def test_single_cost_and_last_layer(cost_func, last_layer):
model = NeuralNetwork(
optimizer=StaticGradientDescent(default_parameters['learning_rate']),
loss=cost_func,
layers=[
Dense(layer_size=50,
activation_func=ReLu(),
weight_initializer=XavierInitializer()),
Dense(layer_size=10,
activation_func=last_layer,
weight_initializer=XavierInitializer())
],
callbacks=[
LoggerCallback(),
PlotCallback(
f'./lab_3/cost/func={cost_func.get_name()}&last_layer={last_layer.get_name()}'
)
])
model.fit(x_train=X_train,
y_train=y_train,
x_val=X_val,
y_val=y_val,
epochs=default_parameters['epochs'],
batch_size=default_parameters['batch_size'])
model.test(X_test, y_test)
def get_variables(cls, all_units, vf_share_layers=False, name=None):
name = name or "keras_mlp_policy"
if "policy" in name:
last_kernel_init_value = 0.01
elif "vf" in name:
last_kernel_init_value = 1.0
else:
raise NotImplementedError
variables = {}
with tf.name_scope(name):
for i, (size_in, size_out) in \
enumerate(zip(all_units, all_units[1:])):
name = f"dense_{i}"
variables[name] = \
Dense.get_variables(
size_in,
size_out,
name=name,
kernel_initializer=(
normc_initializer(1.0) if i < len(all_units) - 2
else normc_initializer(last_kernel_init_value)))
if vf_share_layers:
name = f"dense_vf"
variables[name] = \
Dense.get_variables(
size_in,
1,
name=name,
kernel_initializer=normc_initializer(1.0))
# tricky to remove the name count of the dummy instance
# since the default name is used here
# graph = tf.get_default_graph()
# K.PER_GRAPH_LAYER_NAME_UIDS[graph][
# ("", dummy_instance.name)] -= 1
return variables
def test_single_initializer(initializer):
model = NeuralNetwork(
optimizer=AdamOptimizer(
learning_rate=default_parameters['learning_rate']),
loss=CrossEntropy(),
layers=[
Flatten(),
Dense(layer_size=50,
activation_func=ReLu(),
weight_initializer=initializer),
Dense(layer_size=10,
activation_func=Softmax(),
weight_initializer=initializer)
],
callbacks=[
LoggerCallback(),
PlotCallback(f'./lab_3/initializers/{initializer.get_name()}')
])
model.fit(x_train=X_train,
y_train=y_train,
x_val=X_val,
y_val=y_val,
epochs=default_parameters['epochs'],
batch_size=default_parameters['batch_size'])
model.test(X_test, y_test)
def switch(layer, t):
layer_type = layer[0]
#make shift switch statement
if layer_type == "conv":
c = Conv2D(t, int(layer[1]), (int(layer[2]), int(layer[3])))
return c.conv
elif layer_type == "activation":
a = Activation(t)
return a.activated_tensor
elif layer_type == "pool":
p = MaxPool(t)
return p.pooled
elif layer_type == "dense":
d = Dense(int(layer[1]), t)
return d.fpass
elif layer_type == "output":
d = Dense(int(layer[1]), t)
return d.predict()
else:
print("fail")
def test_single_initializer_with_convo(initializer):
model = NeuralNetwork(
optimizer=AdamOptimizer(
learning_rate=default_parameters['learning_rate'] * 10),
loss=CrossEntropy(),
layers=[
Convolution2D(num_of_filters=8,
kernel=(3, 3),
activation_func=ReLu()),
MaxPooling2D(pool_size=(2, 2), stride=(2, 2)),
Flatten(),
Dense(layer_size=50,
activation_func=ReLu(),
weight_initializer=initializer),
Dense(layer_size=10,
activation_func=Softmax(),
weight_initializer=initializer)
],
callbacks=[
LoggerCallback(),
PlotCallback(f'./lab_3/initializers/{initializer.get_name()}')
])
model.fit(x_train=X_train,
y_train=y_train,
x_val=X_val,
y_val=y_val,
epochs=default_parameters['epochs'],
batch_size=default_parameters['batch_size'])
model.test(X_test, y_test)
def test_single_activation_function(activation):
model = NeuralNetwork(
optimizer=StaticGradientDescent(
learning_rate=default_parameters['learning_rate']),
loss=CrossEntropy(),
layers=[
Dense(layer_size=50,
activation_func=activation,
weight_initializer=XavierInitializer()),
Dense(layer_size=10,
activation_func=Softmax(),
weight_initializer=XavierInitializer())
],
callbacks=[
LoggerCallback(),
PlotCallback(f'./results/activations/{activation.get_name()}')
])
model.fit(x_train=X_train,
y_train=y_train,
x_val=X_val,
y_val=y_val,
epochs=default_parameters['epochs'],
batch_size=default_parameters['batch_size'])
model.test(X_test, y_test)
def __init__(self,
rnn_units,
rnn_output_units,
rnn_output_activation,
mlp_units_exclude_first,
mlp_activation,
custom_params=None,
vf_share_layers=False,
use_linear_baseline=False):
keras_models.Model.__init__(self, name="keras_tesp_policy")
custom_params = custom_params or {}
self.cell = GRUCell(
units=rnn_units,
custom_params=custom_params.get("cell"))
self.dense_projection = Dense(
units=rnn_output_units,
custom_params=custom_params.get("projection"))
self.mlp_policy = KerasMLP(
layer_units_exclude_first=mlp_units_exclude_first,
activation=mlp_activation,
vf_share_layers=vf_share_layers,
name="keras_mlp_policy")
if not vf_share_layers and not use_linear_baseline:
self.mlp_vf = KerasMLP(
layer_units_exclude_first=mlp_units_exclude_first[:-1] + [1],
activation=mlp_activation,
vf_share_layers=False,
name="keras_mlp_vf")
def get_variables(cls,
rnn_input_units,
rnn_units,
rnn_output_units,
projection_kernel_init_value,
# mlp_input_units,
# mlp_units,
# vf_share_layers=False
name=None):
name = name or "keras_tesp_policy"
variables = {}
with tf.name_scope(name):
variables["cell"] = GRUCell.get_variables(
rnn_input_units, rnn_units)
variables["projection"] = Dense.get_variables(
rnn_units,
rnn_output_units,
name="dense_projection",
kernel_initializer=normc_initializer(projection_kernel_init_value))
# variables["mlp"] = KerasMLP.get_dummy_variables(
# [mlp_input_units + rnn_output_units] + mlp_units,
# vf_share_layers=vf_share_layers, **mlp_kwargs)
# graph = tf.get_default_graph()
# K.PER_GRAPH_LAYER_NAME_UIDS[graph][
# ("", dummy_instance.name)] -= 1
return variables
def __init__(self,
layer_units_exclude_first,
activation,
custom_params=None,
vf_share_layers=False,
name=None):
"""
layer_units: list, a list of the number of units of all layers
except the input layer
"""
name = name or "keras_mlp_policy"
if "policy" in name:
last_kernel_init_value = 0.01
elif "vf" in name:
last_kernel_init_value = 0.01
else:
raise NotImplementedError
keras_models.Model.__init__(self, name=name)
custom_params = custom_params or {}
for i, size in enumerate(layer_units_exclude_first):
name = f"dense_{i}"
layer = Dense(
size,
custom_params=custom_params.get(name),
activation=(activation if
i < len(layer_units_exclude_first) - 1 else None),
kernel_initializer=(normc_initializer(1.0) if
i < len(layer_units_exclude_first) - 1 else
normc_initializer(last_kernel_init_value)),
name=name)
setattr(self, name, layer)
if vf_share_layers:
name = f"dense_vf"
layer = Dense(1,
custom_params=custom_params.get(name),
activation=None,
kernel_initializer=normc_initializer(1.0),
name=name)
setattr(self, name, layer)
self._vf_share_layers = vf_share_layers
def make_cnn(input_dim, num_of_classes):
conv1 = Convolution(input_dim=input_dim,
pad=2,
stride=2,
num_filters=10,
filter_size=3,
seed=1)
relu1 = Relu()
maxpool1 = Maxpool(input_dim=conv1.output_dim, filter_size=2, stride=1)
flatten = Flatten(seed=1)
dense1 = Dense(input_dim=np.prod(maxpool1.output_dim),
output_dim=num_of_classes,
seed=1)
layers = [conv1, relu1, maxpool1, flatten, dense1]
return layers
import numpy as np
from core.model import Model
from layers.input import Input
from layers.dense import Dense
from util.cost_functions import L2
if __name__ == '__main__':
# demo MLP
data_x = np.array([1, 2])
data_y = np.array([0.2, 0.4])
train_x = np.reshape(data_x, (len(data_x), 1, 1))
train_y = np.reshape(data_y, (len(data_y), 1, 1))
model = Model()
model.add(Input(1))
model.add(Dense(3))
model.add(Dense(1))
model.compile(cost=L2(),
optimizer='sgd',
num_epochs=30000,
batch_size=1,
lr=0.1)
model.train(train_x, train_y)
test_data_x = np.array([1])
test_x = np.reshape(test_data_x, (len(test_data_x), 1))
print model.predict(test_x)
# Flatten(),
# Dense(layer_size=50, activation_func=ReLu(), weight_initializer=HeInitializer()),
# Dense(layer_size=10, activation_func=Softmax(), weight_initializer=HeInitializer())
# ]
# },
{
'test_name':
'normal_C3x3-F2_MP2x2_F_D50_D10',
'layers': [
Convolution2D(num_of_filters=2,
kernel=(3, 3),
activation_func=ReLu()),
MaxPooling2D(pool_size=(2, 2), stride=(2, 2)),
Flatten(),
Dense(layer_size=50,
activation_func=ReLu(),
weight_initializer=HeInitializer()),
Dense(layer_size=10,
activation_func=Softmax(),
weight_initializer=HeInitializer())
]
},
{
'test_name':
'normal_C3x3-F4_MP2x2_F_D50_D10',
'layers': [
Convolution2D(num_of_filters=4,
kernel=(3, 3),
activation_func=ReLu()),
MaxPooling2D(pool_size=(2, 2), stride=(2, 2)),
Flatten(),
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
import numpy as np
from layers.activation import Activation
from layers.dense import Dense
from model.model import Model
X_train = np.array([[[0, 0]], [[0, 1]], [[1, 0]], [[1, 1]]])
y_train = np.array([
[[0]], # F
[[1]], # M
[[1]], # M
[[0]] # F
])
model = Model()
model.add(Dense(2, 2))
model.add(Activation('tanh'))
model.add(Dense(2, 2))
model.add(Activation('tanh'))
model.add(Activation('softmax'))
model.fit(X_train, y_train, epochs=1000, learning_rate=0.1)
frank = np.array(X_train) # 155 pounts, 68 inches
pred = model.predict(frank)
print(np.array(pred))
def build_model(self,
in_dim,
h_dim,
out_dim,
model=None,
weight_decay=0.01,
optimizer='adadelta'):
x_v = T.tensor3('x_v', dtype=data_type) # batch_size x dim
x_s = T.tensor3('x_s', dtype=data_type) # Nt-1 x batch_size x dim
y = T.tensor3('y', dtype=data_type) # Nt x batch_size x dim
mask_v = T.matrix(
'mask_v', dtype=data_type) # Nt x batch_size mask_sent mask_video
mask_s = T.matrix('mask_s', dtype=data_type)
deterministic = T.scalar('deterministic', dtype=data_type)
one = T.constant(1, dtype=data_type)
zero = T.constant(0, dtype=data_type)
lr = T.scalar('lr', dtype=data_type)
mask_gen = T.matrix('mask_gen', dtype=data_type)
mb_BOS = T.vector('mb_BOS', dtype=data_type)
#maxlen = T.scalar('ml',dtype='int64')
# layers
l_lstm_f = LSTM(4096, h_dim)
l_lstm_b = LSTM(4096, h_dim)
l_lstm_v = LSTM_SA(h_dim * 2, h_dim, 4096, h_dim * 2)
l_word_em = Embedding(out_dim, in_dim)
l_lstm_t = LSTM_SA(in_dim, h_dim, h_dim, h_dim)
l_map = Dense(h_dim, out_dim)
layers = [l_lstm_f, l_lstm_b, l_lstm_v, l_word_em, l_lstm_t, l_map]
# drop_layers
l_drop_xv = drop(0.2, deterministic=deterministic)
l_drop_em = drop(0.2, deterministic=deterministic)
l_drop_t = drop(0.5, deterministic=deterministic)
l_drop_f = drop(0.5, deterministic=deterministic)
l_drop_b = drop(0.5, deterministic=deterministic)
x_v = l_drop_xv.get_outputs(x_v)
# forward pass
out_lstm_f, _ = l_lstm_f.get_outputs(x_v, mask_v)
out_lstm_f = l_drop_f.get_outputs(out_lstm_f)
out_lstm_b, _ = l_lstm_b.get_outputs(x_v[::-1], mask_v[::-1])
out_lstm_b = l_drop_b.get_outputs(out_lstm_b)
in_lstm_v = T.concatenate([out_lstm_f, out_lstm_b[::-1]], axis=2)
out_lstm_v, c_v = l_lstm_v.get_outputs(in_lstm_v,
context=x_v,
mask_x=mask_v,
mask_c=mask_v)
out_word_em = l_word_em.get_outputs(x_s)
out_word_em = l_drop_em.get_outputs(out_word_em)
out_lstm_t, _ = l_lstm_t.get_outputs(out_word_em,
mask_x=mask_s,
context=out_lstm_v,
mask_c=mask_v,
h0=out_lstm_v[-1],
c0=c_v[-1])
out_lstm_t = l_drop_t.get_outputs(out_lstm_t)
out_map = l_map.get_outputs(out_lstm_t)
pred, _ = theano.scan(NN.softmax, sequences=out_map)
# cost caculating
cost_o, _ = theano.scan(NN.categorical_crossentropy,
sequences=[pred, y])
cost_o = cost_o * mask_s
cost_o = cost_o.sum() / mask_s.sum()
params_re = []
for l in layers:
params_re += l.regulariable
cost_w = 0.5 * weight_decay * l2(params_re) / mask_s.sum()
cost = cost_o + cost_w
self.params = l_lstm_f.params + l_lstm_b.params + l_lstm_v.params + l_word_em.params + l_lstm_t.params + l_map.params
p_ctx = T.dot(out_lstm_v, l_lstm_t.Wpc) + l_lstm_t.bc
def _step(x_t, h_tm1, c_tm1, p_ctx, context):
m_t = T.ones_like(x_t).astype(data_type)
o_em_t = l_word_em.W[x_t.astype('int64')]
h_t_t, c_t_t = l_lstm_t._step(m_t, o_em_t, h_tm1, c_tm1, p_ctx,
context, mask_v)
o_map_t = l_map._step(h_t_t)
prob_p = NN.softmax(o_map_t)
word_t = T.argmax(prob_p, axis=1).astype(
data_type) # return an integer, the index of max value
return word_t, h_t_t, c_t_t
[words_idx, out_val_t,
c_t], _ = theano.scan(_step,
outputs_info=[
dict(initial=mb_BOS),
dict(initial=out_lstm_v[-1]),
dict(initial=c_v[-1])
],
non_sequences=[p_ctx, out_lstm_v],
n_steps=self.t_maxlen)
val_model = theano.function(inputs=[x_v, mask_v, mb_BOS],
givens={deterministic: one},
outputs=words_idx)
self.optimizer = optimizers.get(optimizer)
grads = self.optimizer.get_gradients(cost, self.params)
updates = self.optimizer.get_updates(cost, self.params)
train_model = theano.function(inputs=[x_v, x_s, y, mask_v, mask_s],
outputs=[cost, cost_o],
updates=updates,
givens={deterministic: zero})
return train_model, val_model