def gpu_to_cpu_model(model):
for layer in model.layers:
for member, value in layer.__dict__.items():
if is_shared_var(value):
layer.__dict__[member] = T._shared(np.array(value.get_value(), floatX),
name=value.name, borrow=False)
for i in xrange(len(layer.params)):
if is_shared_var(layer.params[i]):
layer.params[i] = T._shared(np.array(layer.params[i].get_value(), floatX),
name=layer.params[i].name, borrow=False)
return model
def gpu_to_cpu_model(model):
for layer in model.layers:
for member, value in list(layer.__dict__.items()):
if is_shared_var(value):
layer.__dict__[member] = T._shared(np.array(value.get_value(), floatX),
name=value.name, borrow=False)
for i in range(len(layer.params)):
if is_shared_var(layer.params[i]):
layer.params[i] = T._shared(np.array(layer.params[i].get_value(), floatX),
name=layer.params[i].name, borrow=False)
return model
def setup(self):
self.log.info( '..begin setting up train object')
#===================[ build params and deltas list ]==================#
params = []
deltas = []
for i, layer in enumerate(self.model.layers):
layer_name = "{}_{}".format(layer.__class__.__name__, i)
if hasattr(layer, 'params'):
for param in layer.params:
# checked that the param to be updated is shared variable
if is_shared_var(param):
param.name = str(i) + '_' + str(param.name)
param.name += '_' + layer.__class__.__name__
params += [param]
deltas += [shared_zeros(shape=param.shape.eval())]
#=====================[ training params updates ]=====================#
self.log.info("..update params: " + str(params))
train_y_pred, train_layers_stats = self.model.train_fprop(self.model.input_var)
train_cost = self.train_cost(self.model.output_var, train_y_pred).astype(floatX)
gparams = T.grad(train_cost, params)
train_updates = self.learning_method.update(deltas, params, gparams)
#=================[ append updates from each layer ]==================#
for i, layer in enumerate(self.model.layers):
layer_name = "{}_{}".format(layer.__class__.__name__, i)
if hasattr(layer, 'updates') and len(layer.updates) > 0:
self.log.info("..{}: has shared variable updates".format(layer_name))
train_updates += layer.updates
#----[ append updates of stats from each layer to train updates ]-----#
self.train_stats_names, train_stats_vars = split_list(train_layers_stats)
train_stats_vars = [var.astype(floatX) for var in train_stats_vars]
self.train_stats_shared = generate_shared_list(train_stats_vars)
train_stats_updates = merge_lists(self.train_stats_shared, train_stats_vars)
if self.verbose:
train_updates += train_stats_updates
#-------------------------[ train functions ]-------------------------#
self.log.info('..begin compiling functions')
self.training = theano.function(inputs=merge_var(self.model.input_var, self.model.output_var),
outputs=train_cost,
updates=train_updates,
on_unused_input='warn',
allow_input_downcast=True)
self.log.info('..training function compiled')
#=============================[ testing ]=============================#
test_y_pred, test_layers_stats = self.model.test_fprop(self.model.input_var)
#-----[ append updates of stats from each layer to test updates ]-----#
self.test_stats_names, test_stats_vars = split_list(test_layers_stats)
test_stats_vars = [var.astype(floatX) for var in test_stats_vars]
self.test_stats_shared = generate_shared_list(test_stats_vars)
test_stats_updates = []
if self.verbose:
test_stats_updates = merge_lists(self.test_stats_shared, test_stats_vars)
#-------------------------[ test functions ]--------------------------#
test_stopping_error = self.valid_cost(self.model.output_var, test_y_pred).astype(floatX)
test_cost = self.train_cost(self.model.output_var, test_y_pred).astype(floatX)
self.testing = theano.function(inputs=merge_var(self.model.input_var, self.model.output_var),
outputs=(test_stopping_error, test_cost),
updates=test_stats_updates,
on_unused_input='warn',
allow_input_downcast=True)
self.log.info('..testing function compiled')
def setup(self):
self.log.info('..begin setting up train object')
#===================[ build params and deltas list ]==================#
params = []
deltas = []
for i, layer in enumerate(self.model.layers):
layer_name = "{}_{}".format(layer.__class__.__name__, i)
if hasattr(layer, 'params'):
for param in layer.params:
# checked that the param to be updated is shared variable
if is_shared_var(param):
param.name = str(i) + '_' + str(param.name)
param.name += '_' + layer.__class__.__name__
params += [param]
deltas += [shared_zeros(shape=param.shape.eval())]
#=====================[ training params updates ]=====================#
self.log.info("..update params: " + str(params))
train_y_pred, train_layers_stats = self.model.train_fprop(
self.model.input_var)
train_cost = self.train_cost(self.model.output_var,
train_y_pred).astype(floatX)
gparams = T.grad(train_cost, params)
train_updates = self.learning_method.update(deltas, params, gparams)
#=================[ append updates from each layer ]==================#
for i, layer in enumerate(self.model.layers):
layer_name = "{}_{}".format(layer.__class__.__name__, i)
if hasattr(layer, 'updates') and len(layer.updates) > 0:
self.log.info(
"..{}: has shared variable updates".format(layer_name))
train_updates += layer.updates
#----[ append updates of stats from each layer to train updates ]-----#
self.train_stats_names, train_stats_vars = split_list(
train_layers_stats)
train_stats_vars = [var.astype(floatX) for var in train_stats_vars]
self.train_stats_shared = generate_shared_list(train_stats_vars)
train_stats_updates = merge_lists(self.train_stats_shared,
train_stats_vars)
if self.verbose:
train_updates += train_stats_updates
#-------------------------[ train functions ]-------------------------#
self.log.info('..begin compiling functions')
self.training = theano.function(inputs=merge_var(
self.model.input_var, self.model.output_var),
outputs=train_cost,
updates=train_updates,
on_unused_input='warn',
allow_input_downcast=True)
self.log.info('..training function compiled')
#=============================[ testing ]=============================#
test_y_pred, test_layers_stats = self.model.test_fprop(
self.model.input_var)
#-----[ append updates of stats from each layer to test updates ]-----#
self.test_stats_names, test_stats_vars = split_list(test_layers_stats)
test_stats_vars = [var.astype(floatX) for var in test_stats_vars]
self.test_stats_shared = generate_shared_list(test_stats_vars)
test_stats_updates = []
if self.verbose:
test_stats_updates = merge_lists(self.test_stats_shared,
test_stats_vars)
#-------------------------[ test functions ]--------------------------#
test_stopping_error = self.valid_cost(self.model.output_var,
test_y_pred).astype(floatX)
test_cost = self.train_cost(self.model.output_var,
test_y_pred).astype(floatX)
self.testing = theano.function(
inputs=merge_var(self.model.input_var, self.model.output_var),
outputs=(test_stopping_error, test_cost),
updates=test_stats_updates,
on_unused_input='warn',
allow_input_downcast=True)
self.log.info('..testing function compiled')
