def compile_test_fn(self):
""" Build a testing routine for speech models.
Returns:
val_fn (theano.function): Function that takes in acoustic inputs,
and calculates the loss. Returns network outputs and ctc cost
"""
logger.info("Building val_fn")
f_inputs = [self.acoustic_input, self.ctc_in_lens]
f_outputs = []
for branch in self.branch_outputs:
labels, label_lens = self.branch_labels[branch.name]
if K.backend() == 'tensorflow':
network_output = branch.output
ctc_cost = K.mean(
K.ctc_batch_cost(labels, network_output, self.ctc_in_lens,
label_lens))
else:
network_output = branch.output.dimshuffle((1, 0, 2))
ctc_cost = ctc_th.gpu_ctc(network_output, labels,
self.ctc_in_lens).mean()
f_inputs.extend([labels, label_lens])
f_outputs.extend([network_output, ctc_cost])
f_inputs.append(K.learning_phase())
self.val_fn = K.function(f_inputs, f_outputs)
return self.val_fn
def compile_train_fn(self, learning_rate=2e-4):
""" Build the CTC training routine for speech models.
Args:
learning_rate (float)
Returns:
train_fn (theano.function): Function that takes in acoustic inputs,
and updates the model. Returns network outputs and ctc cost
"""
logger.info("Building train_fn")
f_inputs = [self.acoustic_input, self.ctc_in_lens]
f_outputs = []
f_updates = []
for branch in self.branch_outputs:
labels, label_lens = self.branch_labels[branch.name]
f_inputs.append(labels)
f_inputs.append(label_lens)
if K.backend() == 'tensorflow':
network_output = branch.output
ctc_cost = K.mean(
K.ctc_batch_cost(labels, network_output, self.ctc_in_lens,
label_lens))
else:
network_output = branch.output.dimshuffle((1, 0, 2))
ctc_cost = ctc_th.gpu_ctc(network_output, labels,
self.ctc_in_lens).mean()
f_outputs.extend([network_output, ctc_cost])
trainable_vars = self.branch_vars[branch.name]
optmz = Adam(lr=learning_rate, clipnorm=100)
f_updates.extend(optmz.get_updates(trainable_vars, [], ctc_cost))
f_inputs.append(K.learning_phase())
self.train_fn = K.function(f_inputs, f_outputs, f_updates)
return self.train_fn
def setup_gpu_op(self, activations, labels, input_length, compute_grad=True):
gpu_ctc_cost = gpu_ctc(activations, labels, input_length)
outputs = [gpu_ctc_cost]
if compute_grad:
# Symbolic gradient of CTC cost
gpu_ctc_grad = T.grad(T.mean(gpu_ctc_cost), activations)
outputs += [gpu_ctc_grad]
return theano.function([], outputs, mode=mode_with_gpu)
def setup_gpu_op(self, activations, labels, input_length, compute_grad=True):
gpu_ctc_cost = gpu_ctc(activations, labels, input_length)
outputs = [gpu_ctc_cost]
if compute_grad:
# Symbolic gradient of CTC cost
gpu_ctc_grad = tt.grad(tt.mean(gpu_ctc_cost), activations)
outputs += [gpu_ctc_grad]
return theano.function([], outputs, mode=mode_with_gpu)
def check_grads_disabled(self, activations, labels, input_length):
"""
Check if optimization to disable gradients is working
"""
gpu_ctc_cost = gpu_ctc(activations, labels, input_length)
gpu_ctc_function = theano.function([], [gpu_ctc_cost])
for node in gpu_ctc_function.maker.fgraph.apply_nodes:
if isinstance(node.op, GpuConnectionistTemporalClassification):
assert (node.op.compute_grad is False)
def check_grads_disabled(self, activations, labels, input_length):
"""
Check if optimization to disable gradients is working
"""
gpu_ctc_cost = gpu_ctc(activations, labels, input_length)
gpu_ctc_function = theano.function([], [gpu_ctc_cost])
for node in gpu_ctc_function.maker.fgraph.apply_nodes:
if isinstance(node.op, GpuConnectionistTemporalClassification):
assert node.op.compute_grad is False
def wrapper(acts):
# Create auxiliary symbolic variables
t_activation_times = theano.shared(in_lengths, name="activation_times")
t_labels = theano.shared(labels, name="labels")
return gpu_ctc(acts, t_labels, t_activation_times)
def wrapper(acts):
# Create auxiliary symbolic variables
t_activation_times = theano.shared(in_lengths,
name="activation_times")
t_labels = theano.shared(labels, name="labels")
return gpu_ctc(acts, t_labels, t_activation_times)
