def _ctc_loss_v3(labels, logits, label_length, logit_length, use_gpu):
with test_util.device(use_gpu=use_gpu):
sparse_labels = ctc_ops.dense_labels_to_sparse(labels, label_length)
with backprop.GradientTape() as t:
t.watch(logits)
ref_loss = ctc_ops.ctc_loss_v3(labels=sparse_labels,
logits=logits,
label_length=label_length,
logit_length=logit_length,
blank_index=0)
ref_grad = t.gradient(ref_loss, [logits])
return ref_loss, ref_grad
def _forwardAndBackward(self, sparse_labels, logits_time_major, seed):
np.random.seed(seed)
params = self._genInputParams(logits_time_major=logits_time_major,
sparse_labels=sparse_labels)
labels, logits, labels_lengths, logits_lengths = params
output_shape = (labels_lengths.shape[0], )
upstream_gradients = self._randomFloats(output_shape)
with backprop.GradientTape() as tape:
tape.watch(logits)
loss = ctc_ops.ctc_loss_v3(labels,
logits,
labels_lengths,
logits_lengths,
logits_time_major=logits_time_major,
blank_index=0)
gradient_injector_output = loss * upstream_gradients
return loss, tape.gradient(gradient_injector_output, logits)