Source code for transformers.models.speech_to_text.modeling_speech_to_text

# coding=utf-8
# Copyright 2021 The Fairseq Authors and The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" PyTorch Speech2Text model. """


import math
import random
from typing import Optional, Tuple

import torch
from torch import nn
from torch.nn import CrossEntropyLoss

from ...activations import ACT2FN
from ...file_utils import (
    add_code_sample_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    replace_return_docstrings,
)
from ...modeling_outputs import (
    BaseModelOutput,
    BaseModelOutputWithPastAndCrossAttentions,
    Seq2SeqLMOutput,
    Seq2SeqModelOutput,
)
from ...modeling_utils import PreTrainedModel
from ...utils import logging
from .configuration_speech_to_text import Speech2TextConfig


logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "Speech2TextConfig"
_TOKENIZER_FOR_DOC = "Speech2TextTokenizer"
_CHECKPOINT_FOR_DOC = "facebook/s2t-small-librispeech-asr"


SPEECH_TO_TEXT_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "facebook/s2t-small-librispeech-asr",
    # See all Speech2Text models at https://huggingface.co/models?filter=speech_to_text
]


# 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

    assert pad_token_id is not None, "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, 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), float("-inf"))
    mask_cond = torch.arange(mask.size(-1))
    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), 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.bool(), torch.finfo(dtype).min)


class Conv1dSubsampler(nn.Module):
    """
    Convolutional subsampler: a stack of 1D convolution (along temporal dimension) followed by non-linear activation
    via gated linear units (https://arxiv.org/abs/1911.08460)
    """

    def __init__(self, config):
        super(Conv1dSubsampler, self).__init__()
        self.config = config
        self.num_layers = config.num_conv_layers
        self.in_channels = config.input_feat_per_channel * config.input_channels
        self.mid_channels = config.conv_channels
        self.out_channels = config.d_model
        self.kernel_sizes = config.conv_kernel_sizes

        self.conv_layers = nn.ModuleList(
            nn.Conv1d(
                self.in_channels if i == 0 else self.mid_channels // 2,
                self.mid_channels if i < self.num_layers - 1 else self.out_channels * 2,
                kernel_size=k,
                stride=2,
                padding=k // 2,
            )
            for i, k in enumerate(self.kernel_sizes)
        )

    def forward(self, input_features):
        hidden_states = input_features.transpose(1, 2).contiguous()  # -> B x (C x D) x T
        for conv in self.conv_layers:
            hidden_states = conv(hidden_states)
            hidden_states = nn.functional.glu(hidden_states, dim=1)
        hidden_states = hidden_states.transpose(1, 2).contiguous()  # -> T x B x (C x D)
        return hidden_states


class Speech2TextSinusoidalPositionalEmbedding(nn.Module):
    """This module produces sinusoidal positional embeddings of any length."""

    def __init__(self, num_positions: int, embedding_dim: int, padding_idx: Optional[int] = None):
        super().__init__()
        self.offset = 2
        self.embedding_dim = embedding_dim
        self.padding_idx = padding_idx
        self.make_weights(num_positions + self.offset, embedding_dim, padding_idx)

    def make_weights(self, num_embeddings: int, embedding_dim: int, padding_idx: Optional[int] = None):
        emb_weights = self.get_embedding(num_embeddings, embedding_dim, padding_idx)
        if hasattr(self, "weights"):
            # in forward put the weights on the correct dtype and device of the param
            emb_weights = emb_weights.to(dtype=self.weights.dtype, device=self.weights.device)

        self.weights = nn.Parameter(emb_weights)
        self.weights.requires_grad = False
        self.weights.detach_()

    @staticmethod
    def get_embedding(num_embeddings: int, embedding_dim: int, padding_idx: Optional[int] = None):
        """
        Build sinusoidal embeddings. This matches the implementation in tensor2tensor, but differs slightly from the
        description in Section 3.5 of "Attention Is All You Need".
        """
        half_dim = embedding_dim // 2
        emb = math.log(10000) / (half_dim - 1)
        emb = torch.exp(torch.arange(half_dim, dtype=torch.float) * -emb)
        emb = torch.arange(num_embeddings, dtype=torch.float).unsqueeze(1) * emb.unsqueeze(0)
        emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1).view(num_embeddings, -1)
        if embedding_dim % 2 == 1:
            # zero pad
            emb = torch.cat([emb, torch.zeros(num_embeddings, 1)], dim=1)
        if padding_idx is not None:
            emb[padding_idx, :] = 0
        return emb

    @torch.no_grad()
    def forward(self, input_ids: torch.Tensor, past_key_values_length: int = 0):
        bsz, seq_len = input_ids.size()
        # Create the position ids from the input token ids. Any padded tokens remain padded.
        position_ids = self.create_position_ids_from_input_ids(input_ids, self.padding_idx, past_key_values_length).to(
            input_ids.device
        )

        # expand embeddings if needed
        max_pos = self.padding_idx + 1 + seq_len
        if max_pos > self.weights.size(0):
            self.make_weights(max_pos + self.offset, self.embedding_dim, self.padding_idx)

        return self.weights.index_select(0, position_ids.view(-1)).view(bsz, seq_len, -1).detach()

    def create_position_ids_from_input_ids(
        self, input_ids: torch.Tensor, padding_idx: int, past_key_values_length: Optional[int] = 0
    ):
        """
        Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding
        symbols are ignored. This is modified from fairseq's `utils.make_positions`.

        Args:
            x: torch.Tensor x:
        Returns: torch.Tensor
        """
        # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
        mask = input_ids.ne(padding_idx).int()
        incremental_indices = (torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length) * mask
        return incremental_indices.long() + padding_idx


# Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->Speech2Text
class Speech2TextAttention(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
        assert (
            self.head_dim * num_heads == self.embed_dim
        ), f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} 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=bias)
        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)

    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()

    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, embed_dim = hidden_states.size()

        # get query proj
        query_states = self.q_proj(hidden_states) * self.scaling
        # get key, value proj
        if is_cross_attention and past_key_value is not None:
            # 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.view(*proj_shape)
        value_states = value_states.view(*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 {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 {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 {attn_output.size()}"
            )

        attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
        attn_output = attn_output.transpose(1, 2)
        attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)

        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights_reshaped, past_key_value


class Speech2TextEncoderLayer(nn.Module):
    def __init__(self, config: Speech2TextConfig):
        super().__init__()
        self.embed_dim = config.d_model
        self.self_attn = Speech2TextAttention(
            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,
    ):
        """
        Args:
            hidden_states (:obj:`torch.FloatTensor`): input to the layer of shape :obj:`(seq_len, batch, embed_dim)`
            attention_mask (:obj:`torch.FloatTensor`): attention mask of size
                :obj:`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            layer_head_mask (:obj:`torch.FloatTensor`): mask for attention heads in a given layer of size
                :obj:`(config.encoder_attention_heads,)`.
            output_attentions (:obj:`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


class Speech2TextDecoderLayer(nn.Module):
    def __init__(self, config: Speech2TextConfig):
        super().__init__()
        self.embed_dim = config.d_model

        self.self_attn = Speech2TextAttention(
            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 = Speech2TextAttention(
            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,
    ):
        """
        Args:
            hidden_states (:obj:`torch.FloatTensor`): input to the layer of shape :obj:`(seq_len, batch, embed_dim)`
            attention_mask (:obj:`torch.FloatTensor`): attention mask of size
                :obj:`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            encoder_hidden_states (:obj:`torch.FloatTensor`): cross attention input to the layer of shape :obj:`(seq_len, batch, embed_dim)`
            encoder_attention_mask (:obj:`torch.FloatTensor`): encoder attention mask of size
                :obj:`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            layer_head_mask (:obj:`torch.FloatTensor`): mask for attention heads in a given layer of size
                :obj:`(encoder_attention_heads,)`.
            cross_attn_layer_head_mask (:obj:`torch.FloatTensor`): mask for cross-attention heads in a given layer of
                size `(decoder_attention_heads,)`.
            past_key_value (:obj:`Tuple(torch.FloatTensor)`): cached past key and value projection states
            output_attentions (:obj:`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 Speech2TextPreTrainedModel(PreTrainedModel):
    config_class = Speech2TextConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True

    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, (Speech2TextDecoder, Speech2TextEncoder)):
            module.gradient_checkpointing = value

    def _get_feat_extract_output_lengths(self, input_lengths: torch.LongTensor):
        """
        Computes the output length of the convolutional layers
        """
        for i in range(self.config.num_conv_layers):
            input_lengths = (input_lengths - 1) // 2 + 1

        return input_lengths

    def _get_feature_vector_attention_mask(self, feature_vector_length, attention_mask):
        # generate creates 3D attention mask, because of the shape of input_features
        # convert it to 2D if thats the case
        if len(attention_mask.shape) > 2:
            attention_mask = attention_mask[:, :, -1]

        subsampled_lengths = self._get_feat_extract_output_lengths(attention_mask.sum(-1))
        bsz = attention_mask.size()[0]
        attention_mask = torch.zeros(
            (bsz, feature_vector_length), dtype=attention_mask.dtype, device=attention_mask.device
        )

        # these two operations makes sure that all values
        # before the output lengths indices are attended to
        attention_mask[(torch.arange(bsz, device=attention_mask.device), subsampled_lengths - 1)] = 1
        attention_mask = attention_mask.flip([-1]).cumsum(-1).flip([-1]).long()
        return attention_mask


SPEECH_TO_TEXT_START_DOCSTRING = r"""
    This model inherits from :class:`~transformers.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 (:class:`~transformers.Speech2TextConfig`):
            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
            :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
"""

SPEECH_TO_TEXT_INPUTS_DOCSTRING = r"""
    Args:
        input_features (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length, feature_size)`):
            Float values of fbank 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 :obj:`List[float]` or a
            :obj:`numpy.ndarray`, *e.g.* via the soundfile library (``pip install soundfile``). To prepare the array
            into :obj:`input_features`, the :class:`~transformers.Speech2TextTokenizer` should be used for extracting
            the fbank features, padding and conversion into a tensor of type :obj:`torch.FloatTensor`. See
            :meth:`~transformers.Speech2TextTokenizer.__call__`
        attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
            Mask to avoid performing convolution and 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.html#attention-mask>`__
        decoder_input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
            Indices of decoder input sequence tokens in the vocabulary.

            Indices can be obtained using :class:`~transformers.SpeechToTextTokenizer`. See
            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
            details.

            `What are decoder input IDs? <../glossary.html#decoder-input-ids>`__

            SpeechToText uses the :obj:`eos_token_id` as the starting token for :obj:`decoder_input_ids` generation. If
            :obj:`past_key_values` is used, optionally only the last :obj:`decoder_input_ids` have to be input (see
            :obj:`past_key_values`).
        decoder_attention_mask (:obj:`torch.LongTensor` of shape :obj:`(batch_size, target_sequence_length)`, `optional`):
            Default behavior: generate a tensor that ignores pad tokens in :obj:`decoder_input_ids`. Causal mask will
            also be used by default. <<<<<<< HEAD

            If you want to change padding behavior, you should read
            :func:`modeling_speech_to_text._prepare_decoder_inputs` and modify to your needs. See diagram 1 in `the
            paper <https://arxiv.org/abs/1910.13461>`__ for more information on the default strategy.
        head_mask (:obj:`torch.Tensor` of shape :obj:`(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 (:obj:`torch.Tensor` of shape :obj:`(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 (:obj:`torch.Tensor` of shape :obj:`(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 (:obj:`tuple(tuple(torch.FloatTensor)`, `optional`):
            Tuple consists of (:obj:`last_hidden_state`, `optional`: :obj:`hidden_states`, `optional`:
            :obj:`attentions`) :obj:`last_hidden_state` of shape :obj:`(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 (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
            shape :obj:`(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 :obj:`past_key_values` input) to speed up sequential decoding.

            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
            instead of all :obj:`decoder_input_ids`` of shape :obj:`(batch_size, sequence_length)`.
        decoder_inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, target_sequence_length, hidden_size)`, `optional`):
            Optionally, instead of passing :obj:`decoder_input_ids` you can choose to directly pass an embedded
            representation. If :obj:`past_key_values` is used, optionally only the last :obj:`decoder_inputs_embeds`
            have to be input (see :obj:`past_key_values`). This is useful if you want more control over how to convert
            :obj:`decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix.

            If :obj:`decoder_input_ids` and :obj:`decoder_inputs_embeds` are both unset, :obj:`decoder_inputs_embeds`
            takes the value of :obj:`inputs_embeds`.
        use_cache (:obj:`bool`, `optional`):
            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
            decoding (see :obj:`past_key_values`).
        output_attentions (:obj:`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 (:obj:`bool`, `optional`):
            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
            more detail.
        return_dict (:obj:`bool`, `optional`):
            Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
"""


class Speech2TextEncoder(Speech2TextPreTrainedModel):
    """
    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
    :class:`Speech2TextEncoderLayer`.

    Args:
        config: Speech2TextConfig
        embed_tokens (nn.Embedding): output embedding
    """

    def __init__(self, config: Speech2TextConfig):
        super().__init__(config)

        self.dropout = config.dropout
        self.layerdrop = config.encoder_layerdrop

        embed_dim = config.d_model
        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.conv = Conv1dSubsampler(config)

        self.embed_positions = Speech2TextSinusoidalPositionalEmbedding(
            self.max_source_positions,
            embed_dim,
            self.padding_idx,
        )
        self.layers = nn.ModuleList([Speech2TextEncoderLayer(config) for _ in range(config.encoder_layers)])
        self.layer_norm = nn.LayerNorm(config.d_model)

        self.init_weights()
        self.gradient_checkpointing = False

    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 (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length, feature_size)`):
                Float values of fbank 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 :obj:`List[float]` or a
                :obj:`numpy.ndarray`, *e.g.* via the soundfile library (``pip install soundfile``). To prepare the
                array into :obj:`input_features`, the :class:`~transformers.Speech2TextTokenizer` should be used for
                extracting the fbank features, padding and conversion into a tensor of type :obj:`torch.FloatTensor`.
                See :meth:`~transformers.Speech2TextTokenizer.__call__`
            attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
                Mask to avoid performing convolution and 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.html#attention-mask>`__
            head_mask (:obj:`torch.Tensor` of shape :obj:`(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 (:obj:`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 (:obj:`bool`, `optional`):
                Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors
                for more detail.
            return_dict (:obj:`bool`, `optional`):
                Whether or not to return a :class:`~transformers.file_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 = self.conv(input_features)
        inputs_embeds = self.embed_scale * inputs_embeds

        # subsample attention mask if necessary
        if attention_mask is not None:
            attention_mask = self._get_feature_vector_attention_mask(inputs_embeds.shape[1], attention_mask)
            padding_mask = attention_mask.ne(1).long()
        else:
            padding_mask = torch.zeros_like(inputs_embeds, dtype=torch.long)

        embed_pos = self.embed_positions(padding_mask)

        hidden_states = inputs_embeds + embed_pos
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)

        # expand attention_mask
        if attention_mask is not None:
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
            attention_mask = _expand_mask(attention_mask, inputs_embeds.dtype)

        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,
                        attention_mask,
                        (head_mask[idx] if head_mask is not None else None),
                    )
                else:
                    layer_outputs = encoder_layer(
                        hidden_states,
                        attention_mask,
                        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 Speech2TextDecoder(Speech2TextPreTrainedModel):
    """
    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a :class:`Speech2TextDecoderLayer`

    Args:
        config: Speech2TextConfig
        embed_tokens (nn.Embedding): output embedding
    """

    def __init__(self, config: Speech2TextConfig):
        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.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 = Speech2TextSinusoidalPositionalEmbedding(
            self.max_target_positions,
            config.d_model,
            self.padding_idx,
        )

        self.layers = nn.ModuleList([Speech2TextDecoderLayer(config) for _ in range(config.decoder_layers)])

        self.layer_norm = nn.LayerNorm(config.d_model)

        self.init_weights()
        self.gradient_checkpointing = False

    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, past_key_values_length=past_key_values_length
            ).to(self.device)

        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,
        encoder_attention_mask=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 (:obj:`torch.LongTensor` of shape :obj:`(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 :class:`~transformers.Speech2TextTokenizer`. See
                :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__`
                for details.

                `What are input IDs? <../glossary.html#input-ids>`__
            attention_mask (:obj:`torch.Tensor` of shape :obj:`(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.html#attention-mask>`__
            encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(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.
            encoder_attention_mask (:obj:`torch.LongTensor` of shape :obj:`(batch_size, encoder_sequence_length)`, `optional`):
                Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values
                selected in ``[0, 1]``:

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

                `What are attention masks? <../glossary.html#attention-mask>`__
            head_mask (:obj:`torch.Tensor` of shape :obj:`(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 (:obj:`torch.Tensor` of shape :obj:`(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 (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
                Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2
                tensors of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional
                tensors of shape :obj:`(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 :obj:`past_key_values` input) to speed up sequential
                decoding.

                If :obj:`past_key_values` are used, the user can optionally input only the last
                :obj:`decoder_input_ids` (those that don't have their past key value states given to this model) of
                shape :obj:`(batch_size, 1)` instead of all :obj:`decoder_input_ids`` of shape :obj:`(batch_size,
                sequence_length)`.
            inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
                Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded
                representation. This is useful if you want more control over how to convert :obj:`input_ids` indices
                into associated vectors than the model's internal embedding lookup matrix.
            output_attentions (:obj:`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 (:obj:`bool`, `optional`):
                Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors
                for more detail.
            return_dict (:obj:`bool`, `optional`):
                Whether or not to return a :class:`~transformers.file_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) * self.embed_scale

        attention_mask = self._prepare_decoder_attention_mask(
            attention_mask, input_shape, inputs_embeds, past_key_values_length
        )

        # expand encoder attention mask
        if encoder_hidden_states is not None and encoder_attention_mask is not None:
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
            encoder_attention_mask = _expand_mask(encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])

        # embed positions
        positions = self.embed_positions(input_ids, 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)

        # 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 {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:

                if use_cache:
                    logger.warning(
                        "`use_cache = True` is incompatible with gradient checkpointing. Setting `use_cache = False`..."
                    )
                    use_cache = False

                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,
                    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,
                )
            else:

                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=attention_mask,
                    encoder_hidden_states=encoder_hidden_states,
                    encoder_attention_mask=encoder_attention_mask,
                    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,
        )


[docs]@add_start_docstrings( "The bare Speech2Text Model outputting raw hidden-states without any specific head on top.", SPEECH_TO_TEXT_START_DOCSTRING, ) class Speech2TextModel(Speech2TextPreTrainedModel): def __init__(self, config: Speech2TextConfig): super().__init__(config) self.encoder = Speech2TextEncoder(config) self.decoder = Speech2TextDecoder(config) self.init_weights() 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
[docs] @add_start_docstrings_to_model_forward(SPEECH_TO_TEXT_INPUTS_DOCSTRING) @add_code_sample_docstrings( tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint=_CHECKPOINT_FOR_DOC, output_type=Seq2SeqModelOutput, config_class=_CONFIG_FOR_DOC, ) def forward( self, input_features=None, attention_mask=None, decoder_input_ids=None, decoder_attention_mask=None, head_mask=None, decoder_head_mask=None, cross_attn_head_mask=None, encoder_outputs=None, past_key_values=None, decoder_inputs_embeds=None, use_cache=None, output_attentions=None, output_hidden_states=None, return_dict=None, ): 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: encoder_outputs = self.encoder( input_features, attention_mask=attention_mask, 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, ) # downsample encoder attention mask if attention_mask is not None: encoder_attention_mask = self._get_feature_vector_attention_mask( encoder_outputs[0].shape[1], attention_mask ) else: encoder_attention_mask = 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], encoder_attention_mask=encoder_attention_mask, 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, )
[docs]@add_start_docstrings( "The Speech2Text Model with a language modeling head. Can be used for summarization.", SPEECH_TO_TEXT_START_DOCSTRING, ) class Speech2TextForConditionalGeneration(Speech2TextPreTrainedModel): base_model_prefix = "model" _keys_to_ignore_on_load_missing = [ r"encoder\.version", r"decoder\.version", r"model.encoder.embed_positions.weights", r"model.decoder.embed_positions.weights", ] _keys_to_ignore_on_save = [ r"model.encoder.embed_positions.weights", r"model.decoder.embed_positions.weights", ] def __init__(self, config: Speech2TextConfig): super().__init__(config) self.model = Speech2TextModel(config) self.lm_head = nn.Linear(config.d_model, self.config.vocab_size, bias=False) self.init_weights() 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.lm_head def set_output_embeddings(self, new_embeddings): self.lm_head = new_embeddings
[docs] @add_start_docstrings_to_model_forward(SPEECH_TO_TEXT_INPUTS_DOCSTRING) @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC) def forward( self, input_features=None, attention_mask=None, decoder_input_ids=None, decoder_attention_mask=None, head_mask=None, decoder_head_mask=None, cross_attn_head_mask=None, encoder_outputs=None, past_key_values=None, decoder_inputs_embeds=None, labels=None, use_cache=None, output_attentions=None, output_hidden_states=None, return_dict=None, ): r""" labels (:obj:`torch.LongTensor` of shape :obj:`(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:: >>> import torch >>> from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration >>> from datasets import load_dataset >>> import soundfile as sf >>> model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-librispeech-asr") >>> processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-librispeech-asr") >>> def map_to_array(batch): >>> speech, _ = sf.read(batch["file"]) >>> batch["speech"] = speech >>> return batch >>> ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation") >>> ds = ds.map(map_to_array) >>> input_features = processor(ds["speech"][0], sampling_rate=16000, return_tensors="pt").input_features # Batch size 1 >>> generated_ids = model.generate(input_ids=input_features) >>> transcription = processor.batch_decode(generated_ids) """ 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: 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.lm_head(outputs[0]) loss = None if labels is not None: loss_fct = CrossEntropyLoss() loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-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 prepare_inputs_for_generation( self, decoder_input_ids, past=None, attention_mask=None, head_mask=None, decoder_head_mask=None, cross_attn_head_mask=None, use_cache=None, encoder_outputs=None, **kwargs ): # cut decoder_input_ids if past is used if past is not None: decoder_input_ids = decoder_input_ids[:, -1:] return { "encoder_outputs": encoder_outputs, "past_key_values": past, "decoder_input_ids": decoder_input_ids, "attention_mask": attention_mask, "head_mask": head_mask, "decoder_head_mask": decoder_head_mask, "cross_attn_head_mask": cross_attn_head_mask, "use_cache": use_cache, # change this to avoid caching (presumably for debugging) } @staticmethod def _reorder_cache(past, beam_idx): reordered_past = () for layer_past in past: reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),) return reordered_past