def set_update_function(recurrent_model,
output_model,
recurrent_optimizer,
output_optimizer):
# set input data (time_step*num_samples*features)
input_seq = tensor.tensor3(name='input_seq', dtype=floatX)
# set target data (time_step*num_samples*output_size)
target_seq = tensor.tensor3(name='target_seq', dtype=floatX)
# truncate grad
truncate_grad_step = tensor.scalar(name='truncate_grad_step', dtype='int32')
# get hidden data
hidden_seq = get_tensor_output(input=[input_seq, None, None, truncate_grad_step], layers=recurrent_model, is_training=True)
# get prediction data
output_seq = get_tensor_output(input=hidden_seq, layers=output_model, is_training=True)
# get cost (here mask_seq is like weight, sum over feature)
sequence_cost = tensor.sqr(output_seq-target_seq)
sample_cost = tensor.sum(sequence_cost, axis=(0, 2))
# get model updates
recurrent_cost = sample_cost.mean()
recurrent_updates_dict = get_model_updates(layers=recurrent_model,
cost=recurrent_cost,
optimizer=recurrent_optimizer,
use_grad_clip=1.0)
output_cost = sample_cost.mean()
output_updates_dict = get_model_updates(layers=output_model,
cost=output_cost,
optimizer=output_optimizer,
use_grad_clip=1.0)
update_function_inputs = [input_seq,
target_seq,
truncate_grad_step]
update_function_outputs = [hidden_seq,
output_seq,
sample_cost]
update_function = theano.function(inputs=update_function_inputs,
outputs=update_function_outputs,
updates=merge_dicts([recurrent_updates_dict, output_updates_dict]),
on_unused_input='ignore')
return update_function
def set_generator_update_function(
generator_rnn_model, generator_mean_model, generator_std_model, generator_optimizer, grad_clipping
):
# input data (time length * num_samples * input_dims)
source_data = tensor.tensor3(name="source_data", dtype=floatX)
target_data = tensor.tensor3(name="target_data", dtype=floatX)
# set generator input data list
generator_input_data_list = [source_data]
# get generator hidden data
hidden_data = generator_rnn_model[0].forward(generator_input_data_list, is_training=True)[0]
hidden_data = hidden_data.dimshuffle(0, 2, 1, 3).flatten(3)
# get generator output data
output_mean_data = get_tensor_output(input=hidden_data, layers=generator_mean_model, is_training=True)
# output_std_data = get_tensor_output(input=hidden_data,
# layers=generator_std_model,
# is_training=True)
output_std_data = 0.22
# get generator cost (time_length x num_samples x hidden_size)
generator_cost = 0.5 * tensor.inv(2.0 * tensor.sqr(output_std_data)) * tensor.sqr(output_mean_data - target_data)
generator_cost += tensor.log(output_std_data) + 0.5 * tensor.log(2.0 * numpy.pi)
generator_cost = tensor.sum(generator_cost, axis=2)
# set generator update
generator_updates_cost = generator_cost.mean()
generator_updates_dict = get_model_updates(
layers=generator_rnn_model + generator_mean_model,
cost=generator_updates_cost,
optimizer=generator_optimizer,
use_grad_clip=grad_clipping,
)
gradient_dict = get_model_gradients(generator_rnn_model + generator_mean_model, generator_updates_cost)
gradient_norm = 0.0
for grad in gradient_dict:
gradient_norm += tensor.sum(grad ** 2)
gradient_norm = tensor.sqrt(gradient_norm)
# set generator update inputs
generator_updates_inputs = [source_data, target_data]
# set generator update outputs
generator_updates_outputs = [generator_cost, gradient_norm]
# set generator update function
generator_updates_function = theano.function(
inputs=generator_updates_inputs,
outputs=generator_updates_outputs,
updates=generator_updates_dict,
on_unused_input="ignore",
)
return generator_updates_function
def set_generator_update_function(generator_rnn_model,
generator_mean_model,
generator_std_model,
generator_optimizer,
grad_clipping):
# input data (time length * num_samples * input_dims)
source_data = tensor.tensor3(name='source_data',
dtype=floatX)
target_data = tensor.tensor3(name='target_data',
dtype=floatX)
# set generator input data list
generator_input_data_list = [source_data,]
# get generator hidden data
hidden_data = generator_rnn_model[0].forward(generator_input_data_list, is_training=True)[0]
# get generator output data
output_mean_data = get_tensor_output(input=hidden_data,
layers=generator_mean_model,
is_training=True)
output_std_data = get_tensor_output(input=hidden_data,
layers=generator_std_model,
is_training=True)
generator_cost = -0.5*tensor.inv(2.0*tensor.sqr(output_std_data))*tensor.sqr(output_mean_data-target_data)
generator_cost += -0.5*tensor.log(2.0*tensor.sqr(output_std_data)*numpy.pi)
# set generator update
generator_updates_cost = generator_cost.mean()
generator_updates_dict = get_model_updates(layers=generator_rnn_model+generator_mean_model+generator_std_model,
cost=generator_updates_cost,
optimizer=generator_optimizer,
use_grad_clip=grad_clipping)
gradient_dict = get_model_gradients(generator_rnn_model+generator_mean_model+generator_std_model, generator_updates_cost)
gradient_norm = 0.
for grad in gradient_dict:
gradient_norm += tensor.sum(grad**2)
gradient_norm = tensor.sqrt(gradient_norm)
# set generator update inputs
generator_updates_inputs = [source_data,
target_data,]
# set generator update outputs
generator_updates_outputs = [generator_cost, gradient_norm]
# set generator update function
generator_updates_function = theano.function(inputs=generator_updates_inputs,
outputs=generator_updates_outputs,
updates=generator_updates_dict,
on_unused_input='ignore')
return generator_updates_function
def set_updater_function(generator_model, generator_optimizer, generator_grad_clipping):
# input sequence data (time_length * num_samples * input_dims)
input_sequence = tensor.tensor3(name="input_sequence", dtype=floatX)
target_sequence = tensor.tensor3(name="target_sequence", dtype=floatX)
lambda_regularizer = tensor.scalar(name="lambda_regularizer", dtype=floatX)
# set generator input data list
generator_input_data_list = [input_sequence]
# get generator output data
generator_output = generator_model[0].forward(generator_input_data_list, is_training=True)
output_sequence = generator_output[0]
data_hidden = generator_output[1]
data_cell = generator_output[2]
model_hidden = generator_output[3]
model_cell = generator_output[4]
generator_random = generator_output[-1]
# get square error
sample_cost = tensor.sqr(target_sequence - output_sequence).sum(axis=2)
# get positive phase hidden
positive_hid = data_hidden[1:]
# get negative phase hidden
negative_hid = model_hidden[1:]
# get phase diff cost
regularizer_cost = tensor.sqr(positive_hid - negative_hid).sum(axis=2)
# set generator update
updater_cost = sample_cost.mean() + regularizer_cost.mean() * lambda_regularizer
updater_dict = get_model_updates(layers=generator_model, cost=updater_cost, optimizer=generator_optimizer)
# get generator gradient norm2
generator_gradient_dict = get_model_gradients(generator_model, updater_cost)
generator_gradient_norm = 0.0
for grad in generator_gradient_dict:
generator_gradient_norm += tensor.sum(grad ** 2)
generator_gradient_norm = tensor.sqrt(generator_gradient_norm)
# set updater inputs
updater_inputs = [input_sequence, target_sequence, lambda_regularizer]
# set updater outputs
updater_outputs = [sample_cost, regularizer_cost, generator_gradient_norm]
# set updater function
updater_function = theano.function(
inputs=updater_inputs,
outputs=updater_outputs,
updates=merge_dicts([updater_dict, generator_random]),
on_unused_input="ignore",
)
return updater_function
def set_update_function(recurrent_model,
output_model,
optimizer,
grad_clip=1.0):
# set input data (time_length * num_samples * input_dims)
input_data = tensor.tensor3(name='input_data', dtype=floatX)
# set input mask (time_length * num_samples)
input_mask = tensor.matrix(name='input_mask', dtype=floatX)
# set init hidden/cell data (num_samples * hidden_dims)
init_hidden = tensor.matrix(name='init_hidden', dtype=floatX)
init_cell = tensor.matrix(name='init_cell', dtype=floatX)
# truncate grad
truncate_grad_step = tensor.scalar(name='truncate_grad_step', dtype='int32')
# set target data (time_length * num_samples * output_dims)
target_data = tensor.tensor3(name='target_data', dtype=floatX)
# get hidden data
input_list = [input_data, None, None, None, truncate_grad_step]
hidden_data = get_tensor_output(input=input_list,
layers=recurrent_model,
is_training=True)[0]
# get prediction data
output_data = get_tensor_output(input=hidden_data,
layers=output_model,
is_training=True)
# get cost (here mask_seq is like weight, sum over feature, and time)
sample_cost = tensor.sqr(output_data-target_data)
sample_cost = tensor.sum(sample_cost, axis=(0, 2))
# get model updates
model_cost = sample_cost.mean()
model_updates_dict = get_model_updates(layers=recurrent_model+output_model,
cost=model_cost,
optimizer=optimizer,
use_grad_clip=grad_clip)
update_function_inputs = [input_data,
input_mask,
init_hidden,
init_cell,
target_data,
truncate_grad_step]
update_function_outputs = [hidden_data,
output_data,
sample_cost]
update_function = theano.function(inputs=update_function_inputs,
outputs=update_function_outputs,
updates=model_updates_dict,
on_unused_input='ignore')
return update_function
def set_updater_function(generator_rnn_model,
generator_emb_matrix,
generator_optimizer,
generator_grad_clipping):
# input/target sequence data (time_length * batch_size, list of idx)
input_sequence = tensor.matrix(name='input_sequence',
dtype=floatX)
target_sequence = tensor.matrix(name='target_sequence',
dtype=floatX)
# generator_emb_matrix.shape = (num_idx, feature_size)
input_emb_sequence = generator_emb_matrix[input_sequence]
target_emb_sequence = generator_emb_matrix[target_sequence]
# set generator input data list
generator_input_data_list = [input_emb_sequence,]
# get generator output data
generator_output = generator_rnn_model[0].forward(generator_input_data_list, is_training=True)
generator_sample = generator_output[0]
generator_random = generator_output[-1]
# get square error
square_error = tensor.sqr(target_sequence-generator_sample).sum(axis=2)
# set generator update
tf_updates_cost = square_error.mean()
tf_updates_dict = get_model_updates(layers=generator_rnn_model,
cost=tf_updates_cost,
optimizer=generator_optimizer)
generator_gradient_dict = get_model_gradients(layers=generator_rnn_model,
cost=tf_updates_cost)
generator_gradient_norm = 0.
for grad in generator_gradient_dict:
generator_gradient_norm += tensor.sum(grad**2)
generator_gradient_norm = tensor.sqrt(generator_gradient_norm)
# set tf update inputs
tf_updates_inputs = [input_sequence,
target_sequence]
# set tf update outputs
tf_updates_outputs = [square_error,
generator_gradient_norm,]
# set tf update function
tf_updates_function = theano.function(inputs=tf_updates_inputs,
outputs=tf_updates_outputs,
updates=merge_dicts([tf_updates_dict, generator_random]),
on_unused_input='ignore')
return tf_updates_function
def set_tf_update_function(generator_model,
generator_optimizer,
generator_grad_clipping):
# input sequence data (time_length * num_samples * input_dims)
input_sequence = tensor.tensor3(name='input_sequence',
dtype=floatX)
target_sequence = tensor.tensor3(name='target_sequence',
dtype=floatX)
# set generator input data list
generator_input_data_list = [input_sequence,]
# get generator output data
generator_output = generator_model[0].forward(generator_input_data_list, is_training=True)
output_sequence = generator_output[0]
generator_random = generator_output[-1]
# get square error
square_error = tensor.sqr(target_sequence-output_sequence).sum(axis=2)
# set generator update
tf_updates_cost = square_error.mean()
tf_updates_dict = get_model_updates(layers=generator_model,
cost=tf_updates_cost,
optimizer=generator_optimizer)
generator_gradient_dict = get_model_gradients(generator_model, tf_updates_cost)
# get generator gradient norm2
generator_gradient_norm = 0.
for grad in generator_gradient_dict:
generator_gradient_norm += tensor.sum(grad**2)
generator_gradient_norm = tensor.sqrt(generator_gradient_norm)
# set tf update inputs
tf_updates_inputs = [input_sequence,
target_sequence]
# set tf update outputs
tf_updates_outputs = [square_error,
generator_gradient_norm,]
# set tf update function
tf_updates_function = theano.function(inputs=tf_updates_inputs,
outputs=tf_updates_outputs,
updates=merge_dicts([tf_updates_dict,
generator_random]),
on_unused_input='ignore')
return tf_updates_function
def set_gan_update_function(generator_model,
discriminator_feature_model,
discriminator_output_model,
generator_optimizer,
discriminator_optimizer,
generator_grad_clipping,
discriminator_grad_clipping):
# input sequence data (time_length * num_samples * input_dims)
input_sequence = tensor.tensor3(name='input_sequence',
dtype=floatX)
target_sequence = tensor.tensor3(name='target_sequence',
dtype=floatX)
# set generator input data list
generator_input_data_list = [input_sequence,
1]
# get generator output data
generator_output = generator_model[0].forward(generator_input_data_list,
is_training=True)
output_sequence = generator_output[0]
data_hidden = generator_output[1]
data_cell = generator_output[2]
model_hidden = generator_output[3]
model_cell = generator_output[4]
generator_random = generator_output[-1]
# get conditional hidden
condition_hid = data_hidden[:-1]
condition_hid = theano.gradient.disconnected_grad(condition_hid)
condition_feature = get_tensor_output(condition_hid,
discriminator_feature_model,
is_training=True)
# get positive phase hidden
positive_hid = data_hidden[1:]
positive_feature = get_tensor_output(positive_hid,
discriminator_feature_model,
is_training=True)
# get negative phase hidden
negative_hid = model_hidden[1:]
negative_feature = get_tensor_output(negative_hid,
discriminator_feature_model,
is_training=True)
# get positive/negative phase pairs
positive_pair = tensor.concatenate([condition_feature, positive_feature], axis=2)
negative_pair = tensor.concatenate([condition_feature, negative_feature], axis=2)
# get positive pair score
positive_score = get_tensor_output(positive_pair,
discriminator_output_model,
is_training=True)
# get negative pair score
negative_score = get_tensor_output(negative_pair,
discriminator_output_model,
is_training=True)
# get generator cost (increase negative score)
generator_gan_cost = tensor.nnet.binary_crossentropy(output=negative_score,
target=tensor.ones_like(negative_score))
# get discriminator cost (increase positive score, decrease negative score)
discriminator_gan_cost = (tensor.nnet.binary_crossentropy(output=positive_score,
target=tensor.ones_like(positive_score)) +
tensor.nnet.binary_crossentropy(output=negative_score,
target=tensor.zeros_like(negative_score)))
# set generator update
generator_updates_cost = generator_gan_cost.mean()
generator_updates_dict = get_model_updates(layers=generator_model,
cost=generator_updates_cost,
optimizer=generator_optimizer,
use_grad_clip=generator_grad_clipping)
# get generator gradient norm2
generator_gradient_dict = get_model_gradients(generator_model, generator_updates_cost)
generator_gradient_norm = 0.
for grad in generator_gradient_dict:
generator_gradient_norm += tensor.sum(grad**2)
generator_gradient_norm = tensor.sqrt(generator_gradient_norm)
# set discriminator update
discriminator_updates_cost = discriminator_gan_cost.mean()
discriminator_updates_dict = get_model_updates(layers=discriminator_feature_model+discriminator_output_model,
cost=discriminator_updates_cost,
optimizer=discriminator_optimizer,
use_grad_clip=discriminator_grad_clipping)
discriminator_gradient_dict = get_model_gradients(discriminator_feature_model+discriminator_output_model,
discriminator_updates_cost)
# get discriminator gradient norm2
discriminator_gradient_norm = 0.
for grad in discriminator_gradient_dict:
discriminator_gradient_norm += tensor.sum(grad**2)
discriminator_gradient_norm = tensor.sqrt(discriminator_gradient_norm)
# get mean square error
square_error = tensor.sqr(target_sequence-output_sequence).sum(axis=2)
# set gan update inputs
gan_updates_inputs = [input_sequence,
target_sequence]
# set gan update outputs
gan_updates_outputs = [generator_gan_cost,
discriminator_gan_cost,
positive_score,
negative_score,
square_error,
generator_gradient_norm,
discriminator_gradient_norm,]
# set gan update function
gan_updates_function = theano.function(inputs=gan_updates_inputs,
outputs=gan_updates_outputs,
updates=merge_dicts([generator_updates_dict,
discriminator_updates_dict,
generator_random]),
on_unused_input='ignore')
return gan_updates_function
def set_gan_update_function(generator_model,
discriminator_model,
generator_optimizer,
discriminator_optimizer,
generator_grad_clipping,
discriminator_grad_clipping):
# input sequence data (time_length * num_samples * input_dims)
input_sequence = tensor.tensor3(name='input_sequence',
dtype=floatX)
target_sequence = tensor.tensor3(name='target_sequence',
dtype=floatX)
# set generator input data list
generator_input_data_list = [input_sequence,]
# get generator output data
output_data_set = generator_model[0].forward(generator_input_data_list, is_training=True)
output_sequence = output_data_set[0]
data_hidden = output_data_set[1]
data_cell = output_data_set[2]
model_hidden = output_data_set[3]
model_cell = output_data_set[4]
condition_hidden = data_hidden[:-1]
condition_cell = data_cell[:-1]
condition_hidden = theano.gradient.disconnected_grad(condition_hidden)
condition_cell = theano.gradient.disconnected_grad(condition_cell)
true_hidden = data_hidden[1:]
true_cell = data_cell[1:]
false_hidden = model_hidden[1:]
false_cell = model_cell[1:]
true_pair_hidden = tensor.concatenate([condition_hidden, true_hidden], axis=2)
true_pair_cell = tensor.concatenate([condition_cell, true_cell], axis=2)
false_pair_hidden = tensor.concatenate([condition_hidden, false_hidden], axis=2)
false_pair_cell = tensor.concatenate([condition_cell, false_cell], axis=2)
discriminator_true_score = get_tensor_output(true_pair_hidden, discriminator_model, is_training=True)
discriminator_false_score = get_tensor_output(false_pair_hidden, discriminator_model, is_training=True)
generator_gan_cost = tensor.nnet.binary_crossentropy(output=discriminator_false_score,
target=tensor.ones_like(discriminator_false_score))
discriminator_gan_cost = (tensor.nnet.binary_crossentropy(output=discriminator_true_score,
target=tensor.ones_like(discriminator_true_score)) +
tensor.nnet.binary_crossentropy(output=discriminator_false_score,
target=tensor.zeros_like(discriminator_false_score)))
# set generator update
generator_updates_cost = generator_gan_cost.mean()
generator_updates_dict = get_model_updates(layers=generator_model,
cost=generator_updates_cost,
optimizer=generator_optimizer,
use_grad_clip=generator_grad_clipping)
generator_gradient_dict = get_model_gradients(generator_model, generator_updates_cost)
generator_gradient_norm = 0.
for grad in generator_gradient_dict:
generator_gradient_norm += tensor.sum(grad**2)
generator_gradient_norm = tensor.sqrt(generator_gradient_norm)
# set discriminator update
discriminator_updates_cost = discriminator_gan_cost.mean()
discriminator_updates_dict = get_model_updates(layers=discriminator_model,
cost=discriminator_updates_cost,
optimizer=discriminator_optimizer,
use_grad_clip=discriminator_grad_clipping)
discriminator_gradient_dict = get_model_gradients(discriminator_model, discriminator_updates_cost)
discriminator_gradient_norm = 0.
for grad in discriminator_gradient_dict:
discriminator_gradient_norm += tensor.sum(grad**2)
discriminator_gradient_norm = tensor.sqrt(discriminator_gradient_norm)
square_error = tensor.sqr(target_sequence-output_sequence).sum(axis=2)
# set gan update inputs
gan_updates_inputs = [input_sequence,
target_sequence]
# set gan update outputs
gan_updates_outputs = [generator_gan_cost,
discriminator_gan_cost,
discriminator_true_score,
discriminator_false_score,
square_error,
generator_gradient_norm,
discriminator_gradient_norm,]
# set gan update function
gan_updates_function = theano.function(inputs=gan_updates_inputs,
outputs=gan_updates_outputs,
updates=merge_dicts([generator_updates_dict, discriminator_updates_dict]),
on_unused_input='ignore')
return gan_updates_function
def set_gan_update_function(generator_rnn_model,
generator_output_model,
discriminator_rnn_model,
discriminator_output_model,
generator_optimizer,
discriminator_optimizer,
generator_grad_clipping,
discriminator_grad_clipping):
# input sequence data (time_length * num_samples * input_dims)
input_sequence = tensor.tensor3(name='input_sequence',
dtype=floatX)
target_sequence = tensor.tensor3(name='target_sequence',
dtype=floatX)
# set generator input data list
generator_input_data_list = [input_sequence,]
# get generator output data
generator_output = generator_rnn_model[0].forward(generator_input_data_list, is_training=True)
generator_hidden = generator_output[0]
generator_cell = generator_output[1]
generator_sample = get_tensor_output(generator_hidden, generator_output_model, is_training=True)
condition_generator_hidden = theano.gradient.disconnected_grad(generator_hidden)
positive_pair = tensor.concatenate([condition_generator_hidden, target_sequence], axis=2)
negative_pair = tensor.concatenate([condition_generator_hidden, generator_sample], axis=2)
# set generator input data list
discriminator_input_data_list = [positive_pair,]
discriminator_output = discriminator_rnn_model[0].forward(discriminator_input_data_list, is_training=True)
positive_hidden = discriminator_output[0]
positive_cell = discriminator_output[1]
positive_score = get_tensor_output(positive_hidden, discriminator_output_model, is_training=True)
discriminator_input_data_list = [negative_pair,]
discriminator_output = discriminator_rnn_model[0].forward(discriminator_input_data_list, is_training=True)
negative_hidden = discriminator_output[0]
negative_cell = discriminator_output[1]
negative_score = get_tensor_output(negative_hidden, discriminator_output_model, is_training=True)
generator_gan_cost = tensor.nnet.binary_crossentropy(output=negative_score,
target=tensor.ones_like(negative_score))
discriminator_gan_cost = (tensor.nnet.binary_crossentropy(output=positive_score,
target=tensor.ones_like(positive_score)) +
tensor.nnet.binary_crossentropy(output=negative_score,
target=tensor.zeros_like(negative_score)))
# set generator update
generator_updates_cost = generator_gan_cost.mean()
generator_updates_dict = get_model_updates(layers=generator_rnn_model+generator_output_model,
cost=generator_updates_cost,
optimizer=generator_optimizer,
use_grad_clip=generator_grad_clipping)
generator_gradient_dict = get_model_gradients(generator_rnn_model+generator_output_model, generator_updates_cost)
generator_gradient_norm = 0.
for grad in generator_gradient_dict:
generator_gradient_norm += tensor.sum(grad**2)
generator_gradient_norm = tensor.sqrt(generator_gradient_norm)
# set discriminator update
discriminator_updates_cost = discriminator_gan_cost.mean()
discriminator_updates_dict = get_model_updates(layers=discriminator_rnn_model+discriminator_output_model,
cost=discriminator_updates_cost,
optimizer=discriminator_optimizer,
use_grad_clip=discriminator_grad_clipping)
discriminator_gradient_dict = get_model_gradients(discriminator_rnn_model+discriminator_output_model, discriminator_updates_cost)
discriminator_gradient_norm = 0.
for grad in discriminator_gradient_dict:
discriminator_gradient_norm += tensor.sum(grad**2)
discriminator_gradient_norm = tensor.sqrt(discriminator_gradient_norm)
square_error = tensor.sqr(target_sequence-generator_sample).sum(axis=2)
# set gan update inputs
gan_updates_inputs = [input_sequence,
target_sequence]
# set gan update outputs
gan_updates_outputs = [generator_gan_cost,
discriminator_gan_cost,
positive_score,
negative_score,
square_error,
generator_gradient_norm,
discriminator_gradient_norm,]
# set gan update function
gan_updates_function = theano.function(inputs=gan_updates_inputs,
outputs=gan_updates_outputs,
updates=merge_dicts([generator_updates_dict, discriminator_updates_dict]),
on_unused_input='ignore')
return gan_updates_function
def set_reg_update_function(generator_model,
generator_optimizer,
generator_grad_clipping):
# input sequence data (time_length * num_samples * input_dims)
input_sequence = tensor.tensor3(name='input_sequence',
dtype=floatX)
target_sequence = tensor.tensor3(name='target_sequence',
dtype=floatX)
# set generator input data list
generator_input_data_list = [input_sequence,
1]
# get generator output data
generator_output = generator_model[0].forward(generator_input_data_list,
is_training=True)
output_sequence = generator_output[0]
data_hidden = generator_output[1]
data_cell = generator_output[2]
model_hidden = generator_output[3]
model_cell = generator_output[4]
generator_random = generator_output[-1]
# get positive phase hidden
positive_hid = data_hidden[1:]
positive_hid = theano.gradient.disconnected_grad(positive_hid)
# get negative phase hidden
negative_hid = model_hidden[1:]
# get phase diff cost
phase_diff = tensor.sqr(positive_hid-negative_hid).sum(axis=2)
# set generator update
generator_updates_cost = phase_diff.mean()
generator_updates_dict = get_model_updates(layers=generator_model,
cost=generator_updates_cost,
optimizer=generator_optimizer,
use_grad_clip=generator_grad_clipping)
# get generator gradient norm2
generator_gradient_dict = get_model_gradients(generator_model, generator_updates_cost)
generator_gradient_norm = 0.
for grad in generator_gradient_dict:
generator_gradient_norm += tensor.sum(grad**2)
generator_gradient_norm = tensor.sqrt(generator_gradient_norm)
# get mean square error
square_error = tensor.sqr(target_sequence-output_sequence).sum(axis=2)
# set reg update inputs
reg_updates_inputs = [input_sequence,
target_sequence]
# set reg update outputs
reg_updates_outputs = [phase_diff,
square_error,
generator_gradient_norm,]
# set reg update function
reg_updates_function = theano.function(inputs=reg_updates_inputs,
outputs=reg_updates_outputs,
updates=merge_dicts([generator_updates_dict,
generator_random]),
on_unused_input='ignore')
return reg_updates_function
def set_generator_update_function(generator_rnn_model,
generator_output_models,
generator_optimizer,
grad_clipping):
# set source data (time_length * num_samples * input_dims)
source_data = tensor.tensor3(name='source_data',
dtype=floatX)
# set target data (time_length * num_samples * input_dims)
target_data = tensor.bmatrix(name='target_data')
# set generator input data list
generator_input_data_list = [source_data,]
# get generator output data
hidden_data = generator_rnn_model[0].forward(generator_input_data_list, is_training=True)[0]
# for each rnn layer
output_data_list = []
for l, output_model in enumerate(generator_output_models):
output_data = get_tensor_output(input=hidden_data[l],
layers=output_model,
is_training=True)
output_data_list.append(output_data)
output_data = tensor.concatenate(output_data_list[::-1], axis=1)
output_sign_data = output_data[:,0]
output_sign_data = 2.0*output_sign_data-tensor.ones_like(output_sign_data)
output_value_data = output_data[:,1:]
output_value_data = output_value_data*tensor.pow(2.0, tensor.arange(output_value_data.shape[1]))
output_value_data = output_sign_data*output_value_data.sum(axis=1)
target_sign_data = target_data[:,0]
target_sign_data = 2.0*target_sign_data-tensor.ones_like(target_sign_data)
target_value_data = target_data[:,1:]
target_value_data = target_value_data*tensor.pow(2.0, tensor.arange(target_value_data.shape[1]))
target_value_data = target_sign_data*target_value_data.sum(axis=1)
mse_cost = tensor.sqr(output_value_data, target_value_data)
bce_cost = tensor.nnet.binary_crossentropy(output_data, target_data).sum(axis=1)
# set generator update
generator_updates_cost = generator_cost.mean()
generator_updates_dict = get_model_updates(layers=generator_rnn_model+,
cost=generator_updates_cost,
optimizer=generator_optimizer,
use_grad_clip=grad_clipping)
# gradient_dict = get_model_gradients(generator_rnn_model, generator_updates_cost)
# gradient_norm = 0.
# for grad in gradient_dict:
# gradient_norm += tensor.sum(grad**2)
# gradient_norm = tensor.sqrt(gradient_norm)
# set generator update inputs
generator_updates_inputs = [init_input_data,
init_hidden_data,
init_cell_data,
sampling_length]
# set generator update outputs
generator_updates_outputs = [sample_cost_data, generator_cost, ]#gradient_norm]
# set generator update function
generator_updates_function = theano.function(inputs=generator_updates_inputs,
outputs=generator_updates_outputs,
updates=merge_dicts([generator_updates_dict, update_data]),
on_unused_input='ignore')
return generator_updates_function
def set_update_function(recurrent_model,
output_model,
controller_optimizer,
model_optimizer,
grad_clip=1.0):
# set input data (time_length * num_samples * input_dims)
input_data = tensor.tensor3(name='input_data', dtype=floatX)
# set target data (time_length * num_samples * output_dims)
target_data = tensor.tensor3(name='target_data', dtype=floatX)
time_length = input_data.shape[0]
num_samples = input_data.shape[1]
# cost control parameter
controller = theano.shared(value=1.0,
name='controller')
# get hidden data
input_list = [input_data, ]
hidden_data = get_lstm_outputs(input_list=input_list,
layers=recurrent_model,
is_training=True)[-1]
# get prediction data
output_data = get_tensor_output(input=hidden_data,
layers=output_model,
is_training=True)
# get cost (here mask_seq is like weight, sum over feature, and time)
sample_cost = tensor.sqr(output_data-target_data)
sample_cost = tensor.sum(input=sample_cost, axis=2).reshape((time_length, num_samples))
# time_step = tensor.arange(start=0, stop=time_length, dtype=floatX).reshape((time_length, 1))
# time_step = tensor.repeat(time_step, num_samples, axis=1)
# cost_weight (time_length * num_samples)
# cost_weight = tensor.transpose(-controller*time_step)
# cost_weight = tensor.nnet.softmax(cost_weight)
# cost_weight = tensor.transpose(cost_weight).reshape((time_length, num_samples))
# weighted_sample_cost = cost_weight*sample_cost
# get model updates
# model_cost = weighted_sample_cost.sum(axis=0).mean()
model_cost = sample_cost.max(axis=0).mean()
model_updates_dict = get_model_updates(layers=recurrent_model+output_model,
cost=model_cost,
optimizer=model_optimizer,
use_grad_clip=grad_clip)
# controller_cost = weighted_sample_cost.var(axis=0).mean()
#
# controller_updates_dict = OrderedDict()
# controller_grad = tensor.grad(cost=controller_cost, wrt=controller)
# for param, update in controller_optimizer(controller, controller_grad).iteritems():
# controller_updates_dict[param] = update
update_function_inputs = [input_data,
target_data]
update_function_outputs = [hidden_data,
output_data,
sample_cost]
# update_function_updates = merge_dicts([model_updates_dict, controller_updates_dict])
update_function_updates = model_updates_dict
update_function = theano.function(inputs=update_function_inputs,
outputs=update_function_outputs,
updates=update_function_updates,
on_unused_input='ignore')
return update_function
def set_generator_update_function(generator_rnn_model,
discriminator_rnn_model,
discriminator_output_model,
generator_optimizer,
grad_clipping):
# init input data (num_samples *input_dims)
init_input_data = tensor.matrix(name='init_input_data',
dtype=floatX)
# init hidden data (num_layers * num_samples *input_dims)
init_hidden_data = tensor.tensor3(name='init_hidden_data',
dtype=floatX)
# init cell data (num_layers * num_samples *input_dims)
init_cell_data = tensor.tensor3(name='init_cell_data',
dtype=floatX)
# sampling length
sampling_length = tensor.scalar(name='sampling_length',
dtype='int32')
# set generator input data list
generator_input_data_list = [init_input_data,
init_hidden_data,
init_cell_data,
sampling_length]
# get generator output data
output_data = generator_rnn_model[0].forward(generator_input_data_list, is_training=True)[0]
# set discriminator input data list
discriminator_input_data_list = [output_data,]
# get discriminator hidden data
discriminator_hidden_data = get_lstm_outputs(input_list=discriminator_input_data_list,
layers=discriminator_rnn_model,
is_training=True)[-1]
# get discriminator output data
sample_cost_data = get_tensor_output(input=discriminator_hidden_data,
layers=discriminator_output_model,
is_training=True)[-1]
# get cost based on discriminator (binary cross-entropy over all data)
# sum over generator cost over time_length and output_dims, then mean over samples
generator_cost = tensor.nnet.binary_crossentropy(output=sample_cost_data,
target=tensor.ones_like(sample_cost_data)).sum(axis=1)
# set generator update
generator_updates_cost = generator_cost.mean()
generator_updates_dict = get_model_updates(layers=generator_rnn_model,
cost=generator_updates_cost,
optimizer=generator_optimizer,
use_grad_clip=grad_clipping)
# set generator update inputs
generator_updates_inputs = [init_input_data,
init_hidden_data,
init_cell_data,
sampling_length]
# set generator update outputs
generator_updates_outputs = [sample_cost_data, generator_cost]
# set generator update function
generator_updates_function = theano.function(inputs=generator_updates_inputs,
outputs=generator_updates_outputs,
updates=generator_updates_dict,
on_unused_input='ignore')
return generator_updates_function
def set_gan_update_function(generator_rnn_model,
discriminator_rnn_model,
discriminator_output_model,
generator_optimizer,
discriminator_optimizer,
generator_grad_clipping,
discriminator_grad_clipping):
# input for loop forward
input_sequence = tensor.tensor3(name='input_sequence',
dtype=floatX)
time_length = tensor.scalar(name='time_length',
dtype='int32')
# get init data for looping
generator_output = generator_rnn_model[0].loop_forward([input_sequence[0], time_length])
generator_sequence = generator_output[0]
generator_rand_update = generator_output[-1]
discriminator_output = discriminator_rnn_model[0].forward([input_sequence, ], is_training=True)
positive_hidden = discriminator_output[0]
positive_score = get_tensor_output(positive_hidden, discriminator_output_model, is_training=True)
discriminator_output = discriminator_rnn_model[0].forward([generator_sequence, ], is_training=True)
negative_hidden = discriminator_output[0]
negative_score = get_tensor_output(negative_hidden, discriminator_output_model, is_training=True)
generator_gan_cost = tensor.nnet.binary_crossentropy(output=negative_score,
target=tensor.ones_like(negative_score))
discriminator_gan_cost = (tensor.nnet.binary_crossentropy(output=positive_score,
target=tensor.ones_like(positive_score)) +
tensor.nnet.binary_crossentropy(output=negative_score,
target=tensor.zeros_like(negative_score)))
# set generator update
generator_updates_cost = generator_gan_cost.mean()
generator_updates_dict = get_model_updates(layers=generator_rnn_model,
cost=generator_updates_cost,
optimizer=generator_optimizer,
use_grad_clip=generator_grad_clipping)
generator_gradient_dict = get_model_gradients(layers=generator_rnn_model,
cost=generator_updates_cost)
generator_gradient_norm = 0.
for grad in generator_gradient_dict:
generator_gradient_norm += tensor.sum(grad**2)
generator_gradient_norm = tensor.sqrt(generator_gradient_norm)
# set discriminator update
discriminator_updates_cost = discriminator_gan_cost.mean()
discriminator_updates_dict = get_model_updates(layers=discriminator_rnn_model+discriminator_output_model,
cost=discriminator_updates_cost,
optimizer=discriminator_optimizer,
use_grad_clip=discriminator_grad_clipping)
discriminator_gradient_dict = get_model_gradients(layers=discriminator_rnn_model+discriminator_output_model,
cost=discriminator_updates_cost)
discriminator_gradient_norm = 0.
for grad in discriminator_gradient_dict:
discriminator_gradient_norm += tensor.sum(grad**2)
discriminator_gradient_norm = tensor.sqrt(discriminator_gradient_norm)
# set gan update inputs
gan_updates_inputs = [input_sequence,
time_length]
# set gan update outputs
gan_updates_outputs = [generator_gan_cost,
discriminator_gan_cost,
positive_score,
negative_score,
generator_gradient_norm,
discriminator_gradient_norm,]
# set gan update function
gan_updates_function = theano.function(inputs=gan_updates_inputs,
outputs=gan_updates_outputs,
updates=merge_dicts([generator_updates_dict,
discriminator_updates_dict,
generator_rand_update]),
on_unused_input='ignore')
return gan_updates_function