def load_models(net, 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 = net.get_size() # latent_size
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(net)
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()
[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))
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(
VariableFilter(theano_name=brick.name + '_apply_states')(
def load_model(self, load_path):
load_pre = saveload.Load(load_path)
self.extensions.append(load_pre)
def training(self,
fea2obj,
batch_size,
learning_rate=0.005,
steprule='adagrad',
wait_epochs=5,
kl_weight_init=None,
klw_ep=50,
klw_inc_rate=0,
num_epochs=None):
networkfile = self._config['net']
n_epochs = num_epochs or int(self._config['nepochs'])
reg_weight = float(self._config['loss_weight'])
reg_type = self._config['loss_reg']
numtrain = int(
self._config['num_train']) if 'num_train' in self._config else None
train_stream, num_samples_train = get_comb_stream(
fea2obj, 'train', batch_size, shuffle=True, num_examples=numtrain)
dev_stream, num_samples_dev = get_comb_stream(fea2obj,
'dev',
batch_size=None,
shuffle=False)
logger.info('sources: %s -- number of train/dev samples: %d/%d',
train_stream.sources, num_samples_train, num_samples_dev)
t2idx = fea2obj['targets'].t2idx
klw_init = kl_weight_init or float(
self._config['kld_weight']) if 'kld_weight' in self._config else 1
logger.info('kl_weight_init: %d', klw_init)
kl_weight = shared_floatx(klw_init, 'kl_weight')
entropy_weight = shared_floatx(1., 'entropy_weight')
cost, p_at_1, _, KLD, logpy_xz, pat1_recog, misclassify_rate = build_model_new(
fea2obj, len(t2idx), self._config, kl_weight, entropy_weight)
cg = ComputationGraph(cost)
weights = VariableFilter(roles=[WEIGHT])(cg.parameters)
logger.info('Model weights are: %s', weights)
if 'L2' in reg_type:
cost += reg_weight * l2_norm(weights)
logger.info('applying %s with weight: %f ', reg_type, reg_weight)
dropout = -0.1
if dropout > 0:
cg = apply_dropout(cg, weights, dropout)
cost = cg.outputs[0]
cost.name = 'cost'
logger.info('Our Algorithm is : %s, and learning_rate: %f', steprule,
learning_rate)
if 'adagrad' in steprule:
cnf_step_rule = AdaGrad(learning_rate)
elif 'adadelta' in steprule:
cnf_step_rule = AdaDelta(decay_rate=0.95)
elif 'decay' in steprule:
cnf_step_rule = RMSProp(learning_rate=learning_rate,
decay_rate=0.90)
cnf_step_rule = CompositeRule([cnf_step_rule, StepClipping(1)])
elif 'momentum' in steprule:
cnf_step_rule = Momentum(learning_rate=learning_rate, momentum=0.9)
elif 'adam' in steprule:
cnf_step_rule = Adam(learning_rate=learning_rate)
else:
logger.info('The steprule param is wrong! which is: %s', steprule)
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=cnf_step_rule,
on_unused_sources='warn')
#algorithm.add_updates(updates)
gradient_norm = aggregation.mean(algorithm.total_gradient_norm)
step_norm = aggregation.mean(algorithm.total_step_norm)
monitored_vars = [
cost, gradient_norm, step_norm, p_at_1, KLD, logpy_xz, kl_weight,
pat1_recog
]
train_monitor = TrainingDataMonitoring(variables=monitored_vars,
after_batch=True,
before_first_epoch=True,
prefix='tra')
dev_monitor = DataStreamMonitoring(variables=[
cost, p_at_1, KLD, logpy_xz, pat1_recog, misclassify_rate
],
after_epoch=True,
before_first_epoch=True,
data_stream=dev_stream,
prefix="dev")
extensions = [
dev_monitor,
train_monitor,
Timing(),
TrackTheBest('dev_cost'),
FinishIfNoImprovementAfter('dev_cost_best_so_far',
epochs=wait_epochs),
Printing(after_batch=False), #, ProgressBar()
FinishAfter(after_n_epochs=n_epochs),
saveload.Load(networkfile + '.toload.pkl'),
] + track_best('dev_cost', networkfile + '.best.pkl')
#extensions.append(SharedVariableModifier(kl_weight,
# lambda n, klw: numpy.cast[theano.config.floatX] (klw_inc_rate + klw), after_epoch=False, every_n_epochs=klw_ep, after_batch=False))
# extensions.append(SharedVariableModifier(entropy_weight,
# lambda n, crw: numpy.cast[theano.config.floatX](crw - klw_inc_rate), after_epoch=False, every_n_epochs=klw_ep, after_batch=False))
logger.info('number of parameters in the model: %d',
tensor.sum([p.size for p in cg.parameters]).eval())
logger.info('Lookup table sizes: %s',
[p.size.eval() for p in cg.parameters if 'lt' in p.name])
main_loop = MainLoop(data_stream=train_stream,
algorithm=algorithm,
model=Model(cost),
extensions=extensions)
main_loop.run()
def run():
# Load Model
net_size = 256 #Hard-code instead of loading model (takes too long to set up network)
#net = vaegan.VAEGAN()
#network_saver = saver.NetworkSaver('vaegan/models/', net=net)
#network_saver.load()
# DATA
train_stream = get_stream(hdf5_file, 'train', batch_size) #TODO jonathan ?
test_stream = get_stream(hdf5_file, 'test', batch_size) #TODO jonathan ?
# MODEL
x = T.TensorType('floatX', [False] * 3)('features')
y = T.tensor3('targets', dtype='floatX')
train_flag = [theano.shared(0)]
x = x.swapaxes(0, 1)
y = y.swapaxes(0, 1)
# More Config
out_size = len(output_columns) - 1 # code_mode=RL-MDN
latent_size = net_size
in_size = latent_size + len(input_columns)
# NN fprop
y_hat, cost, cells = nn_fprop(x, y, in_size, out_size, hidden_size,
num_recurrent_layers, train_flag)
# COST
cg = ComputationGraph(cost)
extra_updates = []
# RMS Prop training optimizer
step_rules = [
RMSProp(learning_rate=learning_rate, decay_rate=decay_rate),
StepClipping(step_clipping)
]
parameters_to_update = cg.parameters
algorithm = GradientDescent(cost=cg.outputs[0],
parameters=parameters_to_update,
step_rule=CompositeRule(step_rules))
algorithm.add_updates(
extra_updates) # TODO jonathan what is this, is this needed?
# Extensions
gradient_norm = aggregation.mean(algorithm.total_gradient_norm)
step_norm = aggregation.mean(algorithm.total_step_norm)
monitored_vars = [
cost, step_rules[0].learning_rate, gradient_norm, step_norm
]
test_monitor = DataStreamMonitoring(variables=[cost],
after_epoch=True,
before_first_epoch=True,
data_stream=test_stream,
prefix="test")
train_monitor = TrainingDataMonitoring(variables=monitored_vars,
after_epoch=True,
before_first_epoch=True,
prefix='train')
set_train_flag = SetTrainFlag(after_epoch=True,
before_epoch=True,
flag=train_flag)
# plot = Plot('Plotting example', channels=[['cost']], after_batch=True, open_browser=True)
extensions = [
set_train_flag,
test_monitor,
train_monitor,
Timing(),
Printing(after_epoch=True),
FinishAfter(after_n_epochs=nepochs),
saveload.Load(load_path),
saveload.Checkpoint(last_path, every_n_epochs=10000),
] + track_best('test_cost',
save_path) #+ track_best('train_cost', last_path)
if learning_rate_decay not in (0, 1):
extensions.append(
SharedVariableModifier(step_rules[0].learning_rate,
lambda n, lr: np.cast[theano.config.floatX]
(learning_rate_decay * lr),
after_epoch=False,
every_n_epochs=lr_decay_every_n_epochs,
after_batch=False))
print 'number of parameters in the model: ' + str(
T.sum([p.size for p in cg.parameters]).eval())
# Finally build the main loop and train the model
mainLoop = MainLoop(data_stream=train_stream,
algorithm=algorithm,
model=Model(cost),
extensions=extensions)
mainLoop.run()
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
