Spaces:
Runtime error
Runtime error
Update whisper_jax/layers.py
Browse files- whisper_jax/layers.py +1 -66
whisper_jax/layers.py
CHANGED
|
@@ -56,16 +56,6 @@ NdInitializer = Callable[[PRNGKey, Shape, DType, InitializerAxis, InitializerAxi
|
|
| 56 |
default_embed_init = nn.initializers.variance_scaling(1.0, "fan_in", "normal", out_axis=0)
|
| 57 |
|
| 58 |
|
| 59 |
-
def nd_dense_init(scale, mode, distribution):
|
| 60 |
-
"""Initializer with in_axis, out_axis set at call time."""
|
| 61 |
-
|
| 62 |
-
def init_fn(key, shape, dtype, in_axis, out_axis):
|
| 63 |
-
fn = variance_scaling(scale, mode, distribution, in_axis, out_axis)
|
| 64 |
-
return fn(key, shape, dtype)
|
| 65 |
-
|
| 66 |
-
return init_fn
|
| 67 |
-
|
| 68 |
-
|
| 69 |
def dot_product_attention(
|
| 70 |
query: Array,
|
| 71 |
key: Array,
|
|
@@ -78,11 +68,9 @@ def dot_product_attention(
|
|
| 78 |
float32_logits: bool = False,
|
| 79 |
):
|
| 80 |
"""Computes dot-product attention given query, key, and value.
|
| 81 |
-
|
| 82 |
This is the core function for applying attention based on
|
| 83 |
https://arxiv.org/abs/1706.03762. It calculates the attention weights given
|
| 84 |
query and key and combines the values using the attention weights.
|
| 85 |
-
|
| 86 |
Args:
|
| 87 |
query: queries for calculating attention with shape of `[batch, q_length,
|
| 88 |
num_heads, qk_depth_per_head]`.
|
|
@@ -99,7 +87,6 @@ def dot_product_attention(
|
|
| 99 |
dtype: the dtype of the computation (default: float32)
|
| 100 |
float32_logits: bool, if True then compute logits in float32 to avoid
|
| 101 |
numerical issues with bfloat16.
|
| 102 |
-
|
| 103 |
Returns:
|
| 104 |
Output of shape `[batch, length, num_heads, v_depth_per_head]`.
|
| 105 |
"""
|
|
@@ -145,7 +132,6 @@ dynamic_vector_slice_in_dim = jax.vmap(lax.dynamic_slice_in_dim, in_axes=(None,
|
|
| 145 |
|
| 146 |
class MultiHeadDotProductAttention(nn.Module):
|
| 147 |
"""Multi-head dot-product attention.
|
| 148 |
-
|
| 149 |
Attributes:
|
| 150 |
num_heads: number of attention heads. Features (i.e. inputs_q.shape[-1])
|
| 151 |
should be divisible by the number of heads.
|
|
@@ -176,22 +162,18 @@ class MultiHeadDotProductAttention(nn.Module):
|
|
| 176 |
deterministic: bool = False,
|
| 177 |
) -> Array:
|
| 178 |
"""Applies multi-head dot product attention on the input data.
|
| 179 |
-
|
| 180 |
Projects the inputs into multi-headed query, key, and value vectors,
|
| 181 |
applies dot-product attention and project the results to an output vector.
|
| 182 |
-
|
| 183 |
There are two modes: decoding and non-decoding (e.g., training). The mode is
|
| 184 |
determined by `decode` argument. For decoding, this method is called twice,
|
| 185 |
first to initialize the cache and then for an actual decoding process. The
|
| 186 |
two calls are differentiated by the presence of 'cached_key' in the variable
|
| 187 |
dict. In the cache initialization stage, the cache variables are initialized
|
| 188 |
as zeros and will be filled in the subsequent decoding process.
|
| 189 |
-
|
| 190 |
In the cache initialization call, `inputs_q` has a shape [batch, length,
|
| 191 |
q_features] and `inputs_kv`: [batch, length, kv_features]. During the
|
| 192 |
incremental decoding stage, query, key and value all have the shape [batch,
|
| 193 |
1, qkv_features] corresponding to a single step.
|
| 194 |
-
|
| 195 |
Args:
|
| 196 |
inputs_q: input queries of shape `[batch, q_length, q_features]`.
|
| 197 |
inputs_kv: key/values of shape `[batch, kv_length, kv_features]`.
|
|
@@ -199,7 +181,6 @@ class MultiHeadDotProductAttention(nn.Module):
|
|
| 199 |
bias: attention bias of shape `[batch, num_heads, q_length, kv_length]`.
|
| 200 |
decode: Whether to prepare and use an autoregressive cache.
|
| 201 |
deterministic: Disables dropout if set to True.
|
| 202 |
-
|
| 203 |
Returns:
|
| 204 |
output of shape `[batch, length, q_features]`.
|
| 205 |
"""
|
|
@@ -360,7 +341,6 @@ def _canonicalize_tuple(x):
|
|
| 360 |
# ------------------------------------------------------------------------------
|
| 361 |
class DenseGeneral(nn.Module):
|
| 362 |
"""A linear transformation (without bias) with flexible axes.
|
| 363 |
-
|
| 364 |
Attributes:
|
| 365 |
features: tuple with numbers of output features.
|
| 366 |
axis: tuple with axes to apply the transformation on.
|
|
@@ -380,10 +360,8 @@ class DenseGeneral(nn.Module):
|
|
| 380 |
@nn.compact
|
| 381 |
def __call__(self, inputs: Array) -> Array:
|
| 382 |
"""Applies a linear transformation to the inputs along multiple dimensions.
|
| 383 |
-
|
| 384 |
Args:
|
| 385 |
inputs: The nd-array to be transformed.
|
| 386 |
-
|
| 387 |
Returns:
|
| 388 |
The transformed input.
|
| 389 |
"""
|
|
@@ -432,7 +410,6 @@ def _convert_to_activation_function(fn_or_string: Union[str, Callable]) -> Calla
|
|
| 432 |
|
| 433 |
class MlpBlock(nn.Module):
|
| 434 |
"""Transformer MLP / feed-forward block.
|
| 435 |
-
|
| 436 |
Attributes:
|
| 437 |
intermediate_dim: Shared dimension of hidden layers.
|
| 438 |
activations: Type of activations for each layer. Each element is either
|
|
@@ -482,7 +459,6 @@ class MlpBlock(nn.Module):
|
|
| 482 |
|
| 483 |
class Embed(nn.Module):
|
| 484 |
"""A parameterized function from integers [0, n) to d-dimensional vectors.
|
| 485 |
-
|
| 486 |
Attributes:
|
| 487 |
num_embeddings: number of embeddings.
|
| 488 |
features: number of feature dimensions for each embedding.
|
|
@@ -513,10 +489,8 @@ class Embed(nn.Module):
|
|
| 513 |
|
| 514 |
def __call__(self, inputs: Array) -> Array:
|
| 515 |
"""Embeds the inputs along the last dimension.
|
| 516 |
-
|
| 517 |
Args:
|
| 518 |
inputs: input data, all dimensions are considered batch dimensions.
|
| 519 |
-
|
| 520 |
Returns:
|
| 521 |
Output which is embedded input data. The output shape follows the input,
|
| 522 |
with an additional `features` dimension appended.
|
|
@@ -536,11 +510,9 @@ class Embed(nn.Module):
|
|
| 536 |
|
| 537 |
def attend(self, query: Array) -> Array:
|
| 538 |
"""Attend over the embedding using a query array.
|
| 539 |
-
|
| 540 |
Args:
|
| 541 |
query: array with last dimension equal the feature depth `features` of the
|
| 542 |
embedding.
|
| 543 |
-
|
| 544 |
Returns:
|
| 545 |
An array with final dim `num_embeddings` corresponding to the batched
|
| 546 |
inner-product of the array of query vectors against each embedding.
|
|
@@ -553,7 +525,6 @@ class Embed(nn.Module):
|
|
| 553 |
|
| 554 |
class RelativePositionBiases(nn.Module):
|
| 555 |
"""Adds T5-style relative positional embeddings to the attention logits.
|
| 556 |
-
|
| 557 |
Attributes:
|
| 558 |
num_buckets: Number of buckets to bucket distances between key and query
|
| 559 |
positions into.
|
|
@@ -574,7 +545,6 @@ class RelativePositionBiases(nn.Module):
|
|
| 574 |
@staticmethod
|
| 575 |
def _relative_position_bucket(relative_position, bidirectional=True, num_buckets=32, max_distance=128):
|
| 576 |
"""Translate relative position to a bucket number for relative attention.
|
| 577 |
-
|
| 578 |
The relative position is defined as memory_position - query_position, i.e.
|
| 579 |
the distance in tokens from the attending position to the attended-to
|
| 580 |
position. If bidirectional=False, then positive relative positions are
|
|
@@ -585,13 +555,11 @@ class RelativePositionBiases(nn.Module):
|
|
| 585 |
positions <=-max_distance map to the same bucket. This should allow for
|
| 586 |
more graceful generalization to longer sequences than the model has been
|
| 587 |
trained on.
|
| 588 |
-
|
| 589 |
Args:
|
| 590 |
relative_position: an int32 array
|
| 591 |
bidirectional: a boolean - whether the attention is bidirectional
|
| 592 |
num_buckets: an integer
|
| 593 |
max_distance: an integer
|
| 594 |
-
|
| 595 |
Returns:
|
| 596 |
a Tensor with the same shape as relative_position, containing int32
|
| 597 |
values in the range [0, num_buckets)
|
|
@@ -619,13 +587,11 @@ class RelativePositionBiases(nn.Module):
|
|
| 619 |
@nn.compact
|
| 620 |
def __call__(self, qlen, klen, bidirectional=True):
|
| 621 |
"""Produce relative position embedding attention biases.
|
| 622 |
-
|
| 623 |
Args:
|
| 624 |
qlen: attention query length.
|
| 625 |
klen: attention key length.
|
| 626 |
bidirectional: whether to allow positive memory-query relative position
|
| 627 |
embeddings.
|
| 628 |
-
|
| 629 |
Returns:
|
| 630 |
output: `(1, len, q_len, k_len)` attention bias
|
| 631 |
"""
|
|
@@ -749,11 +715,9 @@ def make_attention_mask(
|
|
| 749 |
dtype: DType = jnp.float32,
|
| 750 |
) -> Array:
|
| 751 |
"""Mask-making helper for attention weights.
|
| 752 |
-
|
| 753 |
In case of 1d inputs (i.e., `[batch, len_q]`, `[batch, len_kv]`, the
|
| 754 |
attention weights will be `[batch, heads, len_q, len_kv]` and this
|
| 755 |
function will produce `[batch, 1, len_q, len_kv]`.
|
| 756 |
-
|
| 757 |
Args:
|
| 758 |
query_input: a batched, flat input of query_length size
|
| 759 |
key_input: a batched, flat input of key_length size
|
|
@@ -761,7 +725,6 @@ def make_attention_mask(
|
|
| 761 |
extra_batch_dims: number of extra batch dims to add singleton axes for, none
|
| 762 |
by default
|
| 763 |
dtype: mask return dtype
|
| 764 |
-
|
| 765 |
Returns:
|
| 766 |
A `[batch, 1, len_q, len_kv]` shaped mask for 1d attention.
|
| 767 |
"""
|
|
@@ -781,21 +744,17 @@ def make_attention_mask(
|
|
| 781 |
|
| 782 |
def make_causal_mask(x: Array, extra_batch_dims: int = 0, dtype: DType = jnp.float32) -> Array:
|
| 783 |
"""Make a causal mask for self-attention.
|
| 784 |
-
|
| 785 |
In case of 1d inputs (i.e., `[batch, len]`, the self-attention weights
|
| 786 |
will be `[batch, heads, len, len]` and this function will produce a
|
| 787 |
causal mask of shape `[batch, 1, len, len]`.
|
| 788 |
-
|
| 789 |
Note that a causal mask does not depend on the values of x; it only depends on
|
| 790 |
the shape. If x has padding elements, they will not be treated in a special
|
| 791 |
manner.
|
| 792 |
-
|
| 793 |
Args:
|
| 794 |
x: input array of shape `[batch, len]`
|
| 795 |
extra_batch_dims: number of batch dims to add singleton axes for, none by
|
| 796 |
default
|
| 797 |
dtype: mask return dtype
|
| 798 |
-
|
| 799 |
Returns:
|
| 800 |
A `[batch, 1, len, len]` shaped causal mask for 1d attention.
|
| 801 |
"""
|
|
@@ -805,11 +764,9 @@ def make_causal_mask(x: Array, extra_batch_dims: int = 0, dtype: DType = jnp.flo
|
|
| 805 |
|
| 806 |
def combine_masks(*masks: Optional[Array], dtype: DType = jnp.float32):
|
| 807 |
"""Combine attention masks.
|
| 808 |
-
|
| 809 |
Args:
|
| 810 |
*masks: set of attention mask arguments to combine, some can be None.
|
| 811 |
dtype: final mask dtype
|
| 812 |
-
|
| 813 |
Returns:
|
| 814 |
Combined mask, reduced by logical and, returns None if no masks given.
|
| 815 |
"""
|
|
@@ -827,10 +784,8 @@ def combine_masks(*masks: Optional[Array], dtype: DType = jnp.float32):
|
|
| 827 |
|
| 828 |
def combine_biases(*masks: Optional[Array]):
|
| 829 |
"""Combine attention biases.
|
| 830 |
-
|
| 831 |
Args:
|
| 832 |
*masks: set of attention bias arguments to combine, some can be None.
|
| 833 |
-
|
| 834 |
Returns:
|
| 835 |
Combined mask, reduced by summation, returns None if no masks given.
|
| 836 |
"""
|
|
@@ -853,40 +808,30 @@ def make_decoder_mask(
|
|
| 853 |
decoder_segment_ids: Optional[Array] = None,
|
| 854 |
) -> Array:
|
| 855 |
"""Compute the self-attention mask for a decoder.
|
| 856 |
-
|
| 857 |
Decoder mask is formed by combining a causal mask, a padding mask and an
|
| 858 |
optional packing mask. If decoder_causal_attention is passed, it makes the
|
| 859 |
masking non-causal for positions that have value of 1.
|
| 860 |
-
|
| 861 |
A prefix LM is applied to a dataset which has a notion of "inputs" and
|
| 862 |
"targets", e.g., a machine translation task. The inputs and targets are
|
| 863 |
concatenated to form a new target. `decoder_target_tokens` is the concatenated
|
| 864 |
decoder output tokens.
|
| 865 |
-
|
| 866 |
The "inputs" portion of the concatenated sequence can attend to other "inputs"
|
| 867 |
tokens even for those at a later time steps. In order to control this
|
| 868 |
behavior, `decoder_causal_attention` is necessary. This is a binary mask with
|
| 869 |
a value of 1 indicating that the position belonged to "inputs" portion of the
|
| 870 |
original dataset.
|
| 871 |
-
|
| 872 |
Example:
|
| 873 |
-
|
| 874 |
Suppose we have a dataset with two examples.
|
| 875 |
-
|
| 876 |
ds = [{"inputs": [6, 7], "targets": [8]},
|
| 877 |
{"inputs": [3, 4], "targets": [5]}]
|
| 878 |
-
|
| 879 |
After the data preprocessing with packing, the two examples are packed into
|
| 880 |
one example with the following three fields (some fields are skipped for
|
| 881 |
simplicity).
|
| 882 |
-
|
| 883 |
decoder_target_tokens = [[6, 7, 8, 3, 4, 5, 0]]
|
| 884 |
decoder_segment_ids = [[1, 1, 1, 2, 2, 2, 0]]
|
| 885 |
decoder_causal_attention = [[1, 1, 0, 1, 1, 0, 0]]
|
| 886 |
-
|
| 887 |
where each array has [batch, length] shape with batch size being 1. Then,
|
| 888 |
this function computes the following mask.
|
| 889 |
-
|
| 890 |
mask = [[[[1, 1, 0, 0, 0, 0, 0],
|
| 891 |
[1, 1, 0, 0, 0, 0, 0],
|
| 892 |
[1, 1, 1, 0, 0, 0, 0],
|
|
@@ -894,14 +839,11 @@ def make_decoder_mask(
|
|
| 894 |
[0, 0, 0, 1, 1, 0, 0],
|
| 895 |
[0, 0, 0, 1, 1, 1, 0],
|
| 896 |
[0, 0, 0, 0, 0, 0, 0]]]]
|
| 897 |
-
|
| 898 |
mask[b, 1, :, :] represents the mask for the example `b` in the batch.
|
| 899 |
Because mask is for a self-attention layer, the mask's shape is a square of
|
| 900 |
shape [query length, key length].
|
| 901 |
-
|
| 902 |
mask[b, 1, i, j] = 1 means that the query token at position i can attend to
|
| 903 |
the key token at position j.
|
| 904 |
-
|
| 905 |
Args:
|
| 906 |
decoder_target_tokens: decoder output tokens. [batch, length]
|
| 907 |
dtype: dtype of the output mask.
|
|
@@ -910,7 +852,6 @@ def make_decoder_mask(
|
|
| 910 |
bidirectionally. [batch, length]
|
| 911 |
decoder_segment_ids: decoder segmentation info for packed examples. [batch,
|
| 912 |
length]
|
| 913 |
-
|
| 914 |
Returns:
|
| 915 |
the combined decoder mask.
|
| 916 |
"""
|
|
@@ -976,7 +917,6 @@ def _conv_dimension_numbers(input_shape):
|
|
| 976 |
|
| 977 |
class _Conv(nn.Module):
|
| 978 |
"""Convolution Module wrapping `lax.conv_general_dilated[_local]`.
|
| 979 |
-
|
| 980 |
Attributes:
|
| 981 |
features: number of convolution filters.
|
| 982 |
kernel_size: shape of the convolutional kernel. For 1D convolution,
|
|
@@ -1032,19 +972,16 @@ class _Conv(nn.Module):
|
|
| 1032 |
@property
|
| 1033 |
def shared_weights(self) -> bool: # type: ignore
|
| 1034 |
"""Defines whether weights are shared or not between different pixels.
|
| 1035 |
-
|
| 1036 |
Returns:
|
| 1037 |
`True` to use shared weights in convolution (regular convolution).
|
| 1038 |
`False` to use different weights at different pixels, a.k.a.
|
| 1039 |
"locally connected layer", "unshared convolution", or "local convolution".
|
| 1040 |
-
|
| 1041 |
"""
|
| 1042 |
...
|
| 1043 |
|
| 1044 |
@nn.compact
|
| 1045 |
def __call__(self, inputs: Array) -> Array:
|
| 1046 |
"""Applies a (potentially unshared) convolution to the inputs.
|
| 1047 |
-
|
| 1048 |
Args:
|
| 1049 |
inputs: input data with dimensions (*batch_dims, spatial_dims...,
|
| 1050 |
features). This is the channels-last convention, i.e. NHWC for a 2d
|
|
@@ -1057,7 +994,6 @@ class _Conv(nn.Module):
|
|
| 1057 |
better performance than this default flattening approach. If the input
|
| 1058 |
lacks a batch dimension it will be added for the convolution and removed
|
| 1059 |
n return, an allowance made to enable writing single-example code.
|
| 1060 |
-
|
| 1061 |
Returns:
|
| 1062 |
The convolved data.
|
| 1063 |
"""
|
|
@@ -1214,7 +1150,6 @@ class _Conv(nn.Module):
|
|
| 1214 |
|
| 1215 |
class Conv(_Conv):
|
| 1216 |
"""Convolution Module wrapping `lax.conv_general_dilated`.
|
| 1217 |
-
|
| 1218 |
Attributes:
|
| 1219 |
features: number of convolution filters.
|
| 1220 |
kernel_size: shape of the convolutional kernel. For 1D convolution,
|
|
@@ -1252,4 +1187,4 @@ class Conv(_Conv):
|
|
| 1252 |
|
| 1253 |
@property
|
| 1254 |
def shared_weights(self) -> bool:
|
| 1255 |
-
return True
|
|
|
|
| 56 |
default_embed_init = nn.initializers.variance_scaling(1.0, "fan_in", "normal", out_axis=0)
|
| 57 |
|
| 58 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 59 |
def dot_product_attention(
|
| 60 |
query: Array,
|
| 61 |
key: Array,
|
|
|
|
| 68 |
float32_logits: bool = False,
|
| 69 |
):
|
| 70 |
"""Computes dot-product attention given query, key, and value.
|
|
|
|
| 71 |
This is the core function for applying attention based on
|
| 72 |
https://arxiv.org/abs/1706.03762. It calculates the attention weights given
|
| 73 |
query and key and combines the values using the attention weights.
|
|
|
|
| 74 |
Args:
|
| 75 |
query: queries for calculating attention with shape of `[batch, q_length,
|
| 76 |
num_heads, qk_depth_per_head]`.
|
|
|
|
| 87 |
dtype: the dtype of the computation (default: float32)
|
| 88 |
float32_logits: bool, if True then compute logits in float32 to avoid
|
| 89 |
numerical issues with bfloat16.
|
|
|
|
| 90 |
Returns:
|
| 91 |
Output of shape `[batch, length, num_heads, v_depth_per_head]`.
|
| 92 |
"""
|
|
|
|
| 132 |
|
| 133 |
class MultiHeadDotProductAttention(nn.Module):
|
| 134 |
"""Multi-head dot-product attention.
|
|
|
|
| 135 |
Attributes:
|
| 136 |
num_heads: number of attention heads. Features (i.e. inputs_q.shape[-1])
|
| 137 |
should be divisible by the number of heads.
|
|
|
|
| 162 |
deterministic: bool = False,
|
| 163 |
) -> Array:
|
| 164 |
"""Applies multi-head dot product attention on the input data.
|
|
|
|
| 165 |
Projects the inputs into multi-headed query, key, and value vectors,
|
| 166 |
applies dot-product attention and project the results to an output vector.
|
|
|
|
| 167 |
There are two modes: decoding and non-decoding (e.g., training). The mode is
|
| 168 |
determined by `decode` argument. For decoding, this method is called twice,
|
| 169 |
first to initialize the cache and then for an actual decoding process. The
|
| 170 |
two calls are differentiated by the presence of 'cached_key' in the variable
|
| 171 |
dict. In the cache initialization stage, the cache variables are initialized
|
| 172 |
as zeros and will be filled in the subsequent decoding process.
|
|
|
|
| 173 |
In the cache initialization call, `inputs_q` has a shape [batch, length,
|
| 174 |
q_features] and `inputs_kv`: [batch, length, kv_features]. During the
|
| 175 |
incremental decoding stage, query, key and value all have the shape [batch,
|
| 176 |
1, qkv_features] corresponding to a single step.
|
|
|
|
| 177 |
Args:
|
| 178 |
inputs_q: input queries of shape `[batch, q_length, q_features]`.
|
| 179 |
inputs_kv: key/values of shape `[batch, kv_length, kv_features]`.
|
|
|
|
| 181 |
bias: attention bias of shape `[batch, num_heads, q_length, kv_length]`.
|
| 182 |
decode: Whether to prepare and use an autoregressive cache.
|
| 183 |
deterministic: Disables dropout if set to True.
|
|
|
|
| 184 |
Returns:
|
| 185 |
output of shape `[batch, length, q_features]`.
|
| 186 |
"""
|
|
|
|
| 341 |
# ------------------------------------------------------------------------------
|
| 342 |
class DenseGeneral(nn.Module):
|
| 343 |
"""A linear transformation (without bias) with flexible axes.
|
|
|
|
| 344 |
Attributes:
|
| 345 |
features: tuple with numbers of output features.
|
| 346 |
axis: tuple with axes to apply the transformation on.
|
|
|
|
| 360 |
@nn.compact
|
| 361 |
def __call__(self, inputs: Array) -> Array:
|
| 362 |
"""Applies a linear transformation to the inputs along multiple dimensions.
|
|
|
|
| 363 |
Args:
|
| 364 |
inputs: The nd-array to be transformed.
|
|
|
|
| 365 |
Returns:
|
| 366 |
The transformed input.
|
| 367 |
"""
|
|
|
|
| 410 |
|
| 411 |
class MlpBlock(nn.Module):
|
| 412 |
"""Transformer MLP / feed-forward block.
|
|
|
|
| 413 |
Attributes:
|
| 414 |
intermediate_dim: Shared dimension of hidden layers.
|
| 415 |
activations: Type of activations for each layer. Each element is either
|
|
|
|
| 459 |
|
| 460 |
class Embed(nn.Module):
|
| 461 |
"""A parameterized function from integers [0, n) to d-dimensional vectors.
|
|
|
|
| 462 |
Attributes:
|
| 463 |
num_embeddings: number of embeddings.
|
| 464 |
features: number of feature dimensions for each embedding.
|
|
|
|
| 489 |
|
| 490 |
def __call__(self, inputs: Array) -> Array:
|
| 491 |
"""Embeds the inputs along the last dimension.
|
|
|
|
| 492 |
Args:
|
| 493 |
inputs: input data, all dimensions are considered batch dimensions.
|
|
|
|
| 494 |
Returns:
|
| 495 |
Output which is embedded input data. The output shape follows the input,
|
| 496 |
with an additional `features` dimension appended.
|
|
|
|
| 510 |
|
| 511 |
def attend(self, query: Array) -> Array:
|
| 512 |
"""Attend over the embedding using a query array.
|
|
|
|
| 513 |
Args:
|
| 514 |
query: array with last dimension equal the feature depth `features` of the
|
| 515 |
embedding.
|
|
|
|
| 516 |
Returns:
|
| 517 |
An array with final dim `num_embeddings` corresponding to the batched
|
| 518 |
inner-product of the array of query vectors against each embedding.
|
|
|
|
| 525 |
|
| 526 |
class RelativePositionBiases(nn.Module):
|
| 527 |
"""Adds T5-style relative positional embeddings to the attention logits.
|
|
|
|
| 528 |
Attributes:
|
| 529 |
num_buckets: Number of buckets to bucket distances between key and query
|
| 530 |
positions into.
|
|
|
|
| 545 |
@staticmethod
|
| 546 |
def _relative_position_bucket(relative_position, bidirectional=True, num_buckets=32, max_distance=128):
|
| 547 |
"""Translate relative position to a bucket number for relative attention.
|
|
|
|
| 548 |
The relative position is defined as memory_position - query_position, i.e.
|
| 549 |
the distance in tokens from the attending position to the attended-to
|
| 550 |
position. If bidirectional=False, then positive relative positions are
|
|
|
|
| 555 |
positions <=-max_distance map to the same bucket. This should allow for
|
| 556 |
more graceful generalization to longer sequences than the model has been
|
| 557 |
trained on.
|
|
|
|
| 558 |
Args:
|
| 559 |
relative_position: an int32 array
|
| 560 |
bidirectional: a boolean - whether the attention is bidirectional
|
| 561 |
num_buckets: an integer
|
| 562 |
max_distance: an integer
|
|
|
|
| 563 |
Returns:
|
| 564 |
a Tensor with the same shape as relative_position, containing int32
|
| 565 |
values in the range [0, num_buckets)
|
|
|
|
| 587 |
@nn.compact
|
| 588 |
def __call__(self, qlen, klen, bidirectional=True):
|
| 589 |
"""Produce relative position embedding attention biases.
|
|
|
|
| 590 |
Args:
|
| 591 |
qlen: attention query length.
|
| 592 |
klen: attention key length.
|
| 593 |
bidirectional: whether to allow positive memory-query relative position
|
| 594 |
embeddings.
|
|
|
|
| 595 |
Returns:
|
| 596 |
output: `(1, len, q_len, k_len)` attention bias
|
| 597 |
"""
|
|
|
|
| 715 |
dtype: DType = jnp.float32,
|
| 716 |
) -> Array:
|
| 717 |
"""Mask-making helper for attention weights.
|
|
|
|
| 718 |
In case of 1d inputs (i.e., `[batch, len_q]`, `[batch, len_kv]`, the
|
| 719 |
attention weights will be `[batch, heads, len_q, len_kv]` and this
|
| 720 |
function will produce `[batch, 1, len_q, len_kv]`.
|
|
|
|
| 721 |
Args:
|
| 722 |
query_input: a batched, flat input of query_length size
|
| 723 |
key_input: a batched, flat input of key_length size
|
|
|
|
| 725 |
extra_batch_dims: number of extra batch dims to add singleton axes for, none
|
| 726 |
by default
|
| 727 |
dtype: mask return dtype
|
|
|
|
| 728 |
Returns:
|
| 729 |
A `[batch, 1, len_q, len_kv]` shaped mask for 1d attention.
|
| 730 |
"""
|
|
|
|
| 744 |
|
| 745 |
def make_causal_mask(x: Array, extra_batch_dims: int = 0, dtype: DType = jnp.float32) -> Array:
|
| 746 |
"""Make a causal mask for self-attention.
|
|
|
|
| 747 |
In case of 1d inputs (i.e., `[batch, len]`, the self-attention weights
|
| 748 |
will be `[batch, heads, len, len]` and this function will produce a
|
| 749 |
causal mask of shape `[batch, 1, len, len]`.
|
|
|
|
| 750 |
Note that a causal mask does not depend on the values of x; it only depends on
|
| 751 |
the shape. If x has padding elements, they will not be treated in a special
|
| 752 |
manner.
|
|
|
|
| 753 |
Args:
|
| 754 |
x: input array of shape `[batch, len]`
|
| 755 |
extra_batch_dims: number of batch dims to add singleton axes for, none by
|
| 756 |
default
|
| 757 |
dtype: mask return dtype
|
|
|
|
| 758 |
Returns:
|
| 759 |
A `[batch, 1, len, len]` shaped causal mask for 1d attention.
|
| 760 |
"""
|
|
|
|
| 764 |
|
| 765 |
def combine_masks(*masks: Optional[Array], dtype: DType = jnp.float32):
|
| 766 |
"""Combine attention masks.
|
|
|
|
| 767 |
Args:
|
| 768 |
*masks: set of attention mask arguments to combine, some can be None.
|
| 769 |
dtype: final mask dtype
|
|
|
|
| 770 |
Returns:
|
| 771 |
Combined mask, reduced by logical and, returns None if no masks given.
|
| 772 |
"""
|
|
|
|
| 784 |
|
| 785 |
def combine_biases(*masks: Optional[Array]):
|
| 786 |
"""Combine attention biases.
|
|
|
|
| 787 |
Args:
|
| 788 |
*masks: set of attention bias arguments to combine, some can be None.
|
|
|
|
| 789 |
Returns:
|
| 790 |
Combined mask, reduced by summation, returns None if no masks given.
|
| 791 |
"""
|
|
|
|
| 808 |
decoder_segment_ids: Optional[Array] = None,
|
| 809 |
) -> Array:
|
| 810 |
"""Compute the self-attention mask for a decoder.
|
|
|
|
| 811 |
Decoder mask is formed by combining a causal mask, a padding mask and an
|
| 812 |
optional packing mask. If decoder_causal_attention is passed, it makes the
|
| 813 |
masking non-causal for positions that have value of 1.
|
|
|
|
| 814 |
A prefix LM is applied to a dataset which has a notion of "inputs" and
|
| 815 |
"targets", e.g., a machine translation task. The inputs and targets are
|
| 816 |
concatenated to form a new target. `decoder_target_tokens` is the concatenated
|
| 817 |
decoder output tokens.
|
|
|
|
| 818 |
The "inputs" portion of the concatenated sequence can attend to other "inputs"
|
| 819 |
tokens even for those at a later time steps. In order to control this
|
| 820 |
behavior, `decoder_causal_attention` is necessary. This is a binary mask with
|
| 821 |
a value of 1 indicating that the position belonged to "inputs" portion of the
|
| 822 |
original dataset.
|
|
|
|
| 823 |
Example:
|
|
|
|
| 824 |
Suppose we have a dataset with two examples.
|
|
|
|
| 825 |
ds = [{"inputs": [6, 7], "targets": [8]},
|
| 826 |
{"inputs": [3, 4], "targets": [5]}]
|
|
|
|
| 827 |
After the data preprocessing with packing, the two examples are packed into
|
| 828 |
one example with the following three fields (some fields are skipped for
|
| 829 |
simplicity).
|
|
|
|
| 830 |
decoder_target_tokens = [[6, 7, 8, 3, 4, 5, 0]]
|
| 831 |
decoder_segment_ids = [[1, 1, 1, 2, 2, 2, 0]]
|
| 832 |
decoder_causal_attention = [[1, 1, 0, 1, 1, 0, 0]]
|
|
|
|
| 833 |
where each array has [batch, length] shape with batch size being 1. Then,
|
| 834 |
this function computes the following mask.
|
|
|
|
| 835 |
mask = [[[[1, 1, 0, 0, 0, 0, 0],
|
| 836 |
[1, 1, 0, 0, 0, 0, 0],
|
| 837 |
[1, 1, 1, 0, 0, 0, 0],
|
|
|
|
| 839 |
[0, 0, 0, 1, 1, 0, 0],
|
| 840 |
[0, 0, 0, 1, 1, 1, 0],
|
| 841 |
[0, 0, 0, 0, 0, 0, 0]]]]
|
|
|
|
| 842 |
mask[b, 1, :, :] represents the mask for the example `b` in the batch.
|
| 843 |
Because mask is for a self-attention layer, the mask's shape is a square of
|
| 844 |
shape [query length, key length].
|
|
|
|
| 845 |
mask[b, 1, i, j] = 1 means that the query token at position i can attend to
|
| 846 |
the key token at position j.
|
|
|
|
| 847 |
Args:
|
| 848 |
decoder_target_tokens: decoder output tokens. [batch, length]
|
| 849 |
dtype: dtype of the output mask.
|
|
|
|
| 852 |
bidirectionally. [batch, length]
|
| 853 |
decoder_segment_ids: decoder segmentation info for packed examples. [batch,
|
| 854 |
length]
|
|
|
|
| 855 |
Returns:
|
| 856 |
the combined decoder mask.
|
| 857 |
"""
|
|
|
|
| 917 |
|
| 918 |
class _Conv(nn.Module):
|
| 919 |
"""Convolution Module wrapping `lax.conv_general_dilated[_local]`.
|
|
|
|
| 920 |
Attributes:
|
| 921 |
features: number of convolution filters.
|
| 922 |
kernel_size: shape of the convolutional kernel. For 1D convolution,
|
|
|
|
| 972 |
@property
|
| 973 |
def shared_weights(self) -> bool: # type: ignore
|
| 974 |
"""Defines whether weights are shared or not between different pixels.
|
|
|
|
| 975 |
Returns:
|
| 976 |
`True` to use shared weights in convolution (regular convolution).
|
| 977 |
`False` to use different weights at different pixels, a.k.a.
|
| 978 |
"locally connected layer", "unshared convolution", or "local convolution".
|
|
|
|
| 979 |
"""
|
| 980 |
...
|
| 981 |
|
| 982 |
@nn.compact
|
| 983 |
def __call__(self, inputs: Array) -> Array:
|
| 984 |
"""Applies a (potentially unshared) convolution to the inputs.
|
|
|
|
| 985 |
Args:
|
| 986 |
inputs: input data with dimensions (*batch_dims, spatial_dims...,
|
| 987 |
features). This is the channels-last convention, i.e. NHWC for a 2d
|
|
|
|
| 994 |
better performance than this default flattening approach. If the input
|
| 995 |
lacks a batch dimension it will be added for the convolution and removed
|
| 996 |
n return, an allowance made to enable writing single-example code.
|
|
|
|
| 997 |
Returns:
|
| 998 |
The convolved data.
|
| 999 |
"""
|
|
|
|
| 1150 |
|
| 1151 |
class Conv(_Conv):
|
| 1152 |
"""Convolution Module wrapping `lax.conv_general_dilated`.
|
|
|
|
| 1153 |
Attributes:
|
| 1154 |
features: number of convolution filters.
|
| 1155 |
kernel_size: shape of the convolutional kernel. For 1D convolution,
|
|
|
|
| 1187 |
|
| 1188 |
@property
|
| 1189 |
def shared_weights(self) -> bool:
|
| 1190 |
+
return True
|