def lr_decay_fn(learning_rate, global_step):
"""Learning rate decay function.
Args:
learning_rate (float or Tensor): The original learning rate.
global_step (int or scalar int Tensor): optimization step counter.
Returns:
scalar float Tensor: decayed learning rate.
"""
offset_global_step = tf.maximum(
tf.minimum(tf.cast(global_step, tf.int32), end_step) - start_step,
0)
if decay_fn == tf.train.piecewise_constant:
decayed_lr = decay_fn(x=offset_global_step, **fn_kwargs)
else:
fn_kwargs_ = {
"learning_rate": learning_rate,
"global_step": offset_global_step
}
fn_kwargs_.update(fn_kwargs)
decayed_lr = utils.call_function_with_redundant_kwargs(
decay_fn, fn_kwargs_)
decayed_lr = tf.maximum(decayed_lr, hparams["min_learning_rate"])
return decayed_lr
def _connect(self, encoder_results, features, labels, mode):
"""Transforms encoder final state into decoder initial state.
"""
enc_state = encoder_results["final_state"]
possible_kwargs = {
"inputs": enc_state,
"batch_size": get_batch_size(enc_state)
}
outputs = utils.call_function_with_redundant_kwargs(
self._connector._build, possible_kwargs)
return outputs
