experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Update documentation of tf.tpu.experimental.(shared_)embedding_column to reflect that a user should pass in the same function if they use the same dynamic learning rate for multiple tables.

Browse Source 
PiperOrigin-RevId: 299967731
Change-Id: I97c9ecf89c932774ec8daafdc65c3376570a886a
This commit is contained in:
Bruce Fontaine 2020-03-09 16:46:04 -07:00 committed by TensorFlower Gardener
parent 9ebf4a223b
commit 842246a4fd
2 changed files with 16 additions and 4 deletions

10
tensorflow/python/tpu/feature_column.py
View File

@ -86,7 +86,10 @@ def embedding_column(categorical_column,
and any sequence longer will be truncated. This must be positive for
sequence features and 0 for non-sequence features.
learning_rate_fn: A function that takes global step and returns learning
rate for the embedding table.
rate for the embedding table. If you intend to use the same learning rate
for multiple embedding tables, please ensure that you pass the exact same
python function to all calls of embedding_column, otherwise performence
may suffer.
use_safe_embedding_lookup: If true, uses safe_embedding_lookup_sparse
instead of embedding_lookup_sparse. safe_embedding_lookup_sparse ensures
there are no empty rows and all weights and ids are positive at the
@ -196,7 +199,10 @@ def shared_embedding_columns(categorical_columns,
sequence shorter then this will be padded with 0 embeddings and any
sequence longer will be truncated.
learning_rate_fn: A function that takes global step and returns learning
rate for the embedding table.
rate for the embedding table. If you intend to use the same learning rate
for multiple embedding tables, please ensure that you pass the exact same
python function to all calls of shared_embedding_columns, otherwise
performence may suffer.
use_safe_embedding_lookup: If true, uses safe_embedding_lookup_sparse
instead of embedding_lookup_sparse. safe_embedding_lookup_sparse ensures
there are no empty rows and all weights and ids are positive at the

10
tensorflow/python/tpu/feature_column_v2.py
View File

@ -107,7 +107,10 @@ def embedding_column_v2(categorical_column,
and any sequence longer will be truncated. This must be positive for
sequence features and 0 for non-sequence features.
learning_rate_fn: A function that takes global step and returns learning
rate for the embedding table.
rate for the embedding table. If you intend to use the same learning rate
for multiple embedding tables, please ensure that you pass the exact same
python function to all calls of embedding_column, otherwise performence
may suffer.
embedding_lookup_device: The device on which to run the embedding lookup.
Valid options are "cpu", "tpu_tensor_core", and "tpu_embedding_core".
If specifying "tpu_tensor_core", a tensor_core_shape must be supplied.
@ -266,7 +269,10 @@ def shared_embedding_columns_v2(categorical_columns,
sequence shorter then this will be padded with 0 embeddings and any
sequence longer will be truncated.
learning_rate_fn: A function that takes global step and returns learning
rate for the embedding table.
rate for the embedding table. If you intend to use the same learning rate
for multiple embedding tables, please ensure that you pass the exact same
python function to all calls of shared_embedding_columns, otherwise
performence may suffer.
embedding_lookup_device: The device on which to run the embedding lookup.
Valid options are "cpu", "tpu_tensor_core", and "tpu_embedding_core". If
specifying "tpu_tensor_core", a tensor_core_shape must be supplied.