def get_activations(self):
"""Get activations for features.
This should be called within `computation` that is passed to
`tpu.replicate` and friends.
Returns:
A dictionary mapping from `String` of feature name to `Tensor`
of activation.
"""
recv_activations = tpu_ops.recv_tpu_embedding_activations(
num_outputs=len(self._table_to_config_dict),
config=self._config_proto.SerializeToString())
activations = collections.OrderedDict()
for table_id, table in enumerate(self._table_to_features_dict):
features = self._table_to_features_dict[table]
for lookup_id, feature in enumerate(features):
start_row = lookup_id * self._batch_size_per_core
end_row = start_row + self._batch_size_per_core
activations[feature] = gen_tpu_ops.tpu_embedding_activations(
self._dummy_table_variables[table_id],
recv_activations[table_id][start_row:end_row, :],
table_id=table_id,
lookup_id=lookup_id)
return activations
def get_activations(self):
"""Get activations for features.
This should be called within `computation` that is passed to
`tpu.replicate` and friends.
Returns:
A dictionary mapping from `String` of feature name to `Tensor`
of activation.
"""
recv_activations = tpu_ops.recv_tpu_embedding_activations(
num_outputs=len(self._table_to_config_dict),
config=self._config_proto.SerializeToString())
activations = collections.OrderedDict()
for table_id, table in enumerate(self._table_to_features_dict):
features = self._table_to_features_dict[table]
for lookup_id, feature in enumerate(features):
start_row = lookup_id * self._batch_size_per_core
end_row = start_row + self._batch_size_per_core
activations[feature] = gen_tpu_ops.tpu_embedding_activations(
self._dummy_table_variables[table_id],
recv_activations[table_id][start_row:end_row, :],
table_id=table_id,
lookup_id=lookup_id)
return activations
