def load_models(model_path=save_path,
in_size=len(input_columns),
out_size=len(output_columns) -
1 if cost_mode == 'RL-MDN' else len(output_columns),
hidden_size=hidden_size,
num_recurrent_layers=num_recurrent_layers,
model=layer_models[0]):
initials = []
if not os.path.isfile(model_path):
print 'Could not find model file.'
sys.exit(0)
print 'Loading model from {0}...'.format(model_path)
x = tensor.tensor3('features', dtype=theano.config.floatX)
y = tensor.tensor3('targets', dtype='floatX')
train_flag = [theano.shared(0)]
_, latent_size = load_encoder()
in_size = latent_size + len(input_columns)
y_hat, cost, cells = nn_fprop(x, y, in_size, out_size, hidden_size,
num_recurrent_layers, train_flag)
main_loop = MainLoop(algorithm=None,
data_stream=None,
model=Model(cost),
extensions=[saveload.Load(model_path)])
for extension in main_loop.extensions:
extension.main_loop = main_loop
main_loop._run_extensions('before_training')
bin_model = main_loop.model
print 'Model loaded. Building prediction function...'
hiddens = []
for i in range(num_recurrent_layers):
brick = [
b for b in bin_model.get_top_bricks()
if b.name == layer_models[i] + str(i)
][0]
hiddens.extend(
VariableFilter(theano_name=brick.name + '_apply_states')(
bin_model.variables))
hiddens.extend(
VariableFilter(theano_name=brick.name + '_apply_cells')(cells))
initials.extend(
VariableFilter(roles=[roles.INITIAL_STATE])(brick.parameters))
predict_func = theano.function([x], hiddens + [y_hat])
encoder, code_size = load_encoder()
return predict_func, initials, encoder, code_size
step_rules = [RMSProp(learning_rate=learning_rate, decay_rate=decay_rate),
StepClipping(step_clipping)]
algorithm = GradientDescent(cost=cost, parameters=cg.parameters,
step_rule=CompositeRule(step_rules))
# Extensions
gradient_norm = aggregation.mean(algorithm.total_gradient_norm)
step_norm = aggregation.mean(algorithm.total_step_norm)
monitored_vars = [cost, gradient_norm, step_norm]
dev_monitor = DataStreamMonitoring(variables=[cost], after_epoch=True,
before_first_epoch=True, data_stream=dev_stream, prefix="dev")
train_monitor = TrainingDataMonitoring(variables=monitored_vars, after_batch=True,
before_first_epoch=True, prefix='tra')
extensions = [dev_monitor, train_monitor, Timing(), Printing(after_batch=True),
FinishAfter(after_n_epochs=nepochs),
saveload.Load(load_path),
saveload.Checkpoint(last_path),
] + track_best('dev_cost', save_path)
if learning_rate_decay not in (0, 1):
extensions.append(SharedVariableModifier(step_rules[0].learning_rate,
lambda n, lr: numpy.cast[theano.config.floatX](learning_rate_decay * lr), after_epoch=True, after_batch=False))
print 'number of parameters in the model: ' + str(tensor.sum([p.size for p in cg.parameters]).eval())
# Finally build the main loop and train the model
main_loop = MainLoop(data_stream=train_stream, algorithm=algorithm,
model=Model(cost), extensions=extensions)
main_loop.run()
step_rules = [RMSProp(learning_rate=learning_rate, decay_rate=decay_rate),
StepClipping(step_clipping)]
algorithm = GradientDescent(cost=cost, parameters=cg.parameters,
step_rule=CompositeRule(step_rules))
# Extensions
gradient_norm = aggregation.mean(algorithm.total_gradient_norm)
step_norm = aggregation.mean(algorithm.total_step_norm)
monitored_vars = [cost, gradient_norm, step_norm]
dev_monitor = DataStreamMonitoring(variables=[cost], after_epoch=True,
before_first_epoch=True, data_stream=dev_stream, prefix="dev")
train_monitor = TrainingDataMonitoring(variables=monitored_vars, after_batch=True,
before_first_epoch=True, prefix='tra')
extensions = [dev_monitor, train_monitor, Timing(), Printing(after_batch=True),
FinishAfter(after_n_epochs=nepochs),
saveload.Load(load_path),
saveload.Checkpoint(last_path),
] + track_best('dev_cost', save_path)
if learning_rate_decay not in (0, 1):
extensions.append(SharedVariableModifier(step_rules[0].learning_rate,
lambda n, lr: numpy.cast[theano.config.floatX](learning_rate_decay * lr), after_epoch=True, after_batch=False))
print('number of parameters in the model: ' + str(tensor.sum([p.size for p in cg.parameters]).eval()))
# Finally build the main loop and train the model
main_loop = MainLoop(data_stream=train_stream, algorithm=algorithm,
model=Model(cost), extensions=extensions)
main_loop.run()
args.primetext = ix_to_char[numpy.random.randint(vocab_size)]
primetext = ''.join(
[ch for ch in args.primetext if ch in char_to_ix.keys()])
if len(primetext) == 0:
raise Exception('primetext characters are not in the vocabulary')
x_curr = numpy.expand_dims(numpy.array(
[char_to_ix[ch] for ch in primetext], dtype='uint8'),
axis=1)
print 'Loading model from {0}...'.format(args.model)
x = tensor.matrix('features', dtype='uint8')
y = tensor.matrix('targets', dtype='uint8')
y_hat, cost, cells = nn_fprop(x, y, vocab_size, hidden_size, num_layers,
model)
main_loop = MainLoop(algorithm=None,
data_stream=None,
model=Model(cost),
extensions=[saveload.Load(args.model)])
for extension in main_loop.extensions:
extension.main_loop = main_loop
main_loop._run_extensions('before_training')
bin_model = main_loop.model
activations = []
initial_states = []
for i in range(num_layers):
brick = [
b for b in bin_model.get_top_bricks() if b.name == model + str(i)
][0]
activations.extend(
VariableFilter(theano_name=brick.name + '_apply_states')(
bin_model.variables))
activations.extend(
# Define primetext
ix_to_char, char_to_ix, vocab_size = get_metadata(hdf5_file)
if not args.primetext or len(args.primetext) == 0:
args.primetext = ix_to_char[numpy.random.randint(vocab_size)]
primetext = ''.join([ch for ch in args.primetext if ch in char_to_ix.keys()])
if len(primetext) == 0:
raise Exception('primetext characters are not in the vocabulary')
x_curr = numpy.expand_dims(
numpy.array([char_to_ix[ch] for ch in primetext], dtype='uint8'), axis=1)
print('Loading model from {0}...'.format(args.model))
x = tensor.matrix('features', dtype='uint8')
y = tensor.matrix('targets', dtype='uint8')
y_hat, cost, cells = nn_fprop(x, y, vocab_size, hidden_size, num_layers, model)
main_loop = MainLoop(algorithm=None, data_stream=None, model=Model(cost))
with open(args.model) as f:
main_loop.model.set_parameter_values(load_parameters(f))
bin_model = main_loop.model
activations = []
initial_states = []
for i in range(num_layers):
brick = [b for b in bin_model.get_top_bricks() if b.name==model+str(i)][0]
activations.extend(VariableFilter(theano_name=brick.name+'_apply_states')(bin_model.variables))
activations.extend(VariableFilter(theano_name=brick.name+'_apply_cells')(cells))
initial_states.extend(VariableFilter(roles=[roles.INITIAL_STATE])(brick.parameters))
#take activations of last element
activations = [act[-1].flatten() for act in activations]
states_as_params = [tensor.vector(dtype=initial.dtype) for initial in initial_states]
#Get prob. distribution of the last element in the last seq of the batch
y = tensor.tensor3('targets', dtype='floatX')
x = x.swapaxes(0, 1)
y = y.swapaxes(0, 1)
in_size = num_features
out_size = num_features
y_hat, cost, cells = nn_fprop(x,
y,
in_size,
out_size,
hidden_size[network_mode],
num_layers,
layer_models[network_mode][0],
'MDN',
training=False)
main_loop = MainLoop(algorithm=None,
data_stream=None,
model=Model(cost),
extensions=[saveload.Load(save_path[network_mode])])
for extension in main_loop.extensions:
extension.main_loop = main_loop
main_loop._run_extensions('before_training')
bin_model = main_loop.model
print 'Model loaded. Building prediction function...'
hiddens = []
initials = []
for i in range(num_layers):
brick = [
b for b in bin_model.get_top_bricks()
if b.name == layer_models[network_mode][i] + str(i) + '-'
][0]
hiddens.extend(
def load_models(
models=hierarchy_models,
in_size=len(hierarchy_input_columns[level_number_in_hierarchy]),
out_size=len(hierarchy_output_columns[level_number_in_hierarchy]),
hidden_size=hidden_size,
num_layers=num_layers,
model=layer_models[0]):
predict_funcs = []
initials = []
for hierarchy_index in range(len(models)):
saved_model = models[hierarchy_index]
print 'Loading model from {0}...'.format(models[hierarchy_index])
x = tensor.tensor3('features', dtype=theano.config.floatX)
y = tensor.tensor3('targets', dtype=theano.config.floatX)
y_hat, cost, cells = nn_fprop(x,
y,
in_size,
out_size,
hidden_size,
num_layers,
model,
training=False)
main_loop = MainLoop(algorithm=None,
data_stream=None,
model=Model(cost),
extensions=[saveload.Load(saved_model)])
for extension in main_loop.extensions:
extension.main_loop = main_loop
main_loop._run_extensions('before_training')
bin_model = main_loop.model
print 'Model loaded. Building prediction function...'
hiddens = []
initials.append([])
for i in range(num_layers - specialized_layer_num):
brick = [
b for b in bin_model.get_top_bricks()
if b.name == layer_models[i] + str(i) + '-' + str(-1)
][0]
hiddens.extend(
VariableFilter(theano_name=brick.name + '_apply_states')(
bin_model.variables))
hiddens.extend(
VariableFilter(theano_name=brick.name + '_apply_cells')(cells))
initials[hierarchy_index].extend(
VariableFilter(roles=[roles.INITIAL_STATE])(brick.parameters))
specialized_count = len(game_tasks) if task_specialized else 0
for task in range(specialized_count):
for i in range(num_layers - specialized_layer_num, num_layers):
brick = [
b for b in bin_model.get_top_bricks()
if b.name == layer_models[i] + str(i) + '-' + str(task)
][0]
hiddens.extend(
VariableFilter(theano_name=brick.name + '_apply_states')(
bin_model.variables))
hiddens.extend(
VariableFilter(theano_name=brick.name +
'_apply_cells')(cells))
initials[hierarchy_index].extend(
VariableFilter(roles=[roles.INITIAL_STATE])(
brick.parameters))
output_count = len(game_tasks) if task_specialized else 1
predict_funcs.append([])
for task in range(output_count):
predict_funcs[hierarchy_index].append(
theano.function([x], hiddens + [y_hat[task]]))
return predict_funcs, initials
# Define primetext
ix_to_char, char_to_ix, vocab_size = get_metadata(hdf5_file)
if not args.primetext or len(args.primetext) == 0:
args.primetext = ix_to_char[numpy.random.randint(vocab_size)]
primetext = ''.join([ch for ch in args.primetext if ch in char_to_ix.keys()])
if len(primetext) == 0:
raise Exception('primetext characters are not in the vocabulary')
x_curr = numpy.expand_dims(
numpy.array([char_to_ix[ch] for ch in primetext], dtype='uint8'), axis=1)
print 'Loading model from {0}...'.format(args.model)
x = tensor.matrix('features', dtype='uint8')
y = tensor.matrix('targets', dtype='uint8')
y_hat, cost, cells = nn_fprop(x, y, vocab_size, hidden_size, num_layers, model)
main_loop = MainLoop(algorithm=None, data_stream=None, model=Model(cost),
extensions=[saveload.Load(args.model)])
for extension in main_loop.extensions:
extension.main_loop = main_loop
main_loop._run_extensions('before_training')
bin_model = main_loop.model
activations = []
initial_states = []
for i in range(num_layers):
brick = [b for b in bin_model.get_top_bricks() if b.name==model+str(i)][0]
activations.extend(VariableFilter(theano_name=brick.name+'_apply_states')(bin_model.variables))
activations.extend(VariableFilter(theano_name=brick.name+'_apply_cells')(cells))
initial_states.extend(VariableFilter(roles=[roles.INITIAL_STATE])(brick.parameters))
#take activations of last element
activations = [act[-1].flatten() for act in activations]
states_as_params = [tensor.vector(dtype=initial.dtype) for initial in initial_states]
