def _load_model(self, model_path = None, tic = None):
print("loading model begin({})".format(time.time() - tic))
assert model_path is not None
assert tic is not None
print("loading model from {}".format(model_path))
resume_model(model_path, self.net)
print("loading model done({})".format(time.time() - tic))
def _load_model(self, model_path = None, tic = None):
print("loading model begin({})".format(time.time() - tic))
assert model_path is not None
assert tic is not None
print("loading model from {}".format(model_path))
resume_model(model_path, self.net)
print("loading model done({})".format(time.time() - tic))
broadcastable=param.broadcastable)
v = momentum * m - learning_rate * T.grad(net.loss, param)
updates.append((m, v))
updates.append((param, param + v))
else:
print("using normal sgd and learning_rate:{}".format(
learning_rate.get_value()))
for name, param in net.params.items():
print(name, type(param))
grad = T.grad(net.loss, param)
updates.append((param, param - learning_rate * grad))
# resume model
if start_epoch > 0:
resume_path = "../snapshot/{}.pkl".format(start_epoch)
resume_model(resume_path, net)
# build train function
print("building training function({})".format(time.time() - begin))
train = theano.function(inputs=[],
outputs=net.loss,
updates=updates,
givens={
x: x_shared,
x_mask: x_mask_shared,
y: y_shared,
y_clip: y_clip_shared
})
# build valdiating function
print("building validating function({})".format(time.time() - begin))
for name, param in net.params.items():
m = theano.shared(param.get_value()*0., broadcastable=param.broadcastable)
v = momentum * m - learning_rate * T.grad(net.loss, param)
updates.append((m, v))
updates.append((param, param + v))
else:
print("using normal sgd and learning_rate:{}".format(learning_rate.get_value()))
for name, param in net.params.items():
print(name, type(param))
grad = T.grad(net.loss, param)
updates.append((param, param - learning_rate * grad))
# resume model
if start_epoch > 0:
resume_path = "../snapshot/{}.pkl".format(start_epoch)
resume_model(resume_path, net)
# build train function
print("building training function({})".format(time.time() - begin))
train = theano.function(
inputs = [],
outputs = net.loss,
updates = updates,
givens = {
x : x_shared,
x_mask : x_mask_shared,
y : y_shared,
y_clip : y_clip_shared
}
)