diff options
Diffstat (limited to 'tensorflow/python/training/warm_starting_util.py')
| -rw-r--r-- | tensorflow/python/training/warm_starting_util.py | 100 |
1 files changed, 87 insertions, 13 deletions
diff --git a/tensorflow/python/training/warm_starting_util.py b/tensorflow/python/training/warm_starting_util.py index c0dd46bfa5..bea9bb6dff 100644 --- a/tensorflow/python/training/warm_starting_util.py +++ b/tensorflow/python/training/warm_starting_util.py @@ -41,6 +41,7 @@ class VocabInfo( "old_vocab", "old_vocab_size", "backup_initializer", + "axis", ])): """Vocabulary information for warm-starting. @@ -62,6 +63,42 @@ class VocabInfo( backup_initializer: [Optional] A variable initializer used for variables corresponding to new vocabulary entries and OOV. If not provided, these entries will be zero-initialized. + axis: [Optional] Denotes what axis the vocabulary corresponds to. The + default, 0, corresponds to the most common use case (embeddings or + linear weights for binary classification / regression). An axis of 1 + could be used for warm-starting output layers with class vocabularies. + + For example: + + embeddings_vocab_info = tf.VocabInfo( + new_vocab='embeddings_vocab', + new_vocab_size=100, + num_oov_buckets=1, + old_vocab='pretrained_embeddings_vocab', + old_vocab_size=10000, + backup_initializer=tf.truncated_normal_initializer( + mean=0.0, stddev=(1 / math.sqrt(embedding_dim))), + axis=0) + + softmax_output_layer_kernel_vocab_info = tf.VocabInfo( + new_vocab='class_vocab', + new_vocab_size=5, + num_oov_buckets=0, # No OOV for classes. + old_vocab='old_class_vocab', + old_vocab_size=8, + backup_initializer=tf.glorot_uniform_initializer(), + axis=1) + + softmax_output_layer_bias_vocab_info = tf.VocabInfo( + new_vocab='class_vocab', + new_vocab_size=5, + num_oov_buckets=0, # No OOV for classes. + old_vocab='old_class_vocab', + old_vocab_size=8, + backup_initializer=tf.zeros_initializer(), + axis=0) + + Currently, only axis=0 and axis=1 are supported. """ def __new__(cls, @@ -70,7 +107,12 @@ class VocabInfo( num_oov_buckets, old_vocab, old_vocab_size=-1, - backup_initializer=None): + backup_initializer=None, + axis=0): + if axis != 0 and axis != 1: + raise ValueError("The only supported values for the axis argument are 0 " + "and 1. Provided axis: {}".format(axis)) + return super(VocabInfo, cls).__new__( cls, new_vocab, @@ -79,6 +121,7 @@ class VocabInfo( old_vocab, old_vocab_size, backup_initializer, + axis, ) @@ -149,7 +192,8 @@ def _warm_start_var_with_vocab(var, previous_vocab_size=-1, current_oov_buckets=0, prev_tensor_name=None, - initializer=None): + initializer=None, + axis=0): """Warm-starts given variable from `prev_tensor_name` tensor in `prev_ckpt`. Use this method when the `var` is backed by vocabulary. This method stitches @@ -180,6 +224,7 @@ def _warm_start_var_with_vocab(var, None, we lookup tensor with same name as given `var`. initializer: Variable initializer to be used for missing entries. If None, missing entries will be zero-initialized. + axis: Axis of the variable that the provided vocabulary corresponds to. Raises: ValueError: If required args are not provided. @@ -204,6 +249,8 @@ def _warm_start_var_with_vocab(var, # Assume tensor name remains the same. prev_tensor_name = _infer_var_name(var) + # TODO(eddz): Fix functionality for rank-1 Variables (like FC biases). + total_v_first_axis = sum([v.get_shape().as_list()[0] for v in var]) for v in var: v_shape = v.get_shape().as_list() slice_info = v._get_save_slice_info() @@ -213,19 +260,45 @@ def _warm_start_var_with_vocab(var, full_shape=slice_info.full_shape, var_offset=slice_info.var_offset) - # TODO(eddz): Support cases where class vocabularies need remapping too. + if axis == 0: + new_row_vocab_size = current_vocab_size + new_col_vocab_size = v_shape[1] + old_row_vocab_size = previous_vocab_size + old_row_vocab_file = prev_vocab_path + new_row_vocab_file = current_vocab_path + old_col_vocab_file = None + new_col_vocab_file = None + num_row_oov_buckets = current_oov_buckets + num_col_oov_buckets = 0 + elif axis == 1: + # Note that we must compute this value across all partitions, whereas + # in the axis = 0 case, we can simply use v_shape[1] because we don't + # allow partitioning across axis = 1. + new_row_vocab_size = total_v_first_axis + new_col_vocab_size = current_vocab_size + old_row_vocab_size = -1 + old_row_vocab_file = None + new_row_vocab_file = None + old_col_vocab_file = prev_vocab_path + new_col_vocab_file = current_vocab_path + num_row_oov_buckets = 0 + num_col_oov_buckets = current_oov_buckets + else: + raise ValueError("The only supported values for the axis argument are 0 " + "and 1. Provided axis: {}".format(axis)) + init = checkpoint_ops._load_and_remap_matrix_initializer( ckpt_path=checkpoint_utils._get_checkpoint_filename(prev_ckpt), old_tensor_name=prev_tensor_name, - new_row_vocab_size=current_vocab_size, - new_col_vocab_size=v_shape[1], - old_row_vocab_size=previous_vocab_size, - old_row_vocab_file=prev_vocab_path, - new_row_vocab_file=current_vocab_path, - old_col_vocab_file=None, - new_col_vocab_file=None, - num_row_oov_buckets=current_oov_buckets, - num_col_oov_buckets=0, + new_row_vocab_size=new_row_vocab_size, + new_col_vocab_size=new_col_vocab_size, + old_row_vocab_size=old_row_vocab_size, + old_row_vocab_file=old_row_vocab_file, + new_row_vocab_file=new_row_vocab_file, + old_col_vocab_file=old_col_vocab_file, + new_col_vocab_file=new_col_vocab_file, + num_row_oov_buckets=num_row_oov_buckets, + num_col_oov_buckets=num_col_oov_buckets, initializer=initializer) new_init_val = ops.convert_to_tensor( init(shape=v_shape, partition_info=partition_info)) @@ -374,7 +447,8 @@ def warm_start(ckpt_to_initialize_from, previous_vocab_size=vocab_info.old_vocab_size, current_oov_buckets=vocab_info.num_oov_buckets, prev_tensor_name=prev_var_name, - initializer=vocab_info.backup_initializer) + initializer=vocab_info.backup_initializer, + axis=vocab_info.axis) else: # For the special value of vars_to_warm_start = None, # we only warm-start variables with explicitly specified vocabularies. |