def _loss_for_shard(logits, targets, shard):
logits = common_layers.approximate_split(logits, mp.n,
0)[shard]
targets = common_layers.approximate_split(targets, mp.n,
0)[shard]
return common_layers.padded_cross_entropy(
logits, targets, hparams.label_smoothing)
def _loss_for_shard(logits, targets, shard):
logits = common_layers.approximate_split(logits, mp.n, 0)[shard]
targets = common_layers.approximate_split(targets, mp.n, 0)[shard]
return common_layers.padded_cross_entropy(
logits, targets, hparams.label_smoothing)
