def evaluate(model, load_path):
with open(load_path + '/trained_params_best.npz') as f:
loaded = np.load(f)
blocks_model = Model(model.cost)
params_dicts = blocks_model.get_parameter_dict()
params_names = params_dicts.keys()
for param_name in params_names:
param = params_dicts[param_name]
# '/f_6_.W' --> 'f_6_.W'
slash_index = param_name.find('/')
param_name = param_name[slash_index + 1:]
assert param.get_value().shape == loaded[param_name].shape
param.set_value(loaded[param_name])
train_data_stream, valid_data_stream = get_cmv_v2_streams(100)
# T x B x F
data = train_data_stream.get_epoch_iterator().next()
cg = ComputationGraph(model.cost)
f = theano.function(cg.inputs, [model.location, model.scale],
on_unused_input='ignore',
allow_input_downcast=True)
res = f(data[1], data[0])
for i in range(10):
visualize_attention(data[0][:, i, :],
res[0][:, i, :],
res[1][:, i, :],
prefix=str(i))
def evaluate(model, load_path):
with open(load_path + '/trained_params_best.npz') as f:
loaded = np.load(f)
blocks_model = Model(model.cost)
params_dicts = blocks_model.get_parameter_dict()
params_names = params_dicts.keys()
for param_name in params_names:
param = params_dicts[param_name]
# '/f_6_.W' --> 'f_6_.W'
slash_index = param_name.find('/')
param_name = param_name[slash_index + 1:]
assert param.get_value().shape == loaded[param_name].shape
param.set_value(loaded[param_name])
train_data_stream, valid_data_stream = get_cmv_v2_streams(100)
# T x B x F
data = train_data_stream.get_epoch_iterator().next()
cg = ComputationGraph(model.cost)
f = theano.function(cg.inputs, [model.location, model.scale],
on_unused_input='ignore',
allow_input_downcast=True)
res = f(data[1], data[0])
for i in range(10):
visualize_attention(data[0][:, i, :],
res[0][:, i, :], res[1][:, i, :], prefix=str(i))
def train(model, batch_size=100, num_epochs=1000):
cost = model.cost
monitorings = model.monitorings
# Setting Loggesetr
timestr = time.strftime("%Y_%m_%d_at_%H_%M")
save_path = 'results/CMV_V2_' + timestr
log_path = os.path.join(save_path, 'log.txt')
os.makedirs(save_path)
fh = logging.FileHandler(filename=log_path)
fh.setLevel(logging.DEBUG)
logger.addHandler(fh)
# Training
blocks_model = Model(cost)
all_params = blocks_model.parameters
print "Number of found parameters:" + str(len(all_params))
print all_params
clipping = StepClipping(threshold=np.cast[floatX](10))
adam = Adam(learning_rate=model.lr_var)
step_rule = CompositeRule([adam, clipping])
training_algorithm = GradientDescent(cost=cost,
parameters=all_params,
step_rule=step_rule)
monitored_variables = [
model.lr_var, cost,
aggregation.mean(training_algorithm.total_gradient_norm)
] + monitorings
blocks_model = Model(cost)
params_dicts = blocks_model.get_parameter_dict()
for name, param in params_dicts.iteritems():
to_monitor = training_algorithm.gradients[param].norm(2)
to_monitor.name = name + "_grad_norm"
monitored_variables.append(to_monitor)
to_monitor = param.norm(2)
to_monitor.name = name + "_norm"
monitored_variables.append(to_monitor)
train_data_stream, valid_data_stream = get_cmv_v2_streams(batch_size)
train_monitoring = TrainingDataMonitoring(variables=monitored_variables,
prefix="train",
after_epoch=True)
valid_monitoring = DataStreamMonitoring(variables=monitored_variables,
data_stream=valid_data_stream,
prefix="valid",
after_epoch=True)
main_loop = MainLoop(
algorithm=training_algorithm,
data_stream=train_data_stream,
model=blocks_model,
extensions=[
train_monitoring, valid_monitoring,
FinishAfter(after_n_epochs=num_epochs),
SaveParams('valid_misclassificationrate_apply_error_rate',
blocks_model, save_path),
SaveLog(save_path, after_epoch=True),
ProgressBar(),
LRDecay(model.lr_var, [0.001, 0.0001, 0.00001, 0.000001],
[8, 15, 30, 1000],
after_epoch=True),
Printing()
])
main_loop.run()
def train(model, batch_size=100, num_epochs=1000):
cost = model.cost
monitorings = model.monitorings
# Setting Loggesetr
timestr = time.strftime("%Y_%m_%d_at_%H_%M")
save_path = 'results/CMV_V2_' + timestr
log_path = os.path.join(save_path, 'log.txt')
os.makedirs(save_path)
fh = logging.FileHandler(filename=log_path)
fh.setLevel(logging.DEBUG)
logger.addHandler(fh)
# Training
blocks_model = Model(cost)
all_params = blocks_model.parameters
print "Number of found parameters:" + str(len(all_params))
print all_params
clipping = StepClipping(threshold=np.cast[floatX](10))
adam = Adam(learning_rate=model.lr_var)
step_rule = CompositeRule([adam, clipping])
training_algorithm = GradientDescent(
cost=cost, parameters=all_params,
step_rule=step_rule)
monitored_variables = [
model.lr_var,
cost,
aggregation.mean(training_algorithm.total_gradient_norm)] + monitorings
blocks_model = Model(cost)
params_dicts = blocks_model.get_parameter_dict()
for name, param in params_dicts.iteritems():
to_monitor = training_algorithm.gradients[param].norm(2)
to_monitor.name = name + "_grad_norm"
monitored_variables.append(to_monitor)
to_monitor = param.norm(2)
to_monitor.name = name + "_norm"
monitored_variables.append(to_monitor)
train_data_stream, valid_data_stream = get_cmv_v2_streams(batch_size)
train_monitoring = TrainingDataMonitoring(
variables=monitored_variables,
prefix="train",
after_epoch=True)
valid_monitoring = DataStreamMonitoring(
variables=monitored_variables,
data_stream=valid_data_stream,
prefix="valid",
after_epoch=True)
main_loop = MainLoop(
algorithm=training_algorithm,
data_stream=train_data_stream,
model=blocks_model,
extensions=[
train_monitoring,
valid_monitoring,
FinishAfter(after_n_epochs=num_epochs),
SaveParams('valid_misclassificationrate_apply_error_rate',
blocks_model, save_path),
SaveLog(save_path, after_epoch=True),
ProgressBar(),
LRDecay(model.lr_var,
[0.001, 0.0001, 0.00001, 0.000001],
[8, 15, 30, 1000],
after_epoch=True),
Printing()])
main_loop.run()
