truth'
]
fig, axn = plt.subplots(map.shape[1],
map.shape[0],
sharex=True,
sharey=True,
figsize=(map.shape[0] * 1.3,
map.shape[1] * 1.3))
cbar_ax = fig.add_axes([.91, .3, .03, .4])
fig.suptitle(
'GAN\'s compare[enemy(red): positive num, allies(green): negative num, both(yellow): 0]',
fontsize=16)
for i, ax in enumerate(axn.flat):
model_index = i % map.shape[0]
unit_index = int(i / map.shape[0])
if i < map.shape[0]:
ax.set_title(model_names[model_index], fontsize=10)
matrix = map[model_index][unit_index]
if plt_threshhold == 0:
sns.heatmap(matrix,
ax=ax,
cbar=i == 0,
annot=True,
fmt='.1f',
cmap=plt.cm.YlGnBu,
cbar_ax=None if i else cbar_ax)
else:
cbar_kws = {
'ticks': [-3, -2, -1, 0, 1, 2, 3],
'drawedges': True
}
sns.heatmap(matrix,
mask=(matrix < plt_threshhold),
ax=ax,
cbar=i == 0,
annot=False,
square=True,
fmt='.1f',
xticklabels=False,
yticklabels=False,
vmin=-3.5,
vmax=3.5,
cbar_kws=cbar_kws,
cmap=plt.get_cmap('RdYlGn', 7),
cbar_ax=None if i else cbar_ax)
count = 0
for ax in axn.flat:
model_index = count % map.shape[0]
unit_index = int(count / map.shape[0])
if model_index == 0:
ax.set(ylabel='replay {}'.format(unit_index))
ax.axhline(y=0, color='k', linewidth=1)
ax.axhline(y=32, color='k', linewidth=2)
ax.axvline(x=0, color='k', linewidth=1)
ax.axvline(x=32, color='k', linewidth=2)
count += 1
plt.subplots_adjust(hspace=0.03, wspace=0.03)
plt.savefig(path)
plt.close('all')
def build_generator(self):
input_channel = 82
output_channel = 66
input_shape = (input_channel, 32, 32)
img_input = Input(shape=input_shape, name='input')
depth = input_channel
# In: 100
# Out: dim x dim x depth
c1 = Conv2D(depth * 2, (3, 3),
strides=(1, 1),
activation='relu',
input_shape=input_shape,
padding='same',
data_format='channels_first')(img_input)
b1 = BatchNormalization(axis=-1,
momentum=0.99,
epsilon=0.001,
center=True,
scale=True,
beta_initializer='zeros',
gamma_initializer='ones',
moving_mean_initializer='zeros',
moving_variance_initializer='ones',
beta_regularizer=None,
gamma_regularizer=None,
beta_constraint=None,
gamma_constraint=None)(c1)
act1 = Activation('relu')(b1)
c2 = Conv2D(depth * 2, (3, 3),
strides=(2, 2),
activation='relu',
padding='same',
data_format='channels_first')(act1)
b2 = BatchNormalization(axis=-1,
momentum=0.99,
epsilon=0.001,
center=True,
scale=True,
beta_initializer='zeros',
gamma_initializer='ones',
moving_mean_initializer='zeros',
moving_variance_initializer='ones',
beta_regularizer=None,
gamma_regularizer=None,
beta_constraint=None,
gamma_constraint=None)(c2)
act2 = Activation('relu')(b2)
c3 = Conv2D(depth * 4, (3, 3),
strides=(1, 1),
activation='relu',
padding='same',
data_format='channels_first')(act2)
b3 = BatchNormalization(axis=-1,
momentum=0.99,
epsilon=0.001,
center=True,
scale=True,
beta_initializer='zeros',
gamma_initializer='ones',
moving_mean_initializer='zeros',
moving_variance_initializer='ones',
beta_regularizer=None,
gamma_regularizer=None,
beta_constraint=None,
gamma_constraint=None)(c3)
act3 = Activation('relu')(b3)
c4 = Conv2D(depth * 4, (3, 3),
strides=(2, 2),
activation='relu',
padding='same',
data_format='channels_first')(act3)
b4 = BatchNormalization(axis=-1,
momentum=0.99,
epsilon=0.001,
center=True,
scale=True,
beta_initializer='zeros',
gamma_initializer='ones',
moving_mean_initializer='zeros',
moving_variance_initializer='ones',
beta_regularizer=None,
gamma_regularizer=None,
beta_constraint=None,
gamma_constraint=None)(c4)
act4 = Activation('relu')(b4)
c5 = Conv2D(depth * 8, (3, 3),
strides=(1, 1),
activation='relu',
padding='same',
data_format='channels_first')(act4)
b5 = BatchNormalization(axis=-1,
momentum=0.99,
epsilon=0.001,
center=True,
scale=True,
beta_initializer='zeros',
gamma_initializer='ones',
moving_mean_initializer='zeros',
moving_variance_initializer='ones',
beta_regularizer=None,
gamma_regularizer=None,
beta_constraint=None,
gamma_constraint=None)(c5)
act5 = Activation('relu')(b5)
c6 = Conv2D(depth * 8, (3, 3),
strides=(1, 1),
activation='relu',
padding='same',
data_format='channels_first')(act5)
b6 = BatchNormalization(axis=-1,
momentum=0.99,
epsilon=0.001,
center=True,
scale=True,
beta_initializer='zeros',
gamma_initializer='ones',
moving_mean_initializer='zeros',
moving_variance_initializer='ones',
beta_regularizer=None,
gamma_regularizer=None,
beta_constraint=None,
gamma_constraint=None)(c6)
act6 = Activation('relu')(b6)
c7 = Conv2D(depth * 8, (3, 3),
strides=(2, 2),
activation='relu',
padding='same',
data_format='channels_first')(act6)
b7 = BatchNormalization(axis=-1,
momentum=0.99,
epsilon=0.001,
center=True,
scale=True,
beta_initializer='zeros',
gamma_initializer='ones',
moving_mean_initializer='zeros',
moving_variance_initializer='ones',
beta_regularizer=None,
gamma_regularizer=None,
beta_constraint=None,
gamma_constraint=None)(c7)
act7 = Activation('relu')(b7)
ct1 = Conv2DTranspose(depth * 8, (3, 3),
strides=(2, 2),
activation='relu',
padding='same',
data_format='channels_first')(act7)
act8 = Activation('relu')(ct1)
act8_output = Lambda(lambda x: x, name='act8_output')(act8)
act8_output = keras.layers.Add()([act6, act8_output])
ct2 = Conv2DTranspose(depth * 8, (3, 3),
strides=(1, 1),
activation='relu',
padding='same',
data_format='channels_first')(act8_output)
act9 = Activation('relu')(ct2)
act9_output = Lambda(lambda x: x, name='act9_output')(act9)
act9_output = keras.layers.Add()([act5, act9_output])
ct3 = Conv2DTranspose(depth * 4, (3, 3),
strides=(1, 1),
activation='relu',
padding='same',
data_format='channels_first')(act9_output)
act10 = Activation('relu')(ct3)
act10_output = Lambda(lambda x: x, name='act10_output')(act10)
act10_output = keras.layers.Add()([act4, act10_output])
ct4 = Conv2DTranspose(depth * 4, (3, 3),
strides=(2, 2),
activation='relu',
padding='same',
data_format='channels_first')(act10_output)
act11 = Activation('relu')(ct4)
act11_output = Lambda(lambda x: x, name='act11_output')(act11)
act11_output = keras.layers.Add()([act3, act11_output])
ct5 = Conv2DTranspose(depth * 2, (3, 3),
strides=(1, 1),
activation='relu',
padding='same',
data_format='channels_first')(act11_output)
act12 = Activation('relu')(ct5)
act12_output = Lambda(lambda x: x, name='act12_output')(act12)
act12_output = keras.layers.Add()([act2, act12_output])
ct6 = Conv2DTranspose(depth * 2, (3, 3),
strides=(2, 2),
activation='relu',
padding='same',
data_format='channels_first')(act12_output)
act13 = Activation('relu')(ct6)
act13_output = Lambda(lambda x: x, name='act13_output')(act13)
act13_output = keras.layers.Add()([act1, act13_output])
ct7 = Conv2DTranspose(depth, (3, 3),
strides=(1, 1),
activation='relu',
padding='same',
data_format='channels_first')(act13_output)
act14 = Activation('relu')(ct7)
act14_output = Lambda(lambda x: x, name='output')(act14)
act14_output = keras.layers.Add()([img_input, act14_output])
ct8 = Conv2DTranspose(output_channel, (3, 3),
strides=(1, 1),
activation='relu',
padding='same',
data_format='channels_first')(act14_output)
act15 = Activation('relu')(ct8)
img_output = act15
model = Model(inputs=[img_input], outputs=[img_output])
model.summary()
return model
def build_discriminator(self):
D = Sequential()
discriminator_input_channel = 66
depth = discriminator_input_channel
dropout = 0.4
input_shape = (discriminator_input_channel, 32, 32)
D.add(
Conv2D(depth * 1, (3, 3),
strides=(2, 2),
input_shape=input_shape,
padding='same',
data_format='channels_first'))
D.add(LeakyReLU(alpha=0.2))
D.add(Dropout(dropout))
D.add(
Conv2D(depth * 2, (3, 3),
strides=(2, 2),
padding='same',
data_format='channels_first'))
D.add(LeakyReLU(alpha=0.2))
D.add(Dropout(dropout))
D.add(
Conv2D(depth * 4, (3, 3),
strides=(2, 2),
padding='same',
data_format='channels_first'))
D.add(LeakyReLU(alpha=0.2))
D.add(Dropout(dropout))
D.add(Flatten())
D.add(Dense(1))
D.add(Activation('sigmoid'))
D.summary()
return D
def weighted_pyramidal_loss(self, weights):
def pyramidal_loss(y_true, y_pred):
yt_2 = keras.layers.AveragePooling2D((2, 2))(y_true) * 2
yt_4 = keras.layers.AveragePooling2D((4, 4))(y_true) * 4
yt_8 = keras.layers.AveragePooling2D((8, 8))(y_true) * 8
yt_16 = keras.layers.AveragePooling2D((16, 16))(y_true) * 16
yt_32 = keras.layers.AveragePooling2D((32, 32))(y_true) * 32
yp_2 = keras.layers.AveragePooling2D((2, 2))(y_pred) * 2
yp_4 = keras.layers.AveragePooling2D((4, 4))(y_pred) * 4
yp_8 = keras.layers.AveragePooling2D((8, 8))(y_pred) * 8
yp_16 = keras.layers.AveragePooling2D((16, 16))(y_pred) * 16
yp_32 = keras.layers.AveragePooling2D((32, 32))(y_pred) * 32
loss_0 = keras.losses.mean_squared_error(y_true, y_pred)
loss_2 = keras.losses.mean_squared_error(yt_2, yp_2)
loss_4 = keras.losses.mean_squared_error(yt_4, yp_4)
loss_8 = keras.losses.mean_squared_error(yt_8, yp_8)
loss_16 = keras.losses.mean_squared_error(yt_16, yp_16)
loss_32 = keras.losses.mean_squared_error(yt_32, yp_32)
loss_0 = tf.reduce_mean(loss_0, axis=[1, 2])
loss_2 = tf.reduce_mean(loss_2, axis=[1, 2])
loss_4 = tf.reduce_mean(loss_4, axis=[1, 2])
loss_8 = tf.reduce_mean(loss_8, axis=[1, 2])
loss_16 = tf.reduce_mean(loss_16, axis=[1, 2])
loss_32 = tf.reduce_mean(loss_32, axis=[1, 2])
loss = weights[0] * loss_0 + \
weights[1] * loss_2 + \
weights[2] * loss_4 + \
weights[3] * loss_8 + \
weights[4] * loss_16 + \
weights[5] * loss_32
return loss
return pyramidal_loss
def get_count_of_gpu(self):
device_list = device_lib.list_local_devices()
gpu_count = 0
for d in device_list:
if d.device_type == 'GPU':
gpu_count += 1
return int(gpu_count)
def createFolder(self, directory):
try:
if not os.path.exists(directory):
os.makedirs(directory)
except OSError:
print('Error: Creating directory. ' + directory)
def get_MSE_Value(self, x, y):
MSE_GAN = 0
MSE_DIV = x.shape[0] * x.shape[1]
for n in range(x.shape[0]):
for c in range(x.shape[1]):
MSE_GAN += mean_squared_error(x[n][c], y[n][c])
MSE = MSE_GAN / MSE_DIV
return MSE
def test_model_custom_target_tensors():
a = Input(shape=(3, ), name='input_a')
b = Input(shape=(3, ), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
y = K.placeholder([10, 4], name='y')
y1 = K.placeholder([10, 3], name='y1')
y2 = K.placeholder([7, 5], name='y2')
model = Model([a, b], [a_2, b_2])
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
# test list of target tensors
with pytest.raises(ValueError):
model.compile(optimizer,
loss,
metrics=[],
loss_weights=loss_weights,
sample_weight_mode=None,
target_tensors=[y, y1, y2])
model.compile(optimizer,
loss,
metrics=[],
loss_weights=loss_weights,
sample_weight_mode=None,
target_tensors=[y, y1])
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np], {
y: np.random.random((10, 4)),
y1: np.random.random((10, 3))
})
# test dictionary of target_tensors
with pytest.raises(ValueError):
model.compile(optimizer,
loss,
metrics=[],
loss_weights=loss_weights,
sample_weight_mode=None,
target_tensors={'does_not_exist': y2})
# test dictionary of target_tensors
model.compile(optimizer,
loss,
metrics=[],
loss_weights=loss_weights,
sample_weight_mode=None,
target_tensors={
'dense_1': y,
'dropout': y1
})
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np], {
y: np.random.random((10, 4)),
y1: np.random.random((10, 3))
})
if K.backend() == 'tensorflow':
import tensorflow as tf
# test with custom TF placeholder as target
pl_target_a = tf.placeholder('float32', shape=(None, 4))
model.compile(optimizer='rmsprop',
loss='mse',
target_tensors={'dense_1': pl_target_a})
model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
class PolicyValueNet():
"""策略价值网络"""
#def __init__(self, board_width, board_height, model_file=None):
def __init__(self, policy_infer_size, model_file=None):
#self.board_width = board_width
#self.board_height = board_height
self.policy_infer_size = policy_infer_size
self.l2_const = 1e-4 # coef of l2 penalty
self.create_policy_value_net()
self._loss_train_op()
self.load_model_done = True
if model_file and os.path.exists(model_file):
self.load_model_done = False
self.load_model(model_file)
def load_model(self, model_file):
"""重新加载模型(仅用于selfplay时load new model)"""
try:
#net_params = pickle.load(open(model_file, 'rb'), encoding='bytes') #iso-8859-1')
net_params = utils.pickle_load(model_file)
self.model.set_weights(net_params)
self.load_model_done = True
except:
logging.error("load_model fail! {}\t{}".format(
model_file, utils.get_trace()))
self.load_model_done = False
if os.path.exists(
model_file
) and self.load_model_done is False: #鏂囦欢瀛樺湪鍗村鍦ㄥけ璐ユ椂缁堟杩愯
exit(-1)
return self.load_model_done
def create_policy_value_net(self):
"""创建policy-value网络"""
# 输入层
#in_x = network = Input((4, self.board_width, self.board_height))
in_x = network = Input((4, 1, self.policy_infer_size))
# conv layers
network = Conv2D(filters=32,
kernel_size=(3, 3),
padding="same",
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
network = Conv2D(filters=64,
kernel_size=(3, 3),
padding="same",
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
network = Conv2D(filters=128,
kernel_size=(3, 3),
padding="same",
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
# 走子策略 action policy layers
policy_net = Conv2D(filters=4,
kernel_size=(1, 1),
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
policy_net = Flatten()(policy_net)
# infer self.board_width * self.board_height action_probs
#self.policy_net = Dense(self.board_width * self.board_height, activation="softmax", kernel_regularizer=l2(self.l2_const))(policy_net)
self.policy_net = Dense(self.policy_infer_size,
activation="softmax",
kernel_regularizer=l2(
self.l2_const))(policy_net)
# 盘面价值 state value layers
value_net = Conv2D(filters=2,
kernel_size=(1, 1),
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
value_net = Flatten()(value_net)
value_net = Dense(64, kernel_regularizer=l2(self.l2_const))(value_net)
# infer one current state score
self.value_net = Dense(1,
activation="tanh",
kernel_regularizer=l2(self.l2_const))(value_net)
# 创建网络模型
self.model = Model(in_x, [self.policy_net, self.value_net])
# 返回走子策略和价值概率
def policy_value(state_input):
state_input_union = np.array(state_input)
#print(state_input_union)
results = self.model.predict_on_batch(state_input_union)
return results
self.policy_value = policy_value
def policy_value_fn(self, board):
"""使用模型预测棋盘所有actionid的价值概率"""
# 棋盘所有可移动action_ids
legal_positions = board.availables
#print(legal_positions)
# 当前玩家角度的actions过程
current_actions = board.current_actions()
#print(current_actions)
# 使用模型预测走子策略和价值概率
#print(self.policy_infer_size)
#act_probs, value = self.policy_value(current_actions.reshape(-1, 4, self.board_width, self.board_height))
act_probs, value = self.policy_value(
current_actions.reshape(-1, 4, 1, self.policy_infer_size))
act_probs = zip(legal_positions, act_probs.flatten()[legal_positions])
# 返回[(action, 概率)] 以及当前玩家的后续走子value
return act_probs, value[0][0]
def _loss_train_op(self):
"""初始化损失
3个损失函数因子
loss = (z - v)^2 + pi^T * log(p) + c||theta||^2
loss = value损失函数 + policy损失函数 + 惩罚项
"""
# 定义优化器和损失函数
opt = Adam()
losses = ['categorical_crossentropy', 'mean_squared_error']
self.model.compile(optimizer=opt, loss=losses)
def self_entropy(probs):
return -np.mean(np.sum(probs * np.log(probs + 1e-10), axis=1))
def train_step(state_input, mcts_probs, winner, learning_rate):
"""输出训练过程中的结果"""
state_input_union = np.array(state_input)
mcts_probs_union = np.array(mcts_probs)
winner_union = np.array(winner)
# 评估
loss = self.model.evaluate(state_input_union,
[mcts_probs_union, winner_union],
batch_size=len(state_input),
verbose=0)
# 预测
action_probs, _ = self.model.predict_on_batch(state_input_union)
entropy = self_entropy(action_probs)
K.set_value(self.model.optimizer.lr, learning_rate)
self.model.fit(state_input_union, [mcts_probs_union, winner_union],
batch_size=len(state_input),
verbose=0)
return loss[0], entropy
self.train_step = train_step
def get_policy_param(self):
"""获得模型参数"""
net_params = self.model.get_weights()
return net_params
def save_model(self, model_file):
"""保存模型参数到文件"""
net_params = self.get_policy_param()
#pickle.dump(net_params, open(model_file, 'wb'), protocol=4)
utils.pickle_dump(net_params, model_file)
# GAN网络编译
gan_core = Model(inputs=gan_x,
outputs=[
gan_output, features_full, features_medium, features_low,
pred_full, pred_medium, pred_low
])
gan_core.name = "gan_core"
optimizer = Adam(learning_rate, 0.5, decay=decay_rate)
loss_gan = ['mae', 'mae', 'mae', 'mae', 'mse', 'mse', 'mse']
loss_weights_gan = [1, 3.33, 3.33, 3.33, 0.33, 0.33, 0.33]
# gan_core = multi_gpu_model(gan_core_org)
gan_core.compile(optimizer=optimizer,
loss_weights=loss_weights_gan,
loss=loss_gan)
# --------------------------------
# 编译判别器Discriminator
# --------------------------------
discriminator_full.model.trainable = True
discriminator_medium.model.trainable = True
discriminator_low.model.trainable = True
def zero_loss(y_true, y_pred):
return K.zeros_like(y_true)
def test_pandas_dataframe():
input_a = Input(shape=(3,), name='input_a')
input_b = Input(shape=(3,), name='input_b')
x = Dense(4, name='dense_1')(input_a)
y = Dense(3, name='desne_2')(input_b)
model_1 = Model(inputs=input_a, outputs=x)
model_2 = Model(inputs=[input_a, input_b], outputs=[x, y])
optimizer = 'rmsprop'
loss = 'mse'
model_1.compile(optimizer=optimizer, loss=loss)
model_2.compile(optimizer=optimizer, loss=loss)
input_a_df = pd.DataFrame(np.random.random((10, 3)))
input_b_df = pd.DataFrame(np.random.random((10, 3)))
output_a_df = pd.DataFrame(np.random.random((10, 4)))
output_b_df = pd.DataFrame(np.random.random((10, 3)))
model_1.fit(input_a_df,
output_a_df)
model_2.fit([input_a_df, input_b_df],
[output_a_df, output_b_df])
model_1.fit([input_a_df],
[output_a_df])
model_1.fit({'input_a': input_a_df},
output_a_df)
model_2.fit({'input_a': input_a_df, 'input_b': input_b_df},
[output_a_df, output_b_df])
model_1.predict(input_a_df)
model_2.predict([input_a_df, input_b_df])
model_1.predict([input_a_df])
model_1.predict({'input_a': input_a_df})
model_2.predict({'input_a': input_a_df, 'input_b': input_b_df})
model_1.predict_on_batch(input_a_df)
model_2.predict_on_batch([input_a_df, input_b_df])
model_1.predict_on_batch([input_a_df])
model_1.predict_on_batch({'input_a': input_a_df})
model_2.predict_on_batch({'input_a': input_a_df, 'input_b': input_b_df})
model_1.evaluate(input_a_df,
output_a_df)
model_2.evaluate([input_a_df, input_b_df],
[output_a_df, output_b_df])
model_1.evaluate([input_a_df],
[output_a_df])
model_1.evaluate({'input_a': input_a_df},
output_a_df)
model_2.evaluate({'input_a': input_a_df, 'input_b': input_b_df},
[output_a_df, output_b_df])
model_1.train_on_batch(input_a_df,
output_a_df)
model_2.train_on_batch([input_a_df, input_b_df],
[output_a_df, output_b_df])
model_1.train_on_batch([input_a_df],
[output_a_df])
model_1.train_on_batch({'input_a': input_a_df},
output_a_df)
model_2.train_on_batch({'input_a': input_a_df, 'input_b': input_b_df},
[output_a_df, output_b_df])
model_1.test_on_batch(input_a_df,
output_a_df)
model_2.test_on_batch([input_a_df, input_b_df],
[output_a_df, output_b_df])
model_1.test_on_batch([input_a_df],
[output_a_df])
model_1.test_on_batch({'input_a': input_a_df},
output_a_df)
model_2.test_on_batch({'input_a': input_a_df, 'input_b': input_b_df},
[output_a_df, output_b_df])
class FinancialNewsAnalysisModel(object):
model = None
def __init__(self, nb_time_step, dim_data, batch_size=1, model_path=None):
self.model_path = model_path
self.model_path = model_path
self.batch_size = batch_size
self.size_of_input_data_dim = dim_data
self.size_of_input_timesteps = nb_time_step
self.build()
self.weight_loaded = False
if model_path is not None:
self.load_weights()
def build(self):
dim_data = self.size_of_input_data_dim
nb_time_step = self.size_of_input_timesteps
news_input = Input(shape=(nb_time_step, dim_data))
lstm = LSTM(output_dim=nb_hidden_units, dropout_U=dropout, dropout_W=dropout,
W_regularizer=l2(l2_norm_alpha), b_regularizer=l2(l2_norm_alpha), activation='tanh')
bi_lstm = Bidirectional(lstm, input_shape=(nb_time_step, dim_data), merge_mode='concat')
all_news_rep = bi_lstm(news_input)
news_predictions = Dense(1, activation='linear')(all_news_rep)
self.model = Model(news_input, news_predictions, name="deep rnn for financial news analysis")
def reset(self):
for l in self.model.layers:
if type(l) is LSTM:
l.reset_status()
def compile_model(self, lr=0.0001, loss_weights=0.1):
optimizer = Adam(lr=lr)
loss = 'mse'
# loss = custom_objective
self.model.compile(optimizer=optimizer, loss=loss)
#metrics=['mse'])
plot(self.model, to_file='model.png')
def fit_model(self, X, y, X_val=None, y_val=None, epoch=500):
early_stopping = EarlyStopping(monitor='val_loss', patience=100, verbose=0)
if X_val is None:
self.model.fit(X, y, batch_size=self.batch_size, nb_epoch=epoch, validation_split=0.2,
shuffle=True, callbacks=[early_stopping])
else:
self.model.fit(X, y, batch_size=self.batch_size, nb_epoch=epoch, validation_data=(X_val, y_val),
shuffle=True, callbacks=[early_stopping])
def save(self):
self.model.save_weights(self.model_path, overwrite=True)
def load_weights(self):
if os.path.exists(self.model_path):
self.model.load_weights(self.model_path)
self.weight_loaded = True
def print_weights(self, weights=None, detail=False):
weights = weights or self.model.get_weights()
for w in weights:
print("w%s: sum(w)=%s, ave(w)=%s" % (w.shape, np.sum(w), np.average(w)))
if detail:
for w in weights:
print("%s: %s" % (w.shape, w))
def model_eval(self, X, y):
y_hat = self.model.predict(X, batch_size=1)
count_true = 0
count_all = y.shape[0]
for i in range(y.shape[0]):
count_true = count_true + 1 if y[i,0]*y_hat[i,0]>0 else count_true
print y[i,0],y_hat[i,0]
print count_all,count_true
class RenderGAN:
def __init__(self,
tag3d_network,
z_dim_offset=50, z_dim_labels=50, z_dim_bits=12,
discriminator_units=32,
discriminator_depth=2,
generator_units=32,
generator_depth=2,
data_shape=(1, 64, 64),
labels_shape=(24,),
nb_bits=12,
generator_optimizer=Adam(lr=0.0004, beta_1=0.5),
discriminator_optimizer=Adam(lr=0.0004, beta_1=0.5)
):
self.tag3d_network = tag3d_network
self.z_dim_offset = z_dim_offset
self.z_dim_labels = z_dim_labels
self.z_dim_bits = z_dim_bits
self.z_dim = z_dim_offset + z_dim_labels + z_dim_bits
self.discriminator_units = discriminator_units
self.discriminator_depth = discriminator_depth
self.generator_units = generator_units
self.generator_depth = generator_depth
self.preprocess_units = self.generator_units // 2
self.data_shape = data_shape
self.labels_shape = labels_shape
self.nb_bits = nb_bits
self.generator_optimizer = generator_optimizer
self.discriminator_optimizer = discriminator_optimizer
self._build()
def _build_discriminator(self):
x = Input(shape=self.data_shape, name='data')
d = render_gan_discriminator(x, n=self.discriminator_units,
conv_repeat=self.discriminator_depth, dense=[512])
self.discriminator = Model([x], [d])
def _build_generator_given_z_offset_and_labels(self):
labels = Input(shape=self.labels_shape, name='input_labels')
z_offset = Input(shape=(self.z_dim_offset,), name='input_z_offset')
outputs = OrderedDict()
labels_without_bits = Subtensor(self.nb_bits, self.labels_shape[0], axis=1)(labels)
raw_tag3d, tag3d_depth_map = self.tag3d_network(labels)
tag3d = ScaleUnitIntervalTo(-1, 1)(raw_tag3d)
outputs['tag3d'] = tag3d
outputs['tag3d_depth_map'] = tag3d_depth_map
segmentation = Segmentation(threshold=-0.08, smooth_threshold=0.2,
sigma=1.5, name='segmentation')
tag3d_downsampled = PyramidReduce()(tag3d)
tag3d_segmented = segmentation(raw_tag3d)
outputs['tag3d_segmented'] = tag3d_segmented
tag3d_segmented_blur = GaussianBlur(sigma=0.66)(tag3d_segmented)
out_offset_front = get_offset_front([z_offset, ZeroGradient()(labels_without_bits)],
self.generator_units)
light_depth_map = get_preprocess(tag3d_depth_map, self.preprocess_units,
nb_conv_layers=2)
light_outs = get_lighting_generator([out_offset_front, light_depth_map],
self.generator_units)
offset_depth_map = get_preprocess(tag3d_depth_map, self.preprocess_units,
nb_conv_layers=2)
offset_middle_light = get_preprocess(concat(light_outs), self.preprocess_units,
resize=['down', 'down'])
offset_middle_tag3d = get_preprocess(tag3d_downsampled,
self.preprocess_units // 2,
resize=['down', ''],
nb_conv_layers=2)
out_offset_middle = get_offset_middle(
[out_offset_front, offset_depth_map,
offset_middle_light, offset_middle_tag3d],
self.generator_units)
offset_back_tag3d_downsampled = get_preprocess(tag3d_downsampled,
self.preprocess_units // 2,
nb_conv_layers=2)
offset_back_feature_map, out_offset_back = get_offset_back(
[out_offset_middle, offset_back_tag3d_downsampled], self.generator_units)
blur_factor = get_blur_factor(out_offset_middle, min=0.25, max=1.)
outputs['blur_factor'] = blur_factor
tag3d_blur = BlendingBlur(sigma=2.0)([tag3d, blur_factor])
outputs['tag3d_blur'] = tag3d_blur
outputs['light_black'] = light_outs[0]
outputs['light_white'] = light_outs[1]
outputs['light_shift'] = light_outs[2]
tag3d_lighten = AddLighting(
scale_factor=0.90, shift_factor=0.90)([tag3d_blur] + light_outs)
tag3d_lighten = InBounds(clip=True, weight=15)(tag3d_lighten)
outputs['tag3d_lighten'] = tag3d_lighten
outputs['background_offset'] = out_offset_back
blending = Background(name='blending')([out_offset_back, tag3d_lighten,
tag3d_segmented_blur])
outputs['fake_without_noise'] = blending
details = get_details(
[blending, tag3d_segmented_blur, tag3d, out_offset_back,
offset_back_feature_map] + light_outs, self.generator_units)
outputs['details_offset'] = details
details_high_pass = HighPass(3.5, nb_steps=3)(details)
outputs['details_high_pass'] = details_high_pass
fake = InBounds(-2.0, 2.0)(
merge([details_high_pass, blending], mode='sum'))
outputs['fake'] = fake
for name in outputs.keys():
outputs[name] = name_tensor(outputs[name], name)
self.generator_given_z_and_labels = Model([z_offset, labels], [fake])
self.sample_generator_given_z_and_labels_output_names = list(outputs.keys())
self.sample_generator_given_z_and_labels = Model([z_offset, labels],
list(outputs.values()))
@property
def pos_z_bits(self):
return (0, self.z_dim_bits)
@property
def pos_z_labels(self):
return (self.z_dim_bits, self.z_dim_bits + self.z_dim_labels)
@property
def pos_z_offset(self):
return (self.z_dim_labels + self.z_dim_bits,
self.z_dim_labels + self.z_dim_bits + self.z_dim_offset)
def _build_generator_given_z(self):
z = Input(shape=(self.z_dim,), name='z')
z_bits = Subtensor(*self.pos_z_bits, axis=1)(z)
z_labels = Subtensor(*self.pos_z_labels, axis=1)(z)
z_offset = Subtensor(*self.pos_z_offset, axis=1)(z)
bits = ThresholdBits()(z_bits)
nb_labels_without_bits = self.labels_shape[0] - self.nb_bits
generated_labels = get_label_generator(
z_labels, self.generator_units, nb_output_units=nb_labels_without_bits)
labels_normed = NormSinCosAngle(0)(generated_labels)
labels = concat([bits, labels_normed], name='labels')
fake = self.generator_given_z_and_labels([z_offset, labels])
self.generator_given_z = Model([z], [fake])
sample_tensors = self.sample_generator_given_z_and_labels([z_offset, labels])
sample_tensors = [name_tensor(t, n)
for t, n in zip(sample_tensors,
self.sample_generator_given_z_and_labels.output_names)]
self.sample_generator_given_z_output_names = ['labels'] + \
self.sample_generator_given_z_and_labels_output_names
self.sample_generator_given_z = Model([z], [labels] + sample_tensors)
def _build_gan(self):
self.generator_given_z.compile(self.generator_optimizer, 'binary_crossentropy')
self.discriminator.compile(self.discriminator_optimizer, 'binary_crossentropy')
self.gan = GAN(self.generator_given_z, self.discriminator)
def _build(self):
self._build_discriminator()
self._build_generator_given_z_offset_and_labels()
self._build_generator_given_z()
self._build_gan()
def save_weights(self, fname_format, overwrite=False, attrs={}):
def save(name):
model = self.__dict__[name]
fname = fname_format.format(name=name)
os.makedirs(os.path.dirname(fname), exist_ok=True)
save_model(model, fname, overwrite=False, attrs=attrs)
save("sample_generator_given_z")
save("sample_generator_given_z_and_labels")
save("generator_given_z_and_labels")
save("generator_given_z")
save("discriminator")
def test_model_with_external_loss():
# None loss, only regularization loss.
a = Input(shape=(3, ), name='input_a')
a_2 = Dense(4,
name='dense_1',
kernel_regularizer='l1',
bias_regularizer='l2')(a)
dp = Dropout(0.5, name='dropout')
a_3 = dp(a_2)
model = Model(a, [a_2, a_3])
optimizer = 'rmsprop'
loss = None
model.compile(optimizer, loss, metrics=['mae'])
input_a_np = np.random.random((10, 3))
# test train_on_batch
out = model.train_on_batch(input_a_np, None)
out = model.test_on_batch(input_a_np, None)
# fit
out = model.fit(input_a_np, None)
# evaluate
out = model.evaluate(input_a_np, None)
# No dropout, external loss.
a = Input(shape=(3, ), name='input_a')
a_2 = Dense(4, name='dense_1')(a)
a_3 = Dense(4, name='dense_2')(a)
model = Model(a, [a_2, a_3])
model.add_loss(K.mean(a_3 + a_2))
optimizer = 'rmsprop'
loss = None
model.compile(optimizer, loss, metrics=['mae'])
# test train_on_batch
out = model.train_on_batch(input_a_np, None)
out = model.test_on_batch(input_a_np, None)
# fit
out = model.fit(input_a_np, None)
# evaluate
out = model.evaluate(input_a_np, None)
# Test fit with no external data at all.
if K.backend() == 'tensorflow':
import tensorflow as tf
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
a_2 = Dense(4, name='dense_1')(a)
a_2 = Dropout(0.5, name='dropout')(a_2)
model = Model(a, a_2)
model.add_loss(K.mean(a_2))
model.compile(optimizer='rmsprop',
loss=None,
metrics=['mean_squared_error'])
# test train_on_batch
out = model.train_on_batch(None, None)
out = model.test_on_batch(None, None)
out = model.predict_on_batch(None)
# test fit
with pytest.raises(ValueError):
out = model.fit(None, None, epochs=1, batch_size=10)
out = model.fit(None, None, epochs=1, steps_per_epoch=1)
# test fit with validation data
with pytest.raises(ValueError):
out = model.fit(None,
None,
epochs=1,
steps_per_epoch=None,
validation_steps=2)
out = model.fit(None,
None,
epochs=1,
steps_per_epoch=2,
validation_steps=2)
# test evaluate
with pytest.raises(ValueError):
out = model.evaluate(None, None, batch_size=10)
out = model.evaluate(None, None, steps=3)
# test predict
with pytest.raises(ValueError):
out = model.predict(None, batch_size=10)
out = model.predict(None, steps=3)
assert out.shape == (10 * 3, 4)
# Test multi-output model without external data.
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
a_1 = Dense(4, name='dense_1')(a)
a_2 = Dropout(0.5, name='dropout')(a_1)
model = Model(a, [a_1, a_2])
model.add_loss(K.mean(a_2))
model.compile(optimizer='rmsprop',
loss=None,
metrics=['mean_squared_error'])
# test train_on_batch
out = model.train_on_batch(None, None)
out = model.test_on_batch(None, None)
out = model.predict_on_batch(None)
# test fit
with pytest.raises(ValueError):
out = model.fit(None, None, epochs=1, batch_size=10)
out = model.fit(None, None, epochs=1, steps_per_epoch=1)
# test fit with validation data
with pytest.raises(ValueError):
out = model.fit(None,
None,
epochs=1,
steps_per_epoch=None,
validation_steps=2)
out = model.fit(None,
None,
epochs=1,
steps_per_epoch=2,
validation_steps=2)
# test evaluate
with pytest.raises(ValueError):
out = model.evaluate(None, None, batch_size=10)
out = model.evaluate(None, None, steps=3)
# test predict
with pytest.raises(ValueError):
out = model.predict(None, batch_size=10)
out = model.predict(None, steps=3)
assert len(out) == 2
assert out[0].shape == (10 * 3, 4)
assert out[1].shape == (10 * 3, 4)
def decoder_stochastic_wrn(label_sizes,
nb_bits=12,
data_shape=(1, 64, 64),
activation=lambda: Activation('relu'),
normalization=lambda: BatchNormalization(axis=1, mode=0),
weight_init='he_normal',
weight_decay=0.0001,
dropout_probability=0.,
wrn_depth=58,
wrn_k=2,
death_rate=0.5,
optimizer='adam'):
def norm_act_block():
def f(inputs):
x = normalization()(inputs)
x = activation()(x)
return x
return f
def conv2(nb_filter, nb_row, nb_col, subsample=(1, 1), bias=True):
return Convolution2D(nb_filter, nb_row, nb_col,
init=weight_init,
border_mode='same',
subsample=subsample,
bias=bias,
W_regularizer=l2(weight_decay))
def dropout(p=None):
if p is None:
p = dropout_probability
def f(inputs):
if p > 0.:
inputs = Dropout(p)(inputs)
return inputs
return f
def equal_gate_shape(input_shapes):
assert(input_shapes[0] == input_shapes[1])
return input_shapes[1]
def residual_block(nb_filter, stochastic=False, stochastic_layers=None):
def f(inputs):
x = norm_act_block()(inputs)
x = conv2(nb_filter, 3, 3)(x)
x = dropout()(x)
x = norm_act_block()(x)
x = conv2(nb_filter, 3, 3)(x)
if inputs._keras_shape != x._keras_shape:
inputs = conv2(nb_filter, 1, 1, bias=False)(inputs)
if not stochastic:
return merge((inputs, x), mode='sum')
scale = ScaleInTestPhase(death_rate)
x = scale(x)
out = merge([inputs, x], mode="sum", output_shape=x._keras_shape[1:])
rs = RandomSwitch(death_rate)
stochastic_layers.append((rs.death_rate, scale.death_rate))
return rs([out, inputs])
return f
def residual_reduction_block(nb_filter):
def f(inputs):
x = norm_act_block()(inputs)
x = conv2(nb_filter, 3, 3, subsample=(2, 2))(x)
x = dropout()(x)
x = norm_act_block()(x)
x = conv2(nb_filter, 3, 3)(x)
inputs_bottleneck = conv2(nb_filter, 1, 1, subsample=(2, 2),
bias=False)(inputs)
s = merge((inputs_bottleneck, x), mode='sum')
return s
return f
def skip_connection(inputs, residual, stochastic=False,
stochastic_layers=None):
inputs_filters = inputs._keras_shape[1]
residual_filters = residual._keras_shape[1]
subsample = np.array(inputs._keras_shape[2:]) // np.array(residual._keras_shape[2:])
inputs = dropout()(inputs)
if (inputs_filters != residual_filters) or np.any(subsample > 1):
skip = conv2(residual_filters, 1, 1, subsample=subsample, bias=False)(inputs)
else:
skip = inputs
if not stochastic:
return merge((skip, residual), mode='sum')
else:
scale = ScaleInTestPhase(death_rate)
skip = scale(skip)
out = merge([skip, residual], mode="sum")
rs = RandomSwitch(death_rate)
stochastic_layers.append((rs.death_rate, scale.death_rate))
return rs([out, residual])
n = (wrn_depth - 4) // 6
stochastic_layers = []
input = Input(shape=data_shape)
m_stem = conv2(16, 3, 3, subsample=(2, 2))(input)
m_stem = norm_act_block()(m_stem)
m_b1 = residual_block(nb_filter=16 * wrn_k,
stochastic_layers=stochastic_layers)(m_stem)
for _ in range(n - 1):
m_b1 = residual_block(nb_filter=16 * wrn_k, stochastic=True,
stochastic_layers=stochastic_layers)(m_b1)
m_b1 = skip_connection(input, m_b1, stochastic=True,
stochastic_layers=stochastic_layers)
m_b2 = residual_reduction_block(nb_filter=32 * wrn_k)(m_b1)
for _ in range(n - 1):
m_b2 = residual_block(nb_filter=32 * wrn_k, stochastic=True,
stochastic_layers=stochastic_layers)(m_b2)
m_b2 = skip_connection(m_b1, m_b2, stochastic=True,
stochastic_layers=stochastic_layers)
m_b3 = residual_reduction_block(nb_filter=64 * wrn_k)(m_b2)
for _ in range(n - 1):
m_b3 = residual_block(nb_filter=64 * wrn_k, stochastic=True,
stochastic_layers=stochastic_layers)(m_b3)
m_b3 = skip_connection(m_b2, m_b3, stochastic=True,
stochastic_layers=stochastic_layers)
m_b3 = skip_connection(input, m_b3, stochastic=True,
stochastic_layers=stochastic_layers)
x = norm_act_block()(m_b3)
x = AveragePooling2D(pool_size=(8, 8))(x)
x = dropout()(x)
for i, (tb, ts) in enumerate(stochastic_layers, start=0):
K.set_value(tb, i / len(stochastic_layers) * death_rate)
K.set_value(ts, i / len(stochastic_layers) * death_rate)
outputs, losses = decoder_end_block(x, label_sizes, nb_bits, activation, weight_decay)
model = Model(input, list(outputs.values()))
model.compile(optimizer, loss=list(losses.values()),
loss_weights={k: decoder_loss_weights(k) for k in losses.keys()})
return model
class SeqGAN:
#
# Initialization
# ----------------------------------------------------------------------------
def __init__(self, g, d, m, g_optimizer, d_optimizer):
# Model of generator
self.g = g
# Model of discriminator
self.d = d
# Model of ???
self.m = m
self.z, self.seq_input = self.g.inputs
self.fake_prob, = self.g.outputs
self.history = None
with trainable(m, False):
# m_input = merge([self.seq_input, self.fake_prob], mode='concat', concat_axis=1)
m_input = concatenate([self.seq_input, self.fake_prob], axis=1)
self.m_realness = self.m(m_input)
self.model_fit_g = Model(inputs=[self.z, self.seq_input],
outputs=[self.m_realness])
self.model_fit_g.compile(optimizer=g_optimizer,
loss=K.binary_crossentropy)
self.d.compile(optimizer=d_optimizer, loss=K.binary_crossentropy)
#
# Return the shape of input noise variables z
# (batch_size: the number of data read per batch)
# ----------------------------------------------------------------------------
def z_shape(self, batch_size=64):
# layer, _, _ = self.z._keras_history
# _keras_history是一个tuple类型，第一个参数代表previous Layer； 第二个参数node_index; 第三个参数是tensor_index
# （ 因为可能有多个tensor output，如果是只有一个tensor output的话，tensor_index = 0）
layer, node_index, tensor_index = self.z._keras_history
return (batch_size,
) + layer.output_shape[1:] # the first dimension is data size
#
# Sample input noise variables z by uniform distributions
# (batch_size: the number of data read per batch)
# ----------------------------------------------------------------------------
def sample_z(self, batch_size=64):
shape = self.z_shape(batch_size)
return np.random.uniform(-1, 1, shape)
#
# Generate the fake samples with: the input noise variables z & input sequence seq_input
# ----------------------------------------------------------------------------
def generate(self, z, seq_input, batch_size=32):
return self.g.predict([z, seq_input], batch_size=batch_size)
#
# Training
# ----------------------------------------------------------------------------
def train_on_batch(self, seq_input, real, d_target=None):
nb_real = len(real)
nb_fake = len(seq_input)
if d_target is None:
d_target = np.concatenate(
[np.zeros((nb_fake, 1)),
np.ones((nb_real, 1))])
fake_prob = self.generate(self.sample_z(nb_fake), seq_input)
fake = np.concatenate([seq_input, prob_to_sentence(fake_prob)], axis=1)
fake_and_real = np.concatenate([fake, real], axis=0)
d_loss = self.d.train_on_batch(x=fake_and_real, y=d_target)
d_realness = self.d.predict(fake)
m_loss = self.m.train_on_batch(x=np.concatenate([seq_input, fake_prob],
axis=1),
y=d_realness)
g_loss = self.model_fit_g.train_on_batch(
x=[self.sample_z(nb_fake), seq_input], y=np.ones((nb_fake, 1)))
return g_loss, d_loss, m_loss
#
# Training
# ----------------------------------------------------------------------------
def fit_generator(self,
generator,
nb_epoch,
nb_batches_per_epoch,
callbacks=[],
batch_size=None,
verbose=False):
if batch_size is None:
batch_size = 2 * len(next(generator)[0])
out_labels = ['g', 'd', 'm']
self.history = cbks.History()
callbacks = [cbks.BaseLogger()] + callbacks + [self.history]
if verbose:
callbacks += [cbks.ProgbarLogger()]
callbacks = cbks.CallbackList(callbacks)
callbacks.set_model(self)
callbacks.set_params({
'nb_epoch': nb_epoch,
'nb_sample': nb_batches_per_epoch * batch_size,
'verbose': verbose,
'metrics': out_labels,
})
callbacks.on_train_begin()
for e in range(nb_epoch):
callbacks.on_epoch_begin(e)
for batch_index, (seq_input, real) in enumerate(generator):
callbacks.on_batch_begin(batch_index)
batch_logs = dict()
batch_logs['batch'] = batch_index
batch_logs['size'] = len(real) + len(seq_input)
outs = self.train_on_batch(seq_input, real)
for l, o in zip(out_labels, outs):
batch_logs[l] = o
callbacks.on_batch_end(batch_index, batch_logs)
if batch_index + 1 == nb_batches_per_epoch:
break
callbacks.on_epoch_end(e)
callbacks.on_train_end()
def test_model_methods():
a = Input(shape=(3,), name='input_a')
b = Input(shape=(3,), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
model = Model([a, b], [a_2, b_2])
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
# training/testing doesn't work before compiling.
with pytest.raises(RuntimeError):
model.train_on_batch([input_a_np, input_b_np], [output_a_np, output_b_np])
model.compile(optimizer, loss, metrics=[], loss_weights=loss_weights,
sample_weight_mode=None)
# test train_on_batch
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.train_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.train_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np})
# test fit
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np], epochs=1, batch_size=4)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np], epochs=1, batch_size=4)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
epochs=1, batch_size=4)
# test validation_split
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
epochs=1, batch_size=4, validation_split=0.5)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np],
epochs=1, batch_size=4, validation_split=0.5)
# test validation data
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
epochs=1, batch_size=4,
validation_data=([input_a_np, input_b_np], [output_a_np, output_b_np]))
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np],
epochs=1, batch_size=4, validation_split=0.5,
validation_data=({'input_a': input_a_np, 'input_b': input_b_np}, [output_a_np, output_b_np]))
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
epochs=1, batch_size=4, validation_split=0.5,
validation_data=(
{'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np}))
# test_on_batch
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.test_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.test_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np})
# predict_on_batch
out = model.predict_on_batch([input_a_np, input_b_np])
out = model.predict_on_batch({'input_a': input_a_np, 'input_b': input_b_np})
# predict, evaluate
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
# with sample_weight
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
sample_weight = [None, np.random.random((10,))]
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
# test accuracy metric
model.compile(optimizer, loss, metrics=['acc'],
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
# this should also work
model.compile(optimizer, loss, metrics={'dense_1': 'acc'},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# and this as well
model.compile(optimizer, loss, metrics={'dense_1': ['acc']},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# test starting from non-zero initial epoch
trained_epochs = []
trained_batches = []
# define tracer callback
def on_epoch_begin(epoch, logs):
trained_epochs.append(epoch)
def on_batch_begin(batch, logs):
trained_batches.append(batch)
tracker_cb = LambdaCallback(on_epoch_begin=on_epoch_begin,
on_batch_begin=on_batch_begin)
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np], epochs=5, batch_size=4,
initial_epoch=2, callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test starting from non-zero initial epoch for generator too
trained_epochs = []
def gen_data(batch_sz):
while True:
yield ([np.random.random((batch_sz, 3)), np.random.random((batch_sz, 3))],
[np.random.random((batch_sz, 4)), np.random.random((batch_sz, 3))])
out = model.fit_generator(gen_data(4), steps_per_epoch=3, epochs=5,
initial_epoch=2, callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test with a custom metric function
def mse(y_true, y_pred):
return K.mean(K.pow(y_true - y_pred, 2))
model.compile(optimizer, loss, metrics=[mse],
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out_len = 1 + 2 * (1 + 1) # total loss + 2 outputs * (loss + metric)
assert len(out) == out_len
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == out_len
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4, epochs=1)
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
# enable verbose for evaluate_generator
out = model.evaluate_generator(gen_data(4), steps=3, verbose=1)
# empty batch
with pytest.raises(ValueError):
def gen_data():
while True:
yield (np.asarray([]), np.asarray([]))
out = model.evaluate_generator(gen_data(), steps=1)
# x is not a list of numpy arrays.
with pytest.raises(ValueError):
out = model.predict([None])
# x does not match _feed_input_names.
with pytest.raises(ValueError):
out = model.predict([input_a_np, None, input_b_np])
with pytest.raises(ValueError):
out = model.predict([None, input_a_np, input_b_np])
# all input/output/weight arrays should have the same number of samples.
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np[:2]],
[output_a_np, output_b_np],
sample_weight=sample_weight)
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np[:2]],
sample_weight=sample_weight)
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=[sample_weight[1], sample_weight[1][:2]])
# `sample_weight` is neither a dict nor a list.
with pytest.raises(TypeError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=tuple(sample_weight))
# `validation_data` is neither a tuple nor a triple.
with pytest.raises(ValueError):
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
epochs=1, batch_size=4,
validation_data=([input_a_np, input_b_np],))
# `loss` does not match outputs.
with pytest.raises(ValueError):
model.compile(optimizer, loss=['mse', 'mae', 'mape'])
# `loss_weights` does not match output_names.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', loss_weights={'lstm': 0.5})
# `loss_weights` does not match outputs.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', loss_weights=[0.5])
# `loss_weights` is invalid type.
with pytest.raises(TypeError):
model.compile(optimizer, loss='mse', loss_weights=(0.5, 0.5))
# `sample_weight_mode` does not match output_names.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', sample_weight_mode={'lstm': 'temporal'})
# `sample_weight_mode` does not match output_names.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', sample_weight_mode=['temporal'])
# `sample_weight_mode` matches output_names partially.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', sample_weight_mode={'dense_1': 'temporal'})
# `loss` does not exist.
with pytest.raises(ValueError):
model.compile(optimizer, loss=[])
model.compile(optimizer, loss=['mse', 'mae'])
model.compile(optimizer, loss='mse', loss_weights={'dense_1': 0.2, 'dropout': 0.8})
model.compile(optimizer, loss='mse', loss_weights=[0.2, 0.8])
# the rank of weight arrays should be 1.
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=[None, np.random.random((10, 20, 30))])
model.compile(optimizer, loss='mse', sample_weight_mode={'dense_1': None, 'dropout': 'temporal'})
model.compile(optimizer, loss='mse', sample_weight_mode=[None, 'temporal'])
# the rank of output arrays should be at least 3D.
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
model.compile(optimizer, loss, metrics=[], loss_weights=loss_weights,
sample_weight_mode=None)
trained_epochs = []
trained_batches = []
out = model.fit_generator(generator=RandomSequence(3), steps_per_epoch=3, epochs=5,
initial_epoch=0, validation_data=RandomSequence(4),
validation_steps=3, callbacks=[tracker_cb])
assert trained_epochs == [0, 1, 2, 3, 4]
assert trained_batches == list(range(3)) * 5
# steps_per_epoch will be equal to len of sequence if it's unspecified
trained_epochs = []
trained_batches = []
out = model.fit_generator(generator=RandomSequence(3), epochs=5,
initial_epoch=0, validation_data=RandomSequence(4),
callbacks=[tracker_cb])
assert trained_epochs == [0, 1, 2, 3, 4]
assert trained_batches == list(range(12)) * 5
# fit_generator will throw an exception if steps is unspecified for regular generator
with pytest.raises(ValueError):
def gen_data():
while True:
yield (np.asarray([]), np.asarray([]))
out = model.fit_generator(generator=gen_data(), epochs=5,
initial_epoch=0, validation_data=gen_data(),
callbacks=[tracker_cb])
# Check if generator is only accessed an expected number of times
gen_counters = [0, 0]
def gen_data(i):
while True:
gen_counters[i] += 1
yield ([np.random.random((1, 3)), np.random.random((1, 3))],
[np.random.random((1, 4)), np.random.random((1, 3))])
out = model.fit_generator(generator=gen_data(0), epochs=3,
steps_per_epoch=2,
validation_data=gen_data(1),
validation_steps=1,
max_queue_size=2,
workers=2)
# Need range check here as filling of the queue depends on sleep in the enqueuers
assert 6 <= gen_counters[0] <= 8
# 12 = (epoch * workers * validation steps * max_queue_size)
assert 3 <= gen_counters[1] <= 12
gen_counters = [0]
out = model.fit_generator(generator=RandomSequence(3), epochs=3,
validation_data=gen_data(0),
validation_steps=1,
max_queue_size=2,
workers=2)
# 12 = (epoch * workers * validation steps * max_queue_size)
# Need range check here as filling of the queue depends on sleep in the enqueuers
assert 3 <= gen_counters[0] <= 12
# predict_generator output shape behavior should be consistent
def expected_shape(batch_size, n_batches):
return (batch_size * n_batches, 4), (batch_size * n_batches, 3)
# Multiple outputs and one step.
batch_size = 5
sequence_length = 1
shape_0, shape_1 = expected_shape(batch_size, sequence_length)
out = model.predict_generator(RandomSequence(batch_size,
sequence_length=sequence_length))
assert np.shape(out[0]) == shape_0 and np.shape(out[1]) == shape_1
# Multiple outputs and multiple steps.
batch_size = 5
sequence_length = 2
shape_0, shape_1 = expected_shape(batch_size, sequence_length)
out = model.predict_generator(RandomSequence(batch_size,
sequence_length=sequence_length))
assert np.shape(out[0]) == shape_0 and np.shape(out[1]) == shape_1
# Create a model with a single output.
single_output_model = Model([a, b], a_2)
single_output_model.compile(optimizer, loss, metrics=[], sample_weight_mode=None)
# Single output and one step.
batch_size = 5
sequence_length = 1
shape_0, _ = expected_shape(batch_size, sequence_length)
out = single_output_model.predict_generator(RandomSequence(batch_size,
sequence_length=sequence_length))
assert np.shape(out) == shape_0
# Single output and multiple steps.
batch_size = 5
sequence_length = 2
shape_0, _ = expected_shape(batch_size, sequence_length)
out = single_output_model.predict_generator(RandomSequence(batch_size,
sequence_length=sequence_length))
assert np.shape(out) == shape_0
class AdditionNPIModel(NPIStep):
model = None
f_enc = None
def __init__(self, system: RuntimeSystem, model_path: str=None, program_set: AdditionProgramSet=None):
self.system = system
self.model_path = model_path
self.program_set = program_set
self.batch_size = 1
self.build()
self.weight_loaded = False
self.load_weights()
def build(self):
enc_size = self.size_of_env_observation()
argument_size = IntegerArguments.size_of_arguments
input_enc = InputLayer(batch_input_shape=(self.batch_size, enc_size), name='input_enc')
input_arg = InputLayer(batch_input_shape=(self.batch_size, argument_size), name='input_arg')
input_prg = Embedding(input_dim=PROGRAM_VEC_SIZE, output_dim=PROGRAM_KEY_VEC_SIZE, input_length=1,
batch_input_shape=(self.batch_size, 1))
f_enc = Sequential(name='f_enc')
f_enc.add(Merge([input_enc, input_arg], mode='concat'))
f_enc.add(Dense(256))
f_enc.add(Dense(32))
f_enc.add(Activation('relu', name='relu_enc'))
self.f_enc = f_enc
program_embedding = Sequential(name='program_embedding')
program_embedding.add(input_prg)
f_enc_convert = Sequential(name='f_enc_convert')
f_enc_convert.add(f_enc)
f_enc_convert.add(RepeatVector(1))
f_lstm = Sequential(name='f_lstm')
f_lstm.add(Merge([f_enc_convert, program_embedding], mode='concat'))
# f_lstm.add(Activation('relu', name='relu_lstm_0'))
f_lstm.add(LSTM(256, return_sequences=False, stateful=True))
f_lstm.add(Activation('relu', name='relu_lstm_1'))
f_lstm.add(RepeatVector(1))
f_lstm.add(LSTM(256, return_sequences=False, stateful=True))
f_lstm.add(Activation('relu', name='relu_lstm_2'))
# plot(f_lstm, to_file='f_lstm.png', show_shapes=True)
f_end = Sequential(name='f_end')
f_end.add(f_lstm)
f_end.add(Dense(10))
f_end.add(Dense(1))
f_end.add(Activation('hard_sigmoid', name='hard_sigmoid_end'))
# plot(f_end, to_file='f_end.png', show_shapes=True)
f_prog = Sequential(name='f_prog')
f_prog.add(f_lstm)
f_prog.add(Dense(PROGRAM_KEY_VEC_SIZE))
f_prog.add(Dense(PROGRAM_VEC_SIZE))
f_prog.add(Activation('softmax', name='softmax_prog'))
# plot(f_prog, to_file='f_prog.png', show_shapes=True)
f_args = []
for ai in range(1, IntegerArguments.max_arg_num+1):
f_arg = Sequential(name='f_arg%s' % ai)
f_arg.add(f_lstm)
f_arg.add(Dense(32))
f_arg.add(Dense(IntegerArguments.depth))
f_arg.add(Activation('softmax', name='softmax_arg%s' % ai))
f_args.append(f_arg)
# plot(f_arg, to_file='f_arg.png', show_shapes=True)
self.model = Model([input_enc.input, input_arg.input, input_prg.input],
[f_end.output, f_prog.output] + [fa.output for fa in f_args],
name="npi")
self.compile_model()
plot(self.model, to_file='model.png', show_shapes=True)
def reset(self):
super(AdditionNPIModel, self).reset()
for l in self.model.layers:
if type(l) is LSTM:
l.reset_states()
def compile_model(self, lr=0.0001, arg_weight=1.):
arg_num = IntegerArguments.max_arg_num
optimizer = Adam(lr=lr)
loss = ['binary_crossentropy', 'categorical_crossentropy'] + ['categorical_crossentropy'] * arg_num
self.model.compile(optimizer=optimizer, loss=loss, loss_weights=[0.25, 0.25] + [arg_weight] * arg_num)
def fit(self, steps_list, epoch=3000):
"""
:param int epoch:
:param typing.List[typing.Dict[q=dict, steps=typing.List[StepInOut]]] steps_list:
:return:
"""
def filter_question(condition_func):
sub_steps_list = []
for steps_dict in steps_list:
question = steps_dict['q']
if condition_func(question['in1'], question['in2']):
sub_steps_list.append(steps_dict)
return sub_steps_list
# self.print_weights()
if not self.weight_loaded:
self.train_f_enc(filter_question(lambda a, b: 10 <= a < 100 and 10 <= b < 100), epoch=100)
self.f_enc.trainable = False
q_type = "training questions of a+b < 10"
print(q_type)
pr = 0.8
all_ok = self.fit_to_subset(filter_question(lambda a, b: a+b < 10), epoch=epoch, pass_rate=pr)
print("%s is pass_rate >= %s: %s" % (q_type, pr, all_ok))
q_type = "training questions of a<10 and b< 10 and 10 <= a+b"
print(q_type)
pr = 0.8
all_ok = self.fit_to_subset(filter_question(lambda a, b: a<10 and b<10 and a + b >= 10), epoch=epoch, pass_rate=pr)
print("%s is pass_rate >= %s: %s" % (q_type, pr, all_ok))
q_type = "training questions of a<10 and b<10"
print(q_type)
pr = 0.8
all_ok = self.fit_to_subset(filter_question(lambda a, b: a < 10 and b < 10), epoch=epoch, pass_rate=pr)
print("%s is pass_rate >= %s: %s" % (q_type, pr, all_ok))
q_type = "training questions of a<100 and b<100"
print(q_type)
pr = 0.8
all_ok = self.fit_to_subset(filter_question(lambda a, b: a < 100 and b < 100), epoch=epoch, pass_rate=pr)
print("%s is pass_rate >= %s: %s" % (q_type, pr, all_ok))
while True:
print("test all type of questions")
cc, wc = self.test_to_subset(create_questions(1000))
print("Accuracy %s(OK=%d, NG=%d)" % (cc/(cc+wc), cc, wc))
if wc == 0:
break
q_type = "training questions of ALL"
print(q_type)
pr = 1.0
self.fit_to_subset(filter_question(lambda a, b: True), epoch=epoch, pass_rate=pr)
all_ok = self.fit_to_subset(filter_question(lambda a, b: True), epoch=epoch, pass_rate=pr, skip_correct=True)
print("%s is pass_rate >= %s: %s" % (q_type, pr, all_ok))
def fit_to_subset(self, steps_list, epoch=3000, pass_rate=1.0, skip_correct=False):
learning_rate = 0.0001
for i in range(30):
all_ok = self.do_learn(steps_list, 30, learning_rate=learning_rate, pass_rate=pass_rate, arg_weight=1.,
skip_correct=skip_correct)
if all_ok:
return True
learning_rate *= 0.95
return False
def test_to_subset(self, questions):
addition_env = AdditionEnv(FIELD_ROW, FIELD_WIDTH, FIELD_DEPTH)
npi_runner = TerminalNPIRunner(None, self)
correct_count = wrong_count = 0
for idx, question in enumerate(questions):
question = copy(question)
if self.question_test(addition_env, npi_runner, question):
correct_count += 1
else:
wrong_count += 1
return correct_count, wrong_count
@staticmethod
def dict_to_str(d):
return str(tuple([(k, d[k]) for k in sorted(d)]))
def do_learn(self, steps_list, epoch, learning_rate=None, pass_rate=1.0, arg_weight=1., skip_correct=False):
if learning_rate is not None:
self.update_learning_rate(learning_rate, arg_weight)
addition_env = AdditionEnv(FIELD_ROW, FIELD_WIDTH, FIELD_DEPTH)
npi_runner = TerminalNPIRunner(None, self)
last_weights = None
correct_count = Counter()
no_change_count = 0
last_loss = 1000
for ep in range(1, epoch+1):
correct_new = wrong_new = 0
losses = []
ok_rate = []
np.random.shuffle(steps_list)
for idx, steps_dict in enumerate(steps_list):
question = copy(steps_dict['q'])
question_key = self.dict_to_str(question)
if self.question_test(addition_env, npi_runner, question):
if correct_count[question_key] == 0:
correct_new += 1
correct_count[question_key] += 1
print("GOOD!: ep=%2d idx=%3d :%s CorrectCount=%s" % (ep, idx, self.dict_to_str(question), correct_count[question_key]))
ok_rate.append(1)
if skip_correct or int(math.sqrt(correct_count[question_key])) ** 2 != correct_count[question_key]:
continue
else:
ok_rate.append(0)
if correct_count[question_key] > 0:
print("Degraded: ep=%2d idx=%3d :%s CorrectCount=%s -> 0" % (ep, idx, self.dict_to_str(question), correct_count[question_key]))
correct_count[question_key] = 0
wrong_new += 1
steps = steps_dict['steps']
xs = []
ys = []
ws = []
for step in steps:
xs.append(self.convert_input(step.input))
y, w = self.convert_output(step.output)
ys.append(y)
ws.append(w)
self.reset()
for i, (x, y, w) in enumerate(zip(xs, ys, ws)):
loss = self.model.train_on_batch(x, y, sample_weight=w)
if not np.isfinite(loss):
print("Loss is not finite!, Last Input=%s" % ([i, (x, y, w)]))
self.print_weights(last_weights, detail=True)
raise RuntimeError("Loss is not finite!")
losses.append(loss)
last_weights = self.model.get_weights()
if losses:
cur_loss = np.average(losses)
print("ep=%2d: ok_rate=%.2f%% (+%s -%s): ave loss %s (%s samples)" %
(ep, np.average(ok_rate)*100, correct_new, wrong_new, cur_loss, len(steps_list)))
# self.print_weights()
if correct_new + wrong_new == 0:
no_change_count += 1
else:
no_change_count = 0
if math.fabs(1 - cur_loss/last_loss) < 0.001 and no_change_count > 5:
print("math.fabs(1 - cur_loss/last_loss) < 0.001 and no_change_count > 5:")
return False
last_loss = cur_loss
print("=" * 80)
self.save()
if np.average(ok_rate) >= pass_rate:
return True
return False
def update_learning_rate(self, learning_rate, arg_weight=1.):
print("Re-Compile Model lr=%s aw=%s" % (learning_rate, arg_weight))
self.compile_model(learning_rate, arg_weight=arg_weight)
def train_f_enc(self, steps_list, epoch=50):
print("training f_enc")
f_add0 = Sequential(name='f_add0')
f_add0.add(self.f_enc)
f_add0.add(Dense(FIELD_DEPTH))
f_add0.add(Activation('softmax', name='softmax_add0'))
f_add1 = Sequential(name='f_add1')
f_add1.add(self.f_enc)
f_add1.add(Dense(FIELD_DEPTH))
f_add1.add(Activation('softmax', name='softmax_add1'))
env_model = Model(self.f_enc.inputs, [f_add0.output, f_add1.output], name="env_model")
env_model.compile(optimizer='adam', loss=['categorical_crossentropy']*2)
for ep in range(epoch):
losses = []
for idx, steps_dict in enumerate(steps_list):
prev = None
for step in steps_dict['steps']:
x = self.convert_input(step.input)[:2]
env_values = step.input.env.reshape((4, -1))
in1 = np.clip(env_values[0].argmax() - 1, 0, 9)
in2 = np.clip(env_values[1].argmax() - 1, 0, 9)
carry = np.clip(env_values[2].argmax() - 1, 0, 9)
y_num = in1 + in2 + carry
now = (in1, in2, carry)
if prev == now:
continue
prev = now
y0 = to_one_hot_array((y_num % 10)+1, FIELD_DEPTH)
y1 = to_one_hot_array((y_num // 10)+1, FIELD_DEPTH)
y = [yy.reshape((self.batch_size, -1)) for yy in [y0, y1]]
loss = env_model.train_on_batch(x, y)
losses.append(loss)
print("ep %3d: loss=%s" % (ep, np.average(losses)))
def question_test(self, addition_env, npi_runner, question):
addition_env.reset()
self.reset()
try:
run_npi(addition_env, npi_runner, self.program_set.ADD, question)
if question['correct']:
return True
except StopIteration:
pass
return False
def convert_input(self, p_in: StepInput):
x_pg = np.array((p_in.program.program_id,))
x = [xx.reshape((self.batch_size, -1)) for xx in (p_in.env, p_in.arguments.values, x_pg)]
return x
def convert_output(self, p_out: StepOutput):
y = [np.array((p_out.r,))]
weights = [[1.]]
if p_out.program:
arg_values = p_out.arguments.values
arg_num = len(p_out.program.args or [])
y += [p_out.program.to_one_hot(PROGRAM_VEC_SIZE)]
weights += [[1.]]
else:
arg_values = IntegerArguments().values
arg_num = 0
y += [np.zeros((PROGRAM_VEC_SIZE, ))]
weights += [[1e-10]]
for v in arg_values: # split by each args
y += [v]
weights += [[1.]] * arg_num + [[1e-10]] * (len(arg_values) - arg_num)
weights = [np.array(w) for w in weights]
return [yy.reshape((self.batch_size, -1)) for yy in y], weights
def step(self, env_observation: np.ndarray, pg: Program, arguments: IntegerArguments) -> StepOutput:
x = self.convert_input(StepInput(env_observation, pg, arguments))
results = self.model.predict(x, batch_size=1) # if batch_size==1, returns single row
r, pg_one_hot, arg_values = results[0], results[1], results[2:]
program = self.program_set.get(pg_one_hot.argmax())
ret = StepOutput(r, program, IntegerArguments(values=np.stack(arg_values)))
return ret
def save(self):
self.model.save_weights(self.model_path, overwrite=True)
def load_weights(self):
if os.path.exists(self.model_path):
self.model.load_weights(self.model_path)
self.weight_loaded = True
def print_weights(self, weights=None, detail=False):
weights = weights or self.model.get_weights()
for w in weights:
print("w%s: sum(w)=%s, ave(w)=%s" % (w.shape, np.sum(w), np.average(w)))
if detail:
for w in weights:
print("%s: %s" % (w.shape, w))
@staticmethod
def size_of_env_observation():
return FIELD_ROW * FIELD_DEPTH
class PolicyValueNet():
"""policy-value network """
def __init__(self, board_width, board_height, model_file=None):
self.board_width = board_width
self.board_height = board_height
self.l2_const = 1e-4 # coef of l2 penalty
self.create_policy_value_net()
if model_file:
print("[Notice] load model from file")
self.model = load_model(model_file)
else:
print("[Notice] create model")
self.create_policy_value_net()
self._loss_train_op()
def create_policy_value_net(self):
"""create the policy value network """
in_x = network = Input((4, self.board_width, self.board_height))
# conv layers
network = Conv2D(filters=32,
kernel_size=(3, 3),
padding="same",
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
network = Conv2D(filters=64,
kernel_size=(3, 3),
padding="same",
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
network = Conv2D(filters=128,
kernel_size=(3, 3),
padding="same",
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
# action policy layers
policy_net = Conv2D(filters=4,
kernel_size=(1, 1),
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
policy_net = Flatten()(policy_net)
self.policy_net = Dense(self.board_width * self.board_height,
activation="softmax",
kernel_regularizer=l2(
self.l2_const))(policy_net)
# state value layers
value_net = Conv2D(filters=2,
kernel_size=(1, 1),
data_format="channels_first",
activation="relu",
kernel_regularizer=l2(self.l2_const))(network)
value_net = Flatten()(value_net)
value_net = Dense(64, kernel_regularizer=l2(self.l2_const))(value_net)
self.value_net = Dense(1,
activation="tanh",
kernel_regularizer=l2(self.l2_const))(value_net)
self.model = Model(in_x, [self.policy_net, self.value_net])
def policy_value(state_input):
state_input_union = np.array(state_input)
results = self.model.predict_on_batch(state_input_union)
return results
self.policy_value = policy_value
def policy_value_fn(self, board):
"""
input: board
output: a list of (action, probability) tuples for each available action and the score of the board state
"""
legal_positions = board.availables
current_state = board.current_state()
act_probs, value = self.policy_value(
current_state.reshape(-1, 4, self.board_width, self.board_height))
act_probs = zip(legal_positions, act_probs.flatten()[legal_positions])
return act_probs, value[0][0]
def _loss_train_op(self):
"""
Three loss terms：
loss = (z - v)^2 + pi^T * log(p) + c||theta||^2
"""
# get the train op
opt = Adam()
losses = ['categorical_crossentropy', 'mean_squared_error']
self.model.compile(optimizer=opt, loss=losses)
def self_entropy(probs):
return -np.mean(np.sum(probs * np.log(probs + 1e-10), axis=1))
def train_step(state_input, mcts_probs, winner, learning_rate):
state_input_union = np.array(state_input)
mcts_probs_union = np.array(mcts_probs)
winner_union = np.array(winner)
loss = self.model.evaluate(state_input_union,
[mcts_probs_union, winner_union],
batch_size=len(state_input),
verbose=0)
action_probs, _ = self.model.predict_on_batch(state_input_union)
entropy = self_entropy(action_probs)
K.set_value(self.model.optimizer.lr, learning_rate)
self.model.fit(state_input_union, [mcts_probs_union, winner_union],
batch_size=len(state_input),
verbose=0)
return loss[0], entropy
self.train_step = train_step
def get_policy_param(self):
net_params = self.model.get_weights()
return net_params
def save_model(self, model_file):
""" save model to file """
print("save model file")
self.model.save(model_file)
def test_pandas_dataframe():
input_a = Input(shape=(3, ), name='input_a')
input_b = Input(shape=(3, ), name='input_b')
x = Dense(4, name='dense_1')(input_a)
y = Dense(3, name='desne_2')(input_b)
model_1 = Model(inputs=input_a, outputs=x)
model_2 = Model(inputs=[input_a, input_b], outputs=[x, y])
optimizer = 'rmsprop'
loss = 'mse'
model_1.compile(optimizer=optimizer, loss=loss)
model_2.compile(optimizer=optimizer, loss=loss)
input_a_df = pd.DataFrame(np.random.random((10, 3)))
input_b_df = pd.DataFrame(np.random.random((10, 3)))
output_a_df = pd.DataFrame(np.random.random((10, 4)))
output_b_df = pd.DataFrame(np.random.random((10, 3)))
model_1.fit(input_a_df, output_a_df)
model_2.fit([input_a_df, input_b_df], [output_a_df, output_b_df])
model_1.fit([input_a_df], [output_a_df])
model_1.fit({'input_a': input_a_df}, output_a_df)
model_2.fit({
'input_a': input_a_df,
'input_b': input_b_df
}, [output_a_df, output_b_df])
model_1.predict(input_a_df)
model_2.predict([input_a_df, input_b_df])
model_1.predict([input_a_df])
model_1.predict({'input_a': input_a_df})
model_2.predict({'input_a': input_a_df, 'input_b': input_b_df})
model_1.predict_on_batch(input_a_df)
model_2.predict_on_batch([input_a_df, input_b_df])
model_1.predict_on_batch([input_a_df])
model_1.predict_on_batch({'input_a': input_a_df})
model_2.predict_on_batch({'input_a': input_a_df, 'input_b': input_b_df})
model_1.evaluate(input_a_df, output_a_df)
model_2.evaluate([input_a_df, input_b_df], [output_a_df, output_b_df])
model_1.evaluate([input_a_df], [output_a_df])
model_1.evaluate({'input_a': input_a_df}, output_a_df)
model_2.evaluate({
'input_a': input_a_df,
'input_b': input_b_df
}, [output_a_df, output_b_df])
model_1.train_on_batch(input_a_df, output_a_df)
model_2.train_on_batch([input_a_df, input_b_df],
[output_a_df, output_b_df])
model_1.train_on_batch([input_a_df], [output_a_df])
model_1.train_on_batch({'input_a': input_a_df}, output_a_df)
model_2.train_on_batch({
'input_a': input_a_df,
'input_b': input_b_df
}, [output_a_df, output_b_df])
model_1.test_on_batch(input_a_df, output_a_df)
model_2.test_on_batch([input_a_df, input_b_df], [output_a_df, output_b_df])
model_1.test_on_batch([input_a_df], [output_a_df])
model_1.test_on_batch({'input_a': input_a_df}, output_a_df)
model_2.test_on_batch({
'input_a': input_a_df,
'input_b': input_b_df
}, [output_a_df, output_b_df])
class VAE(AE):
"""
Variational Autoencoder. This consists of an encoder and a decoder plus an interpolateable latent space.
"""
def __init__(self,
encoder=None,
decoder=None,
autoencoder=None,
latent_dim=None):
super(VAE, self).__init__(encoder=None, decoder=None)
# Encoder and decoder must be provided.
assert (encoder != None and decoder != None)
# From loading.
if encoder != None and decoder != None and autoencoder != None:
self.encoder = encoder
self.decoder = decoder
self.autoencoder = autoencoder
self.latent_dim = decoder.inputs[0].shape.as_list()[-1]
return
# Set the latent dimensions.
self.latent_dim = latent_dim
assert self.latent_dim != None
# Encoder.
encoder_input = encoder.inputs[0]
encoder_output = encoder.outputs[0]
z_mean = layers.Dense(self.latent_dim, name='z_mean')(encoder_output)
z_log_var = layers.Dense(self.latent_dim,
name='z_log_var')(encoder_output)
z = layers.Lambda(sampling, output_shape=(self.latent_dim, ),
name='z')([z_mean, z_log_var])
self.encoder = Model(encoder_input, [z_mean, z_log_var, z],
name='encoder')
# Decoder.
self.decoder = decoder
# Creating the VAE.
inputs = self.encoder.inputs[0]
outputs = self.decoder(self.encoder(inputs)[2]) # This is z.
self.autoencoder = Model(inputs, outputs, name="vae")
def compile(self,
optimizer,
loss=None,
metrics=None,
loss_weights=None,
sample_weight_mode=None,
weighted_metrics=None,
target_tensors=None,
**kwargs):
"""
Compiles the VAE.
Additionally to the default functionality of *compile*, it adds the VAE-loss.
This loss takes the provided loss and interprets it as a reconstruction-loss.
The VAE loss is similar to
>>> vae_loss = mean(r_loss + kl_loss)
See the literature for details.
"""
self.loss = loss
# Inputs.
inputs = self.encoder.inputs[0]
inputs_dim = int(np.prod(inputs.shape.as_list()[1:]))
# Outputs.
z_mean = self.encoder.outputs[0]
z_log_var = self.encoder.outputs[1]
outputs = self.decoder(self.encoder(inputs)[2]) # This is z.
# Define the loss.
def vae_loss(loss_inputs, loss_outputs):
# Flatten all to accept different dimensions.
loss_inputs = K.flatten(loss_inputs)
loss_outputs = K.flatten(loss_outputs)
# Reconstruction loss.
if isinstance(self.loss, str):
r_loss = losses.get(self.loss)
else:
r_loss = self.loss
r_loss *= inputs_dim
# kl loss.
kl_loss = 1 + z_log_var - K.square(z_mean) - K.exp(z_log_var)
kl_loss = K.sum(kl_loss, axis=-1)
kl_loss *= -0.5
# VAE loss.
vae_loss = K.mean(r_loss + kl_loss)
vae_loss /= inputs_dim
return vae_loss
# Compile model.
loss = vae_loss
self.autoencoder.compile(optimizer, loss, metrics, loss_weights,
sample_weight_mode, weighted_metrics,
**kwargs)
def predict_embed_samples_into_latent(self,
x,
batch_size=None,
verbose=0,
steps=None):
return self.encoder.predict(x, batch_size, verbose, steps)[2]
def test_model_with_input_feed_tensor():
"""We test building a model with a TF variable as input.
We should be able to call fit, evaluate, predict,
by only passing them data for the placeholder inputs
in the model.
"""
import tensorflow as tf
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
b = Input(shape=(3, ), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
model = Model([a, b], [a_2, b_2])
model.summary()
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
model.compile(optimizer,
loss,
metrics=['mean_squared_error'],
loss_weights=loss_weights,
sample_weight_mode=None)
# test train_on_batch
out = model.train_on_batch(input_b_np, [output_a_np, output_b_np])
out = model.train_on_batch({'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.test_on_batch({'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.predict_on_batch({'input_b': input_b_np})
# test fit
out = model.fit({'input_b': input_b_np}, [output_a_np, output_b_np],
epochs=1,
batch_size=10)
out = model.fit(input_b_np, [output_a_np, output_b_np],
epochs=1,
batch_size=10)
# test evaluate
out = model.evaluate({'input_b': input_b_np}, [output_a_np, output_b_np],
batch_size=10)
out = model.evaluate(input_b_np, [output_a_np, output_b_np], batch_size=10)
# test predict
out = model.predict({'input_b': input_b_np}, batch_size=10)
out = model.predict(input_b_np, batch_size=10)
assert len(out) == 2
# Now test a model with a single input
# i.e. we don't pass any data to fit the model.
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
a_2 = Dense(4, name='dense_1')(a)
a_2 = Dropout(0.5, name='dropout')(a_2)
model = Model(a, a_2)
model.summary()
optimizer = 'rmsprop'
loss = 'mse'
model.compile(optimizer, loss, metrics=['mean_squared_error'])
# test train_on_batch
out = model.train_on_batch(None, output_a_np)
out = model.train_on_batch(None, output_a_np)
out = model.test_on_batch(None, output_a_np)
out = model.predict_on_batch(None)
out = model.train_on_batch([], output_a_np)
out = model.train_on_batch({}, output_a_np)
# test fit
out = model.fit(None, output_a_np, epochs=1, batch_size=10)
out = model.fit(None, output_a_np, epochs=1, batch_size=10)
# test evaluate
out = model.evaluate(None, output_a_np, batch_size=10)
out = model.evaluate(None, output_a_np, batch_size=10)
# test predict
out = model.predict(None, steps=3)
out = model.predict(None, steps=3)
assert out.shape == (10 * 3, 4)
# Same, without learning phase
# i.e. we don't pass any data to fit the model.
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
a_2 = Dense(4, name='dense_1')(a)
model = Model(a, a_2)
model.summary()
optimizer = 'rmsprop'
loss = 'mse'
model.compile(optimizer, loss, metrics=['mean_squared_error'])
# test train_on_batch
out = model.train_on_batch(None, output_a_np)
out = model.train_on_batch(None, output_a_np)
out = model.test_on_batch(None, output_a_np)
out = model.predict_on_batch(None)
out = model.train_on_batch([], output_a_np)
out = model.train_on_batch({}, output_a_np)
# test fit
out = model.fit(None, output_a_np, epochs=1, batch_size=10)
out = model.fit(None, output_a_np, epochs=1, batch_size=10)
# test evaluate
out = model.evaluate(None, output_a_np, batch_size=10)
out = model.evaluate(None, output_a_np, batch_size=10)
# test predict
out = model.predict(None, steps=3)
out = model.predict(None, steps=3)
assert out.shape == (10 * 3, 4)
class TL(Model):
"""
Triplet-Loss trained Neural Network.
https://arxiv.org/abs/1503.03832
"""
def __init__(self, base=None, siamese=None):
super(TL, self).__init__()
# Store the base model.
assert (base != None)
self.base = base
# For loading.
if base != None and siamese != None:
self.base = base
self.siamese = siamese
self.latent_dim = self.base.outputs[0].shape[1]
return
# Get the latent dimension.
assert len(self.base.outputs) == 1
assert len(self.base.outputs[0].shape) == 2
self.latent_dim = self.base.outputs[0].shape[1]
# Get the input shape.
input_shape = self.base.inputs[0].shape.as_list()[1:]
# Create the anchor.
input_anchor = layers.Input(shape=input_shape)
output_anchor = input_anchor
output_anchor = self.base(output_anchor)
# Create the positive.
input_positive = layers.Input(shape=input_shape)
output_positive = input_positive
output_positive = self.base(output_positive)
# Create the negative.
input_negative = layers.Input(shape=input_shape)
output_negative = input_negative
output_negative = self.base(output_negative)
# Create a dummy output.
output = layers.concatenate(
[output_anchor, output_positive, output_negative])
# Create the model.
self.siamese = Model([input_anchor, input_positive, input_negative],
output,
name="triplet_model")
def compile(self,
optimizer,
loss=None,
metrics=None,
loss_weights=None,
sample_weight_mode=None,
weighted_metrics=None,
target_tensors=None,
triplet_loss="mse",
**kwargs):
"""
Compiles the TL.
Additionally to the default functionality of *compile*, it adds the triplet-loss.
In order to do so you have to provide it via the parameter *triplet_loss*.
The VAE loss is similar to
>>> vae_loss = max(0.0, pos_dist - neg_dist + alpha)
See the literature for details.
Additional args:
triplet_loss (string): The base-loss for the triplet-loss. Values are either *euclidean* for euclidean norm or *cosine* for cosine similarity.
"""
assert loss == None, "Not expected to provide an explicit loss for TL. Use 'triplet_loss'"
self.triplet_loss = triplet_loss
def triplet_loss_function(y_true, y_pred, alpha=0.4):
anchor = y_pred[:, 0:self.latent_dim]
positive = y_pred[:, self.latent_dim:self.latent_dim * 2]
negative = y_pred[:, self.latent_dim * 2:self.latent_dim * 3]
if triplet_loss == "euclidean":
pos_dist = euclidean_loss(positive, anchor)
neg_dist = euclidean_loss(negative, anchor)
elif triplet_loss == "cosine":
pos_dist = cosine_loss(positive, anchor)
neg_dist = cosine_loss(negative, anchor)
else:
raise Exception("Unexpected: " + triplet_loss)
basic_loss = pos_dist - neg_dist + alpha
loss = K.maximum(basic_loss, 0.0)
return loss
loss = triplet_loss_function
self.siamese.compile(optimizer, loss, metrics, loss_weights,
sample_weight_mode, weighted_metrics, **kwargs)
def fit(self,
x=None,
y=None,
batch_size=None,
minibatch_size=None,
epochs=1,
verbose=1,
callbacks=None,
validation_split=0.,
validation_data=None,
shuffle=True,
class_weight=None,
sample_weight=None,
initial_epoch=0,
steps_per_epoch=None,
validation_steps=None,
**kwargs):
"""
This is basically the same as in vanilla Keras.
Additional args:
minibatch_size (int): The model internally does some sampling. The *minibatch_size* specifies how many candidates to use in order to create a triplet for training.
"""
assert minibatch_size != None, "ERROR! Must provide 'minibatch_size'."
assert steps_per_epoch != None, "ERROR! Must provide 'steps_per_epoch'."
assert validation_steps != None, "ERROR! Must provide 'validation_steps'."
y_dummy = np.zeros((batch_size, self.latent_dim * 3))
# Template generator.
def triplet_loss_generator(x_generator, y_generator, model, sampling):
# Get the classes.
classes = sorted(list(set(y_generator)))
# Sort by classes for easy indexing.
class_indices = {}
for c in classes:
class_indices[c] = []
for index, c in enumerate(y_generator):
class_indices[c].append(index)
# Compute the complements.
class_complements = {}
for c in classes:
class_complements[c] = [c2 for c2 in classes if c2 != c]
# Generator loop.
while True:
x_input_anchors = []
x_input_positives = []
x_input_negatives = []
# Generate a whole batch.
for _ in range(batch_size):
anchor_class = random.choice(classes)
anchor_index = random.choice(class_indices[anchor_class])
anchor_input = x_generator[anchor_index]
#print("anchor_class", anchor_class)
anchor_latent = self.base.predict(
np.expand_dims(anchor_input, axis=0))[0]
# Generate some positive candidates.
positive_candidates = []
while len(positive_candidates) < minibatch_size:
positive_class = anchor_class
positive_index = random.choice(
class_indices[positive_class])
positive_input = x_generator[positive_index]
assert positive_class == y_generator[positive_index]
#print("positive_class", positive_class)
positive_candidates.append(positive_input)
# Find the farthest positive candidate.
positive_candidates = np.array(positive_candidates)
positive_latents = self.base.predict(positive_candidates)
positive_extremum = compute_latent_extremum(
anchor_latent, positive_latents, "argmax",
self.triplet_loss)
positive_input = positive_candidates[positive_extremum]
# Generate some negative candidates.
negative_candidates = []
while len(negative_candidates) < minibatch_size:
negative_class = random.choice(
class_complements[anchor_class])
negative_index = random.choice(
class_indices[negative_class])
negative_input = x_generator[negative_index]
assert negative_class == y_generator[negative_index]
#print("negative_class", negative_class)
negative_candidates.append(negative_input)
# Find the closest negative candidate.
negative_candidates = np.array(negative_candidates)
negative_latents = self.base.predict(negative_candidates)
negative_extremum = compute_latent_extremum(
anchor_latent, negative_latents, "argmin",
self.triplet_loss)
negative_input = negative_candidates[negative_extremum]
# Done.
x_input_anchors.append(anchor_input)
x_input_positives.append(positive_input)
x_input_negatives.append(negative_input)
x_input_anchors = np.array(x_input_anchors)
x_input_positives = np.array(x_input_positives)
x_input_negatives = np.array(x_input_negatives)
x_input = [
x_input_anchors, x_input_positives, x_input_negatives
]
yield x_input, y_dummy
# Create the generators.
training_generator = triplet_loss_generator(x, y, batch_size,
self.siamese)
if validation_data != None:
validation_generator = triplet_loss_generator(
validation_data[0], validation_data[1], batch_size,
self.siamese)
else:
validation_generator = None
# Create the history.
history_keys = ["loss", "val_loss"]
history = {}
for history_key in history_keys:
history[history_key] = []
# Training the model
for epoch in range(epochs):
print("Epoch " + str(epoch + 1) + "/" + str(epochs) + "...")
# Generating data for training.
training_input, training_output = next(training_generator)
if validation_generator != None:
validation_input, validation_output = next(
validation_generator)
model_history = self.siamese.fit(
training_input,
training_output,
validation_data=(validation_input, validation_output),
epochs=1,
steps_per_epoch=steps_per_epoch,
verbose=0,
validation_steps=validation_steps)
# Update the history.
for history_key in history_keys:
history_value = model_history.history[history_key]
history[history_key].append(history_value)
print(history_key, history_value)
return history
def fit_generator(self,
generator,
steps_per_epoch=None,
epochs=1,
verbose=1,
callbacks=None,
validation_data=None,
validation_steps=None,
class_weight=None,
max_queue_size=10,
workers=1,
use_multiprocessing=False,
shuffle=True,
initial_epoch=0):
"""
Coming soon...
"""
print("TODO: implement fit_generator!")
raise Exception("Not implemented!")
return self.siamese.fit_generator(generator, steps_per_epoch, epochs,
verbose, callbacks, validation_data,
validation_steps, class_weight,
max_queue_size, workers,
use_multiprocessing, shuffle,
initial_epoch)
def evaluate(self,
x=None,
y=None,
batch_size=None,
verbose=1,
sample_weight=None,
steps=None):
"""
Evaluates the model. Same as vanilla Keras.
"""
return self.siamese.evaluate(x,
y,
batch_size,
verbose,
sample_weight,
steps=None)
def predict(self, x, batch_size=None, verbose=0, steps=None):
"""
Does a prediction. Same as vanilla Keras.
"""
return self.siamese.predict(x, batch_size, verbose, steps)
def summary(self):
"""
Provides a summary.
"""
print("Basemodel:")
self.base.summary()
print("Siamese model:")
self.siamese.summary()
def save(self, path):
"""
Saves the TL.
This includes the whole Siamese Net plus the base-model.
This code
>>> tl.save("myae.h5")
will create the files *tl.h5*, and *tl-base.h5*.
"""
self.siamese.save(path)
self.base.save(append_to_filepath(path, "-base"))
cnn = Convolution1D(filters=50, kernel_size=3, activation='tanh')(dropouted)
cnn = Convolution1D(filters=50, kernel_size=3, activation='tanh')(cnn)
flattened = Flatten()(cnn)
dense = Dense(100, activation='tanh')(flattened)
predict = Dense(2, activation='softmax')(dense)
model = Model(input=[word, distance_e1, distance_e2], output=predict)
# opt = RMSprop(lr=0.001, rho=0.9, epsilon=1e-06)
# opt = Adagrad(lr=0.01, epsilon=1e-06)
# opt = Adadelta(lr=1.0, rho=0.95, epsilon=1e-06)
# opt = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08)
opt = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='categorical_crossentropy', optimizer=opt)
train_instances = [line.strip() for line in lines]
label_array_t, word_array_t, dis_e1_array_t, dis_e2_array_t = rep.represent_instances(
train_instances)
model.fit([word_array_t, dis_e1_array_t, dis_e2_array_t],
label_array_t,
batch_size=128,
epochs=epoch_size)
model.save(output_file)
label_array_ans = model.predict([word_array_t, dis_e1_array_t, dis_e2_array_t],
batch_size=128)
print(label_array_ans)
print("训练完成！！")
eval_mulclass(label_array_t, label_array_ans)
class GAN(Model):
""" Generative Adversarial Network (GAN). """
def __init__(self, generator, discriminator):
super(GAN, self).__init__()
assert generator != None
assert discriminator != None
assert discriminator.optimizer != None, "Discriminator must be compiled!"
self.generator = generator
self.discriminator = discriminator
# Create the GAN.
z_shape = generator.inputs[0].shape[1:]
gan_input = layers.Input(shape=z_shape)
gan_output = gan_input
gan_output = self.generator(gan_output)
self.discriminator.trainable = False
gan_output = self.discriminator(gan_output)
self.gan = Model(gan_input, gan_output)
def compile(self,
optimizer,
loss=None,
metrics=None,
loss_weights=None,
sample_weight_mode=None,
weighted_metrics=None,
target_tensors=None,
**kwargs):
"""
Compiles the model. Same as vanilla Keras.
"""
self.gan.compile(optimizer, loss, metrics, loss_weights,
sample_weight_mode, weighted_metrics, **kwargs)
def fit(
self,
x=None,
y=None,
batch_size=None,
epochs=1,
sample_interval=None, # TODO document!
verbose=1,
callbacks=None,
validation_split=0.,
validation_data=None,
shuffle=True,
class_weight=None,
sample_weight=None,
initial_epoch=0,
steps_per_epoch=None,
validation_steps=None,
**kwargs):
"""
Trains the GAN.
This is almost the same as in vanilla Keras.
"""
# Adversarial ground truths
valid = np.ones((batch_size, 1))
fake = np.zeros((batch_size, 1))
for epoch in range(epochs):
# Select a random batch of images
idx = np.random.randint(0, x.shape[0], batch_size)
imgs = x[idx]
# Create some noise.
noise = np.random.normal(0, 1, (batch_size, 100))
# Generate a batch of new images.
gen_imgs = self.generator.predict(noise)
# Train the discriminator
d_loss_real = self.discriminator.train_on_batch(imgs, valid)
d_loss_fake = self.discriminator.train_on_batch(gen_imgs, fake)
d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)
# Create some noise.
noise = np.random.normal(0, 1, (batch_size, 100))
# Train the generator (to have the discriminator label samples as valid).
g_loss = self.gan.train_on_batch(noise, valid)
if type(g_loss) == list:
g_loss = g_loss[0]
# Plot the progress.
print("%d [D loss: %f, acc.: %.2f%%] [G loss: %f]" %
(epoch, d_loss[0], 100 * d_loss[1], g_loss),
end="\r")
# If at save interval => save generated image samples
if sample_interval != None and epoch % sample_interval == 0:
self.sample_images(epoch)
def sample_images(self, epoch):
"""
Samples images.
"""
r, c = 5, 5
noise = np.random.normal(0, 1, (r * c, 100))
gen_imgs = self.generator.predict(noise)
# Rescale images 0 - 1
gen_imgs = 0.5 * gen_imgs + 0.5
fig, axs = plt.subplots(r, c)
cnt = 0
for i in range(r):
for j in range(c):
axs[i, j].imshow(gen_imgs[cnt, :, :, 0], cmap='gray')
axs[i, j].axis('off')
cnt += 1
#fig.savefig("images/%d.png" % epoch)
plt.show()
plt.close()
def summary(self):
"""
Provides a summary.
"""
print("Generator:")
self.generator.summary()
print("Discriminator:")
self.discriminator.summary()
print("GAN:")
self.gan.summary()
def save(self, path):
"""
Saves the GAN.
This includes the whole autoencoder plus the encoder and the decoder.
The encoder and decoder use the path plus a respective annotation.
This code
>>> ae.save("myae.h5")
will create the files *myae.h5*, *myae-encoder.h5*, and *myae-decoder.h5*.
"""
self.gan.save(path)
self.generator.save(append_to_filepath(path, "-generator"))
self.discriminator.save(append_to_filepath(path, "-discriminator"))
class AIPlayer(Player):
def __init__(self, buffer_size, sim_count, train=True, model="", tau = 1, compile=False):
self.buffer = ReplayBuffer(buffer_size)
self.temp_state = deque()
self.train = train
self.loss = 0
self.acc = 0
self.batch_count = 0
self.sim_count = sim_count
if model != "":
self.load(model, compile)
else:
self.create_network()
self.tau = tau
@staticmethod
def create_if_nonexistant(config):
models = glob.glob(config.data.model_location + "*.h5")
if len(models) == 0:
ai = AIPlayer(config.buffer_size, config.game.simulation_num_per_move)
ai.save(config.data.model_location+"model_0.h5")
del ai
def set_training(self, train):
self.train = train
@staticmethod
def clear():
K.clear_session()
def load(self, file, compile=False):
try:
del self.network
except Exception:
pass
self.network = load_model(file, custom_objects={"objective_function_for_policy":AIPlayer.objective_function_for_policy,
"objective_function_for_value":AIPlayer.objective_function_for_value}, compile=compile)
def save(self, file):
self.network.save(file)
def create_network(self):
x_in = Input((3, 8, 8))
x = Conv2D(filters=128, kernel_size=(3,3), padding="same", data_format="channels_first")(x_in)
x = BatchNormalization(axis=1)(x)
x = Activation("relu")(x)
for _ in range(10):
x = self._build_residual_block(x)
res_out = x
x = Conv2D(filters=2, kernel_size=1, data_format="channels_first")(res_out)
x = BatchNormalization(axis=1)(x)
x = Activation("relu")(x)
x = Flatten()(x)
policy_out = Dense(8*8+1, activation="softmax", name="policy_out")(x)
x = Conv2D(filters=1, kernel_size=1, data_format="channels_first")(res_out)
x = BatchNormalization(axis=1)(x)
x = Activation("relu")(x)
x = Flatten()(x)
x = Dense(64, activation="relu")(x)
value_out = Dense(1, activation="tanh", name="value_out")(x)
self.network = Model(x_in, [policy_out, value_out], name="reversi_model")
self.compile()
def _build_residual_block(self, x):
in_x = x
x = Conv2D(filters=128, kernel_size=(3,3), padding="same", data_format="channels_first")(x)
x = BatchNormalization(axis=1)(x)
x = Activation("relu")(x)
x = Conv2D(filters=128, kernel_size=(3,3), padding="same", data_format="channels_first")(x)
x = BatchNormalization(axis=1)(x)
x = Add()([in_x, x])
x = Activation("relu")(x)
return x
def compile(self):
losses = [AIPlayer.objective_function_for_policy, AIPlayer.objective_function_for_value]
self.network.compile(optimizer=optimizers.SGD(lr=1e-3, momentum=0.9), loss=losses)
def update_lr(self, lr):
K.set_value(self.network.optimizer.lr, lr)
@staticmethod
def objective_function_for_policy(y_true, y_pred):
# can use categorical_crossentropy??
return K.sum(-y_true * K.log(y_pred + K.epsilon()), axis=-1)
@staticmethod
def objective_function_for_value(y_true, y_pred):
return mean_squared_error(y_true, y_pred)
def update_buffer(self, winner):
if self.train:
while len(self.temp_state) > 0:
t = self.temp_state.pop()
self.buffer.add((t[0], t[1], winner))
def train_batches(self, batch_size, batches=-1, verbose=2):
if batches == -1:
s_buffer = np.array([_[0] for _ in self.buffer.buffer])
p_buffer = np.array([_[1] for _ in self.buffer.buffer])
v_buffer = np.array([_[2] for _ in self.buffer.buffer])
else:
sample_size = batch_size*batches
sample = []
while sample_size > 0:
sample += self.buffer.sample(sample_size)
sample_size -= self.buffer.size()
s_buffer = np.array([_[0] for _ in sample])
p_buffer = np.array([_[1] for _ in sample])
v_buffer = np.array([_[2] for _ in sample])
history = self.network.fit(s_buffer, [p_buffer, v_buffer], batch_size=batch_size, epochs=1, verbose=verbose)
return history
def preprocess_input(self, board, side):
state = np.zeros((3, 8, 8), dtype=np.int)
for i in range(8):
for j in range(8):
if board[i,j] == 1:
state[0,i,j] = 1
elif board[i,j] == -1:
state[1,i,j] = 1
if side == 1:
state[2,i,j] = 1
return state
def evaluate(self, game, side):
current_input = self.preprocess_input(game.board, side)
pred = self.network.predict(current_input[np.newaxis,:])
return pred[1][0]
def pick_move(self, game, side):
possible_moves = game.possible_moves(side)
if len(possible_moves) == 0:
possible_moves.append((-1,-1))
monte_prob = self.monte_carlo(game, side)
if self.train:
self.temp_state.append((self.preprocess_input(game.board, side), np.divide(monte_prob, np.sum(monte_prob))))
monte_prob = np.float_power(monte_prob, 1/self.tau)
monte_prob = np.divide(monte_prob, np.sum(monte_prob))
r = random()
for i, move in enumerate(possible_moves):
r -= monte_prob[Othello.move_id(move)]
if r <= 0:
return move
return possible_moves[-1]
def monte_carlo(self, game, side):
N = defaultdict(lambda: 0)
W = defaultdict(lambda: 0)
Q = defaultdict(lambda: 0)
P = defaultdict(lambda: 0)
possible_moves = game.possible_moves(side)
if len(possible_moves) == 0:
policy = np.zeros((65))
policy[64] = 1
return policy
elif len(possible_moves) == 1:
policy = np.zeros((65))
policy[Othello.move_id(possible_moves[0])] = 1
return policy
current_input = self.preprocess_input(game.board, side)
sid = Othello.state_id(game.board)
pred = self.network.predict(current_input[np.newaxis,:])
policy = pred[0][0]
total = 1e-10
for i, move in enumerate(possible_moves):
total += policy[Othello.move_id(move)]
for move in possible_moves:
P[(sid, Othello.move_id(move))] = policy[Othello.move_id(move)]/total
for i in range(self.sim_count):
#print("Sim #%d"% i)
clone = deepcopy(game)
current_side = side
visited = deque()
while True:
possible_moves = clone.possible_moves(current_side)
if len(possible_moves) == 0:
possible_moves.append((-1,-1))
best_move = None
best_move_value = -2
sid = Othello.state_id(clone.board)
for move in possible_moves:
mid = Othello.move_id(move)
qu_val = Q[(sid, mid)] + P[(sid, mid)]/(N[(sid, mid)]+1)
if qu_val > best_move_value:
best_move_value = qu_val
best_move = move
#print(best_move)
if N[(sid, Othello.move_id(best_move))] == 0:
visited.append((sid, Othello.move_id(best_move)))
clone.play_move(best_move[0], best_move[1], current_side)
current_side *= -1
if clone.game_over():
for node in visited:
N[node] += 1
W[node] += clone.get_winner()*side
Q[node] = W[node]/N[node]
break
current_input = self.preprocess_input(clone.board, current_side)
sid = Othello.state_id(clone.board)
pred = self.network.predict(current_input[np.newaxis,:])
policy = pred[0][0]
value = pred[1][0]
possible_moves = clone.possible_moves(current_side)
if len(possible_moves) == 0:
possible_moves.append((-1,-1))
total = 1e-10
for i, move in enumerate(possible_moves):
total += policy[Othello.move_id(move)]
for move in possible_moves:
P[(sid, Othello.move_id(move))] = policy[Othello.move_id(move)]/total
for node in visited:
N[node] += 1
W[node] += value*side
Q[node] = W[node]/N[node]
#print()
break
else:
visited.append((sid, Othello.move_id(best_move)))
clone.play_move(best_move[0], best_move[1], current_side)
current_side *= -1
if clone.game_over():
for node in visited:
N[node] += 1
W[node] += clone.get_winner()*side
Q[node] = W[node]/N[node]
break
policy = np.zeros((65))
possible_moves = game.possible_moves(side)
sid = Othello.state_id(game.board)
for move in possible_moves:
mid = Othello.move_id(move)
policy[mid] = N[(sid,mid)]
return policy
class AdditionNPIModel(NPIStep):
model = None
f_enc = None
def __init__(self,
system: RuntimeSystem,
model_path: str = None,
program_set: AdditionProgramSet = None):
self.system = system
self.model_path = model_path
self.program_set = program_set
self.batch_size = 1
self.build()
self.weight_loaded = False
self.load_weights()
def build(self):
enc_size = self.size_of_env_observation()
argument_size = IntegerArguments.size_of_arguments
input_enc = InputLayer(batch_input_shape=(self.batch_size, enc_size),
name='input_enc')
input_arg = InputLayer(batch_input_shape=(self.batch_size,
argument_size),
name='input_arg')
input_prg = Embedding(input_dim=PROGRAM_VEC_SIZE,
output_dim=PROGRAM_KEY_VEC_SIZE,
input_length=1,
batch_input_shape=(self.batch_size, 1))
f_enc = Sequential(name='f_enc')
f_enc.add(Merge([input_enc, input_arg], mode='concat'))
f_enc.add(MaxoutDense(128, nb_feature=4))
self.f_enc = f_enc
program_embedding = Sequential(name='program_embedding')
program_embedding.add(input_prg)
f_enc_convert = Sequential(name='f_enc_convert')
f_enc_convert.add(f_enc)
f_enc_convert.add(RepeatVector(1))
f_lstm = Sequential(name='f_lstm')
f_lstm.add(Merge([f_enc_convert, program_embedding], mode='concat'))
f_lstm.add(
LSTM(256,
return_sequences=False,
stateful=True,
W_regularizer=l2(0.0000001)))
f_lstm.add(Activation('relu', name='relu_lstm_1'))
f_lstm.add(RepeatVector(1))
f_lstm.add(
LSTM(256,
return_sequences=False,
stateful=True,
W_regularizer=l2(0.0000001)))
f_lstm.add(Activation('relu', name='relu_lstm_2'))
# plot(f_lstm, to_file='f_lstm.png', show_shapes=True)
f_end = Sequential(name='f_end')
f_end.add(f_lstm)
f_end.add(Dense(1, W_regularizer=l2(0.001)))
f_end.add(Activation('sigmoid', name='sigmoid_end'))
f_prog = Sequential(name='f_prog')
f_prog.add(f_lstm)
f_prog.add(Dense(PROGRAM_KEY_VEC_SIZE, activation="relu"))
f_prog.add(Dense(PROGRAM_VEC_SIZE, W_regularizer=l2(0.0001)))
f_prog.add(Activation('softmax', name='softmax_prog'))
# plot(f_prog, to_file='f_prog.png', show_shapes=True)
f_args = []
for ai in range(1, IntegerArguments.max_arg_num + 1):
f_arg = Sequential(name='f_arg%s' % ai)
f_arg.add(f_lstm)
f_arg.add(Dense(IntegerArguments.depth, W_regularizer=l2(0.0001)))
f_arg.add(Activation('softmax', name='softmax_arg%s' % ai))
f_args.append(f_arg)
# plot(f_arg, to_file='f_arg.png', show_shapes=True)
self.model = Model([input_enc.input, input_arg.input, input_prg.input],
[f_end.output, f_prog.output] +
[fa.output for fa in f_args],
name="npi")
self.compile_model()
plot(self.model, to_file='model.png', show_shapes=True)
def reset(self):
super(AdditionNPIModel, self).reset()
for l in self.model.layers:
if type(l) is LSTM:
l.reset_states()
def compile_model(self, lr=0.0001, arg_weight=1.):
arg_num = IntegerArguments.max_arg_num
optimizer = Adam(lr=lr)
loss = ['binary_crossentropy', 'categorical_crossentropy'
] + ['categorical_crossentropy'] * arg_num
self.model.compile(optimizer=optimizer,
loss=loss,
loss_weights=[0.25, 0.25] + [arg_weight] * arg_num)
def fit(self, steps_list, epoch=3000):
# 过滤一些问题
def filter_question(condition_func):
sub_steps_list = []
for steps_dict in steps_list:
question = steps_dict['q']
if condition_func(question['in1'], question['in2']):
sub_steps_list.append(steps_dict)
return sub_steps_list
if not self.weight_loaded:
self.train_f_enc(
filter_question(lambda a, b: 10 <= a < 100 and 10 <= b < 100),
epoch=100)
self.f_enc.trainable = False
self.update_learning_rate(0.0001)
q_type = "training questions of a<100 and b<100"
print(q_type)
pr = 0.8
all_ok = self.fit_to_subset(
filter_question(lambda a, b: a < 100 and b < 100), pass_rate=pr)
print("%s is pass_rate >= %s: %s" % (q_type, pr, all_ok))
while True:
if self.test_and_learn([10, 100, 1000]):
break
q_type = "training questions of ALL"
print(q_type)
q_num = 100
skip_correct = False
pr = 1.0
questions = filter_question(lambda a, b: True)
np.random.shuffle(questions)
questions = questions[:q_num]
all_ok = self.fit_to_subset(questions,
pass_rate=pr,
skip_correct=skip_correct)
print("%s is pass_rate >= %s: %s" % (q_type, pr, all_ok))
def fit_to_subset(self, steps_list, pass_rate=1.0, skip_correct=False):
for i in range(10):
all_ok = self.do_learn(steps_list,
100,
pass_rate=pass_rate,
skip_correct=skip_correct)
if all_ok:
return True
return False
def test_and_learn(self, num_questions):
for num in num_questions:
print("test all type of %d questions" % num)
cc, wc, wrong_questions = self.test_to_subset(
create_random_questions(num))
acc_rate = cc / (cc + wc)
print("Accuracy %s(OK=%d, NG=%d)" % (acc_rate, cc, wc))
if wc > 0:
self.fit_to_subset(wrong_questions,
pass_rate=1.0,
skip_correct=False)
return False
return True
def test_to_subset(self, questions):
addition_env = AdditionEnv(FIELD_ROW, FIELD_WIDTH, FIELD_DEPTH)
teacher = AdditionTeacher(self.program_set)
npi_runner = TerminalNPIRunner(None, self)
teacher_runner = TerminalNPIRunner(None, teacher)
correct_count = wrong_count = 0
wrong_steps_list = []
for idx, question in enumerate(questions):
question = copy(question)
if self.question_test(addition_env, npi_runner, question):
correct_count += 1
else:
self.question_test(addition_env, teacher_runner, question)
wrong_steps_list.append({
"q": question,
"steps": teacher_runner.step_list
})
wrong_count += 1
return correct_count, wrong_count, wrong_steps_list
@staticmethod
def dict_to_str(d):
return str(tuple([(k, d[k]) for k in sorted(d)]))
def do_learn(self, steps_list, epoch, pass_rate=1.0, skip_correct=False):
addition_env = AdditionEnv(FIELD_ROW, FIELD_WIDTH, FIELD_DEPTH)
npi_runner = TerminalNPIRunner(None, self)
last_weights = None
correct_count = Counter()
no_change_count = 0
last_loss = 1000
for ep in range(1, epoch + 1):
correct_new = wrong_new = 0
losses = []
ok_rate = []
np.random.shuffle(steps_list)
for idx, steps_dict in enumerate(steps_list):
question = copy(steps_dict['q'])
question_key = self.dict_to_str(question)
if self.question_test(addition_env, npi_runner, question):
if correct_count[question_key] == 0:
correct_new += 1
correct_count[question_key] += 1
print("GOOD!: ep=%2d idx=%3d :%s CorrectCount=%s" %
(ep, idx, self.dict_to_str(question),
correct_count[question_key]))
ok_rate.append(1)
cc = correct_count[question_key]
if skip_correct or int(math.sqrt(cc))**2 != cc:
continue
else:
ok_rate.append(0)
if correct_count[question_key] > 0:
print(
"Degraded: ep=%2d idx=%3d :%s CorrectCount=%s -> 0"
% (ep, idx, self.dict_to_str(question),
correct_count[question_key]))
correct_count[question_key] = 0
wrong_new += 1
steps = steps_dict['steps']
xs = []
ys = []
ws = []
for step in steps:
xs.append(self.convert_input(step.input))
y, w = self.convert_output(step.output)
ys.append(y)
ws.append(w)
self.reset()
for i, (x, y, w) in enumerate(zip(xs, ys, ws)):
loss = self.model.train_on_batch(x, y, sample_weight=w)
if not np.isfinite(loss):
print("Loss is not finite!, Last Input=%s" %
([i, (x, y, w)]))
self.print_weights(last_weights, detail=True)
raise RuntimeError("Loss is not finite!")
losses.append(loss)
last_weights = self.model.get_weights()
if losses:
cur_loss = np.average(losses)
print(
"ep=%2d: ok_rate=%.2f%% (+%s -%s): ave loss %s (%s samples)"
% (ep, np.average(ok_rate) * 100, correct_new, wrong_new,
cur_loss, len(steps_list)))
# self.print_weights()
if correct_new + wrong_new == 0:
no_change_count += 1
else:
no_change_count = 0
if math.fabs(1 - cur_loss /
last_loss) < 0.001 and no_change_count > 5:
print(
"math.fabs(1 - cur_loss/last_loss) < 0.001 and no_change_count > 5:"
)
return False
last_loss = cur_loss
print("=" * 80)
self.save()
if np.average(ok_rate) >= pass_rate:
return True
return False
def update_learning_rate(self, learning_rate, arg_weight=1.):
print("Re-Compile Model lr=%s aw=%s" % (learning_rate, arg_weight))
self.compile_model(learning_rate, arg_weight=arg_weight)
def train_f_enc(self, steps_list, epoch=50):
print("training f_enc")
f_add0 = Sequential(name='f_add0')
f_add0.add(self.f_enc)
f_add0.add(Dense(FIELD_DEPTH))
f_add0.add(Activation('softmax', name='softmax_add0'))
f_add1 = Sequential(name='f_add1')
f_add1.add(self.f_enc)
f_add1.add(Dense(FIELD_DEPTH))
f_add1.add(Activation('softmax', name='softmax_add1'))
env_model = Model(self.f_enc.inputs, [f_add0.output, f_add1.output],
name="env_model")
env_model.compile(optimizer='adam',
loss=['categorical_crossentropy'] * 2)
for ep in range(epoch):
losses = []
for idx, steps_dict in enumerate(steps_list):
prev = None
for step in steps_dict['steps']:
x = self.convert_input(step.input)[:2]
env_values = step.input.env.reshape((4, -1))
in1 = np.clip(env_values[0].argmax() - 1, 0, 9)
in2 = np.clip(env_values[1].argmax() - 1, 0, 9)
carry = np.clip(env_values[2].argmax() - 1, 0, 9)
y_num = in1 + in2 + carry
now = (in1, in2, carry)
if prev == now:
continue
prev = now
y0 = to_one_hot_array((y_num % 10) + 1, FIELD_DEPTH)
y1 = to_one_hot_array((y_num // 10) + 1, FIELD_DEPTH)
y = [yy.reshape((self.batch_size, -1)) for yy in [y0, y1]]
loss = env_model.train_on_batch(x, y)
losses.append(loss)
print("ep %3d: loss=%s" % (ep, np.average(losses)))
if np.average(losses) < 1e-06:
break
def question_test(self, addition_env, npi_runner, question):
addition_env.reset()
self.reset()
try:
run_npi(addition_env, npi_runner, self.program_set.ADD, question)
if question['correct']:
return True
except StopIteration:
pass
return False
def convert_input(self, p_in: StepInput):
x_pg = np.array((p_in.program.program_id, ))
x = [
xx.reshape((self.batch_size, -1))
for xx in (p_in.env, p_in.arguments.values, x_pg)
]
return x
def convert_output(self, p_out: StepOutput):
y = [np.array((p_out.r, ))]
weights = [[1.]]
if p_out.program:
arg_values = p_out.arguments.values
arg_num = len(p_out.program.args or [])
y += [p_out.program.to_one_hot(PROGRAM_VEC_SIZE)]
weights += [[1.]]
else:
arg_values = IntegerArguments().values
arg_num = 0
y += [np.zeros((PROGRAM_VEC_SIZE, ))]
weights += [[1e-10]]
for v in arg_values: # split by each args
y += [v]
weights += [[1.]] * arg_num + [[1e-10]] * (len(arg_values) - arg_num)
weights = [np.array(w) for w in weights]
return [yy.reshape((self.batch_size, -1)) for yy in y], weights
def step(self, env_observation: np.ndarray, pg: Program,
arguments: IntegerArguments) -> StepOutput:
x = self.convert_input(StepInput(env_observation, pg, arguments))
results = self.model.predict(
x, batch_size=1) # if batch_size==1, returns single row
r, pg_one_hot, arg_values = results[0], results[1], results[2:]
program = self.program_set.get(pg_one_hot.argmax())
ret = StepOutput(r, program,
IntegerArguments(values=np.stack(arg_values)))
return ret
def save(self):
self.model.save_weights(self.model_path, overwrite=True)
def load_weights(self):
if os.path.exists(self.model_path):
self.model.load_weights(self.model_path)
self.weight_loaded = True
def print_weights(self, weights=None, detail=False):
weights = weights or self.model.get_weights()
for w in weights:
print("w%s: sum(w)=%s, ave(w)=%s" %
(w.shape, np.sum(w), np.average(w)))
if detail:
for w in weights:
print("%s: %s" % (w.shape, w))
@staticmethod
def size_of_env_observation():
return FIELD_ROW * FIELD_DEPTH
class AIPlayer(Player):
def __init__(self,
buffer_size,
sim_count,
train=True,
model="",
tau=1,
compile=False):
self.buffer = ReplayBuffer(buffer_size)
self.temp_state = deque()
self.train = train
self.loss = 0
self.acc = 0
self.batch_count = 0
self.sim_count = sim_count
if model != "":
self.load(model, compile)
else:
self.create_network()
self.tau = tau
@staticmethod
def create_if_nonexistant(config):
models = glob.glob(config.data.model_location + "*.h5")
if len(models) == 0:
ai = AIPlayer(config.buffer_size,
config.game.simulation_num_per_move)
ai.save(config.data.model_location + "model_0.h5")
del ai
def set_training(self, train):
self.train = train
@staticmethod
def clear():
K.clear_session()
def load(self, file, compile=False):
try:
del self.network
except Exception:
pass
self.network = load_model(file,
custom_objects={
"objective_function_for_policy":
AIPlayer.objective_function_for_policy,
"objective_function_for_value":
AIPlayer.objective_function_for_value
},
compile=compile)
def save(self, file):
self.network.save(file)
def create_network(self):
x_in = Input((3, 8, 8))
x = Conv2D(filters=128,
kernel_size=(3, 3),
padding="same",
data_format="channels_first")(x_in)
x = BatchNormalization(axis=1)(x)
x = Activation("relu")(x)
for _ in range(10):
x = self._build_residual_block(x)
res_out = x
x = Conv2D(filters=2, kernel_size=1,
data_format="channels_first")(res_out)
x = BatchNormalization(axis=1)(x)
x = Activation("relu")(x)
x = Flatten()(x)
policy_out = Dense(8 * 8 + 1, activation="softmax",
name="policy_out")(x)
x = Conv2D(filters=1, kernel_size=1,
data_format="channels_first")(res_out)
x = BatchNormalization(axis=1)(x)
x = Activation("relu")(x)
x = Flatten()(x)
x = Dense(64, activation="relu")(x)
value_out = Dense(1, activation="tanh", name="value_out")(x)
self.network = Model(x_in, [policy_out, value_out],
name="reversi_model")
self.compile()
def _build_residual_block(self, x):
in_x = x
x = Conv2D(filters=128,
kernel_size=(3, 3),
padding="same",
data_format="channels_first")(x)
x = BatchNormalization(axis=1)(x)
x = Activation("relu")(x)
x = Conv2D(filters=128,
kernel_size=(3, 3),
padding="same",
data_format="channels_first")(x)
x = BatchNormalization(axis=1)(x)
x = Add()([in_x, x])
x = Activation("relu")(x)
return x
def compile(self):
losses = [
AIPlayer.objective_function_for_policy,
AIPlayer.objective_function_for_value
]
self.network.compile(optimizer=optimizers.SGD(lr=1e-3, momentum=0.9),
loss=losses)
def update_lr(self, lr):
K.set_value(self.network.optimizer.lr, lr)
@staticmethod
def objective_function_for_policy(y_true, y_pred):
# can use categorical_crossentropy??
return K.sum(-y_true * K.log(y_pred + K.epsilon()), axis=-1)
@staticmethod
def objective_function_for_value(y_true, y_pred):
return mean_squared_error(y_true, y_pred)
def update_buffer(self, winner):
if self.train:
while len(self.temp_state) > 0:
t = self.temp_state.pop()
self.buffer.add((t[0], t[1], winner))
def train_batches(self, batch_size, batches=-1, verbose=2):
if batches == -1:
s_buffer = np.array([_[0] for _ in self.buffer.buffer])
p_buffer = np.array([_[1] for _ in self.buffer.buffer])
v_buffer = np.array([_[2] for _ in self.buffer.buffer])
else:
sample_size = batch_size * batches
sample = []
while sample_size > 0:
sample += self.buffer.sample(sample_size)
sample_size -= self.buffer.size()
s_buffer = np.array([_[0] for _ in sample])
p_buffer = np.array([_[1] for _ in sample])
v_buffer = np.array([_[2] for _ in sample])
history = self.network.fit(s_buffer, [p_buffer, v_buffer],
batch_size=batch_size,
epochs=1,
verbose=verbose)
return history
def preprocess_input(self, board, side):
state = np.zeros((3, 8, 8), dtype=np.int)
for i in range(8):
for j in range(8):
if board[i, j] == 1:
state[0, i, j] = 1
elif board[i, j] == -1:
state[1, i, j] = 1
if side == 1:
state[2, i, j] = 1
return state
def evaluate(self, game, side):
current_input = self.preprocess_input(game.board, side)
pred = self.network.predict(current_input[np.newaxis, :])
return pred[1][0]
def pick_move(self, game, side):
possible_moves = game.possible_moves(side)
if len(possible_moves) == 0:
possible_moves.append((-1, -1))
monte_prob = self.monte_carlo(game, side)
if self.train:
self.temp_state.append((self.preprocess_input(game.board, side),
np.divide(monte_prob, np.sum(monte_prob))))
monte_prob = np.float_power(monte_prob, 1 / self.tau)
monte_prob = np.divide(monte_prob, np.sum(monte_prob))
r = random()
for i, move in enumerate(possible_moves):
r -= monte_prob[Othello.move_id(move)]
if r <= 0:
return move
return possible_moves[-1]
def monte_carlo(self, game, side):
N = defaultdict(lambda: 0)
W = defaultdict(lambda: 0)
Q = defaultdict(lambda: 0)
P = defaultdict(lambda: 0)
possible_moves = game.possible_moves(side)
if len(possible_moves) == 0:
policy = np.zeros((65))
policy[64] = 1
return policy
elif len(possible_moves) == 1:
policy = np.zeros((65))
policy[Othello.move_id(possible_moves[0])] = 1
return policy
current_input = self.preprocess_input(game.board, side)
sid = Othello.state_id(game.board)
pred = self.network.predict(current_input[np.newaxis, :])
policy = pred[0][0]
total = 1e-10
for i, move in enumerate(possible_moves):
total += policy[Othello.move_id(move)]
for move in possible_moves:
P[(sid,
Othello.move_id(move))] = policy[Othello.move_id(move)] / total
for i in range(self.sim_count):
#print("Sim #%d"% i)
clone = deepcopy(game)
current_side = side
visited = deque()
while True:
possible_moves = clone.possible_moves(current_side)
if len(possible_moves) == 0:
possible_moves.append((-1, -1))
best_move = None
best_move_value = -2
sid = Othello.state_id(clone.board)
for move in possible_moves:
mid = Othello.move_id(move)
qu_val = Q[(sid,
mid)] + P[(sid, mid)] / (N[(sid, mid)] + 1)
if qu_val > best_move_value:
best_move_value = qu_val
best_move = move
#print(best_move)
if N[(sid, Othello.move_id(best_move))] == 0:
visited.append((sid, Othello.move_id(best_move)))
clone.play_move(best_move[0], best_move[1], current_side)
current_side *= -1
if clone.game_over():
for node in visited:
N[node] += 1
W[node] += clone.get_winner() * side
Q[node] = W[node] / N[node]
break
current_input = self.preprocess_input(
clone.board, current_side)
sid = Othello.state_id(clone.board)
pred = self.network.predict(current_input[np.newaxis, :])
policy = pred[0][0]
value = pred[1][0]
possible_moves = clone.possible_moves(current_side)
if len(possible_moves) == 0:
possible_moves.append((-1, -1))
total = 1e-10
for i, move in enumerate(possible_moves):
total += policy[Othello.move_id(move)]
for move in possible_moves:
P[(sid, Othello.move_id(move)
)] = policy[Othello.move_id(move)] / total
for node in visited:
N[node] += 1
W[node] += value * side
Q[node] = W[node] / N[node]
#print()
break
else:
visited.append((sid, Othello.move_id(best_move)))
clone.play_move(best_move[0], best_move[1], current_side)
current_side *= -1
if clone.game_over():
for node in visited:
N[node] += 1
W[node] += clone.get_winner() * side
Q[node] = W[node] / N[node]
break
policy = np.zeros((65))
possible_moves = game.possible_moves(side)
sid = Othello.state_id(game.board)
for move in possible_moves:
mid = Othello.move_id(move)
policy[mid] = N[(sid, mid)]
return policy
class FinancialTimeSeriesAnalysisModel(object):
model = None
def __init__(self, nb_time_step, dim_data, batch_size=1, model_path=None):
self.model_path = model_path
self.model_path = model_path
self.batch_size = batch_size
self.size_of_input_data_dim = dim_data
self.size_of_input_timesteps = nb_time_step
self.build()
self.weight_loaded = False
if model_path is not None:
self.load_weights()
def build(self):
dim_data = self.size_of_input_data_dim
nb_time_step = self.size_of_input_timesteps
financial_time_series_input = Input(shape=(nb_time_step, dim_data), name='x1')
lstm_layer_1 = LSTM(output_dim=nb_hidden_units, dropout_U=dropout, dropout_W=dropout,
W_regularizer=l2(l2_norm_alpha), b_regularizer=l2(l2_norm_alpha), activation='tanh',
return_sequences=True, name='lstm_layer1')
lstm_layer_21 = LSTM(output_dim=nb_hidden_units, dropout_U=dropout, dropout_W=dropout,
W_regularizer=l2(l2_norm_alpha), b_regularizer=l2(l2_norm_alpha), activation='tanh',
return_sequences=True, name='lstm_layer2_loss1')
lstm_layer_22 = LSTM(output_dim=nb_hidden_units, dropout_U=dropout, dropout_W=dropout,
W_regularizer=l2(l2_norm_alpha), b_regularizer=l2(l2_norm_alpha), activation='tanh',
return_sequences=True, name='lstm_layer2_loss2')
lstm_layer_23 = LSTM(output_dim=nb_hidden_units, dropout_U=dropout, dropout_W=dropout,
W_regularizer=l2(l2_norm_alpha), b_regularizer=l2(l2_norm_alpha), activation='tanh',
return_sequences=True, name='lstm_layer2_loss3')
lstm_layer_24 = LSTM(output_dim=nb_hidden_units, dropout_U=dropout, dropout_W=dropout,
W_regularizer=l2(l2_norm_alpha), b_regularizer=l2(l2_norm_alpha), activation='tanh',
return_sequences=True, name='lstm_layer2_loss4')
lstm_layer_25 = LSTM(output_dim=nb_hidden_units, dropout_U=dropout, dropout_W=dropout,
W_regularizer=l2(l2_norm_alpha), b_regularizer=l2(l2_norm_alpha), activation='tanh',
return_sequences=True, name='lstm_layer2_loss5')
h1 = lstm_layer_1(financial_time_series_input)
h21 = lstm_layer_21(h1)
h22 = lstm_layer_22(h1)
h23 = lstm_layer_23(h1)
h24 = lstm_layer_24(h1)
h25 = lstm_layer_25(h1)
time_series_predictions1 = TimeDistributed(Dense(1), name="p1")(h21) # custom 1
time_series_predictions2 = TimeDistributed(Dense(1), name="p2")(h22) # custom 2
time_series_predictions3 = TimeDistributed(Dense(1), name="p3")(h23) # mse
time_series_predictions4 = TimeDistributed(Dense(1, activation='sigmoid'), name="p4")(h24) # logloss
time_series_predictions5 = TimeDistributed(Dense(nb_labels, activation='softmax'), name="p5")(h25) # cross
self.model = Model(input=financial_time_series_input,
output=[time_series_predictions1, time_series_predictions2,
time_series_predictions3, time_series_predictions4,
time_series_predictions5],
name="multi-task deep rnn for financial time series forecasting")
plot(self.model, to_file='model.png')
def reset(self):
for l in self.model.layers:
if type(l) is LSTM:
l.reset_status()
def compile_model(self, lr=0.0001, arg_weight=1.):
optimizer = Adam(lr=lr)
loss = [custom_objective1, custom_objective2, 'mse', 'binary_crossentropy', 'categorical_crossentropy']
self.model.compile(optimizer=optimizer, loss=loss)
def fit_model(self, X, y, y_label, epoch=300):
early_stopping = EarlyStopping(monitor='val_loss', patience=10, verbose=0)
self.model.fit(X, [y]*3 + [y > 0] + [y_label], batch_size=self.batch_size, nb_epoch=epoch, validation_split=0.2,
shuffle=True, callbacks=[early_stopping])
def save(self):
self.model.save_weights(self.model_path, overwrite=True)
def load_weights(self):
if os.path.exists(self.model_path):
self.model.load_weights(self.model_path)
self.weight_loaded = True
def print_weights(self, weights=None, detail=False):
weights = weights or self.model.get_weights()
for w in weights:
print("w%s: sum(w)=%s, ave(w)=%s" % (w.shape, np.sum(w), np.average(w)))
if detail:
for w in weights:
print("%s: %s" % (w.shape, w))
def model_eval(self, X, y):
y_hat = self.model.predict(X, batch_size=1)[0]
count_true = 0
count_all = y.shape[1]
for i in range(y.shape[1]):
count_true = count_true + 1 if y[0,i,0]*y_hat[0,i,0]>0 else count_true
print(y[0,i,0],y_hat[0,i,0])
print(count_all,count_true)
x = Flatten()(x)
x = Dense(512, activation=None)(x)
x = BatchNormalization()(x)
x = advanced_activations.LeakyReLU(alpha=0.1)(x)
logits = Dense(num_classes, activation=None)(x)
output = Activation('softmax')(logits)
opt = keras.optimizers.Adam(lr=0.003, beta_1=0.9, beta_2=0.999, epsilon=1e-08)
model = Model(input_layer, output)
model.summary()
plot_model(model, show_shapes=True, to_file='teacher_model.png')
model.compile(loss='categorical_crossentropy',
optimizer=opt,
metrics=['accuracy'])
print('Using real-time data augmentation.')
datagen = ImageDataGenerator(
featurewise_center=False, # set input mean to 0 over the dataset
samplewise_center=False, # set each sample mean to 0
featurewise_std_normalization=False, # divide inputs by std of the dataset
samplewise_std_normalization=False, # divide each input by its std
zca_whitening=False, # apply ZCA whitening
zca_epsilon=1e-06, # epsilon for ZCA whitening
rotation_range=20, # randomly rotate images in the range (degrees, 0 to 180)
# randomly shift images horizontally (fraction of total width)
width_shift_range=0.1,
# randomly shift images vertically (fraction of total height)
def test_model_methods():
a = Input(shape=(3, ), name='input_a')
b = Input(shape=(3, ), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
model = Model([a, b], [a_2, b_2])
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
# training/testing doesn't work before compiling.
with pytest.raises(RuntimeError):
model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
model.compile(optimizer,
loss,
metrics=[],
loss_weights=loss_weights,
sample_weight_mode=None)
# test train_on_batch
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.train_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np])
out = model.train_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
})
# test fit
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
epochs=1,
batch_size=4)
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np],
epochs=1,
batch_size=4)
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
},
epochs=1,
batch_size=4)
# test validation_split
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
epochs=1,
batch_size=4,
validation_split=0.5)
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np],
epochs=1,
batch_size=4,
validation_split=0.5)
# test validation data
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
epochs=1,
batch_size=4,
validation_data=([input_a_np,
input_b_np], [output_a_np, output_b_np]))
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np],
epochs=1,
batch_size=4,
validation_split=0.5,
validation_data=({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np]))
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
},
epochs=1,
batch_size=4,
validation_split=0.5,
validation_data=({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
}))
# test_on_batch
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.test_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np])
out = model.test_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
})
# predict_on_batch
out = model.predict_on_batch([input_a_np, input_b_np])
out = model.predict_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
})
# predict, evaluate
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np],
batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
# with sample_weight
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
sample_weight = [None, np.random.random((10, ))]
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
# test accuracy metric
model.compile(optimizer, loss, metrics=['acc'], sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
# this should also work
model.compile(optimizer,
loss,
metrics={'dense_1': 'acc'},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# and this as well
model.compile(optimizer,
loss,
metrics={'dense_1': ['acc']},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# test starting from non-zero initial epoch
trained_epochs = []
trained_batches = []
# define tracer callback
def on_epoch_begin(epoch, logs):
trained_epochs.append(epoch)
def on_batch_begin(batch, logs):
trained_batches.append(batch)
tracker_cb = LambdaCallback(on_epoch_begin=on_epoch_begin,
on_batch_begin=on_batch_begin)
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
epochs=5,
batch_size=4,
initial_epoch=2,
callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test starting from non-zero initial epoch for generator too
trained_epochs = []
def gen_data(batch_sz):
while True:
yield ([
np.random.random((batch_sz, 3)),
np.random.random((batch_sz, 3))
], [
np.random.random((batch_sz, 4)),
np.random.random((batch_sz, 3))
])
out = model.fit_generator(gen_data(4),
steps_per_epoch=3,
epochs=5,
initial_epoch=2,
callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test with a custom metric function
def mse(y_true, y_pred):
return K.mean(K.pow(y_true - y_pred, 2))
model.compile(optimizer, loss, metrics=[mse], sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out_len = 1 + 2 * (1 + 1) # total loss + 2 outputs * (loss + metric)
assert len(out) == out_len
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == out_len
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
batch_size=4,
epochs=1)
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np],
batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
# empty batch
with pytest.raises(ValueError):
def gen_data():
while True:
yield (np.asarray([]), np.asarray([]))
out = model.evaluate_generator(gen_data(), steps=1)
# x is not a list of numpy arrays.
with pytest.raises(ValueError):
out = model.predict([None])
# x does not match _feed_input_names.
with pytest.raises(ValueError):
out = model.predict([input_a_np, None, input_b_np])
with pytest.raises(ValueError):
out = model.predict([None, input_a_np, input_b_np])
# all input/output/weight arrays should have the same number of samples.
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np[:2]],
[output_a_np, output_b_np],
sample_weight=sample_weight)
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np[:2]],
sample_weight=sample_weight)
with pytest.raises(ValueError):
out = model.train_on_batch(
[input_a_np, input_b_np], [output_a_np, output_b_np],
sample_weight=[sample_weight[1], sample_weight[1][:2]])
# `sample_weight` is neither a dict nor a list.
with pytest.raises(TypeError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=tuple(sample_weight))
# `validation_data` is neither a tuple nor a triple.
with pytest.raises(ValueError):
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
epochs=1,
batch_size=4,
validation_data=([input_a_np, input_b_np], ))
# `loss` does not match outputs.
with pytest.raises(ValueError):
model.compile(optimizer, loss=['mse', 'mae', 'mape'])
# `loss_weights` does not match output_names.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', loss_weights={'lstm': 0.5})
# `loss_weights` does not match outputs.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', loss_weights=[0.5])
# `loss_weights` is invalid type.
with pytest.raises(TypeError):
model.compile(optimizer, loss='mse', loss_weights=(0.5, 0.5))
# `sample_weight_mode` does not match output_names.
with pytest.raises(ValueError):
model.compile(optimizer,
loss='mse',
sample_weight_mode={'lstm': 'temporal'})
# `sample_weight_mode` does not match output_names.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', sample_weight_mode=['temporal'])
# `sample_weight_mode` matches output_names partially.
with pytest.raises(ValueError):
model.compile(optimizer,
loss='mse',
sample_weight_mode={'dense_1': 'temporal'})
# `loss` does not exist.
with pytest.raises(ValueError):
model.compile(optimizer, loss=[])
model.compile(optimizer, loss=['mse', 'mae'])
model.compile(optimizer,
loss='mse',
loss_weights={
'dense_1': 0.2,
'dropout': 0.8
})
model.compile(optimizer, loss='mse', loss_weights=[0.2, 0.8])
# the rank of weight arrays should be 1.
with pytest.raises(ValueError):
out = model.train_on_batch(
[input_a_np, input_b_np], [output_a_np, output_b_np],
sample_weight=[None, np.random.random((10, 20, 30))])
model.compile(optimizer,
loss='mse',
sample_weight_mode={
'dense_1': None,
'dropout': 'temporal'
})
model.compile(optimizer, loss='mse', sample_weight_mode=[None, 'temporal'])
# the rank of output arrays should be at least 3D.
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
model.compile(optimizer,
loss,
metrics=[],
loss_weights=loss_weights,
sample_weight_mode=None)
trained_epochs = []
trained_batches = []
out = model.fit_generator(generator=RandomSequence(3),
steps_per_epoch=3,
epochs=5,
initial_epoch=0,
validation_data=RandomSequence(4),
validation_steps=3,
callbacks=[tracker_cb])
assert trained_epochs == [0, 1, 2, 3, 4]
assert trained_batches == list(range(3)) * 5
# steps_per_epoch will be equal to len of sequence if it's unspecified
trained_epochs = []
trained_batches = []
out = model.fit_generator(generator=RandomSequence(3),
epochs=5,
initial_epoch=0,
validation_data=RandomSequence(4),
callbacks=[tracker_cb])
assert trained_epochs == [0, 1, 2, 3, 4]
assert trained_batches == list(range(12)) * 5
# fit_generator will throw an exception if steps is unspecified for regular generator
with pytest.raises(ValueError):
def gen_data():
while True:
yield (np.asarray([]), np.asarray([]))
out = model.fit_generator(generator=gen_data(),
epochs=5,
initial_epoch=0,
validation_data=gen_data(),
callbacks=[tracker_cb])
# predict_generator output shape behavior should be consistent
def expected_shape(batch_size, n_batches):
return (batch_size * n_batches, 4), (batch_size * n_batches, 3)
# Multiple outputs and one step.
batch_size = 5
sequence_length = 1
shape_0, shape_1 = expected_shape(batch_size, sequence_length)
out = model.predict_generator(
RandomSequence(batch_size, sequence_length=sequence_length))
assert np.shape(out[0]) == shape_0 and np.shape(out[1]) == shape_1
# Multiple outputs and multiple steps.
batch_size = 5
sequence_length = 2
shape_0, shape_1 = expected_shape(batch_size, sequence_length)
out = model.predict_generator(
RandomSequence(batch_size, sequence_length=sequence_length))
assert np.shape(out[0]) == shape_0 and np.shape(out[1]) == shape_1
# Create a model with a single output.
single_output_model = Model([a, b], a_2)
single_output_model.compile(optimizer,
loss,
metrics=[],
sample_weight_mode=None)
# Single output and one step.
batch_size = 5
sequence_length = 1
shape_0, _ = expected_shape(batch_size, sequence_length)
out = single_output_model.predict_generator(
RandomSequence(batch_size, sequence_length=sequence_length))
assert np.shape(out) == shape_0
# Single output and multiple steps.
batch_size = 5
sequence_length = 2
shape_0, _ = expected_shape(batch_size, sequence_length)
out = single_output_model.predict_generator(
RandomSequence(batch_size, sequence_length=sequence_length))
assert np.shape(out) == shape_0
def test_model_methods():
a = Input(shape=(3, ), name='input_a')
b = Input(shape=(3, ), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
model = Model([a, b], [a_2, b_2])
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
model.compile(optimizer,
loss,
metrics=[],
loss_weights=loss_weights,
sample_weight_mode=None)
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
# test train_on_batch
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.train_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np])
out = model.train_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
})
# test fit
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
nb_epoch=1,
batch_size=4)
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np],
nb_epoch=1,
batch_size=4)
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
},
nb_epoch=1,
batch_size=4)
# test validation_split
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
nb_epoch=1,
batch_size=4,
validation_split=0.5)
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np],
nb_epoch=1,
batch_size=4,
validation_split=0.5)
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
},
nb_epoch=1,
batch_size=4,
validation_split=0.5)
# test validation data
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
nb_epoch=1,
batch_size=4,
validation_data=([input_a_np,
input_b_np], [output_a_np, output_b_np]))
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np],
nb_epoch=1,
batch_size=4,
validation_split=0.5,
validation_data=({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np]))
out = model.fit({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
},
nb_epoch=1,
batch_size=4,
validation_split=0.5,
validation_data=({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
}))
# test_on_batch
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.test_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
}, [output_a_np, output_b_np])
out = model.test_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
}, {
'dense_1': output_a_np,
'dropout': output_b_np
})
# predict_on_batch
out = model.predict_on_batch([input_a_np, input_b_np])
out = model.predict_on_batch({
'input_a': input_a_np,
'input_b': input_b_np
})
# predict, evaluate
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np],
batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
# with sample_weight
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
sample_weight = [None, np.random.random((10, ))]
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
# test accuracy metric
model.compile(optimizer, loss, metrics=['acc'], sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
# this should also work
model.compile(optimizer,
loss,
metrics={'dense_1': 'acc'},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# and this as well
model.compile(optimizer,
loss,
metrics={'dense_1': ['acc']},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# test starting from non-zero initial epoch
trained_epochs = []
def on_epoch_begin(epoch, logs):
trained_epochs.append(epoch)
tracker_cb = LambdaCallback(on_epoch_begin=on_epoch_begin)
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
nb_epoch=5,
batch_size=4,
initial_epoch=2,
callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test starting from non-zero initial epoch for generator too
trained_epochs = []
def gen_data(batch_sz):
while True:
yield ([
np.random.random((batch_sz, 3)),
np.random.random((batch_sz, 3))
], [
np.random.random((batch_sz, 4)),
np.random.random((batch_sz, 3))
])
out = model.fit_generator(gen_data(4),
samples_per_epoch=10,
nb_epoch=5,
initial_epoch=2,
callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test with a custom metric function
mse = lambda y_true, y_pred: K.mean(K.pow(y_true - y_pred, 2))
def mse_powers(y_true, y_pred):
m = mse(y_true, y_pred)
return {'mse_squared': K.pow(m, 2), 'mse_cubed': K.pow(m, 3)}
model.compile(optimizer,
loss,
metrics=[mse, mse_powers],
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out_len = 1 + 2 * 4 # total loss, per layer: loss + 3 metrics
assert len(out) == out_len
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == out_len
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np],
batch_size=4,
nb_epoch=1)
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np],
batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
def decoder_stochastic_wrn(
label_sizes,
nb_bits=12,
data_shape=(1, 64, 64),
activation=lambda: Activation('relu'),
normalization=lambda: BatchNormalization(axis=1, mode=0),
weight_init='he_normal',
weight_decay=0.0001,
dropout_probability=0.,
wrn_depth=58,
wrn_k=2,
death_rate=0.5,
optimizer='adam'):
def norm_act_block():
def f(inputs):
x = normalization()(inputs)
x = activation()(x)
return x
return f
def conv2(nb_filter, nb_row, nb_col, subsample=(1, 1), bias=True):
return Convolution2D(nb_filter,
nb_row,
nb_col,
init=weight_init,
border_mode='same',
subsample=subsample,
bias=bias,
W_regularizer=l2(weight_decay))
def dropout(p=None):
if p is None:
p = dropout_probability
def f(inputs):
if p > 0.:
inputs = Dropout(p)(inputs)
return inputs
return f
def equal_gate_shape(input_shapes):
assert (input_shapes[0] == input_shapes[1])
return input_shapes[1]
def residual_block(nb_filter, stochastic=False, stochastic_layers=None):
def f(inputs):
x = norm_act_block()(inputs)
x = conv2(nb_filter, 3, 3)(x)
x = dropout()(x)
x = norm_act_block()(x)
x = conv2(nb_filter, 3, 3)(x)
if inputs._keras_shape != x._keras_shape:
inputs = conv2(nb_filter, 1, 1, bias=False)(inputs)
if not stochastic:
return merge((inputs, x), mode='sum')
scale = ScaleInTestPhase(death_rate)
x = scale(x)
out = merge([inputs, x],
mode="sum",
output_shape=x._keras_shape[1:])
rs = RandomSwitch(death_rate)
stochastic_layers.append((rs.death_rate, scale.death_rate))
return rs([out, inputs])
return f
def residual_reduction_block(nb_filter):
def f(inputs):
x = norm_act_block()(inputs)
x = conv2(nb_filter, 3, 3, subsample=(2, 2))(x)
x = dropout()(x)
x = norm_act_block()(x)
x = conv2(nb_filter, 3, 3)(x)
inputs_bottleneck = conv2(nb_filter,
1,
1,
subsample=(2, 2),
bias=False)(inputs)
s = merge((inputs_bottleneck, x), mode='sum')
return s
return f
def skip_connection(inputs,
residual,
stochastic=False,
stochastic_layers=None):
inputs_filters = inputs._keras_shape[1]
residual_filters = residual._keras_shape[1]
subsample = np.array(inputs._keras_shape[2:]) // np.array(
residual._keras_shape[2:])
inputs = dropout()(inputs)
if (inputs_filters != residual_filters) or np.any(subsample > 1):
skip = conv2(residual_filters,
1,
1,
subsample=subsample,
bias=False)(inputs)
else:
skip = inputs
if not stochastic:
return merge((skip, residual), mode='sum')
else:
scale = ScaleInTestPhase(death_rate)
skip = scale(skip)
out = merge([skip, residual], mode="sum")
rs = RandomSwitch(death_rate)
stochastic_layers.append((rs.death_rate, scale.death_rate))
return rs([out, residual])
n = (wrn_depth - 4) // 6
stochastic_layers = []
input = Input(shape=data_shape)
m_stem = conv2(16, 3, 3, subsample=(2, 2))(input)
m_stem = norm_act_block()(m_stem)
m_b1 = residual_block(nb_filter=16 * wrn_k,
stochastic_layers=stochastic_layers)(m_stem)
for _ in range(n - 1):
m_b1 = residual_block(nb_filter=16 * wrn_k,
stochastic=True,
stochastic_layers=stochastic_layers)(m_b1)
m_b1 = skip_connection(input,
m_b1,
stochastic=True,
stochastic_layers=stochastic_layers)
m_b2 = residual_reduction_block(nb_filter=32 * wrn_k)(m_b1)
for _ in range(n - 1):
m_b2 = residual_block(nb_filter=32 * wrn_k,
stochastic=True,
stochastic_layers=stochastic_layers)(m_b2)
m_b2 = skip_connection(m_b1,
m_b2,
stochastic=True,
stochastic_layers=stochastic_layers)
m_b3 = residual_reduction_block(nb_filter=64 * wrn_k)(m_b2)
for _ in range(n - 1):
m_b3 = residual_block(nb_filter=64 * wrn_k,
stochastic=True,
stochastic_layers=stochastic_layers)(m_b3)
m_b3 = skip_connection(m_b2,
m_b3,
stochastic=True,
stochastic_layers=stochastic_layers)
m_b3 = skip_connection(input,
m_b3,
stochastic=True,
stochastic_layers=stochastic_layers)
x = norm_act_block()(m_b3)
x = AveragePooling2D(pool_size=(8, 8))(x)
x = dropout()(x)
for i, (tb, ts) in enumerate(stochastic_layers, start=0):
K.set_value(tb, i / len(stochastic_layers) * death_rate)
K.set_value(ts, i / len(stochastic_layers) * death_rate)
outputs, losses = decoder_end_block(x, label_sizes, nb_bits, activation,
weight_decay)
model = Model(input, list(outputs.values()))
model.compile(
optimizer,
loss=list(losses.values()),
loss_weights={k: decoder_loss_weights(k)
for k in losses.keys()})
return model
class SeqGAN:
def __init__(self, g, d, m, g_optimizer, d_optimizer):
self.g = g
self.d = d
self.m = m
self.z, self.seq_input = self.g.inputs
self.fake_prob, = self.g.outputs
with trainable(m, False):
m_input = merge([self.seq_input, self.fake_prob], mode='concat', concat_axis=1)
self.m_realness = self.m(m_input)
self.model_fit_g = Model([self.z, self.seq_input], [self.m_realness])
self.model_fit_g.compile(g_optimizer, K.binary_crossentropy)
self.d.compile(d_optimizer, loss=K.binary_crossentropy)
def z_shape(self, batch_size=64):
layer, _, _ = self.z._keras_history
return (batch_size,) + layer.output_shape[1:]
def sample_z(self, batch_size=64):
shape = self.z_shape(batch_size)
return np.random.uniform(-1, 1, shape)
def generate(self, z, seq_input, batch_size=32):
return self.g.predict([z, seq_input], batch_size=batch_size)
def train_on_batch(self, seq_input, real, d_target=None):
nb_real = len(real)
nb_fake = len(seq_input)
if d_target is None:
d_target = np.concatenate([
np.zeros((nb_fake, 1)),
np.ones((nb_real, 1))
])
fake_prob = self.generate(self.sample_z(nb_fake), seq_input)
fake = np.concatenate([seq_input, prob_to_sentence(fake_prob)], axis=1)
fake_and_real = np.concatenate([fake, real], axis=0)
d_loss = self.d.train_on_batch(fake_and_real, d_target)
d_realness = self.d.predict(fake)
m_loss = self.m.train_on_batch(
np.concatenate([seq_input, fake_prob], axis=1), d_realness)
g_loss = self.model_fit_g.train_on_batch([self.sample_z(nb_fake), seq_input],
np.ones((nb_fake, 1)))
return g_loss, d_loss, m_loss
def fit_generator(self, generator, nb_epoch, nb_batches_per_epoch, callbacks=[],
batch_size=None,
verbose=False):
if batch_size is None:
batch_size = 2*len(next(generator)[0])
out_labels = ['g', 'd', 'm']
self.history = cbks.History()
callbacks = [cbks.BaseLogger()] + callbacks + [self.history]
if verbose:
callbacks += [cbks.ProgbarLogger()]
callbacks = cbks.CallbackList(callbacks)
callbacks._set_model(self)
callbacks._set_params({
'nb_epoch': nb_epoch,
'nb_sample': nb_batches_per_epoch*batch_size,
'verbose': verbose,
'metrics': out_labels,
})
callbacks.on_train_begin()
for e in range(nb_epoch):
callbacks.on_epoch_begin(e)
for batch_index, (seq_input, real) in enumerate(generator):
callbacks.on_batch_begin(batch_index)
batch_logs = {}
batch_logs['batch'] = batch_index
batch_logs['size'] = len(real) + len(seq_input
)
outs = self.train_on_batch(seq_input, real)
for l, o in zip(out_labels, outs):
batch_logs[l] = o
callbacks.on_batch_end(batch_index, batch_logs)
if batch_index + 1 == nb_batches_per_epoch:
break
callbacks.on_epoch_end(e)
callbacks.on_train_end()
def build_CNN_model(inputType, do_training=False, model_inputs=None, loss_func='binary_crossentropy',
optimize_proc='adam', is_IntermediateModel=False, load_weight_path=None, **kwargs):
"""
:param inputType:
:param do_training:
:param model_inputs:
:param loss_func:
:param optimize_proc:
:param is_IntermediateModel:
:param load_weight_path:
:param kwargs:
:return:
"""
# assert not do_training and model_inputs, "if do_training then must pass in model_inputs dictionary"
EMBEDDING_TYPE = 'embeddingMatrix'
ONEHOT_TYPE = '1hotVector'
defined_input_types = {EMBEDDING_TYPE, ONEHOT_TYPE}
assert inputType in defined_input_types, "unknown input type {0}".format(inputType)
if inputType is ONEHOT_TYPE:
review_input = Input(shape=(modelParameters.MaxLen_w,), dtype='float32',
name="ONEHOT_INPUT")
layer = Embedding(modelParameters.VocabSize_w + modelParameters.INDEX_FROM, embedding_dims,
embeddings_initializer=embedding_init, embeddings_regularizer=embedding_reg,
input_length=modelParameters.MaxLen_w, name='1hot_embeddingLayer')(review_input)
layer = SpatialDropout1D(0.50)(layer)
elif inputType is EMBEDDING_TYPE:
review_input = Input(shape=(modelParameters.MaxLen_w, embedding_dims), dtype="float32", name="EMBEDDING_INPUT")
layer = review_input
else:
raise ValueError("Bad inputType arg to build_CNN_model")
layer = Convolution1D(filters=num_filters1,
kernel_size=filter_length1,
padding=region,
strides=1,
activation=conv_activation1,
kernel_initializer='glorot_uniform',
bias_initializer='zeros',
kernel_regularizer=conv_reg1,
dilation_rate=1,
name='ConvLayer1')(layer)
layer = SpatialDropout1D(0.50)(layer)
layer = MaxPooling1D(pool_size=pool_len1)(layer)
# layer = Convolution1D(filters=num_filters2,
# kernel_size=filter_length2,
# padding=region,
# strides=1,
# activation=conv_activation2,
# kernel_initializer=conv_init2,
# kernel_regularizer=conv_reg2,
# dilation_rate=1,
# name='ConvLayer2')(layer)
#
# layer = SpatialDropout1D(0.50)(layer)
#
# layer = MaxPooling1D(pool_size=pool_len2)(layer)
# layer = Convolution1D(filters=num_filters3,
# kernel_size=filter_length3,
# padding=region,
# activation=conv_activation3,
# kernel_initializer=conv_init3,
# kernel_regularizer=conv_reg3,
# dilation_rate=1,
# name='ConvLayer3')(layer)
#
# layer = SpatialDropout1D(0.50)(layer)
#
# layer = MaxPooling1D(pool_size=pool_len3)(layer)
# #layer = GlobalMaxPool1D()(layer)
#
# layer = Convolution1D(filters=num_filters4,
# kernel_size=filter_length4,
# padding=region,
# activation=conv_activation4,
# kernel_initializer=conv_init4,
# kernel_regularizer=conv_reg4,
# dilation_rate=1,
# name='ConvLayer4')(layer)
#
# #layer = leaky_relu(layer)
#
# layer = SpatialDropout1D(0.50)(layer)
#
# layer = MaxPooling1D(pool_size=pool_len4)(layer)
# #layer = GlobalMaxPool1D()(layer)
#
# # layer = BatchNormalization()(layer)
layer = Flatten()(layer)
layer = Dense(dense_dims0, activation=dense_activation0, kernel_regularizer=dense_reg0,
kernel_initializer='glorot_normal', bias_initializer='zeros',
name='dense0')(layer)
layer = Dropout(0.50)(layer)
layer = Dense(dense_dims1, activation=dense_activation1, kernel_regularizer=dense_reg1,
kernel_initializer='glorot_normal', bias_initializer='zeros',
name='dense1')(layer)
layer = Dropout(0.50)(layer)
# layer = Dense(dense_dims2, activation=dense_activation2, kernel_regularizer=dense_reg2,
# kernel_initializer=dense_init2,
# name='dense2')(layer)
#
#
# layer = Dropout(0.50)(layer)
#
# layer = Dense(dense_dims3, activation=dense_activation3, kernel_regularizer=dense_reg3,
# kernel_initializer=dense_init3,
# name='dense3_outA')(layer)
# #layer = leaky_relu(layer)
#
if is_IntermediateModel:
return Model(inputs=[review_input], outputs=[layer], name="CNN_model")
#
# layer = Dropout(0.5)(layer)
layer = Dense(dense_dims_final, activation=dense_activation_final, kernel_initializer=dense_init_final,
kernel_regularizer=dense_reg0,
name='output_Full')(layer)
CNN_model = Model(inputs=[review_input], outputs=[layer], name="CNN_model")
CNN_model.compile(optimizer=Adam(lr=0.001, decay=0.0), loss=loss_func, metrics=[binary_accuracy])
if load_weight_path is not None:
CNN_model.load_weights(load_weight_path)
hist = ""
if do_training:
weightPath = os.path.join(modelParameters.WEIGHT_PATH, filename)
configPath = os.path.join(modelParameters.WEIGHT_PATH, filename_config)
with open(configPath + ".json", 'wb') as f:
f.write(CNN_model.to_json())
checkpoint = ModelCheckpoint(weightPath + '_W.{epoch:02d}-{val_loss:.4f}.hdf5',
verbose=1, save_best_only=True, save_weights_only=False, monitor='val_loss')
earlyStop = EarlyStopping(patience=3, verbose=1, monitor='val_loss')
LRadjuster = ReduceLROnPlateau(monitor='val_loss', factor=0.30, patience=0, verbose=1, cooldown=1,
min_lr=0.00001, epsilon=1e-2)
call_backs = [checkpoint, earlyStop, LRadjuster]
CNN_model.summary()
hist = CNN_model.fit(*model_inputs['training'],
batch_size=batch_size,
epochs=nb_epoch, verbose=1,
validation_data=model_inputs['dev'],
callbacks=call_backs)
return {"model": CNN_model, "hist": hist}
def test_model_with_input_feed_tensor():
"""We test building a model with a TF variable as input.
We should be able to call fit, evaluate, predict,
by only passing them data for the placeholder inputs
in the model.
"""
import tensorflow as tf
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
b = Input(shape=(3,), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
model = Model([a, b], [a_2, b_2])
model.summary()
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
model.compile(optimizer, loss, metrics=['mean_squared_error'],
loss_weights=loss_weights,
sample_weight_mode=None)
# test train_on_batch
out = model.train_on_batch(input_b_np,
[output_a_np, output_b_np])
out = model.train_on_batch({'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.test_on_batch({'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.predict_on_batch({'input_b': input_b_np})
# test fit
out = model.fit({'input_b': input_b_np},
[output_a_np, output_b_np], epochs=1, batch_size=10)
out = model.fit(input_b_np,
[output_a_np, output_b_np], epochs=1, batch_size=10)
# test evaluate
out = model.evaluate({'input_b': input_b_np},
[output_a_np, output_b_np], batch_size=10)
out = model.evaluate(input_b_np,
[output_a_np, output_b_np], batch_size=10)
# test predict
out = model.predict({'input_b': input_b_np}, batch_size=10)
out = model.predict(input_b_np, batch_size=10)
assert len(out) == 2
# Now test a model with a single input
# i.e. we don't pass any data to fit the model.
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
a_2 = Dense(4, name='dense_1')(a)
a_2 = Dropout(0.5, name='dropout')(a_2)
model = Model(a, a_2)
model.summary()
optimizer = 'rmsprop'
loss = 'mse'
model.compile(optimizer, loss, metrics=['mean_squared_error'])
# test train_on_batch
out = model.train_on_batch(None,
output_a_np)
out = model.train_on_batch(None,
output_a_np)
out = model.test_on_batch(None,
output_a_np)
out = model.predict_on_batch(None)
out = model.train_on_batch([],
output_a_np)
out = model.train_on_batch({},
output_a_np)
# test fit
out = model.fit(None,
output_a_np, epochs=1, batch_size=10)
out = model.fit(None,
output_a_np, epochs=1, batch_size=10)
# test evaluate
out = model.evaluate(None,
output_a_np, batch_size=10)
out = model.evaluate(None,
output_a_np, batch_size=10)
# test predict
out = model.predict(None, steps=3)
out = model.predict(None, steps=3)
assert out.shape == (10 * 3, 4)
# Same, without learning phase
# i.e. we don't pass any data to fit the model.
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
a_2 = Dense(4, name='dense_1')(a)
model = Model(a, a_2)
model.summary()
optimizer = 'rmsprop'
loss = 'mse'
model.compile(optimizer, loss, metrics=['mean_squared_error'])
# test train_on_batch
out = model.train_on_batch(None,
output_a_np)
out = model.train_on_batch(None,
output_a_np)
out = model.test_on_batch(None,
output_a_np)
out = model.predict_on_batch(None)
out = model.train_on_batch([],
output_a_np)
out = model.train_on_batch({},
output_a_np)
# test fit
out = model.fit(None,
output_a_np, epochs=1, batch_size=10)
out = model.fit(None,
output_a_np, epochs=1, batch_size=10)
# test evaluate
out = model.evaluate(None,
output_a_np, batch_size=10)
out = model.evaluate(None,
output_a_np, batch_size=10)
# test predict
out = model.predict(None, steps=3)
out = model.predict(None, steps=3)
assert out.shape == (10 * 3, 4)
numerical_inputs = Input(shape=(11, ), name='num')
numerical_logits = numerical_inputs
numerical_logits = BatchNormalization()(numerical_logits)
numerical_logits = Dense(128, activation='relu')(numerical_logits)
numerical_logits = Dropout(0.5)(numerical_logits)
numerical_logits = BatchNormalization()(numerical_logits)
numerical_logits = Dense(128, activation='relu')(numerical_logits)
numerical_logits = Dense(64, activation='relu')(numerical_logits)
logits = Concatenate()([numerical_logits, categorical_logits])
logits = Dense(64, activation='relu')(logits)
out = Dense(1, activation='sigmoid')(logits)
model = Model(inputs=categorical_inputs + [numerical_inputs], outputs=out)
model.compile(optimizer='adam', loss=binary_crossentropy)
# In[ ]:
def get_input(market_train, indices):
X_num = market_train.loc[indices, num_cols].values
X = {'num': X_num}
for cat in cat_cols:
X[cat] = market_train.loc[indices, cat_cols].values
y = (market_train.loc[indices, 'returnsOpenNextMktres10'] >= 0).values
r = market_train.loc[indices, 'returnsOpenNextMktres10'].values
u = market_train.loc[indices, 'universe']
d = market_train.loc[indices, 'time'].dt.date
return X, y, r, u, d
def test_model_with_external_loss():
# None loss, only regularization loss.
a = Input(shape=(3,), name='input_a')
a_2 = Dense(4, name='dense_1',
kernel_regularizer='l1',
bias_regularizer='l2')(a)
dp = Dropout(0.5, name='dropout')
a_3 = dp(a_2)
model = Model(a, [a_2, a_3])
optimizer = 'rmsprop'
loss = None
model.compile(optimizer, loss, metrics=['mae'])
input_a_np = np.random.random((10, 3))
# test train_on_batch
out = model.train_on_batch(input_a_np, None)
out = model.test_on_batch(input_a_np, None)
# fit
out = model.fit(input_a_np, None)
# evaluate
out = model.evaluate(input_a_np, None)
# No dropout, external loss.
a = Input(shape=(3,), name='input_a')
a_2 = Dense(4, name='dense_1')(a)
a_3 = Dense(4, name='dense_2')(a)
model = Model(a, [a_2, a_3])
model.add_loss(K.mean(a_3 + a_2))
optimizer = 'rmsprop'
loss = None
model.compile(optimizer, loss, metrics=['mae'])
# test train_on_batch
out = model.train_on_batch(input_a_np, None)
out = model.test_on_batch(input_a_np, None)
# fit
out = model.fit(input_a_np, None)
# evaluate
out = model.evaluate(input_a_np, None)
# Test fit with no external data at all.
if K.backend() == 'tensorflow':
import tensorflow as tf
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
a_2 = Dense(4, name='dense_1')(a)
a_2 = Dropout(0.5, name='dropout')(a_2)
model = Model(a, a_2)
model.add_loss(K.mean(a_2))
model.compile(optimizer='rmsprop',
loss=None,
metrics=['mean_squared_error'])
# test train_on_batch
out = model.train_on_batch(None, None)
out = model.test_on_batch(None, None)
out = model.predict_on_batch(None)
# test fit
with pytest.raises(ValueError):
out = model.fit(None, None, epochs=1, batch_size=10)
out = model.fit(None, None, epochs=1, steps_per_epoch=1)
# test fit with validation data
with pytest.raises(ValueError):
out = model.fit(None, None,
epochs=1,
steps_per_epoch=None,
validation_steps=2)
out = model.fit(None, None,
epochs=1,
steps_per_epoch=2,
validation_steps=2)
# test evaluate
with pytest.raises(ValueError):
out = model.evaluate(None, None, batch_size=10)
out = model.evaluate(None, None, steps=3)
# test predict
with pytest.raises(ValueError):
out = model.predict(None, batch_size=10)
out = model.predict(None, steps=3)
assert out.shape == (10 * 3, 4)
# Test multi-output model without external data.
a = Input(tensor=tf.Variable(input_a_np, dtype=tf.float32))
a_1 = Dense(4, name='dense_1')(a)
a_2 = Dropout(0.5, name='dropout')(a_1)
model = Model(a, [a_1, a_2])
model.add_loss(K.mean(a_2))
model.compile(optimizer='rmsprop',
loss=None,
metrics=['mean_squared_error'])
# test train_on_batch
out = model.train_on_batch(None, None)
out = model.test_on_batch(None, None)
out = model.predict_on_batch(None)
# test fit
with pytest.raises(ValueError):
out = model.fit(None, None, epochs=1, batch_size=10)
out = model.fit(None, None, epochs=1, steps_per_epoch=1)
# test fit with validation data
with pytest.raises(ValueError):
out = model.fit(None, None,
epochs=1,
steps_per_epoch=None,
validation_steps=2)
out = model.fit(None, None,
epochs=1,
steps_per_epoch=2,
validation_steps=2)
# test evaluate
with pytest.raises(ValueError):
out = model.evaluate(None, None, batch_size=10)
out = model.evaluate(None, None, steps=3)
# test predict
with pytest.raises(ValueError):
out = model.predict(None, batch_size=10)
out = model.predict(None, steps=3)
assert len(out) == 2
assert out[0].shape == (10 * 3, 4)
assert out[1].shape == (10 * 3, 4)
def test_model_methods():
a = Input(shape=(3,), name='input_a')
b = Input(shape=(3,), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
model = Model([a, b], [a_2, b_2])
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
model.compile(optimizer, loss, metrics=[], loss_weights=loss_weights,
sample_weight_mode=None)
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
# test train_on_batch
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.train_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.train_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np})
# test fit
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np], nb_epoch=1, batch_size=4)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np], nb_epoch=1, batch_size=4)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
nb_epoch=1, batch_size=4)
# test validation_split
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
nb_epoch=1, batch_size=4, validation_split=0.5)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np],
nb_epoch=1, batch_size=4, validation_split=0.5)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
nb_epoch=1, batch_size=4, validation_split=0.5)
# test validation data
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
nb_epoch=1, batch_size=4,
validation_data=([input_a_np, input_b_np], [output_a_np, output_b_np]))
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np],
nb_epoch=1, batch_size=4, validation_split=0.5,
validation_data=({'input_a': input_a_np, 'input_b': input_b_np}, [output_a_np, output_b_np]))
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
nb_epoch=1, batch_size=4, validation_split=0.5,
validation_data=({'input_a': input_a_np, 'input_b': input_b_np}, {'dense_1': output_a_np, 'dropout': output_b_np}))
# test_on_batch
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.test_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.test_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np})
# predict_on_batch
out = model.predict_on_batch([input_a_np, input_b_np])
out = model.predict_on_batch({'input_a': input_a_np, 'input_b': input_b_np})
# predict, evaluate
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
# with sample_weight
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
sample_weight = [None, np.random.random((10,))]
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
# test accuracy metric
model.compile(optimizer, loss, metrics=['acc'],
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
# this should also work
model.compile(optimizer, loss, metrics={'dense_1': 'acc'},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# and this as well
model.compile(optimizer, loss, metrics={'dense_1': ['acc']},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# test with a custom metric function
mse = lambda y_true, y_pred: K.mean(K.pow(y_true - y_pred, 2))
model.compile(optimizer, loss, metrics=[mse],
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4, nb_epoch=1)
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
def test_model_methods():
a = Input(shape=(3,), name='input_a')
b = Input(shape=(3,), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
model = Model([a, b], [a_2, b_2])
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
input_a_df = pd.DataFrame(input_a_np)
input_b_df = pd.DataFrame(input_b_np)
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
output_a_df = pd.DataFrame(output_a_np)
output_b_df = pd.DataFrame(output_b_np)
# training/testing doesn't work before compiling.
with pytest.raises(RuntimeError):
model.train_on_batch([input_a_np, input_b_np], [output_a_np, output_b_np])
model.compile(optimizer, loss, metrics=[], loss_weights=loss_weights,
sample_weight_mode=None)
# test train_on_batch
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.train_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.train_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np})
out = model.train_on_batch([input_a_df, input_b_df],
[output_a_df, output_b_df])
# test fit
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np], epochs=1, batch_size=4)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np], epochs=1, batch_size=4)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
epochs=1, batch_size=4)
out = model.fit([input_a_df, input_b_df],
[output_a_df, output_b_df], epochs=1, batch_size=4)
# test validation_split
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
epochs=1, batch_size=4, validation_split=0.5)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np],
epochs=1, batch_size=4, validation_split=0.5)
# test validation data
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
epochs=1, batch_size=4,
validation_data=([input_a_np, input_b_np], [output_a_np, output_b_np]))
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np],
epochs=1, batch_size=4, validation_split=0.5,
validation_data=({'input_a': input_a_np, 'input_b': input_b_np}, [output_a_np, output_b_np]))
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
epochs=1, batch_size=4, validation_split=0.5,
validation_data=(
{'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np}))
# test_on_batch
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.test_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.test_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np})
out = model.test_on_batch([input_a_df, input_b_df],
[output_a_df, output_b_df])
# predict_on_batch
out = model.predict_on_batch([input_a_np, input_b_np])
out = model.predict_on_batch({'input_a': input_a_np, 'input_b': input_b_np})
out = model.predict_on_batch([input_a_df, input_b_df])
# predict, evaluate
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4)
out = model.evaluate([input_a_df, input_b_df], [output_a_df, output_b_df], batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
out = model.predict([input_a_df, input_b_df], batch_size=4)
# with sample_weight
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
sample_weight = [None, np.random.random((10,))]
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
# test accuracy metric
model.compile(optimizer, loss, metrics=['acc'],
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
# this should also work
model.compile(optimizer, loss, metrics={'dense_1': 'acc'},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# and this as well
model.compile(optimizer, loss, metrics={'dense_1': ['acc']},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# test starting from non-zero initial epoch
trained_epochs = []
# define tracer callback
def on_epoch_begin(epoch, logs):
trained_epochs.append(epoch)
tracker_cb = LambdaCallback(on_epoch_begin=on_epoch_begin)
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np], epochs=5, batch_size=4,
initial_epoch=2, callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test starting from non-zero initial epoch for generator too
trained_epochs = []
def gen_data(batch_sz):
while True:
yield ([np.random.random((batch_sz, 3)), np.random.random((batch_sz, 3))],
[np.random.random((batch_sz, 4)), np.random.random((batch_sz, 3))])
out = model.fit_generator(gen_data(4), steps_per_epoch=3, epochs=5,
initial_epoch=2, callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test with a custom metric function
def mse(y_true, y_pred):
return K.mean(K.pow(y_true - y_pred, 2))
model.compile(optimizer, loss, metrics=[mse],
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out_len = 1 + 2 * (1 + 1) # total loss + 2 outputs * (loss + metric)
assert len(out) == out_len
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == out_len
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4, epochs=1)
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
# empty batch
with pytest.raises(ValueError):
def gen_data():
while True:
yield (np.asarray([]), np.asarray([]))
out = model.evaluate_generator(gen_data(), steps=1)
# x is not a list of numpy arrays.
with pytest.raises(ValueError):
out = model.predict([None])
# x does not match _feed_input_names.
with pytest.raises(ValueError):
out = model.predict([input_a_np, None, input_b_np])
with pytest.raises(ValueError):
out = model.predict([None, input_a_np, input_b_np])
# all input/output/weight arrays should have the same number of samples.
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np[:2]],
[output_a_np, output_b_np],
sample_weight=sample_weight)
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np[:2]],
sample_weight=sample_weight)
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=[sample_weight[1], sample_weight[1][:2]])
# `sample_weight` is neither a dict nor a list.
with pytest.raises(TypeError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=tuple(sample_weight))
# `validation_data` is neither a tuple nor a triple.
with pytest.raises(ValueError):
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
epochs=1, batch_size=4,
validation_data=([input_a_np, input_b_np],))
# `loss` does not match outputs.
with pytest.raises(ValueError):
model.compile(optimizer, loss=['mse', 'mae', 'mape'])
# `loss_weights` does not match output_names.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', loss_weights={'lstm': 0.5})
# `loss_weights` does not match outputs.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', loss_weights=[0.5])
# `loss_weights` is invalid type.
with pytest.raises(TypeError):
model.compile(optimizer, loss='mse', loss_weights=(0.5, 0.5))
# `sample_weight_mode` does not match output_names.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', sample_weight_mode={'lstm': 'temporal'})
# `sample_weight_mode` does not match output_names.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', sample_weight_mode=['temporal'])
# `sample_weight_mode` matches output_names partially.
with pytest.raises(ValueError):
model.compile(optimizer, loss='mse', sample_weight_mode={'dense_1': 'temporal'})
# `loss` does not exist.
with pytest.raises(ValueError):
model.compile(optimizer, loss=[])
model.compile(optimizer, loss=['mse', 'mae'])
model.compile(optimizer, loss='mse', loss_weights={'dense_1': 0.2, 'dropout': 0.8})
model.compile(optimizer, loss='mse', loss_weights=[0.2, 0.8])
# the rank of weight arrays should be 1.
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=[None, np.random.random((10, 20, 30))])
model.compile(optimizer, loss='mse', sample_weight_mode={'dense_1': None, 'dropout': 'temporal'})
model.compile(optimizer, loss='mse', sample_weight_mode=[None, 'temporal'])
# the rank of output arrays should be at least 3D.
with pytest.raises(ValueError):
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
model.compile(optimizer, loss, metrics=[], loss_weights=loss_weights,
sample_weight_mode=None)
trained_epochs = []
out = model.fit_generator(generator=RandomSequence(3), steps_per_epoch=12, epochs=5,
initial_epoch=0, validation_data=RandomSequence(4),
validation_steps=12, callbacks=[tracker_cb])
assert trained_epochs == [0, 1, 2, 3, 4]
class AE(Model):
"""
Autoencoder. This is a simple autoencoder consisting of an encoder and a decoder.
You can use the class like this:
>>> encoder = ...
>>> decoder = ...
>>> ae = Autoencoder(encoder=encoder, decoder=decoder)
>>> ae.compile(...)
>>> ae.fit(...)
"""
def __init__(self, encoder=None, decoder=None, autoencoder=None):
super(AE, self).__init__()
# For calling this as a super-constructor.
parameters = [encoder, decoder]
if all(v is None for v in parameters):
return
# From loading.
if encoder != None and decoder != None and autoencoder != None:
self.encoder = encoder
self.decoder = decoder
self.autoencoder = autoencoder
return
# Check preconditions.
assert len(encoder.outputs) == 1
assert len(decoder.inputs) == 1
assert encoder.outputs[0].shape[1:] == decoder.inputs[0].shape[
1:], str(encoder.outputs[0].shape) + " " + str(
decoder.inputs[0].shape)
self.latent_dim = encoder.outputs[0].shape[1]
self.encoder = encoder
self.decoder = decoder
# Creating the AE.
inputs = self.encoder.inputs[0]
outputs = self.decoder(self.encoder(inputs))
self.autoencoder = Model(inputs, outputs, name='ae')
def compile(self,
optimizer,
loss=None,
metrics=None,
loss_weights=None,
sample_weight_mode=None,
weighted_metrics=None,
target_tensors=None,
**kwargs):
"""
Compiles the model.
This is the same as compilation in Keras.
"""
assert "reconstruction_loss" not in kwargs, "Not expected to use reconstruction_loss in AE."
self.autoencoder.compile(optimizer, loss, metrics, loss_weights,
sample_weight_mode, weighted_metrics,
**kwargs)
def fit(self,
x=None,
y=None,
batch_size=None,
epochs=1,
verbose=1,
callbacks=None,
validation_split=0.,
validation_data=None,
shuffle=True,
class_weight=None,
sample_weight=None,
initial_epoch=0,
steps_per_epoch=None,
validation_steps=None,
**kwargs):
"""
Trains the autoencoder.
"""
return self.autoencoder.fit(x, y, batch_size, epochs, verbose,
callbacks, validation_split,
validation_data, shuffle, class_weight,
sample_weight, initial_epoch,
steps_per_epoch, validation_steps,
**kwargs)
def fit_generator(self,
generator,
steps_per_epoch=None,
epochs=1,
verbose=1,
callbacks=None,
validation_data=None,
validation_steps=None,
class_weight=None,
max_queue_size=10,
workers=1,
use_multiprocessing=False,
shuffle=True,
initial_epoch=0):
"""
Trains the autoencoder with a generator.
"""
return self.autoencoder.fit_generator(
generator,
steps_per_epoch,
epochs,
verbose=verbose,
callbacks=callbacks,
validation_data=validation_data,
validation_steps=validation_steps,
class_weight=class_weight,
max_queue_size=max_queue_size,
workers=workers,
use_multiprocessing=use_multiprocessing,
shuffle=shuffle,
initial_epoch=initial_epoch)
def evaluate(self,
x=None,
y=None,
batch_size=None,
verbose=1,
sample_weight=None,
steps=None):
"""
Evaluates the autoencoder.
"""
return self.autoencoder.evaluate(x,
y,
batch_size,
verbose,
sample_weight,
steps=None)
def predict(self, x, batch_size=None, verbose=0, steps=None):
"""
Does a prediction. This is the same as :func:`~ngdlm.models.AE.predict_reconstruct_from_samples`
"""
return self.predict_reconstruct_from_samples(x, batch_size, verbose,
steps)
def predict_reconstruct_from_samples(self,
x,
batch_size=None,
verbose=0,
steps=None):
"""
Reconstructs samples.
Samples are firstly mapped to latent space using the encoder.
The resulting latent vectors are then mapped to reconstruction space via the decoder.
"""
return self.autoencoder.predict(x, batch_size, verbose, steps)
def predict_embed_samples_into_latent(self,
x,
batch_size=None,
verbose=0,
steps=None):
"""
Embeds samples into latent space using the encoder.
"""
return self.encoder.predict(x, batch_size, verbose, steps)
def predict_reconstruct_from_latent(self,
x,
batch_size=None,
verbose=0,
steps=None):
"""
Maps latent vectors to reconstruction space using the decoder.
"""
return self.decoder.predict(x, batch_size, verbose, steps)
def summary(self):
"""
Provides a summary.
"""
print("Encoder:")
self.encoder.summary()
print("Decoder:")
self.decoder.summary()
print("Autoencoder:")
self.autoencoder.summary()
def save(self, path):
"""
Saves the autoencoder.
This includes the whole autoencoder plus the encoder and the decoder.
The encoder and decoder use the path plus a respective annotation.
This code
>>> ae.save("myae.h5")
will create the files *myae.h5*, *myae-encoder.h5*, and *myae-decoder.h5*.
"""
self.autoencoder.save(path)
self.encoder.save(append_to_filepath(path, "-encoder"))
self.decoder.save(append_to_filepath(path, "-decoder"))
def test_model_methods():
a = Input(shape=(3,), name='input_a')
b = Input(shape=(3,), name='input_b')
a_2 = Dense(4, name='dense_1')(a)
dp = Dropout(0.5, name='dropout')
b_2 = dp(b)
model = Model([a, b], [a_2, b_2])
optimizer = 'rmsprop'
loss = 'mse'
loss_weights = [1., 0.5]
model.compile(optimizer, loss, metrics=[], loss_weights=loss_weights,
sample_weight_mode=None)
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
# test train_on_batch
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.train_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.train_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np})
# test fit
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np], nb_epoch=1, batch_size=4)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np], nb_epoch=1, batch_size=4)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
nb_epoch=1, batch_size=4)
# test validation_split
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
nb_epoch=1, batch_size=4, validation_split=0.5)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np],
nb_epoch=1, batch_size=4, validation_split=0.5)
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
nb_epoch=1, batch_size=4, validation_split=0.5)
# test validation data
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np],
nb_epoch=1, batch_size=4,
validation_data=([input_a_np, input_b_np], [output_a_np, output_b_np]))
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np],
nb_epoch=1, batch_size=4, validation_split=0.5,
validation_data=({'input_a': input_a_np, 'input_b': input_b_np}, [output_a_np, output_b_np]))
out = model.fit({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np},
nb_epoch=1, batch_size=4, validation_split=0.5,
validation_data=({'input_a': input_a_np, 'input_b': input_b_np}, {'dense_1': output_a_np, 'dropout': output_b_np}))
# test_on_batch
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out = model.test_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
[output_a_np, output_b_np])
out = model.test_on_batch({'input_a': input_a_np, 'input_b': input_b_np},
{'dense_1': output_a_np, 'dropout': output_b_np})
# predict_on_batch
out = model.predict_on_batch([input_a_np, input_b_np])
out = model.predict_on_batch({'input_a': input_a_np, 'input_b': input_b_np})
# predict, evaluate
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
# with sample_weight
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
sample_weight = [None, np.random.random((10,))]
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np],
sample_weight=sample_weight)
# test accuracy metric
model.compile(optimizer, loss, metrics=['acc'],
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 5
# this should also work
model.compile(optimizer, loss, metrics={'dense_1': 'acc'},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# and this as well
model.compile(optimizer, loss, metrics={'dense_1': ['acc']},
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == 4
# test starting from non-zero initial epoch
trained_epochs = []
def on_epoch_begin(epoch, logs):
trained_epochs.append(epoch)
tracker_cb = LambdaCallback(on_epoch_begin=on_epoch_begin)
out = model.fit([input_a_np, input_b_np],
[output_a_np, output_b_np], nb_epoch=5, batch_size=4,
initial_epoch=2, callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test starting from non-zero initial epoch for generator too
trained_epochs = []
def gen_data(batch_sz):
while True:
yield ([np.random.random((batch_sz, 3)), np.random.random((batch_sz, 3))],
[np.random.random((batch_sz, 4)), np.random.random((batch_sz, 3))])
out = model.fit_generator(gen_data(4), samples_per_epoch=10, nb_epoch=5,
initial_epoch=2, callbacks=[tracker_cb])
assert trained_epochs == [2, 3, 4]
# test with a custom metric function
mse = lambda y_true, y_pred: K.mean(K.pow(y_true - y_pred, 2))
def mse_powers(y_true, y_pred):
m = mse(y_true, y_pred)
return {
'mse_squared': K.pow(m, 2),
'mse_cubed': K.pow(m, 3)
}
model.compile(optimizer, loss, metrics=[mse, mse_powers],
sample_weight_mode=None)
out = model.train_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
out_len = 1 + 2 * 4 # total loss, per layer: loss + 3 metrics
assert len(out) == out_len
out = model.test_on_batch([input_a_np, input_b_np],
[output_a_np, output_b_np])
assert len(out) == out_len
input_a_np = np.random.random((10, 3))
input_b_np = np.random.random((10, 3))
output_a_np = np.random.random((10, 4))
output_b_np = np.random.random((10, 3))
out = model.fit([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4, nb_epoch=1)
out = model.evaluate([input_a_np, input_b_np], [output_a_np, output_b_np], batch_size=4)
out = model.predict([input_a_np, input_b_np], batch_size=4)
class PolicyValueNet():
"""policy-value network """
def __init__(self, board_width, board_height, model_file=None):
self.board_width = board_width
self.board_height = board_height
self.l2_const = 1e-4 # coef of l2 penalty
self.create_policy_value_net()
self._loss_train_op()
if model_file:
net_params = pickle.load(open(model_file, 'rb'))
self.model.set_weights(net_params)
def create_policy_value_net(self):
"""create the policy value network """
in_x = network = Input((4, self.board_width, self.board_height))
# conv layers
network = Conv2D(filters=32, kernel_size=(3, 3), padding="same", data_format="channels_first", activation="relu", kernel_regularizer=l2(self.l2_const))(network)
network = Conv2D(filters=64, kernel_size=(3, 3), padding="same", data_format="channels_first", activation="relu", kernel_regularizer=l2(self.l2_const))(network)
network = Conv2D(filters=128, kernel_size=(3, 3), padding="same", data_format="channels_first", activation="relu", kernel_regularizer=l2(self.l2_const))(network)
# action policy layers
policy_net = Conv2D(filters=4, kernel_size=(1, 1), data_format="channels_first", activation="relu", kernel_regularizer=l2(self.l2_const))(network)
policy_net = Flatten()(policy_net)
self.policy_net = Dense(self.board_width*self.board_height, activation="softmax", kernel_regularizer=l2(self.l2_const))(policy_net)
# state value layers
value_net = Conv2D(filters=2, kernel_size=(1, 1), data_format="channels_first", activation="relu", kernel_regularizer=l2(self.l2_const))(network)
value_net = Flatten()(value_net)
value_net = Dense(64, kernel_regularizer=l2(self.l2_const))(value_net)
self.value_net = Dense(1, activation="tanh", kernel_regularizer=l2(self.l2_const))(value_net)
self.model = Model(in_x, [self.policy_net, self.value_net])
def policy_value(state_input):
state_input_union = np.array(state_input)
results = self.model.predict_on_batch(state_input_union)
return results
self.policy_value = policy_value
def policy_value_fn(self, board):
"""
input: board
output: a list of (action, probability) tuples for each available action and the score of the board state
"""
legal_positions = board.availables
current_state = board.current_state()
act_probs, value = self.policy_value(current_state.reshape(-1, 4, self.board_width, self.board_height))
act_probs = zip(legal_positions, act_probs.flatten()[legal_positions])
return act_probs, value[0][0]
def _loss_train_op(self):
"""
Three loss terms：
loss = (z - v)^2 + pi^T * log(p) + c||theta||^2
"""
# get the train op
opt = Adam()
losses = ['categorical_crossentropy', 'mean_squared_error']
self.model.compile(optimizer=opt, loss=losses)
def self_entropy(probs):
return -np.mean(np.sum(probs * np.log(probs + 1e-10), axis=1))
def train_step(state_input, mcts_probs, winner, learning_rate):
state_input_union = np.array(state_input)
mcts_probs_union = np.array(mcts_probs)
winner_union = np.array(winner)
loss = self.model.evaluate(state_input_union, [mcts_probs_union, winner_union], batch_size=len(state_input), verbose=0)
action_probs, _ = self.model.predict_on_batch(state_input_union)
entropy = self_entropy(action_probs)
K.set_value(self.model.optimizer.lr, learning_rate)
self.model.fit(state_input_union, [mcts_probs_union, winner_union], batch_size=len(state_input), verbose=0)
return loss[0], entropy
self.train_step = train_step
def get_policy_param(self):
net_params = self.model.get_weights()
return net_params
def save_model(self, model_file):
""" save model params to file """
net_params = self.get_policy_param()
pickle.dump(net_params, open(model_file, 'wb'), protocol=2)
