def gender(X, f, **kwargs):
nb_gender = kwargs.get('nb_gender', 4)
if f is None:
f = N.Sequence([
N.Dimshuffle(pattern=(0, 1, 2, 'x')),
N.Conv(num_filters=32,
filter_size=3,
strides=1,
b_init=None,
pad='valid'),
N.BatchNorm(activation=K.relu),
N.Pool(pool_size=2, mode='avg'),
N.Conv(num_filters=64,
filter_size=3,
strides=1,
b_init=None,
pad='valid'),
N.BatchNorm(activation=K.relu),
N.Pool(pool_size=2, mode='avg'),
N.Flatten(outdim=3),
N.Dense(num_units=512, b_init=None),
N.BatchNorm(axes=(0, 1)),
N.AutoRNN(num_units=128,
rnn_mode='gru',
num_layers=2,
input_mode='linear',
direction_mode='unidirectional'),
N.Flatten(outdim=2),
N.Dense(num_units=nb_gender, activation=K.softmax)
],
debug=True)
return f(X), f
def test_shape(self):
x = K.variable(np.ones((25, 8, 12)))
def test_func(func):
y = func(x)
yT = func.T(func(x))
self.assertEquals(K.eval(y).shape, tuple(y.shape.as_list()))
self.assertEquals(K.eval(yT).shape, (25, 8, 12))
self.assertEquals(K.eval(yT).shape, tuple(yT.shape.as_list()))
test_func(N.Flatten(outdim=2))
test_func(N.Flatten(outdim=1))
test_func(N.Reshape((25, 4, 2, 6, 2)))
test_func(N.Dimshuffle((2, 0, 1)))
def test_computational_graph3(self):
# validate the number of updates found by ComputationGraph
X = K.placeholder(shape=(None, 28, 28, 3))
f = N.Sequence([
N.Conv(32, 3, pad='same', activation=K.linear),
N.BatchNorm(activation=K.relu),
N.Flatten(outdim=2),
N.Dense(16),
N.BatchNorm(),
N.Dense(10)
])
K.set_training(True)
y_train = f(X)
K.set_training(False)
y_score = f(X)
self.assertTrue(
K.get_shape(y_train) == K.get_shape(y_score)
and K.get_shape(y_score) == (None, 10))
cc_train = K.ComputationGraph(y_train)
cc_score = K.ComputationGraph(y_score)
self.assertTrue(len(cc_score.updates) == 0)
self.assertTrue(len(cc_train.updates) == 4)
# create real function
fn_train = K.function(X, y_train)
fn_score = K.function(X, y_score)
shape1 = fn_train(np.random.rand(12, 28, 28, 3)).shape
shape2 = fn_score(np.random.rand(12, 28, 28, 3)).shape
self.assertTrue(shape1 == shape2 and shape1 == (12, 10))
def test_seq(self):
X = K.placeholder((None, 28, 28, 1))
f = N.Sequence([
N.Conv(8, (3, 3), strides=1, pad='same'),
N.Dimshuffle(pattern=(0, 3, 1, 2)),
N.Flatten(outdim=2),
N.Noise(level=0.3, noise_dims=None, noise_type='gaussian'),
N.Dense(128, activation=tf.nn.relu),
N.Dropout(level=0.3, noise_dims=None),
N.Dense(10, activation=tf.nn.softmax)
])
y = f(X)
yT = f.T(y)
f1 = K.function(X, y, defaults={K.is_training(): True})
f2 = K.function(X, yT, defaults={K.is_training(): False})
f = cPickle.loads(cPickle.dumps(f))
y = f(X)
yT = f.T(y)
f3 = K.function(X, y, defaults={K.is_training(): True})
f4 = K.function(X, yT, defaults={K.is_training(): False})
x = np.random.rand(12, 28, 28, 1)
self.assertEquals(f1(x).shape, (2688, 10))
self.assertEquals(f3(x).shape, (2688, 10))
self.assertEqual(np.round(f1(x).sum(), 4), np.round(f3(x).sum(), 4))
self.assertEquals(y.shape.as_list(), (None, 10))
self.assertEquals(f2(x).shape, (12, 28, 28, 1))
self.assertEquals(f4(x).shape, (12, 28, 28, 1))
self.assertEqual(str(f2(x).sum())[:4], str(f4(x).sum())[:4])
self.assertEquals(yT.shape.as_list(), (None, 28, 28, 1))
def cnn(X, y):
nb_classes = y.shape.as_list()[-1]
with N.args_scope(['Conv', dict(b_init=None, activation=K.linear)],
['BatchNorm', dict(activation=K.relu)]):
f = N.Sequence([
N.Dimshuffle(pattern=(0, 2, 3, 1)),
N.Conv(32, (3, 3), pad='same', stride=(1, 1)),
N.BatchNorm(),
N.Conv(32, (3, 3), pad='same', stride=(1, 1),
b_init=0, activation=K.relu),
N.Pool(pool_size=(2, 2), strides=None, mode='max'),
N.Dropout(level=0.25),
#
N.Conv(64, (3, 3), pad='same', stride=(1, 1)),
N.BatchNorm(),
N.Conv(64, (3, 3), pad='same', stride=(1, 1),
b_init=0., activation=K.relu),
N.Pool(pool_size=(2, 2), strides=None, mode='max'),
N.Dropout(level=0.25),
#
N.Flatten(outdim=2),
N.Dense(512, activation=K.relu),
N.Dropout(level=0.5),
N.Dense(nb_classes, activation=K.linear)
], debug=1)
logit = f(X)
prob = tf.nn.softmax(logit)
return {'logit': logit, 'prob': prob}
def test(X, y):
nb_classes = y.shape.as_list()[-1]
f = N.Sequence([
N.Flatten(outdim=2),
N.Dense(512, activation=K.relu),
N.Dropout(level=0.5),
N.Dense(nb_classes, activation=K.linear)
], debug=2)
logit = f(X)
prob = tf.nn.softmax(logit)
return {'logit': logit, 'prob': prob}
def test_slice_ops(self):
X = K.placeholder(shape=(None, 28, 28, 28, 3))
f = N.Sequence([
N.Conv(32, 3, pad='same', activation=K.linear),
N.BatchNorm(activation=tf.nn.relu),
N.Flatten(outdim=4)[:, 8:12, 18:25, 13:],
])
y = f(X)
fn = K.function(X, y)
self.assertTrue(
fn(np.random.rand(12, 28, 28, 28, 3)).shape[1:] == tuple(
y.shape.as_list()[1:]))
self.assertEqual(y.shape.as_list()[1:], [4, 7, 883])
def ladder1(X, y, states, **kwargs):
noise = kwargs.get('noise', 0.3)
# hyperparameters that denote the importance of each layer
denoising_cost = [1000.0, 10.0, 0.10, 0.10, 0.10]
if states is None:
#
f_encoder = N.Sequence([
N.Flatten(outdim=2),
N.Dense(num_units=1024, b_init=None),
N.BatchNorm(
axes=0, noise_level=noise, noise_dims=None, activation=K.relu),
N.Dense(num_units=512, b_init=None),
N.BatchNorm(
axes=0, noise_level=noise, noise_dims=None, activation=K.relu),
N.Dense(num_units=256, b_init=None),
N.BatchNorm(
axes=0, noise_level=noise, noise_dims=None, activation=K.relu),
N.Dense(num_units=128, b_init=None),
N.BatchNorm(
axes=0, noise_level=noise, noise_dims=None, activation=K.relu),
N.Dense(num_units=10, activation=K.softmax),
],
all_layers=True,
debug=True,
name='Encoder')
#
f_decoder = N.Sequence([
N.Dense(num_units=128, b_init=None),
N.BatchNorm(axes=0, activation=K.relu),
N.Dense(num_units=256, b_init=None),
N.BatchNorm(axes=0, activation=K.relu),
N.Dense(num_units=512, b_init=None),
N.BatchNorm(axes=0, activation=K.relu),
N.Dense(num_units=1024, b_init=None),
N.BatchNorm(axes=0, activation=K.relu),
N.Reshape(shape=(-1, 28, 28)),
],
all_layers=True,
debug=True,
name='Decoder')
else:
f_encoder, f_decoder = states
y_encoder_clean = f_encoder(X, noise=-1)[2::2]
y_encoder_corrp = f_encoder(X, noise=1)[2::2]
print(len(y_encoder_clean), len(y_encoder_corrp))
exit()
return (None, None), [f_encoder, f_decoder]
def test_load_save3(self):
X = K.placeholder(shape=(None, 28, 28))
ops = N.Sequence([
N.Dimshuffle(pattern=(0, 1, 2, 'x')),
N.Conv(8, (3, 3), strides=(1, 1), pad='same', activation=K.relu),
K.pool2d,
N.Flatten(outdim=2),
N.Dense(64, activation=K.relu),
N.Dense(10, activation=K.softmax)
])
y = ops(X)
f1 = K.function(X, y)
ops_ = cPickle.loads(
cPickle.dumps(ops, protocol=cPickle.HIGHEST_PROTOCOL))
y_ = ops_(X)
f2 = K.function(X, y_)
x = np.random.rand(32, 28, 28)
self.assertEqual(np.sum(f1(x) - f2(x)), 0.)
f_encoder = N.Sequence([
N.Dropout(level=0.5),
N.Dimshuffle((0, 2, 3, 1)) if is_cifar10 else N.Dimshuffle((0, 1, 2, 'x')),
N.Conv(num_filters=32, filter_size=3, pad='valid'),
N.Pool(),
N.BatchNorm(),
N.Conv(num_filters=64, filter_size=3, pad='same'),
N.BatchNorm(),
N.Conv(num_filters=64, filter_size=3, pad='valid'),
N.BatchNorm(activation=tf.nn.elu),
N.Pool(),
N.Flatten(outdim=2),
N.Dense(num_units=args.dim)
], debug=True, name='EncoderNetwork')
f_decoder = N.Sequence([
N.Dropout(level=LATENT_DROPOUT, noise_type='uniform'),
N.Noise(level=1.0, noise_type='gaussian'),
N.Dimshuffle((0, 'x', 'x', 1)),
N.TransposeConv(num_filters=64, filter_size=3, pad='valid'),
N.Upsample(size=2, axes=(1, 2)),
N.BatchNorm(),
N.TransposeConv(num_filters=64, filter_size=3, pad='same'),
N.BatchNorm(),
with N.args_scope(
['TimeDelayedConv',
dict(time_pool='none', activation=K.relu)],
['Dense', dict(activation=K.linear, b_init=None)]):
f = N.Sequence([
N.Dropout(level=0.3),
N.TimeDelayedConv(n_new_features=512, n_time_context=5),
N.TimeDelayedConv(n_new_features=512, n_time_context=5),
N.TimeDelayedConv(
n_new_features=512, n_time_context=7, name="LatentTDNN"),
N.Dense(512),
N.BatchNorm(activation=K.relu),
N.Dense(1500),
N.BatchNorm(activation=K.relu),
N.StatsPool(axes=1, output_mode='concat'),
N.Flatten(outdim=2, name="StatsPooling"),
N.Dense(512, name="LatentDense"),
N.BatchNorm(activation=K.relu),
N.Dense(512),
N.BatchNorm(activation=K.relu),
N.Dense(num_units=n_classes,
activation=K.linear,
b_init=init_ops.constant_initializer(0))
],
debug=1)
# ====== create outputs ====== #
y_logit = f(X)
y_proba = tf.nn.softmax(y_logit)
z1 = K.ComputationGraph(y_proba).get(roles=N.Dense,
scope='LatentDense',
beginning_scope=False)[0]
# ===========================================================================
# Build model
# ===========================================================================
f = N.Sequence(
[
N.Embedding(max_features, embedding_size),
N.Dropout(0.25),
N.Dimshuffle(pattern=(0, 1, 'x', 2)), # convolution on time dimension
N.Conv(nb_filter,
filter_size=(filter_length, 1),
pad='valid',
stride=(1, 1),
activation=K.relu),
N.Pool(pool_size=(pool_length, 1), mode='max'),
N.Flatten(outdim=3),
N.Merge(
[
N.Dense(lstm_output_size, activation=K.linear,
name='ingate'), # input-gate
N.Dense(lstm_output_size,
activation=K.linear,
name='forgetgate'), # forget-gate
N.Dense(lstm_output_size,
activation=K.linear,
name='cellupdate'), # cell-update
N.Dense(lstm_output_size, activation=K.linear,
name='outgate') # output-gate
],
merge_function=K.concatenate),
N.LSTM(num_units=lstm_output_size, input_mode='skip')[:, -1],
def test_mnist(self):
ds = fuel.load_mnist()
m = model.SequentialClassifier(N.Flatten(outdim=2),
N.Dense(64, activation=K.relu),
N.Dense(10, activation=K.softmax))
m.set_inputs(
K.placeholder(shape=(None, 28, 28), name='X',
dtype='float32')).set_outputs(
K.placeholder(shape=(None, ),
name='y',
dtype='int32'))
# ====== query info ====== #
m.path
self.assertEqual(m.is_initialized, True)
self.assertEqual(m.input_shape, (None, 28, 28))
self.assertEqual(m.output_shape, (None, 10))
# ====== training test ====== #
m.set_training_info(learning_rate=0.001, n_epoch=3)
m.fit(X=(ds['X_train'], ds['y_train']),
X_valid=(ds['X_valid'], ds['y_valid']))
score = m.score(ds['X_test'][:], ds['y_test'][:])
self.assertEqual(
score > 0.8,
True,
msg='Test if the model get reasonable results: %f accuracy' %
score)
# ====== make prediction and transform test ====== #
np.random.seed(12)
_ = np.random.rand(8, 28, 28)
self.assertEqual(m.transform(_).shape, (8, 10))
self.assertEqual(
np.isclose(m.predict_proba(_).sum(-1), 1.).sum() == 8, True)
self.assertEqual(len(m.predict(_)), 8)
# ====== pickling test ====== #
str_old = str(m)
p_old = m.get_params(True)
m = cPickle.loads(cPickle.dumps(m, protocol=cPickle.HIGHEST_PROTOCOL))
str_new = str(m)
p_new = m.get_params(True)
# ====== test same configurations ====== #
self.assertEqual(str_new, str_old)
# ====== test same params ====== #
for i, j in p_new.iteritems():
k = p_old[i]
for a, b in zip(j, k):
self.assertEqual(np.array_equal(a, b), True)
# ====== test set params ====== #
params = m.get_params(deep=True)
params_new = {}
for n, p in params.iteritems():
params_new[n] = [
np.random.rand(*i.shape).astype('float32') for i in p
]
m.set_params(**params_new)
# test if equal new onces
for i, j in m.get_params(deep=True).iteritems():
k = params_new[i]
for a, b in zip(j, k):
self.assertEqual(np.array_equal(a, b), True)
# ====== training test ====== #
print('Re-train the model second time:')
m.fit(X=(ds['X_train'], ds['y_train']),
X_valid=(ds['X_valid'], ds['y_valid']))
score = m.score(ds['X_test'][:], ds['y_test'][:])
self.assertEqual(
score > 0.8,
True,
msg='Test if the model get reasonable results: %f accuracy' %
score)
CNN = [
N.Dimshuffle(pattern=(0, 1, 2, 'x')),
N.Conv(num_filters=32,
filter_size=3,
pad='same',
strides=1,
activation=K.linear),
N.BatchNorm(activation=K.relu),
N.Conv(num_filters=64,
filter_size=3,
pad='same',
strides=1,
activation=K.linear),
N.BatchNorm(activation=K.relu),
N.Pool(pool_size=2, strides=None, pad='valid', mode='max'),
N.Flatten(outdim=3)
] if args['cnn'] else []
f = N.Sequence(
CNN + [
# ====== RNN ====== #
N.AutoRNN(128,
rnn_mode='lstm',
num_layers=3,
direction_mode='bidirectional',
prefer_cudnn=True),
# ====== Dense ====== #
N.Flatten(outdim=2),
# N.Dropout(level=0.2), # adding dropout does not help
N.Dense(num_units=1024, activation=K.relu),
def convolutional_vae(X, saved_states, **kwargs):
""" convolutional_vae
Return
------
[y_encoder, y_decoder]
States
------
[f_inference (encoder), f_generative (decoder)]
"""
n = kwargs.get('n', 10)
batch_size = K.get_shape(X)[0]
if batch_size is None:
raise ValueError(
"You must specify batch_size dimension for the input placeholder.")
# ====== init ====== #
if saved_states is None:
# Encoder
f_inference = N.Sequence([
N.Reshape(shape=(-1, 28, 28, 1)),
N.Conv(num_filters=32,
filter_size=3,
strides=1,
pad='valid',
b_init=init_ops.constant_initializer(0.),
activation=K.elu),
N.Conv(num_filters=64,
filter_size=5,
strides=2,
pad='same',
b_init=init_ops.constant_initializer(0.),
activation=K.elu),
N.Dropout(level=0.1),
N.Flatten(outdim=2),
N.Dense(num_units=n * 2, b_init=None),
N.BatchNorm(axes=0)
],
debug=True,
name='Encoder')
# Decoder
f_generative = N.Sequence([
N.Dimshuffle(pattern=(0, 'x', 'x', 1)),
N.TransposeConv(num_filters=64,
filter_size=3,
strides=1,
pad='valid',
b_init=init_ops.constant_initializer(0.),
activation=K.elu),
N.TransposeConv(num_filters=32,
filter_size=5,
strides=2,
pad='same',
b_init=init_ops.constant_initializer(0.),
activation=K.elu),
N.TransposeConv(num_filters=1,
filter_size=13,
strides=3,
pad='valid',
b_init=None),
N.BatchNorm(activation=K.linear),
N.Flatten(outdim=3)
],
debug=True,
name="Decoder")
else:
f_inference, f_generative = saved_states
# ====== Perfrom ====== #
# Encoder
y_encoder = f_inference(K.cast(X, 'float32'))
mu = y_encoder[:, :n]
sigma = K.softplus(y_encoder[:, n:])
qz = Normal(mu=mu, sigma=sigma, name='Normal_qz')
# Decoder
z = Normal(mu=K.zeros(shape=(batch_size, n)),
sigma=K.ones(shape=(batch_size, n)),
name="Normal_pz")
logits = f_generative(z)
X_reconstruct = Bernoulli(logits=logits)
# inference
params = f_inference.parameters + f_generative.parameters
inference = ed.KLqp(latent_vars={z: qz}, data={X_reconstruct: X})
# ====== get cost for training ====== #
# Bind p(x, z) and q(z | x) to the same placeholder for x.
if K.is_training():
import tensorflow as tf
inference.initialize()
if True:
optimizer = tf.train.AdamOptimizer(0.01, epsilon=1.0)
updates = optimizer.apply_gradients(
optimizer.compute_gradients(inference.loss, var_list=params))
init = tf.global_variables_initializer()
init.run()
f_train = K.function(X, inference.loss, updates)
else:
optimizer = tf.train.AdamOptimizer(0.01, epsilon=1.0)
inference.initialize(optimizer=optimizer, var_list=params)
init = tf.global_variables_initializer()
init.run()
f_train = lambda x: inference.update(feed_dict={X: x})['loss']
samples = K.sigmoid(logits)
return (samples, z, qz), (f_inference, f_generative)
