Skip to content

pad_model_inputs.md

Latest commit

 

History

History
119 lines (101 loc) · 3.26 KB

pad_model_inputs.md

File metadata and controls

119 lines (101 loc) · 3.26 KB
Raw

description: Pad model input and generate corresponding input masks.

text.pad_model_inputs

View source

Pad model input and generate corresponding input masks.

text.pad_model_inputs(
    input, max_seq_length, pad_value=0
)

pad_model_inputs performs the final packaging of a model's inputs commonly found in text models. This includes padding out (or simply truncating) to a fixed-size, max 2-dimensional Tensor and generating mask Tensors (of the same shape) with values of 0 if the corresponding item is a pad value and 1 if it is part of the original input.

Note that a simple truncation strategy (drop everything after max sequence length) is used to force the inputs to the specified shape. This may be incorrect and users should instead apply a Trimmer upstream to safely truncate large inputs.

>>> input_data = tf.ragged.constant([
...            [101, 1, 2, 102, 10, 20, 102],
...            [101, 3, 4, 102, 30, 40, 50, 60, 70, 80],
...            [101, 5, 6, 7, 8, 9, 102, 70],
...        ], np.int32)
>>> data, mask = pad_model_inputs(input=input_data, max_seq_length=9)
>>> print("data: %s, mask: %s" % (data, mask))
  data: tf.Tensor(
  [[101   1   2 102  10  20 102   0   0]
   [101   3   4 102  30  40  50  60  70]
   [101   5   6   7   8   9 102  70   0]], shape=(3, 9), dtype=int32),
  mask: tf.Tensor(
  [[1 1 1 1 1 1 1 0 0]
   [1 1 1 1 1 1 1 1 1]
   [1 1 1 1 1 1 1 1 0]], shape=(3, 9), dtype=int32)

Args

`input` A `RaggedTensor` or `Tensor` with rank >= 1.
`max_seq_length` An int, or scalar `Tensor`. The "input" `Tensor` will be flattened down to 2 dimensions (if needed), and then have its inner dimension either padded out or truncated to this size.
`pad_value` An int or scalar `Tensor` specifying the value used for padding.

Returns

A tuple of (padded_input, pad_mask) where:
`padded_input` A `Tensor` corresponding to `inputs` that has been padded/truncated out to a fixed size and flattened to max 2 dimensions.
`pad_mask` A `Tensor` corresponding to `padded_input` whose values are 0 if the corresponding item is a pad value and 1 if it is not.