# This script is based on https://github.com/huggingface/transformers/blob/v4.29.1/src/transformers/models/whisper/modeling_whisper.py
""" PyTorch Whisper model."""
import math
import random
from typing import Optional, Tuple, Union
import numpy as np
import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import CrossEntropyLoss
from transformers.activations import ACT2FN
from transformers.generation.logits_process import WhisperTimeStampLogitsProcessor
from transformers.modeling_outputs import (
BaseModelOutput,
BaseModelOutputWithPastAndCrossAttentions,
Seq2SeqLMOutput,
Seq2SeqModelOutput,
SequenceClassifierOutput,
)
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
from transformers.models.whisper.configuration_whisper import WhisperConfig
from transformers.models.whisper.tokenization_whisper import TASK_IDS, TO_LANGUAGE_CODE
logger = logging.get_logger(__name__)
_CONFIG_FOR_DOC = "WhisperConfig"
_CHECKPOINT_FOR_DOC = "openai/whisper-tiny"
WHISPER_PRETRAINED_MODEL_ARCHIVE_LIST = [
"openai/whisper-base",
# See all Whisper models at https://huggingface.co/models?filter=whisper
]
# Copied from transformers.models.bart.modeling_bart.shift_tokens_right
def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int):
"""
Shift input ids one token to the right.
"""
shifted_input_ids = input_ids.new_zeros(input_ids.shape)
shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
shifted_input_ids[:, 0] = decoder_start_token_id
if pad_token_id is None:
raise ValueError("self.model.config.pad_token_id has to be defined.")
# replace possible -100 values in labels by `pad_token_id`
shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)
return shifted_input_ids
# Copied from transformers.models.bart.modeling_bart._make_causal_mask
def _make_causal_mask(
input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
):
"""
Make causal mask used for bi-directional self-attention.
"""
bsz, tgt_len = input_ids_shape
mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
mask_cond = torch.arange(mask.size(-1), device=device)
mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
mask = mask.to(dtype)
if past_key_values_length > 0:
mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
# Copied from transformers.models.bart.modeling_bart._expand_mask
def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
"""
Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
"""
bsz, src_len = mask.size()
tgt_len = tgt_len if tgt_len is not None else src_len
expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
inverted_mask = 1.0 - expanded_mask
return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
# Copied from transformers.models.wav2vec2.modeling_wav2vec2._compute_mask_indices
def _compute_mask_indices(
shape: Tuple[int, int],
mask_prob: float,
mask_length: int,
attention_mask: Optional[torch.LongTensor] = None,
min_masks: int = 0,
) -> np.ndarray:
"""
Computes random mask spans for a given shape. Used to implement [SpecAugment: A Simple Data Augmentation Method for
ASR](https://arxiv.org/abs/1904.08779). Note that this method is not optimized to run on TPU and should be run on
CPU as part of the preprocessing during training.
Args:
shape: The shape for which to compute masks. This should be of a tuple of size 2 where
the first element is the batch size and the second element is the length of the axis to span.
mask_prob: The percentage of the whole axis (between 0 and 1) which will be masked. The number of
independently generated mask spans of length `mask_length` is computed by
`mask_prob*shape[1]/mask_length`. Note that due to overlaps, `mask_prob` is an upper bound and the
actual percentage will be smaller.
mask_length: size of the mask
min_masks: minimum number of masked spans
attention_mask: A (right-padded) attention mask which independently shortens the feature axis of
each batch dimension.
"""
batch_size, sequence_length = shape
if mask_length < 1:
raise ValueError("`mask_length` has to be bigger than 0.")
if mask_length > sequence_length:
raise ValueError(
f"`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: {mask_length}"
f" and `sequence_length`: {sequence_length}`"
)
# epsilon is used for probabilistic rounding
epsilon = np.random.rand(1).item()
def compute_num_masked_span(input_length):
"""Given input length, compute how many spans should be masked"""
num_masked_span = int(mask_prob * input_length / mask_length + epsilon)
num_masked_span = max(num_masked_span, min_masks)
# make sure num masked span <= sequence_length
if num_masked_span * mask_length > sequence_length:
num_masked_span = sequence_length // mask_length
# make sure num_masked span is also <= input_length - (mask_length - 1)
if input_length - (mask_length - 1) < num_masked_span:
num_masked_span = max(input_length - (mask_length - 1), 0)
return num_masked_span
# compute number of masked spans in batch
input_lengths = (
attention_mask.sum(-1).detach().tolist()
if attention_mask is not None
else [sequence_length for _ in range(batch_size)]
)
# SpecAugment mask to fill
spec_aug_mask = np.zeros((batch_size, sequence_length), dtype=bool)
spec_aug_mask_idxs = []
max_num_masked_span = compute_num_masked_span(sequence_length)
if max_num_masked_span == 0:
return spec_aug_mask
for input_length in input_lengths:
# compute num of masked spans for this input
num_masked_span = compute_num_masked_span(input_length)
# get random indices to mask
spec_aug_mask_idx = np.random.choice(
np.arange(input_length - (mask_length - 1)), num_masked_span, replace=False
)
# pick first sampled index that will serve as a dummy index to pad vector
# to ensure same dimension for all batches due to probabilistic rounding
# Picking first sample just pads those vectors twice.
if len(spec_aug_mask_idx) == 0:
# this case can only happen if `input_length` is strictly smaller then
# `sequence_length` in which case the last token has to be a padding
# token which we can use as a dummy mask id
dummy_mask_idx = sequence_length - 1
else:
dummy_mask_idx = spec_aug_mask_idx[0]
spec_aug_mask_idx = np.concatenate(
[spec_aug_mask_idx, np.ones(max_num_masked_span - num_masked_span, dtype=np.int32) * dummy_mask_idx]
)
spec_aug_mask_idxs.append(spec_aug_mask_idx)
spec_aug_mask_idxs = np.array(spec_aug_mask_idxs)
# expand masked indices to masked spans
spec_aug_mask_idxs = np.broadcast_to(
spec_aug_mask_idxs[:, :, None], (batch_size, max_num_masked_span, mask_length)
)
spec_aug_mask_idxs = spec_aug_mask_idxs.reshape(batch_size, max_num_masked_span * mask_length)
# add offset to the starting indexes so that indexes now create a span
offsets = np.arange(mask_length)[None, None, :]
offsets = np.broadcast_to(offsets, (batch_size, max_num_masked_span, mask_length)).reshape(
batch_size, max_num_masked_span * mask_length
)
spec_aug_mask_idxs = spec_aug_mask_idxs + offsets
# ensure that we cannot have indices larger than sequence_length
if spec_aug_mask_idxs.max() > sequence_length - 1:
spec_aug_mask_idxs[spec_aug_mask_idxs > sequence_length - 1] = sequence_length - 1
# scatter indices to mask
np.put_along_axis(spec_aug_mask, spec_aug_mask_idxs, 1, -1)
return spec_aug_mask
class WhisperPositionalEmbedding(nn.Embedding):
def __init__(self, num_positions: int, embedding_dim: int, padding_idx: Optional[int] = None):
super().__init__(num_positions, embedding_dim)
def forward(self, input_ids, past_key_values_length=0):
return self.weight[past_key_values_length : past_key_values_length + input_ids.shape[1]]
class WhisperAttention(nn.Module):
"""Multi-headed attention from 'Attention Is All You Need' paper"""
def __init__(
self,
embed_dim: int,
num_heads: int,
dropout: float = 0.0,
is_decoder: bool = False,
bias: bool = True,
):
super().__init__()
self.embed_dim = embed_dim
self.num_heads = num_heads
self.dropout = dropout
self.head_dim = embed_dim // num_heads
if (self.head_dim * num_heads) != self.embed_dim:
raise ValueError(
f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
f" and `num_heads`: {num_heads})."
)
self.scaling = self.head_dim**-0.5
self.is_decoder = is_decoder
self.k_proj = nn.Linear(embed_dim, embed_dim, bias=False)
self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
# Copied from transformers.models.bart.modeling_bart.BartAttention._shape with BART->whisper
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
# Copied from transformers.models.bart.modeling_bart.BartAttention.forward with BART->whisper
def forward(
self,
hidden_states: torch.Tensor,
key_value_states: Optional[torch.Tensor] = None,
past_key_value: Optional[Tuple[torch.Tensor]] = None,
attention_mask: Optional[torch.Tensor] = None,
layer_head_mask: Optional[torch.Tensor] = None,
output_attentions: bool = False,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
"""Input shape: Batch x Time x Channel"""
# if key_value_states are provided this layer is used as a cross-attention layer
# for the decoder
is_cross_attention = key_value_states is not None
bsz, tgt_len, _ = hidden_states.size()
# get query proj
query_states = self.q_proj(hidden_states) * self.scaling
# get key, value proj
# `past_key_value[0].shape[2] == key_value_states.shape[1]`
# is checking that the `sequence_length` of the `past_key_value` is the same as
# the provided `key_value_states` to support prefix tuning
if (
is_cross_attention
and past_key_value is not None
and past_key_value[0].shape[2] == key_value_states.shape[1]
):
# reuse k,v, cross_attentions
key_states = past_key_value[0]
value_states = past_key_value[1]
elif is_cross_attention:
# cross_attentions
key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
elif past_key_value is not None:
# reuse k, v, self_attention
key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
key_states = torch.cat([past_key_value[0], key_states], dim=2)
value_states = torch.cat([past_key_value[1], value_states], dim=2)
else:
# self_attention
key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
if self.is_decoder:
# if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
# Further calls to cross_attention layer can then reuse all cross-attention
# key/value_states (first "if" case)
# if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
# all previous decoder key/value_states. Further calls to uni-directional self-attention
# can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
# if encoder bi-directional self-attention `past_key_value` is always `None`
past_key_value = (key_states, value_states)
proj_shape = (bsz * self.num_heads, -1, self.head_dim)
query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
key_states = key_states.reshape(*proj_shape)
value_states = value_states.reshape(*proj_shape)
src_len = key_states.size(1)
attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
raise ValueError(
f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
f" {attn_weights.size()}"
)
if attention_mask is not None:
if attention_mask.size() != (bsz, 1, tgt_len, src_len):
raise ValueError(
f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
)
attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
attn_weights = nn.functional.softmax(attn_weights, dim=-1)
if layer_head_mask is not None:
if layer_head_mask.size() != (self.num_heads,):
raise ValueError(
f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
f" {layer_head_mask.size()}"
)
attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
if output_attentions:
# this operation is a bit awkward, but it's required to
# make sure that attn_weights keeps its gradient.
# In order to do so, attn_weights have to be reshaped
# twice and have to be reused in the following
attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
else:
attn_weights_reshaped = None
attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
attn_output = torch.bmm(attn_probs, value_states)
if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
raise ValueError(
f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
f" {attn_output.size()}"
)
attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
attn_output = attn_output.transpose(1, 2)
# Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
# partitioned across GPUs when using tensor-parallelism.
attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
attn_output = self.out_proj(attn_output)
return attn_output, attn_weights_reshaped, past_key_value
# Copied from transformers.models.mbart.modeling_mbart.MBartEncoderLayer with MBart->Whisper
class WhisperEncoderLayer(nn.Module):
def __init__(self, config: WhisperConfig):
super().__init__()
self.embed_dim = config.d_model
self.self_attn = WhisperAttention(
embed_dim=self.embed_dim,
num_heads=config.encoder_attention_heads,
dropout=config.attention_dropout,
)
self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
self.dropout = config.dropout
self.activation_fn = ACT2FN[config.activation_function]
self.activation_dropout = config.activation_dropout
self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
self.final_layer_norm = nn.LayerNorm(self.embed_dim)
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: torch.Tensor,
layer_head_mask: torch.Tensor,
output_attentions: bool = False,
) -> torch.Tensor:
"""
Args:
hidden_states (`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
attention_mask (`torch.FloatTensor`): attention mask of size
`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
`(encoder_attention_heads,)`.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
"""
residual = hidden_states
hidden_states = self.self_attn_layer_norm(hidden_states)
hidden_states, attn_weights, _ = self.self_attn(
hidden_states=hidden_states,
attention_mask=attention_mask,
layer_head_mask=layer_head_mask,
output_attentions=output_attentions,
)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
residual = hidden_states
hidden_states = self.final_layer_norm(hidden_states)
hidden_states = self.activation_fn(self.fc1(hidden_states))
hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
hidden_states = self.fc2(hidden_states)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
if hidden_states.dtype == torch.float16 and (
torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any()
):
clamp_value = torch.finfo(hidden_states.dtype).max - 1000
hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
outputs = (hidden_states,)
if output_attentions:
outputs += (attn_weights,)
return outputs
# Copied from transformers.models.mbart.modeling_mbart.MBartDecoderLayer with MBart->Whisper
class WhisperDecoderLayer(nn.Module):
def __init__(self, config: WhisperConfig):
super().__init__()
self.embed_dim = config.d_model
self.self_attn = WhisperAttention(
embed_dim=self.embed_dim,
num_heads=config.decoder_attention_heads,
dropout=config.attention_dropout,
is_decoder=True,
)
self.dropout = config.dropout
self.activation_fn = ACT2FN[config.activation_function]
self.activation_dropout = config.activation_dropout
self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
self.encoder_attn = WhisperAttention(
self.embed_dim,
config.decoder_attention_heads,
dropout=config.attention_dropout,
is_decoder=True,
)
self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
self.final_layer_norm = nn.LayerNorm(self.embed_dim)
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
encoder_hidden_states: Optional[torch.Tensor] = None,
encoder_attention_mask: Optional[torch.Tensor] = None,
layer_head_mask: Optional[torch.Tensor] = None,
cross_attn_layer_head_mask: Optional[torch.Tensor] = None,
past_key_value: Optional[Tuple[torch.Tensor]] = None,
output_attentions: Optional[bool] = False,
use_cache: Optional[bool] = True,
) -> torch.Tensor:
"""
Args:
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
attention_mask (`torch.FloatTensor`): attention mask of size
`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
encoder_hidden_states (`torch.FloatTensor`):
cross attention input to the layer of shape `(batch, seq_len, embed_dim)`
encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size
`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
`(encoder_attention_heads,)`.
cross_attn_layer_head_mask (`torch.FloatTensor`): mask for cross-attention heads in a given layer of
size `(decoder_attention_heads,)`.
past_key_value (`Tuple(torch.FloatTensor)`): cached past key and value projection states
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
"""
residual = hidden_states
hidden_states = self.self_attn_layer_norm(hidden_states)
# Self Attention
# decoder uni-directional self-attention cached key/values tuple is at positions 1,2
self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
# add present self-attn cache to positions 1,2 of present_key_value tuple
hidden_states, self_attn_weights, present_key_value = self.self_attn(
hidden_states=hidden_states,
past_key_value=self_attn_past_key_value,
attention_mask=attention_mask,
layer_head_mask=layer_head_mask,
output_attentions=output_attentions,
)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
# Cross-Attention Block
cross_attn_present_key_value = None
cross_attn_weights = None
if encoder_hidden_states is not None:
residual = hidden_states
hidden_states = self.encoder_attn_layer_norm(hidden_states)
# cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
hidden_states=hidden_states,
key_value_states=encoder_hidden_states,
attention_mask=encoder_attention_mask,
layer_head_mask=cross_attn_layer_head_mask,
past_key_value=cross_attn_past_key_value,
output_attentions=output_attentions,
)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
# add cross-attn to positions 3,4 of present_key_value tuple
present_key_value = present_key_value + cross_attn_present_key_value
# Fully Connected
residual = hidden_states
hidden_states = self.final_layer_norm(hidden_states)
hidden_states = self.activation_fn(self.fc1(hidden_states))
hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
hidden_states = self.fc2(hidden_states)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
outputs = (hidden_states,)
if output_attentions:
outputs += (self_attn_weights, cross_attn_weights)
if use_cache:
outputs += (present_key_value,)
return outputs
class WhisperPreTrainedModel(PreTrainedModel):
config_class = WhisperConfig
base_model_prefix = "model"
main_input_name = "input_features"
supports_gradient_checkpointing = True
_no_split_modules = ["WhisperEncoderLayer", "WhisperDecoderLayer"]
def _init_weights(self, module):
std = self.config.init_std
if isinstance(module, (nn.Linear, nn.Conv1d)):
module.weight.data.normal_(mean=0.0, std=std)
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, nn.Embedding):
module.weight.data.normal_(mean=0.0, std=std)
if module.padding_idx is not None:
module.weight.data[module.padding_idx].zero_()
def _set_gradient_checkpointing(self, module, value=False):
if isinstance(module, (WhisperDecoder, WhisperEncoder)):
module.gradient_checkpointing = value
def _get_feat_extract_output_lengths(self, input_lengths: torch.LongTensor):
"""
Computes the output length of the convolutional layers
"""
input_lengths = (input_lengths - 1) // 2 + 1
return input_lengths
WHISPER_START_DOCSTRING = r"""
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.
Parameters:
config ([`WhisperConfig`]):
Model configuration class with all the parameters of the model. Initializing with a config file does not
load the weights associated with the model, only the configuration. Check out the
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""
WHISPER_INPUTS_DOCSTRING = r"""
Args:
input_features (`torch.FloatTensor` of shape `(batch_size, feature_size, sequence_length)`):
Float values mel features extracted from the raw speech waveform. Raw speech waveform can be obtained by
loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via
the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
[`AutoFeatureExtractor`] should be used for extracting the mel features, padding and conversion into a
tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing *SpecAugment* data augmentation on padding token indices. Mask values selected in
`[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
Indices of decoder input sequence tokens in the vocabulary.
Indices can be obtained using [`WhisperTokenizer`]. See [`PreTrainedTokenizer.encode`] and
[`PreTrainedTokenizer.__call__`] for details.
[What are decoder input IDs?](../glossary#decoder-input-ids)
Whisper uses the `decoder_start_token_id` as the starting token for `decoder_input_ids` generation. If
`past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
`past_key_values`).
decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
be used by default.
If you want to change padding behavior, you should read
[`modeling_whisper._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the BART
paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy.
head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
decoder_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in `[0, 1]`:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*):
Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
`last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of
hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
`decoder_input_ids` of shape `(batch_size, sequence_length)`.
decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*):
Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded
representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be
input (see `past_key_values`). This is useful if you want more control over how to convert
`decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix.
use_cache (`bool`, *optional*):
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
`past_key_values`).
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
WHISPER_ENCODER_INPUTS_DOCSTRING = r"""
Args:
input_features (`torch.FloatTensor` of shape `(batch_size, feature_size, sequence_length)`):
Float values mel features extracted from the raw speech waveform. Raw speech waveform can be obtained by
loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via
the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
[`AutoFeatureExtractor`] should be used for extracting the mel features, padding and conversion into a
tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*):
Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
`last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of
hidden-states at the output of the last layer of the encoder.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
class WhisperEncoder(WhisperPreTrainedModel):
"""
Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
[`WhisperEncoderLayer`].
Args:
config: WhisperConfig
"""
def __init__(self, config: WhisperConfig):
super().__init__(config)
self.dropout = config.dropout
self.layerdrop = config.encoder_layerdrop
embed_dim = config.d_model
self.num_mel_bins = config.num_mel_bins
self.padding_idx = config.pad_token_id
self.max_source_positions = config.max_source_positions
self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0
self.conv1 = nn.Conv1d(self.num_mel_bins, embed_dim, kernel_size=3, padding=1)
self.conv2 = nn.Conv1d(embed_dim, embed_dim, kernel_size=3, stride=2, padding=1)
self.embed_positions = nn.Embedding(self.max_source_positions, embed_dim)
self.layers = nn.ModuleList([WhisperEncoderLayer(config) for _ in range(config.encoder_layers)])
self.layer_norm = nn.LayerNorm(config.d_model)
self.gradient_checkpointing = False
# Initialize weights and apply final processing
self.post_init()
def _freeze_parameters(self):
for param in self.parameters():
param.requires_grad = False
self._requires_grad = False
def get_input_embeddings(self) -> nn.Module:
return self.conv1
def set_input_embeddings(self, value: nn.Module):
self.conv1 = value
def forward(
self,
input_features,
attention_mask=None,
head_mask=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
Args:
input_features (`torch.LongTensor` of shape `(batch_size, feature_size, sequence_length)`):
Float values of mel features extracted from the raw speech waveform. Raw speech waveform can be
obtained by loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a
`numpy.ndarray`, *e.g.* via the soundfile library (`pip install soundfile`). To prepare the array into
`input_features`, the [`AutoFeatureExtractor`] should be used for extracting the mel features, padding
and conversion into a tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
attention_mask (`torch.Tensor`)`, *optional*):
Whisper does not support masking of the `input_features`, this argument is preserved for compatibility,
but it is not used. By default the silence in the input log mel spectrogram are ignored.
head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
for more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
inputs_embeds = nn.functional.gelu(self.conv1(input_features))
inputs_embeds = nn.functional.gelu(self.conv2(inputs_embeds))
inputs_embeds = inputs_embeds.permute(0, 2, 1)
embed_pos = self.embed_positions.weight
hidden_states = inputs_embeds + embed_pos
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
encoder_states = () if output_hidden_states else None
all_attentions = () if output_attentions else None
# check if head_mask has a correct number of layers specified if desired
if head_mask is not None:
assert head_mask.size()[0] == (
len(self.layers)
), f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
for idx, encoder_layer in enumerate(self.layers):
if output_hidden_states:
encoder_states = encoder_states + (hidden_states,)
# add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
dropout_probability = random.uniform(0, 1)
if self.training and (dropout_probability < self.layerdrop): # skip the layer
layer_outputs = (None, None)
else:
if self.gradient_checkpointing and self.training:
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs, output_attentions)
return custom_forward
layer_outputs = torch.utils.checkpoint.checkpoint(
create_custom_forward(encoder_layer),
hidden_states,
None,
(head_mask[idx] if head_mask is not None else None),
)
else:
layer_outputs = encoder_layer(
hidden_states,
None,
layer_head_mask=(head_mask[idx] if head_mask is not None else None),
output_attentions=output_attentions,
)
hidden_states = layer_outputs[0]
if output_attentions:
all_attentions = all_attentions + (layer_outputs[1],)
hidden_states = self.layer_norm(hidden_states)
if output_hidden_states:
encoder_states = encoder_states + (hidden_states,)
if not return_dict:
return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
return BaseModelOutput(
last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
)
class WhisperDecoder(WhisperPreTrainedModel):
"""
Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`WhisperDecoderLayer`]
Args:
config: WhisperConfig
"""
def __init__(self, config: WhisperConfig):
super().__init__(config)
self.dropout = config.dropout
self.layerdrop = config.decoder_layerdrop
self.padding_idx = config.pad_token_id
self.max_target_positions = config.max_target_positions
self.max_source_positions = config.max_source_positions
self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx)
self.embed_positions = WhisperPositionalEmbedding(self.max_target_positions, config.d_model)
self.layers = nn.ModuleList([WhisperDecoderLayer(config) for _ in range(config.decoder_layers)])
self.layer_norm = nn.LayerNorm(config.d_model)
self.gradient_checkpointing = False
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self):
return self.embed_tokens
def set_input_embeddings(self, value):
self.embed_tokens = value
def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
# create causal mask
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
combined_attention_mask = None
if input_shape[-1] > 1:
combined_attention_mask = _make_causal_mask(
input_shape,
inputs_embeds.dtype,
device=inputs_embeds.device,
past_key_values_length=past_key_values_length,
)
if attention_mask is not None:
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])
combined_attention_mask = (
expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
)
return combined_attention_mask
def forward(
self,
input_ids=None,
attention_mask=None,
encoder_hidden_states=None,
head_mask=None,
cross_attn_head_mask=None,
past_key_values=None,
inputs_embeds=None,
use_cache=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
Args:
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
provide it.
Indices can be obtained using [`WhisperTokenizer`]. See [`PreTrainedTokenizer.encode`] and
[`PreTrainedTokenizer.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
of the decoder.
head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
Mask to nullify selected heads of the attention modules in encoder to avoid performing cross-attention
on hidden heads. Mask values selected in `[0, 1]`:
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of
shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing
`input_ids` you can choose to directly pass an embedded representation. This is useful if you want more
control over how to convert `input_ids` indices into associated vectors than the model's internal
embedding lookup matrix.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
for more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
use_cache = use_cache if use_cache is not None else self.config.use_cache
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
# retrieve input_ids and inputs_embeds
if input_ids is not None and inputs_embeds is not None:
raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
elif input_ids is not None:
input_shape = input_ids.size()
input_ids = input_ids.view(-1, input_shape[-1])
elif inputs_embeds is not None:
input_shape = inputs_embeds.size()[:-1]
else:
raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
# past_key_values_length
past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
if inputs_embeds is None:
inputs_embeds = self.embed_tokens(input_ids)
attention_mask = self._prepare_decoder_attention_mask(
attention_mask, input_shape, inputs_embeds, past_key_values_length
)
# embed positions
if input_ids is not None:
positions = self.embed_positions(input_ids, past_key_values_length=past_key_values_length)
else:
positions = self.embed_positions(inputs_embeds, past_key_values_length=past_key_values_length)
hidden_states = inputs_embeds + positions
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
if self.gradient_checkpointing and self.training:
if use_cache:
logger.warning_once(
"`use_cache = True` is incompatible with gradient checkpointing. Setting `use_cache = False`..."
)
use_cache = False
# decoder layers
all_hidden_states = () if output_hidden_states else None
all_self_attns = () if output_attentions else None
all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
next_decoder_cache = () if use_cache else None
# check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
for attn_mask, mask_name in zip([head_mask, cross_attn_head_mask], ["head_mask", "cross_attn_head_mask"]):
if attn_mask is not None:
assert attn_mask.size()[0] == (len(self.layers)), (
f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
f" {head_mask.size()[0]}."
)
for idx, decoder_layer in enumerate(self.layers):
# add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
if output_hidden_states:
all_hidden_states += (hidden_states,)
dropout_probability = random.uniform(0, 1)
if self.training and (dropout_probability < self.layerdrop):
continue
past_key_value = past_key_values[idx] if past_key_values is not None else None
if self.gradient_checkpointing and self.training:
def create_custom_forward(module):
def custom_forward(*inputs):
# None for past_key_value
return module(*inputs, output_attentions, use_cache)
return custom_forward
layer_outputs = torch.utils.checkpoint.checkpoint(
create_custom_forward(decoder_layer),
hidden_states,
attention_mask,
encoder_hidden_states,
None, # encoder attention mask
head_mask[idx] if head_mask is not None else None,
cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None,
None, # past_key_value
)
else:
layer_outputs = decoder_layer(
hidden_states,
attention_mask=attention_mask,
encoder_hidden_states=encoder_hidden_states,
layer_head_mask=(head_mask[idx] if head_mask is not None else None),
cross_attn_layer_head_mask=(
cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None
),
past_key_value=past_key_value,
output_attentions=output_attentions,
use_cache=use_cache,
)
hidden_states = layer_outputs[0]
if use_cache:
next_decoder_cache += (layer_outputs[3 if output_attentions else 1],)
if output_attentions:
all_self_attns += (layer_outputs[1],)
if encoder_hidden_states is not None:
all_cross_attentions += (layer_outputs[2],)
hidden_states = self.layer_norm(hidden_states)
# add hidden states from the last decoder layer
if output_hidden_states:
all_hidden_states += (hidden_states,)
next_cache = next_decoder_cache if use_cache else None
if not return_dict:
return tuple(
v
for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
if v is not None
)
return BaseModelOutputWithPastAndCrossAttentions(
last_hidden_state=hidden_states,
past_key_values=next_cache,
hidden_states=all_hidden_states,
attentions=all_self_attns,
cross_attentions=all_cross_attentions,
)
@add_start_docstrings(
"The bare Whisper Model outputting raw hidden-states without any specific head on top.",
WHISPER_START_DOCSTRING,
)
class WhisperModel(WhisperPreTrainedModel):
_keys_to_ignore_on_load_missing = [r"proj_out.weight"]
def __init__(self, config: WhisperConfig):
super().__init__(config)
self.encoder = WhisperEncoder(config)
self.decoder = WhisperDecoder(config)
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self):
return self.decoder.embed_tokens
def set_input_embeddings(self, value):
self.decoder.embed_tokens = value
def get_encoder(self):
return self.encoder
def get_decoder(self):
return self.decoder
def freeze_encoder(self):
"""
Calling this function will disable the gradient computation for the Whisper encoder so that its parameters will
not be updated during training.
"""
self.encoder._freeze_parameters()
def _mask_input_features(
self,
input_features: torch.FloatTensor,
attention_mask: Optional[torch.LongTensor] = None,
):
"""
Masks extracted features along time axis and/or along feature axis according to
[SpecAugment](https://arxiv.org/abs/1904.08779).
"""
# `config.apply_spec_augment` can set masking to False
if not getattr(self.config, "apply_spec_augment", True):
return input_features
# generate indices & apply SpecAugment along time axis
batch_size, hidden_size, sequence_length = input_features.size()
if self.config.mask_time_prob > 0 and self.training:
# generate indices & apply SpecAugment along time axis
mask_time_indices = _compute_mask_indices(
(batch_size, sequence_length),
mask_prob=self.config.mask_time_prob,
mask_length=self.config.mask_time_length,
attention_mask=attention_mask,
min_masks=self.config.mask_time_min_masks,
)
mask_time_indices = torch.tensor(mask_time_indices, device=input_features.device, dtype=torch.bool)
mask_time_indices = mask_time_indices[:, None].expand(-1, hidden_size, -1)
input_features[mask_time_indices] = 0
if self.config.mask_feature_prob > 0 and self.training:
# generate indices & apply SpecAugment along feature axis
mask_feature_indices = _compute_mask_indices(
(batch_size, hidden_size),
mask_prob=self.config.mask_feature_prob,
mask_length=self.config.mask_feature_length,
min_masks=self.config.mask_feature_min_masks,
)
mask_feature_indices = torch.tensor(mask_feature_indices, device=input_features.device, dtype=torch.bool)
input_features[mask_feature_indices] = 0
return input_features
@add_start_docstrings_to_model_forward(WHISPER_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=Seq2SeqModelOutput, config_class=_CONFIG_FOR_DOC)
def forward(
self,
input_features: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.LongTensor] = None,
decoder_input_ids: Optional[torch.LongTensor] = None,
decoder_attention_mask: Optional[torch.LongTensor] = None,
head_mask: Optional[torch.Tensor] = None,
decoder_head_mask: Optional[torch.Tensor] = None,
cross_attn_head_mask: Optional[torch.Tensor] = None,
encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
decoder_inputs_embeds: Optional[Tuple[torch.FloatTensor]] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple[torch.Tensor], Seq2SeqModelOutput]:
r"""
Returns:
Example:
```python
>>> import torch
>>> from transformers import AutoFeatureExtractor, WhisperModel
>>> from datasets import load_dataset
>>> model = WhisperModel.from_pretrained("openai/whisper-base")
>>> feature_extractor = AutoFeatureExtractor.from_pretrained("openai/whisper-base")
>>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
>>> inputs = feature_extractor(ds[0]["audio"]["array"], return_tensors="pt")
>>> input_features = inputs.input_features
>>> decoder_input_ids = torch.tensor([[1, 1]]) * model.config.decoder_start_token_id
>>> last_hidden_state = model(input_features, decoder_input_ids=decoder_input_ids).last_hidden_state
>>> list(last_hidden_state.shape)
[1, 2, 512]
```"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
use_cache = use_cache if use_cache is not None else self.config.use_cache
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
if encoder_outputs is None:
input_features = self._mask_input_features(input_features, attention_mask=attention_mask)
encoder_outputs = self.encoder(
input_features,
head_mask=head_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
# If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
encoder_outputs = BaseModelOutput(
last_hidden_state=encoder_outputs[0],
hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
)
# decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
decoder_outputs = self.decoder(
input_ids=decoder_input_ids,
attention_mask=decoder_attention_mask,
encoder_hidden_states=encoder_outputs[0],
head_mask=decoder_head_mask,
cross_attn_head_mask=cross_attn_head_mask,
past_key_values=past_key_values,
inputs_embeds=decoder_inputs_embeds,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
if not return_dict:
return decoder_outputs + encoder_outputs
return Seq2SeqModelOutput(
last_hidden_state=decoder_outputs.last_hidden_state,
past_key_values=decoder_outputs.past_key_values,
decoder_hidden_states=decoder_outputs.hidden_states,
decoder_attentions=decoder_outputs.attentions,
cross_attentions=decoder_outputs.cross_attentions,
encoder_last_hidden_state=encoder_outputs.last_hidden_state,
encoder_hidden_states=encoder_outputs.hidden_states,
encoder_attentions=encoder_outputs.attentions,
)
@add_start_docstrings(
"The Whisper Model with a language modeling head. Can be used for automatic speech recognition.",
WHISPER_START_DOCSTRING,
)
class WhisperForConditionalGeneration(WhisperPreTrainedModel):
base_model_prefix = "model"
_keys_to_ignore_on_load_missing = [
r"encoder.version",
r"decoder.version",
r"proj_out.weight",
]
_keys_to_ignore_on_save = [
r"proj_out.weight",
]
def __init__(self, config: WhisperConfig):
super().__init__(config)
self.model = WhisperModel(config)
self.proj_out = nn.Linear(config.d_model, config.vocab_size, bias=False)
# Initialize weights and apply final processing
self.post_init()
def get_encoder(self):
return self.model.get_encoder()
def get_decoder(self):
return self.model.get_decoder()
def resize_token_embeddings(self, new_num_tokens: int) -> nn.Embedding:
new_embeddings = super().resize_token_embeddings(new_num_tokens)
return new_embeddings
def get_output_embeddings(self):
return self.proj_out
def set_output_embeddings(self, new_embeddings):
self.proj_out = new_embeddings
def get_input_embeddings(self) -> nn.Module:
return self.model.get_input_embeddings()
def freeze_encoder(self):
"""
Calling this function will disable the gradient computation for the Whisper encoder so that its parameters will
not be updated during training.
"""
self.model.encoder._freeze_parameters()
@add_start_docstrings_to_model_forward(WHISPER_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
def forward(
self,
input_features: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.LongTensor] = None,
decoder_input_ids: Optional[torch.LongTensor] = None,
decoder_attention_mask: Optional[torch.LongTensor] = None,
head_mask: Optional[torch.Tensor] = None,
decoder_head_mask: Optional[torch.Tensor] = None,
cross_attn_head_mask: Optional[torch.Tensor] = None,
encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
decoder_inputs_embeds: Optional[Tuple[torch.FloatTensor]] = None,
labels: Optional[torch.LongTensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple[torch.Tensor], Seq2SeqLMOutput]:
r"""
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Labels for computing the language modeling loss. Indices should either be in `[0, ..., config.vocab_size]`
or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is
only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
Returns:
Example:
```python
>>> import torch
>>> from transformers import AutoProcessor, WhisperForConditionalGeneration
>>> from datasets import load_dataset
>>> processor = AutoProcessor.from_pretrained("openai/whisper-tiny.en")
>>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
>>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
>>> inputs = processor(ds[0]["audio"]["array"], return_tensors="pt")
>>> input_features = inputs.input_features
>>> generated_ids = model.generate(inputs=input_features)
>>> transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
>>> transcription
' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.'
```"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
if labels is not None:
if decoder_input_ids is None and decoder_inputs_embeds is None:
decoder_input_ids = shift_tokens_right(
labels, self.config.pad_token_id, self.config.decoder_start_token_id
)
outputs = self.model(
input_features,
attention_mask=attention_mask,
decoder_input_ids=decoder_input_ids,
encoder_outputs=encoder_outputs,
decoder_attention_mask=decoder_attention_mask,
head_mask=head_mask,
decoder_head_mask=decoder_head_mask,
cross_attn_head_mask=cross_attn_head_mask,
past_key_values=past_key_values,
decoder_inputs_embeds=decoder_inputs_embeds,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
lm_logits = self.proj_out(outputs[0])
loss = None
if labels is not None:
loss_fct = CrossEntropyLoss()
# move labels to correct device to enable PP
labels = labels.to(lm_logits.device)
loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.reshape(-1))
if not return_dict:
output = (lm_logits,) + outputs[1:]
return ((loss,) + output) if loss is not None else output
return Seq2SeqLMOutput(
loss=loss,
logits=lm_logits,
past_key_values=outputs.past_key_values,
decoder_hidden_states=outputs.decoder_hidden_states,
decoder_attentions=outputs.decoder_attentions,
cross_attentions=outputs.cross_attentions,
encoder_last_hidden_state=outputs.encoder_last_hidden_state,
encoder_hidden_states=outputs.encoder_hidden_states,
encoder_attentions=outputs.encoder_attentions,
)
def generate(
self,
inputs: Optional[torch.Tensor] = None,
generation_config=None,
logits_processor=None,
stopping_criteria=None,
prefix_allowed_tokens_fn=None,
synced_gpus=False,
return_timestamps=None,
task=None,
language=None,
is_multilingual=None,
**kwargs,
):
"""
Generates sequences of token ids for models with a language modeling head.
<Tip warning={true}>
Most generation-controlling parameters are set in `generation_config` which, if not passed, will be set to the
model's default generation configuration. You can override any `generation_config` by passing the corresponding
parameters to generate(), e.g. `.generate(inputs, num_beams=4, do_sample=True)`.
For an overview of generation strategies and code examples, check out the [following
guide](./generation_strategies).
</Tip>
Parameters:
inputs (`torch.Tensor` of varying shape depending on the modality, *optional*):
The sequence used as a prompt for the generation or as model inputs to the encoder. If `None` the
method initializes it with `bos_token_id` and a batch size of 1. For decoder-only models `inputs`
should of in the format of `input_ids`. For encoder-decoder models *inputs* can represent any of
`input_ids`, `input_values`, `input_features`, or `pixel_values`.
generation_config (`~generation.GenerationConfig`, *optional*):
The generation configuration to be used as base parametrization for the generation call. `**kwargs`
passed to generate matching the attributes of `generation_config` will override them. If
`generation_config` is not provided, the default will be used, which had the following loading
priority: 1) from the `generation_config.json` model file, if it exists; 2) from the model
configuration. Please note that unspecified parameters will inherit [`~generation.GenerationConfig`]'s
default values, whose documentation should be checked to parameterize generation.
logits_processor (`LogitsProcessorList`, *optional*):
Custom logits processors that complement the default logits processors built from arguments and
generation config. If a logit processor is passed that is already created with the arguments or a
generation config an error is thrown. This feature is intended for advanced users.
stopping_criteria (`StoppingCriteriaList`, *optional*):
Custom stopping criteria that complement the default stopping criteria built from arguments and a
generation config. If a stopping criteria is passed that is already created with the arguments or a
generation config an error is thrown. This feature is intended for advanced users.
prefix_allowed_tokens_fn (`Callable[[int, torch.Tensor], List[int]]`, *optional*):
If provided, this function constraints the beam search to allowed tokens only at each step. If not
provided no constraint is applied. This function takes 2 arguments: the batch ID `batch_id` and
`input_ids`. It has to return a list with the allowed tokens for the next generation step conditioned
on the batch ID `batch_id` and the previously generated tokens `inputs_ids`. This argument is useful
for constrained generation conditioned on the prefix, as described in [Autoregressive Entity
Retrieval](https://arxiv.org/abs/2010.00904).
synced_gpus (`bool`, *optional*, defaults to `False`):
Whether to continue running the while loop until max_length (needed for ZeRO stage 3)
return_timestamps (`bool`, *optional*):
Whether to return the timestamps with the text. This enables the `WhisperTimestampsLogitsProcessor`.
task (`bool`, *optional*):
Task to use for generation, either "translate" or "transcribe". The `model.config.forced_decoder_ids`
will be updated accordingly.
language (`bool`, *optional*):
Language token to use for generation, can be either in the form of `<|en|>`, `en` or `english`. You can
find all the possible language tokens in the `model.generation_config.lang_to_id` dictionary.
is_multilingual (`bool`, *optional*):
Whether or not the model is multilingual.
kwargs:
Ad hoc parametrization of `generate_config` and/or additional model-specific kwargs that will be
forwarded to the `forward` function of the model. If the model is an encoder-decoder model, encoder
specific kwargs should not be prefixed and decoder specific kwargs should be prefixed with *decoder_*.
Return:
[`~utils.ModelOutput`] or `torch.LongTensor`: A [`~utils.ModelOutput`] (if `return_dict_in_generate=True`
or when `config.return_dict_in_generate=True`) or a `torch.FloatTensor`.
If the model is *not* an encoder-decoder model (`model.config.is_encoder_decoder=False`), the possible
[`~utils.ModelOutput`] types are:
- [`~generation.GreedySearchDecoderOnlyOutput`],
- [`~generation.SampleDecoderOnlyOutput`],
- [`~generation.BeamSearchDecoderOnlyOutput`],
- [`~generation.BeamSampleDecoderOnlyOutput`]
If the model is an encoder-decoder model (`model.config.is_encoder_decoder=True`), the possible
[`~utils.ModelOutput`] types are:
- [`~generation.GreedySearchEncoderDecoderOutput`],
- [`~generation.SampleEncoderDecoderOutput`],
- [`~generation.BeamSearchEncoderDecoderOutput`],
- [`~generation.BeamSampleEncoderDecoderOutput`]
"""
if generation_config is None:
generation_config = self.generation_config
if return_timestamps is not None:
if not hasattr(generation_config, "no_timestamps_token_id"):
raise ValueError(
"You are trying to return timestamps, but the generation config is not properly set."
"Make sure to initialize the generation config with the correct attributes that are needed such as `no_timestamps_token_id`."
"For more details on how to generate the approtiate config, refer to https://github.com/huggingface/transformers/issues/21878#issuecomment-1451902363"
)
generation_config.return_timestamps = return_timestamps
else:
generation_config.return_timestamps = False
if language is not None:
language = language.lower()
generation_config.language = language
if task is not None:
generation_config.task = task
forced_decoder_ids = []
if task is not None or language is not None:
if hasattr(generation_config, "language"):
if generation_config.language in generation_config.lang_to_id.keys():
language_token = generation_config.language
elif generation_config.language in TO_LANGUAGE_CODE.keys():
language_token = f"<|{TO_LANGUAGE_CODE[generation_config.language]}|>"
elif generation_config.language in TO_LANGUAGE_CODE.values():
language_token = f"<|{generation_config.language}|>"
else:
is_language_code = len(generation_config.language) == 2
raise ValueError(
f"Unsupported language: {generation_config.language}. Language should be one of:"
f" {list(TO_LANGUAGE_CODE.values()) if is_language_code else list(TO_LANGUAGE_CODE.keys())}."
)
forced_decoder_ids.append((1, generation_config.lang_to_id[language_token]))
else:
forced_decoder_ids.append((1, None)) # automatically detect the language
if hasattr(generation_config, "task"):
if generation_config.task in TASK_IDS:
forced_decoder_ids.append((2, generation_config.task_to_id[generation_config.task]))
else:
raise ValueError(
f"The `{generation_config.task}`task is not supported. The task should be one of `{TASK_IDS}`"
)
else:
forced_decoder_ids.append((2, generation_config.task_to_id["transcribe"])) # defaults to transcribe
if hasattr(generation_config, "no_timestamps_token_id") and not generation_config.return_timestamps:
idx = forced_decoder_ids[-1][0] + 1 if forced_decoder_ids else 1
forced_decoder_ids.append((idx, generation_config.no_timestamps_token_id))
# Legacy code for backward compatibility
elif hasattr(self.config, "forced_decoder_ids") and self.config.forced_decoder_ids is not None:
forced_decoder_ids = self.config.forced_decoder_ids
elif (
hasattr(self.generation_config, "forced_decoder_ids")
and self.generation_config.forced_decoder_ids is not None
):
forced_decoder_ids = self.generation_config.forced_decoder_ids
if generation_config.return_timestamps:
logits_processor = [WhisperTimeStampLogitsProcessor(generation_config)]
if len(forced_decoder_ids) > 0:
generation_config.forced_decoder_ids = forced_decoder_ids
return super().generate(
inputs,
generation_config,
logits_processor,
stopping_criteria,
prefix_allowed_tokens_fn,
synced_gpus,
**kwargs,
)
def prepare_inputs_for_generation(
self,
decoder_input_ids,
past_key_values=None,
use_cache=None,
encoder_outputs=None,
attention_mask=None,
**kwargs,
):
# cut decoder_input_ids if past is used
if past_key_values is not None:
decoder_input_ids = decoder_input_ids[:, -1:]
return {
"encoder_outputs": encoder_outputs,
"past_key_values": past_key_values,
"decoder_input_ids": decoder_input_ids,
"use_cache": use_cache,
"decoder_attention_mask": None,
}
#
@staticmethod
def _reorder_cache(past_key_values, beam_idx):
reordered_past = ()
for layer_past in past_key_values:
reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
return reordered_past
@add_start_docstrings(
"""
Whisper Encoder Model with a sequence classification head on top (a linear layer over the pooled output) for tasks
like SUPERB Keyword Spotting.
""",
WHISPER_ENCODER_INPUTS_DOCSTRING,
)
class WhisperForAudioClassification(WhisperPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.encoder = WhisperEncoder(config)
num_layers = config.num_hidden_layers + 1 # transformer layers + input embeddings
if config.use_weighted_layer_sum:
self.layer_weights = nn.Parameter(torch.ones(num_layers) / num_layers)
self.projector = nn.Linear(config.hidden_size, config.classifier_proj_size)
self.classifier = nn.Linear(config.classifier_proj_size, config.num_labels)
# Initialize weights and apply final processing
self.post_init()
def freeze_encoder(self):
"""
Calling this function will disable the gradient computation for the Whisper encoder so that its parameters will
not be updated during training. Only the projection layers and classification head will be updated.
"""
self.encoder._freeze_parameters()
def get_input_embeddings(self) -> nn.Module:
return self.encoder.get_input_embeddings()
def set_input_embeddings(self, value: nn.Module):
self.encoder.set_input_embeddings(value)
@add_start_docstrings_to_model_forward(WHISPER_ENCODER_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=SequenceClassifierOutput, config_class=_CONFIG_FOR_DOC)
def forward(
self,
input_features: Optional[torch.LongTensor] = None,
head_mask: Optional[torch.Tensor] = None,
encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
labels: Optional[torch.LongTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
r"""
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
Returns:
Example:
```python
>>> import torch
>>> from transformers import AutoFeatureExtractor, WhisperForAudioClassification
>>> from datasets import load_dataset
>>> feature_extractor = AutoFeatureExtractor.from_pretrained("sanchit-gandhi/whisper-medium-fleurs-lang-id")
>>> model = WhisperForAudioClassification.from_pretrained("sanchit-gandhi/whisper-medium-fleurs-lang-id")
>>> ds = load_dataset("google/fleurs", "all", split="validation", streaming=True)
>>> sample = next(iter(ds))
>>> inputs = feature_extractor(
... sample["audio"]["array"], sampling_rate=sample["audio"]["sampling_rate"], return_tensors="pt"
... )
>>> input_features = inputs.input_features
>>> with torch.no_grad():
... logits = model(input_features).logits
>>> predicted_class_ids = torch.argmax(logits).item()
>>> predicted_label = model.config.id2label[predicted_class_ids]
>>> predicted_label
'af_za'
```"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
if encoder_outputs is None:
encoder_outputs = self.encoder(
input_features,
head_mask=head_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
if self.config.use_weighted_layer_sum:
hidden_states = torch.stack(encoder_outputs, dim=1)
norm_weights = nn.functional.softmax(self.layer_weights, dim=-1)
hidden_states = (hidden_states * norm_weights.view(-1, 1, 1)).sum(dim=1)
else:
hidden_states = encoder_outputs[0]
hidden_states = self.projector(hidden_states)
pooled_output = hidden_states.mean(dim=1)
logits = self.classifier(pooled_output)
loss = None
if labels is not None:
loss_fct = CrossEntropyLoss()
# move labels to correct device to enable PP
labels = labels.to(logits.device)
loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1))
if not return_dict:
output = (logits,) + encoder_outputs[1:]
return ((loss,) + output) if loss is not None else output
return SequenceClassifierOutput(
loss=loss,
logits=logits,
hidden_states=encoder_outputs.hidden_states,
attentions=encoder_outputs.attentions,
)