Initial commit: FunASR Speech Recognition Toolkit
Update API Documentation / build-api-docs (push) Has been cancelled
Update API Documentation / build-api-docs (push) Has been cancelled
Add complete FunASR codebase including models, runtime, and documentation.
This commit is contained in:
freedakgmail
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,524 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- encoding: utf-8 -*-
|
||||
# Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
|
||||
# MIT License (https://opensource.org/licenses/MIT)
|
||||
|
||||
from typing import List
|
||||
from typing import Tuple
|
||||
import logging
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import numpy as np
|
||||
|
||||
from funasr.models.scama import utils as myutils
|
||||
from funasr.models.transformer.decoder import BaseTransformerDecoder
|
||||
|
||||
from funasr.models.sanm.attention import (
|
||||
MultiHeadedAttentionSANMDecoder,
|
||||
MultiHeadedAttentionCrossAtt,
|
||||
)
|
||||
from funasr.models.transformer.embedding import PositionalEncoding
|
||||
from funasr.models.transformer.layer_norm import LayerNorm
|
||||
from funasr.models.sanm.positionwise_feed_forward import PositionwiseFeedForwardDecoderSANM
|
||||
from funasr.models.transformer.utils.repeat import repeat
|
||||
|
||||
from funasr.register import tables
|
||||
|
||||
|
||||
class DecoderLayerSANM(nn.Module):
|
||||
"""Single decoder layer module.
|
||||
|
||||
Args:
|
||||
size (int): Input dimension.
|
||||
self_attn (torch.nn.Module): Self-attention module instance.
|
||||
`MultiHeadedAttention` instance can be used as the argument.
|
||||
src_attn (torch.nn.Module): Self-attention module instance.
|
||||
`MultiHeadedAttention` instance can be used as the argument.
|
||||
feed_forward (torch.nn.Module): Feed-forward module instance.
|
||||
`PositionwiseFeedForward`, `MultiLayeredConv1d`, or `Conv1dLinear` instance
|
||||
can be used as the argument.
|
||||
dropout_rate (float): Dropout rate.
|
||||
normalize_before (bool): Whether to use layer_norm before the first block.
|
||||
concat_after (bool): Whether to concat attention layer's input and output.
|
||||
if True, additional linear will be applied.
|
||||
i.e. x -> x + linear(concat(x, att(x)))
|
||||
if False, no additional linear will be applied. i.e. x -> x + att(x)
|
||||
|
||||
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
size,
|
||||
self_attn,
|
||||
src_attn,
|
||||
feed_forward,
|
||||
dropout_rate,
|
||||
normalize_before=True,
|
||||
concat_after=False,
|
||||
):
|
||||
"""Construct an DecoderLayer object."""
|
||||
super(DecoderLayerSANM, self).__init__()
|
||||
self.size = size
|
||||
self.self_attn = self_attn
|
||||
self.src_attn = src_attn
|
||||
self.feed_forward = feed_forward
|
||||
self.norm1 = LayerNorm(size)
|
||||
if self_attn is not None:
|
||||
self.norm2 = LayerNorm(size)
|
||||
if src_attn is not None:
|
||||
self.norm3 = LayerNorm(size)
|
||||
self.dropout = nn.Dropout(dropout_rate)
|
||||
self.normalize_before = normalize_before
|
||||
self.concat_after = concat_after
|
||||
if self.concat_after:
|
||||
self.concat_linear1 = nn.Linear(size + size, size)
|
||||
self.concat_linear2 = nn.Linear(size + size, size)
|
||||
|
||||
def forward(self, tgt, tgt_mask, memory, memory_mask=None, cache=None):
|
||||
"""Compute decoded features.
|
||||
|
||||
Args:
|
||||
tgt (torch.Tensor): Input tensor (#batch, maxlen_out, size).
|
||||
tgt_mask (torch.Tensor): Mask for input tensor (#batch, maxlen_out).
|
||||
memory (torch.Tensor): Encoded memory, float32 (#batch, maxlen_in, size).
|
||||
memory_mask (torch.Tensor): Encoded memory mask (#batch, maxlen_in).
|
||||
cache (List[torch.Tensor]): List of cached tensors.
|
||||
Each tensor shape should be (#batch, maxlen_out - 1, size).
|
||||
|
||||
Returns:
|
||||
torch.Tensor: Output tensor(#batch, maxlen_out, size).
|
||||
torch.Tensor: Mask for output tensor (#batch, maxlen_out).
|
||||
torch.Tensor: Encoded memory (#batch, maxlen_in, size).
|
||||
torch.Tensor: Encoded memory mask (#batch, maxlen_in).
|
||||
|
||||
"""
|
||||
# tgt = self.dropout(tgt)
|
||||
residual = tgt
|
||||
if self.normalize_before:
|
||||
tgt = self.norm1(tgt)
|
||||
tgt = self.feed_forward(tgt)
|
||||
|
||||
x = tgt
|
||||
if self.self_attn:
|
||||
if self.normalize_before:
|
||||
tgt = self.norm2(tgt)
|
||||
x, _ = self.self_attn(tgt, tgt_mask)
|
||||
x = residual + self.dropout(x)
|
||||
|
||||
if self.src_attn is not None:
|
||||
residual = x
|
||||
if self.normalize_before:
|
||||
x = self.norm3(x)
|
||||
|
||||
x = residual + self.dropout(self.src_attn(x, memory, memory_mask))
|
||||
|
||||
return x, tgt_mask, memory, memory_mask, cache
|
||||
|
||||
def forward_one_step(self, tgt, tgt_mask, memory, memory_mask=None, cache=None):
|
||||
"""Compute decoded features.
|
||||
|
||||
Args:
|
||||
tgt (torch.Tensor): Input tensor (#batch, maxlen_out, size).
|
||||
tgt_mask (torch.Tensor): Mask for input tensor (#batch, maxlen_out).
|
||||
memory (torch.Tensor): Encoded memory, float32 (#batch, maxlen_in, size).
|
||||
memory_mask (torch.Tensor): Encoded memory mask (#batch, maxlen_in).
|
||||
cache (List[torch.Tensor]): List of cached tensors.
|
||||
Each tensor shape should be (#batch, maxlen_out - 1, size).
|
||||
|
||||
Returns:
|
||||
torch.Tensor: Output tensor(#batch, maxlen_out, size).
|
||||
torch.Tensor: Mask for output tensor (#batch, maxlen_out).
|
||||
torch.Tensor: Encoded memory (#batch, maxlen_in, size).
|
||||
torch.Tensor: Encoded memory mask (#batch, maxlen_in).
|
||||
|
||||
"""
|
||||
# tgt = self.dropout(tgt)
|
||||
residual = tgt
|
||||
if self.normalize_before:
|
||||
tgt = self.norm1(tgt)
|
||||
tgt = self.feed_forward(tgt)
|
||||
|
||||
x = tgt
|
||||
if self.self_attn:
|
||||
if self.normalize_before:
|
||||
tgt = self.norm2(tgt)
|
||||
if self.training:
|
||||
cache = None
|
||||
x, cache = self.self_attn(tgt, tgt_mask, cache=cache)
|
||||
x = residual + self.dropout(x)
|
||||
|
||||
if self.src_attn is not None:
|
||||
residual = x
|
||||
if self.normalize_before:
|
||||
x = self.norm3(x)
|
||||
|
||||
x = residual + self.dropout(self.src_attn(x, memory, memory_mask))
|
||||
|
||||
return x, tgt_mask, memory, memory_mask, cache
|
||||
|
||||
def forward_chunk(
|
||||
self, tgt, memory, fsmn_cache=None, opt_cache=None, chunk_size=None, look_back=0
|
||||
):
|
||||
"""Compute decoded features.
|
||||
|
||||
Args:
|
||||
tgt (torch.Tensor): Input tensor (#batch, maxlen_out, size).
|
||||
tgt_mask (torch.Tensor): Mask for input tensor (#batch, maxlen_out).
|
||||
memory (torch.Tensor): Encoded memory, float32 (#batch, maxlen_in, size).
|
||||
memory_mask (torch.Tensor): Encoded memory mask (#batch, maxlen_in).
|
||||
cache (List[torch.Tensor]): List of cached tensors.
|
||||
Each tensor shape should be (#batch, maxlen_out - 1, size).
|
||||
|
||||
Returns:
|
||||
torch.Tensor: Output tensor(#batch, maxlen_out, size).
|
||||
torch.Tensor: Mask for output tensor (#batch, maxlen_out).
|
||||
torch.Tensor: Encoded memory (#batch, maxlen_in, size).
|
||||
torch.Tensor: Encoded memory mask (#batch, maxlen_in).
|
||||
|
||||
"""
|
||||
residual = tgt
|
||||
if self.normalize_before:
|
||||
tgt = self.norm1(tgt)
|
||||
tgt = self.feed_forward(tgt)
|
||||
|
||||
x = tgt
|
||||
if self.self_attn:
|
||||
if self.normalize_before:
|
||||
tgt = self.norm2(tgt)
|
||||
x, fsmn_cache = self.self_attn(tgt, None, fsmn_cache)
|
||||
x = residual + self.dropout(x)
|
||||
|
||||
if self.src_attn is not None:
|
||||
residual = x
|
||||
if self.normalize_before:
|
||||
x = self.norm3(x)
|
||||
|
||||
x, opt_cache = self.src_attn.forward_chunk(x, memory, opt_cache, chunk_size, look_back)
|
||||
x = residual + x
|
||||
|
||||
return x, memory, fsmn_cache, opt_cache
|
||||
|
||||
|
||||
@tables.register("decoder_classes", "FsmnDecoder")
|
||||
class FsmnDecoder(BaseTransformerDecoder):
|
||||
"""
|
||||
Author: Zhifu Gao, Shiliang Zhang, Ming Lei, Ian McLoughlin
|
||||
San-m: Memory equipped self-attention for end-to-end speech recognition
|
||||
https://arxiv.org/abs/2006.01713
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vocab_size: int,
|
||||
encoder_output_size: int,
|
||||
attention_heads: int = 4,
|
||||
linear_units: int = 2048,
|
||||
num_blocks: int = 6,
|
||||
dropout_rate: float = 0.1,
|
||||
positional_dropout_rate: float = 0.1,
|
||||
self_attention_dropout_rate: float = 0.0,
|
||||
src_attention_dropout_rate: float = 0.0,
|
||||
input_layer: str = "embed",
|
||||
use_output_layer: bool = True,
|
||||
pos_enc_class=PositionalEncoding,
|
||||
normalize_before: bool = True,
|
||||
concat_after: bool = False,
|
||||
att_layer_num: int = 6,
|
||||
kernel_size: int = 21,
|
||||
sanm_shfit: int = None,
|
||||
concat_embeds: bool = False,
|
||||
attention_dim: int = None,
|
||||
tf2torch_tensor_name_prefix_torch: str = "decoder",
|
||||
tf2torch_tensor_name_prefix_tf: str = "seq2seq/decoder",
|
||||
embed_tensor_name_prefix_tf: str = None,
|
||||
):
|
||||
"""Initialize FsmnDecoder.
|
||||
|
||||
Args:
|
||||
vocab_size: Size/dimension parameter.
|
||||
encoder_output_size: Size/dimension parameter.
|
||||
attention_heads: TODO.
|
||||
linear_units: TODO.
|
||||
num_blocks: TODO.
|
||||
dropout_rate: TODO.
|
||||
positional_dropout_rate: TODO.
|
||||
self_attention_dropout_rate: TODO.
|
||||
src_attention_dropout_rate: TODO.
|
||||
input_layer: TODO.
|
||||
use_output_layer: TODO.
|
||||
pos_enc_class: TODO.
|
||||
normalize_before: TODO.
|
||||
concat_after: TODO.
|
||||
att_layer_num: TODO.
|
||||
kernel_size: Size/dimension parameter.
|
||||
sanm_shfit: TODO.
|
||||
concat_embeds: TODO.
|
||||
attention_dim: Size/dimension parameter.
|
||||
tf2torch_tensor_name_prefix_torch: TODO.
|
||||
tf2torch_tensor_name_prefix_tf: TODO.
|
||||
embed_tensor_name_prefix_tf: TODO.
|
||||
"""
|
||||
super().__init__(
|
||||
vocab_size=vocab_size,
|
||||
encoder_output_size=encoder_output_size,
|
||||
dropout_rate=dropout_rate,
|
||||
positional_dropout_rate=positional_dropout_rate,
|
||||
input_layer=input_layer,
|
||||
use_output_layer=use_output_layer,
|
||||
pos_enc_class=pos_enc_class,
|
||||
normalize_before=normalize_before,
|
||||
)
|
||||
if attention_dim is None:
|
||||
attention_dim = encoder_output_size
|
||||
|
||||
if input_layer == "embed":
|
||||
self.embed = torch.nn.Sequential(
|
||||
torch.nn.Embedding(vocab_size, attention_dim),
|
||||
)
|
||||
elif input_layer == "linear":
|
||||
self.embed = torch.nn.Sequential(
|
||||
torch.nn.Linear(vocab_size, attention_dim),
|
||||
torch.nn.LayerNorm(attention_dim),
|
||||
torch.nn.Dropout(dropout_rate),
|
||||
torch.nn.ReLU(),
|
||||
pos_enc_class(attention_dim, positional_dropout_rate),
|
||||
)
|
||||
else:
|
||||
raise ValueError(f"only 'embed' or 'linear' is supported: {input_layer}")
|
||||
|
||||
self.normalize_before = normalize_before
|
||||
if self.normalize_before:
|
||||
self.after_norm = LayerNorm(attention_dim)
|
||||
if use_output_layer:
|
||||
self.output_layer = torch.nn.Linear(attention_dim, vocab_size)
|
||||
else:
|
||||
self.output_layer = None
|
||||
|
||||
self.att_layer_num = att_layer_num
|
||||
self.num_blocks = num_blocks
|
||||
if sanm_shfit is None:
|
||||
sanm_shfit = (kernel_size - 1) // 2
|
||||
self.decoders = repeat(
|
||||
att_layer_num,
|
||||
lambda lnum: DecoderLayerSANM(
|
||||
attention_dim,
|
||||
MultiHeadedAttentionSANMDecoder(
|
||||
attention_dim, self_attention_dropout_rate, kernel_size, sanm_shfit=sanm_shfit
|
||||
),
|
||||
MultiHeadedAttentionCrossAtt(
|
||||
attention_heads,
|
||||
attention_dim,
|
||||
src_attention_dropout_rate,
|
||||
encoder_output_size=encoder_output_size,
|
||||
),
|
||||
PositionwiseFeedForwardDecoderSANM(attention_dim, linear_units, dropout_rate),
|
||||
dropout_rate,
|
||||
normalize_before,
|
||||
concat_after,
|
||||
),
|
||||
)
|
||||
if num_blocks - att_layer_num <= 0:
|
||||
self.decoders2 = None
|
||||
else:
|
||||
self.decoders2 = repeat(
|
||||
num_blocks - att_layer_num,
|
||||
lambda lnum: DecoderLayerSANM(
|
||||
attention_dim,
|
||||
MultiHeadedAttentionSANMDecoder(
|
||||
attention_dim,
|
||||
self_attention_dropout_rate,
|
||||
kernel_size,
|
||||
sanm_shfit=sanm_shfit,
|
||||
),
|
||||
None,
|
||||
PositionwiseFeedForwardDecoderSANM(attention_dim, linear_units, dropout_rate),
|
||||
dropout_rate,
|
||||
normalize_before,
|
||||
concat_after,
|
||||
),
|
||||
)
|
||||
|
||||
self.decoders3 = repeat(
|
||||
1,
|
||||
lambda lnum: DecoderLayerSANM(
|
||||
attention_dim,
|
||||
None,
|
||||
None,
|
||||
PositionwiseFeedForwardDecoderSANM(attention_dim, linear_units, dropout_rate),
|
||||
dropout_rate,
|
||||
normalize_before,
|
||||
concat_after,
|
||||
),
|
||||
)
|
||||
if concat_embeds:
|
||||
self.embed_concat_ffn = repeat(
|
||||
1,
|
||||
lambda lnum: DecoderLayerSANM(
|
||||
attention_dim + encoder_output_size,
|
||||
None,
|
||||
None,
|
||||
PositionwiseFeedForwardDecoderSANM(
|
||||
attention_dim + encoder_output_size,
|
||||
linear_units,
|
||||
dropout_rate,
|
||||
adim=attention_dim,
|
||||
),
|
||||
dropout_rate,
|
||||
normalize_before,
|
||||
concat_after,
|
||||
),
|
||||
)
|
||||
else:
|
||||
self.embed_concat_ffn = None
|
||||
self.concat_embeds = concat_embeds
|
||||
self.tf2torch_tensor_name_prefix_torch = tf2torch_tensor_name_prefix_torch
|
||||
self.tf2torch_tensor_name_prefix_tf = tf2torch_tensor_name_prefix_tf
|
||||
self.embed_tensor_name_prefix_tf = embed_tensor_name_prefix_tf
|
||||
|
||||
def forward(
|
||||
self,
|
||||
hs_pad: torch.Tensor,
|
||||
hlens: torch.Tensor,
|
||||
ys_in_pad: torch.Tensor,
|
||||
ys_in_lens: torch.Tensor,
|
||||
chunk_mask: torch.Tensor = None,
|
||||
pre_acoustic_embeds: torch.Tensor = None,
|
||||
) -> Tuple[torch.Tensor, torch.Tensor]:
|
||||
"""Forward decoder.
|
||||
|
||||
Args:
|
||||
hs_pad: encoded memory, float32 (batch, maxlen_in, feat)
|
||||
hlens: (batch)
|
||||
ys_in_pad:
|
||||
input token ids, int64 (batch, maxlen_out)
|
||||
if input_layer == "embed"
|
||||
input tensor (batch, maxlen_out, #mels) in the other cases
|
||||
ys_in_lens: (batch)
|
||||
Returns:
|
||||
(tuple): tuple containing:
|
||||
|
||||
x: decoded token score before softmax (batch, maxlen_out, token)
|
||||
if use_output_layer is True,
|
||||
olens: (batch, )
|
||||
"""
|
||||
tgt = ys_in_pad
|
||||
tgt_mask = myutils.sequence_mask(ys_in_lens, device=tgt.device)[:, :, None]
|
||||
|
||||
memory = hs_pad
|
||||
memory_mask = myutils.sequence_mask(hlens, device=memory.device)[:, None, :]
|
||||
if chunk_mask is not None:
|
||||
memory_mask = memory_mask * chunk_mask
|
||||
if tgt_mask.size(1) != memory_mask.size(1):
|
||||
memory_mask = torch.cat((memory_mask, memory_mask[:, -2:-1, :]), dim=1)
|
||||
|
||||
x = self.embed(tgt)
|
||||
|
||||
if pre_acoustic_embeds is not None and self.concat_embeds:
|
||||
x = torch.cat((x, pre_acoustic_embeds), dim=-1)
|
||||
x, _, _, _, _ = self.embed_concat_ffn(x, None, None, None, None)
|
||||
|
||||
x, tgt_mask, memory, memory_mask, _ = self.decoders(x, tgt_mask, memory, memory_mask)
|
||||
if self.decoders2 is not None:
|
||||
x, tgt_mask, memory, memory_mask, _ = self.decoders2(x, tgt_mask, memory, memory_mask)
|
||||
x, tgt_mask, memory, memory_mask, _ = self.decoders3(x, tgt_mask, memory, memory_mask)
|
||||
if self.normalize_before:
|
||||
x = self.after_norm(x)
|
||||
if self.output_layer is not None:
|
||||
x = self.output_layer(x)
|
||||
|
||||
olens = tgt_mask.sum(1)
|
||||
return x, olens
|
||||
|
||||
def score(
|
||||
self,
|
||||
ys,
|
||||
state,
|
||||
x,
|
||||
x_mask=None,
|
||||
pre_acoustic_embeds: torch.Tensor = None,
|
||||
):
|
||||
"""Score."""
|
||||
ys_mask = myutils.sequence_mask(
|
||||
torch.tensor([len(ys)], dtype=torch.int32), device=x.device
|
||||
)[:, :, None]
|
||||
logp, state = self.forward_one_step(
|
||||
ys.unsqueeze(0),
|
||||
ys_mask,
|
||||
x.unsqueeze(0),
|
||||
memory_mask=x_mask,
|
||||
pre_acoustic_embeds=pre_acoustic_embeds,
|
||||
cache=state,
|
||||
)
|
||||
return logp.squeeze(0), state
|
||||
|
||||
def forward_one_step(
|
||||
self,
|
||||
tgt: torch.Tensor,
|
||||
tgt_mask: torch.Tensor,
|
||||
memory: torch.Tensor,
|
||||
memory_mask: torch.Tensor = None,
|
||||
pre_acoustic_embeds: torch.Tensor = None,
|
||||
cache: List[torch.Tensor] = None,
|
||||
) -> Tuple[torch.Tensor, List[torch.Tensor]]:
|
||||
"""Forward one step.
|
||||
|
||||
Args:
|
||||
tgt: input token ids, int64 (batch, maxlen_out)
|
||||
tgt_mask: input token mask, (batch, maxlen_out)
|
||||
dtype=torch.uint8 in PyTorch 1.2-
|
||||
dtype=torch.bool in PyTorch 1.2+ (include 1.2)
|
||||
memory: encoded memory, float32 (batch, maxlen_in, feat)
|
||||
cache: cached output list of (batch, max_time_out-1, size)
|
||||
Returns:
|
||||
y, cache: NN output value and cache per `self.decoders`.
|
||||
y.shape` is (batch, maxlen_out, token)
|
||||
"""
|
||||
|
||||
x = tgt[:, -1:]
|
||||
tgt_mask = None
|
||||
x = self.embed(x)
|
||||
|
||||
if pre_acoustic_embeds is not None and self.concat_embeds:
|
||||
x = torch.cat((x, pre_acoustic_embeds), dim=-1)
|
||||
x, _, _, _, _ = self.embed_concat_ffn(x, None, None, None, None)
|
||||
|
||||
if cache is None:
|
||||
cache_layer_num = len(self.decoders)
|
||||
if self.decoders2 is not None:
|
||||
cache_layer_num += len(self.decoders2)
|
||||
cache = [None] * cache_layer_num
|
||||
new_cache = []
|
||||
# for c, decoder in zip(cache, self.decoders):
|
||||
for i in range(self.att_layer_num):
|
||||
decoder = self.decoders[i]
|
||||
c = cache[i]
|
||||
x, tgt_mask, memory, memory_mask, c_ret = decoder.forward_one_step(
|
||||
x, tgt_mask, memory, memory_mask, cache=c
|
||||
)
|
||||
new_cache.append(c_ret)
|
||||
|
||||
if self.num_blocks - self.att_layer_num >= 1:
|
||||
for i in range(self.num_blocks - self.att_layer_num):
|
||||
j = i + self.att_layer_num
|
||||
decoder = self.decoders2[i]
|
||||
c = cache[j]
|
||||
x, tgt_mask, memory, memory_mask, c_ret = decoder.forward_one_step(
|
||||
x, tgt_mask, memory, memory_mask, cache=c
|
||||
)
|
||||
new_cache.append(c_ret)
|
||||
|
||||
for decoder in self.decoders3:
|
||||
x, tgt_mask, memory, memory_mask, _ = decoder.forward_one_step(
|
||||
x, tgt_mask, memory, None, cache=None
|
||||
)
|
||||
|
||||
if self.normalize_before:
|
||||
y = self.after_norm(x[:, -1])
|
||||
else:
|
||||
y = x[:, -1]
|
||||
if self.output_layer is not None:
|
||||
y = self.output_layer(y)
|
||||
y = torch.log_softmax(y, dim=-1)
|
||||
|
||||
return y, new_cache
|
||||
@@ -0,0 +1,694 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- encoding: utf-8 -*-
|
||||
# Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
|
||||
# MIT License (https://opensource.org/licenses/MIT)
|
||||
|
||||
from typing import List
|
||||
from typing import Optional
|
||||
from typing import Sequence
|
||||
from typing import Tuple
|
||||
from typing import Union
|
||||
import logging
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
|
||||
import numpy as np
|
||||
from funasr.train_utils.device_funcs import to_device
|
||||
from funasr.models.transformer.utils.nets_utils import make_pad_mask
|
||||
from funasr.models.sanm.attention import MultiHeadedAttention, MultiHeadedAttentionSANM
|
||||
from funasr.models.transformer.embedding import (
|
||||
SinusoidalPositionEncoder,
|
||||
StreamSinusoidalPositionEncoder,
|
||||
)
|
||||
from funasr.models.transformer.layer_norm import LayerNorm
|
||||
from funasr.models.transformer.utils.multi_layer_conv import Conv1dLinear
|
||||
from funasr.models.transformer.utils.multi_layer_conv import MultiLayeredConv1d
|
||||
from funasr.models.transformer.positionwise_feed_forward import (
|
||||
PositionwiseFeedForward, # noqa: H301
|
||||
)
|
||||
from funasr.models.transformer.utils.repeat import repeat
|
||||
from funasr.models.transformer.utils.subsampling import Conv2dSubsampling
|
||||
from funasr.models.transformer.utils.subsampling import Conv2dSubsampling2
|
||||
from funasr.models.transformer.utils.subsampling import Conv2dSubsampling6
|
||||
from funasr.models.transformer.utils.subsampling import Conv2dSubsampling8
|
||||
from funasr.models.transformer.utils.subsampling import TooShortUttError
|
||||
from funasr.models.transformer.utils.subsampling import check_short_utt
|
||||
|
||||
|
||||
from funasr.models.ctc.ctc import CTC
|
||||
|
||||
from funasr.register import tables
|
||||
|
||||
|
||||
class EncoderLayerSANM(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
in_size,
|
||||
size,
|
||||
self_attn,
|
||||
feed_forward,
|
||||
dropout_rate,
|
||||
normalize_before=True,
|
||||
concat_after=False,
|
||||
stochastic_depth_rate=0.0,
|
||||
):
|
||||
"""Construct an EncoderLayer object."""
|
||||
super(EncoderLayerSANM, self).__init__()
|
||||
self.self_attn = self_attn
|
||||
self.feed_forward = feed_forward
|
||||
self.norm1 = LayerNorm(in_size)
|
||||
self.norm2 = LayerNorm(size)
|
||||
self.dropout = nn.Dropout(dropout_rate)
|
||||
self.in_size = in_size
|
||||
self.size = size
|
||||
self.normalize_before = normalize_before
|
||||
self.concat_after = concat_after
|
||||
if self.concat_after:
|
||||
self.concat_linear = nn.Linear(size + size, size)
|
||||
self.stochastic_depth_rate = stochastic_depth_rate
|
||||
self.dropout_rate = dropout_rate
|
||||
|
||||
def forward(self, x, mask, cache=None, mask_shfit_chunk=None, mask_att_chunk_encoder=None):
|
||||
"""Compute encoded features.
|
||||
|
||||
Args:
|
||||
x_input (torch.Tensor): Input tensor (#batch, time, size).
|
||||
mask (torch.Tensor): Mask tensor for the input (#batch, time).
|
||||
cache (torch.Tensor): Cache tensor of the input (#batch, time - 1, size).
|
||||
|
||||
Returns:
|
||||
torch.Tensor: Output tensor (#batch, time, size).
|
||||
torch.Tensor: Mask tensor (#batch, time).
|
||||
|
||||
"""
|
||||
skip_layer = False
|
||||
# with stochastic depth, residual connection `x + f(x)` becomes
|
||||
# `x <- x + 1 / (1 - p) * f(x)` at training time.
|
||||
stoch_layer_coeff = 1.0
|
||||
if self.training and self.stochastic_depth_rate > 0:
|
||||
skip_layer = torch.rand(1).item() < self.stochastic_depth_rate
|
||||
stoch_layer_coeff = 1.0 / (1 - self.stochastic_depth_rate)
|
||||
|
||||
if skip_layer:
|
||||
if cache is not None:
|
||||
x = torch.cat([cache, x], dim=1)
|
||||
return x, mask
|
||||
|
||||
residual = x
|
||||
if self.normalize_before:
|
||||
x = self.norm1(x)
|
||||
|
||||
if self.concat_after:
|
||||
x_concat = torch.cat(
|
||||
(
|
||||
x,
|
||||
self.self_attn(
|
||||
x,
|
||||
mask,
|
||||
mask_shfit_chunk=mask_shfit_chunk,
|
||||
mask_att_chunk_encoder=mask_att_chunk_encoder,
|
||||
),
|
||||
),
|
||||
dim=-1,
|
||||
)
|
||||
if self.in_size == self.size:
|
||||
x = residual + stoch_layer_coeff * self.concat_linear(x_concat)
|
||||
else:
|
||||
x = stoch_layer_coeff * self.concat_linear(x_concat)
|
||||
else:
|
||||
if self.in_size == self.size:
|
||||
x = residual + stoch_layer_coeff * self.dropout(
|
||||
self.self_attn(
|
||||
x,
|
||||
mask,
|
||||
mask_shfit_chunk=mask_shfit_chunk,
|
||||
mask_att_chunk_encoder=mask_att_chunk_encoder,
|
||||
)
|
||||
)
|
||||
else:
|
||||
x = stoch_layer_coeff * self.dropout(
|
||||
self.self_attn(
|
||||
x,
|
||||
mask,
|
||||
mask_shfit_chunk=mask_shfit_chunk,
|
||||
mask_att_chunk_encoder=mask_att_chunk_encoder,
|
||||
)
|
||||
)
|
||||
if not self.normalize_before:
|
||||
x = self.norm1(x)
|
||||
|
||||
residual = x
|
||||
if self.normalize_before:
|
||||
x = self.norm2(x)
|
||||
x = residual + stoch_layer_coeff * self.dropout(self.feed_forward(x))
|
||||
if not self.normalize_before:
|
||||
x = self.norm2(x)
|
||||
|
||||
return x, mask, cache, mask_shfit_chunk, mask_att_chunk_encoder
|
||||
|
||||
def forward_chunk(self, x, cache=None, chunk_size=None, look_back=0):
|
||||
"""Compute encoded features.
|
||||
|
||||
Args:
|
||||
x_input (torch.Tensor): Input tensor (#batch, time, size).
|
||||
mask (torch.Tensor): Mask tensor for the input (#batch, time).
|
||||
cache (torch.Tensor): Cache tensor of the input (#batch, time - 1, size).
|
||||
|
||||
Returns:
|
||||
torch.Tensor: Output tensor (#batch, time, size).
|
||||
torch.Tensor: Mask tensor (#batch, time).
|
||||
|
||||
"""
|
||||
|
||||
residual = x
|
||||
if self.normalize_before:
|
||||
x = self.norm1(x)
|
||||
|
||||
if self.in_size == self.size:
|
||||
attn, cache = self.self_attn.forward_chunk(x, cache, chunk_size, look_back)
|
||||
x = residual + attn
|
||||
else:
|
||||
x, cache = self.self_attn.forward_chunk(x, cache, chunk_size, look_back)
|
||||
|
||||
if not self.normalize_before:
|
||||
x = self.norm1(x)
|
||||
|
||||
residual = x
|
||||
if self.normalize_before:
|
||||
x = self.norm2(x)
|
||||
x = residual + self.feed_forward(x)
|
||||
if not self.normalize_before:
|
||||
x = self.norm2(x)
|
||||
|
||||
return x, cache
|
||||
|
||||
|
||||
@tables.register("encoder_classes", "SANMEncoder")
|
||||
class SANMEncoder(nn.Module):
|
||||
"""
|
||||
Author: Zhifu Gao, Shiliang Zhang, Ming Lei, Ian McLoughlin
|
||||
San-m: Memory equipped self-attention for end-to-end speech recognition
|
||||
https://arxiv.org/abs/2006.01713
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
input_size: int,
|
||||
output_size: int = 256,
|
||||
attention_heads: int = 4,
|
||||
linear_units: int = 2048,
|
||||
num_blocks: int = 6,
|
||||
dropout_rate: float = 0.1,
|
||||
positional_dropout_rate: float = 0.1,
|
||||
attention_dropout_rate: float = 0.0,
|
||||
input_layer: Optional[str] = "conv2d",
|
||||
pos_enc_class=SinusoidalPositionEncoder,
|
||||
normalize_before: bool = True,
|
||||
concat_after: bool = False,
|
||||
positionwise_layer_type: str = "linear",
|
||||
positionwise_conv_kernel_size: int = 1,
|
||||
padding_idx: int = -1,
|
||||
interctc_layer_idx: List[int] = [],
|
||||
interctc_use_conditioning: bool = False,
|
||||
kernel_size: int = 11,
|
||||
sanm_shfit: int = 0,
|
||||
lora_list: List[str] = None,
|
||||
lora_rank: int = 8,
|
||||
lora_alpha: int = 16,
|
||||
lora_dropout: float = 0.1,
|
||||
selfattention_layer_type: str = "sanm",
|
||||
tf2torch_tensor_name_prefix_torch: str = "encoder",
|
||||
tf2torch_tensor_name_prefix_tf: str = "seq2seq/encoder",
|
||||
):
|
||||
"""Initialize SANMEncoder.
|
||||
|
||||
Args:
|
||||
input_size: Size/dimension parameter.
|
||||
output_size: Size/dimension parameter.
|
||||
attention_heads: TODO.
|
||||
linear_units: TODO.
|
||||
num_blocks: TODO.
|
||||
dropout_rate: TODO.
|
||||
positional_dropout_rate: TODO.
|
||||
attention_dropout_rate: TODO.
|
||||
input_layer: TODO.
|
||||
pos_enc_class: TODO.
|
||||
normalize_before: TODO.
|
||||
concat_after: TODO.
|
||||
positionwise_layer_type: TODO.
|
||||
positionwise_conv_kernel_size: Size/dimension parameter.
|
||||
padding_idx: TODO.
|
||||
interctc_layer_idx: TODO.
|
||||
interctc_use_conditioning: TODO.
|
||||
kernel_size: Size/dimension parameter.
|
||||
sanm_shfit: TODO.
|
||||
lora_list: TODO.
|
||||
lora_rank: TODO.
|
||||
lora_alpha: TODO.
|
||||
lora_dropout: TODO.
|
||||
selfattention_layer_type: TODO.
|
||||
tf2torch_tensor_name_prefix_torch: TODO.
|
||||
tf2torch_tensor_name_prefix_tf: TODO.
|
||||
"""
|
||||
super().__init__()
|
||||
self._output_size = output_size
|
||||
|
||||
if input_layer == "linear":
|
||||
self.embed = torch.nn.Sequential(
|
||||
torch.nn.Linear(input_size, output_size),
|
||||
torch.nn.LayerNorm(output_size),
|
||||
torch.nn.Dropout(dropout_rate),
|
||||
torch.nn.ReLU(),
|
||||
pos_enc_class(output_size, positional_dropout_rate),
|
||||
)
|
||||
elif input_layer == "conv2d":
|
||||
self.embed = Conv2dSubsampling(input_size, output_size, dropout_rate)
|
||||
elif input_layer == "conv2d2":
|
||||
self.embed = Conv2dSubsampling2(input_size, output_size, dropout_rate)
|
||||
elif input_layer == "conv2d6":
|
||||
self.embed = Conv2dSubsampling6(input_size, output_size, dropout_rate)
|
||||
elif input_layer == "conv2d8":
|
||||
self.embed = Conv2dSubsampling8(input_size, output_size, dropout_rate)
|
||||
elif input_layer == "embed":
|
||||
self.embed = torch.nn.Sequential(
|
||||
torch.nn.Embedding(input_size, output_size, padding_idx=padding_idx),
|
||||
SinusoidalPositionEncoder(),
|
||||
)
|
||||
elif input_layer is None:
|
||||
if input_size == output_size:
|
||||
self.embed = None
|
||||
else:
|
||||
self.embed = torch.nn.Linear(input_size, output_size)
|
||||
elif input_layer == "pe":
|
||||
self.embed = SinusoidalPositionEncoder()
|
||||
elif input_layer == "pe_online":
|
||||
self.embed = StreamSinusoidalPositionEncoder()
|
||||
else:
|
||||
raise ValueError("unknown input_layer: " + input_layer)
|
||||
self.normalize_before = normalize_before
|
||||
if positionwise_layer_type == "linear":
|
||||
positionwise_layer = PositionwiseFeedForward
|
||||
positionwise_layer_args = (
|
||||
output_size,
|
||||
linear_units,
|
||||
dropout_rate,
|
||||
)
|
||||
elif positionwise_layer_type == "conv1d":
|
||||
positionwise_layer = MultiLayeredConv1d
|
||||
positionwise_layer_args = (
|
||||
output_size,
|
||||
linear_units,
|
||||
positionwise_conv_kernel_size,
|
||||
dropout_rate,
|
||||
)
|
||||
elif positionwise_layer_type == "conv1d-linear":
|
||||
positionwise_layer = Conv1dLinear
|
||||
positionwise_layer_args = (
|
||||
output_size,
|
||||
linear_units,
|
||||
positionwise_conv_kernel_size,
|
||||
dropout_rate,
|
||||
)
|
||||
else:
|
||||
raise NotImplementedError("Support only linear or conv1d.")
|
||||
|
||||
if selfattention_layer_type == "selfattn":
|
||||
encoder_selfattn_layer = MultiHeadedAttention
|
||||
encoder_selfattn_layer_args = (
|
||||
attention_heads,
|
||||
output_size,
|
||||
attention_dropout_rate,
|
||||
)
|
||||
|
||||
elif selfattention_layer_type == "sanm":
|
||||
encoder_selfattn_layer = MultiHeadedAttentionSANM
|
||||
encoder_selfattn_layer_args0 = (
|
||||
attention_heads,
|
||||
input_size,
|
||||
output_size,
|
||||
attention_dropout_rate,
|
||||
kernel_size,
|
||||
sanm_shfit,
|
||||
lora_list,
|
||||
lora_rank,
|
||||
lora_alpha,
|
||||
lora_dropout,
|
||||
)
|
||||
|
||||
encoder_selfattn_layer_args = (
|
||||
attention_heads,
|
||||
output_size,
|
||||
output_size,
|
||||
attention_dropout_rate,
|
||||
kernel_size,
|
||||
sanm_shfit,
|
||||
lora_list,
|
||||
lora_rank,
|
||||
lora_alpha,
|
||||
lora_dropout,
|
||||
)
|
||||
self.encoders0 = repeat(
|
||||
1,
|
||||
lambda lnum: EncoderLayerSANM(
|
||||
input_size,
|
||||
output_size,
|
||||
encoder_selfattn_layer(*encoder_selfattn_layer_args0),
|
||||
positionwise_layer(*positionwise_layer_args),
|
||||
dropout_rate,
|
||||
normalize_before,
|
||||
concat_after,
|
||||
),
|
||||
)
|
||||
|
||||
self.encoders = repeat(
|
||||
num_blocks - 1,
|
||||
lambda lnum: EncoderLayerSANM(
|
||||
output_size,
|
||||
output_size,
|
||||
encoder_selfattn_layer(*encoder_selfattn_layer_args),
|
||||
positionwise_layer(*positionwise_layer_args),
|
||||
dropout_rate,
|
||||
normalize_before,
|
||||
concat_after,
|
||||
),
|
||||
)
|
||||
if self.normalize_before:
|
||||
self.after_norm = LayerNorm(output_size)
|
||||
|
||||
self.interctc_layer_idx = interctc_layer_idx
|
||||
if len(interctc_layer_idx) > 0:
|
||||
assert 0 < min(interctc_layer_idx) and max(interctc_layer_idx) < num_blocks
|
||||
self.interctc_use_conditioning = interctc_use_conditioning
|
||||
self.conditioning_layer = None
|
||||
self.dropout = nn.Dropout(dropout_rate)
|
||||
self.tf2torch_tensor_name_prefix_torch = tf2torch_tensor_name_prefix_torch
|
||||
self.tf2torch_tensor_name_prefix_tf = tf2torch_tensor_name_prefix_tf
|
||||
|
||||
def output_size(self) -> int:
|
||||
"""Output size."""
|
||||
return self._output_size
|
||||
|
||||
def forward(
|
||||
self,
|
||||
xs_pad: torch.Tensor,
|
||||
ilens: torch.Tensor,
|
||||
prev_states: torch.Tensor = None,
|
||||
ctc: CTC = None,
|
||||
) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
|
||||
"""Embed positions in tensor.
|
||||
|
||||
Args:
|
||||
xs_pad: input tensor (B, L, D)
|
||||
ilens: input length (B)
|
||||
prev_states: Not to be used now.
|
||||
Returns:
|
||||
position embedded tensor and mask
|
||||
"""
|
||||
masks = (~make_pad_mask(ilens)[:, None, :]).to(xs_pad.device)
|
||||
xs_pad = xs_pad * self.output_size() ** 0.5
|
||||
if self.embed is None:
|
||||
xs_pad = xs_pad
|
||||
elif (
|
||||
isinstance(self.embed, Conv2dSubsampling)
|
||||
or isinstance(self.embed, Conv2dSubsampling2)
|
||||
or isinstance(self.embed, Conv2dSubsampling6)
|
||||
or isinstance(self.embed, Conv2dSubsampling8)
|
||||
):
|
||||
short_status, limit_size = check_short_utt(self.embed, xs_pad.size(1))
|
||||
if short_status:
|
||||
raise TooShortUttError(
|
||||
f"has {xs_pad.size(1)} frames and is too short for subsampling "
|
||||
+ f"(it needs more than {limit_size} frames), return empty results",
|
||||
xs_pad.size(1),
|
||||
limit_size,
|
||||
)
|
||||
xs_pad, masks = self.embed(xs_pad, masks)
|
||||
else:
|
||||
xs_pad = self.embed(xs_pad)
|
||||
|
||||
# xs_pad = self.dropout(xs_pad)
|
||||
encoder_outs = self.encoders0(xs_pad, masks)
|
||||
xs_pad, masks = encoder_outs[0], encoder_outs[1]
|
||||
intermediate_outs = []
|
||||
if len(self.interctc_layer_idx) == 0:
|
||||
encoder_outs = self.encoders(xs_pad, masks)
|
||||
xs_pad, masks = encoder_outs[0], encoder_outs[1]
|
||||
else:
|
||||
for layer_idx, encoder_layer in enumerate(self.encoders):
|
||||
encoder_outs = encoder_layer(xs_pad, masks)
|
||||
xs_pad, masks = encoder_outs[0], encoder_outs[1]
|
||||
|
||||
if layer_idx + 1 in self.interctc_layer_idx:
|
||||
encoder_out = xs_pad
|
||||
|
||||
# intermediate outputs are also normalized
|
||||
if self.normalize_before:
|
||||
encoder_out = self.after_norm(encoder_out)
|
||||
|
||||
intermediate_outs.append((layer_idx + 1, encoder_out))
|
||||
|
||||
if self.interctc_use_conditioning:
|
||||
ctc_out = ctc.softmax(encoder_out)
|
||||
xs_pad = xs_pad + self.conditioning_layer(ctc_out)
|
||||
|
||||
if self.normalize_before:
|
||||
xs_pad = self.after_norm(xs_pad)
|
||||
|
||||
olens = masks.squeeze(1).sum(1)
|
||||
if len(intermediate_outs) > 0:
|
||||
return (xs_pad, intermediate_outs), olens, None
|
||||
return xs_pad, olens, None
|
||||
|
||||
def _add_overlap_chunk(self, feats: np.ndarray, cache: dict = None):
|
||||
"""Internal: add overlap chunk.
|
||||
|
||||
Args:
|
||||
feats: Feature tensor (e.g., fbank), shape (batch, frames, dim).
|
||||
cache: State cache dict for streaming inference.
|
||||
"""
|
||||
if cache is None:
|
||||
cache = {}
|
||||
if len(cache) == 0:
|
||||
return feats
|
||||
cache["feats"] = to_device(cache["feats"], device=feats.device)
|
||||
overlap_feats = torch.cat((cache["feats"], feats), dim=1)
|
||||
cache["feats"] = overlap_feats[:, -(cache["chunk_size"][0] + cache["chunk_size"][2]) :, :]
|
||||
return overlap_feats
|
||||
|
||||
def forward_chunk(
|
||||
self,
|
||||
xs_pad: torch.Tensor,
|
||||
ilens: torch.Tensor,
|
||||
cache: dict = None,
|
||||
ctc: CTC = None,
|
||||
):
|
||||
"""Forward chunk.
|
||||
|
||||
Args:
|
||||
xs_pad: TODO.
|
||||
ilens: TODO.
|
||||
cache: State cache dict for streaming inference.
|
||||
ctc: TODO.
|
||||
"""
|
||||
if cache is None:
|
||||
cache = {}
|
||||
xs_pad *= self.output_size() ** 0.5
|
||||
if self.embed is None:
|
||||
xs_pad = xs_pad
|
||||
else:
|
||||
xs_pad = self.embed(xs_pad, cache)
|
||||
if cache["tail_chunk"]:
|
||||
xs_pad = to_device(cache["feats"], device=xs_pad.device)
|
||||
else:
|
||||
xs_pad = self._add_overlap_chunk(xs_pad, cache)
|
||||
encoder_outs = self.encoders0(xs_pad, None, None, None, None)
|
||||
xs_pad, masks = encoder_outs[0], encoder_outs[1]
|
||||
intermediate_outs = []
|
||||
if len(self.interctc_layer_idx) == 0:
|
||||
encoder_outs = self.encoders(xs_pad, None, None, None, None)
|
||||
xs_pad, masks = encoder_outs[0], encoder_outs[1]
|
||||
else:
|
||||
for layer_idx, encoder_layer in enumerate(self.encoders):
|
||||
encoder_outs = encoder_layer(xs_pad, None, None, None, None)
|
||||
xs_pad, masks = encoder_outs[0], encoder_outs[1]
|
||||
if layer_idx + 1 in self.interctc_layer_idx:
|
||||
encoder_out = xs_pad
|
||||
|
||||
# intermediate outputs are also normalized
|
||||
if self.normalize_before:
|
||||
encoder_out = self.after_norm(encoder_out)
|
||||
|
||||
intermediate_outs.append((layer_idx + 1, encoder_out))
|
||||
|
||||
if self.interctc_use_conditioning:
|
||||
ctc_out = ctc.softmax(encoder_out)
|
||||
xs_pad = xs_pad + self.conditioning_layer(ctc_out)
|
||||
|
||||
if self.normalize_before:
|
||||
xs_pad = self.after_norm(xs_pad)
|
||||
|
||||
if len(intermediate_outs) > 0:
|
||||
return (xs_pad, intermediate_outs), None, None
|
||||
return xs_pad, ilens, None
|
||||
|
||||
|
||||
class EncoderLayerSANMExport(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
model,
|
||||
):
|
||||
"""Construct an EncoderLayer object."""
|
||||
super().__init__()
|
||||
self.self_attn = model.self_attn
|
||||
self.feed_forward = model.feed_forward
|
||||
self.norm1 = model.norm1
|
||||
self.norm2 = model.norm2
|
||||
self.in_size = model.in_size
|
||||
self.size = model.size
|
||||
|
||||
def forward(self, x, mask):
|
||||
|
||||
"""Forward pass for training.
|
||||
|
||||
Args:
|
||||
x: TODO.
|
||||
mask: TODO.
|
||||
"""
|
||||
residual = x
|
||||
x = self.norm1(x)
|
||||
x = self.self_attn(x, mask)
|
||||
if self.in_size == self.size:
|
||||
x = x + residual
|
||||
residual = x
|
||||
x = self.norm2(x)
|
||||
x = self.feed_forward(x)
|
||||
x = x + residual
|
||||
|
||||
return x, mask
|
||||
|
||||
|
||||
@tables.register("encoder_classes", "SANMEncoderChunkOptExport")
|
||||
@tables.register("encoder_classes", "SANMEncoderExport")
|
||||
class SANMEncoderExport(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
model,
|
||||
max_seq_len=512,
|
||||
feats_dim=560,
|
||||
model_name="encoder",
|
||||
onnx: bool = True,
|
||||
ctc_linear: nn.Module = None,
|
||||
):
|
||||
"""Initialize SANMEncoderExport.
|
||||
|
||||
Args:
|
||||
model: Model instance or model name.
|
||||
max_seq_len: TODO.
|
||||
feats_dim: Size/dimension parameter.
|
||||
model_name: TODO.
|
||||
onnx: TODO.
|
||||
ctc_linear: TODO.
|
||||
"""
|
||||
super().__init__()
|
||||
self.embed = model.embed
|
||||
if isinstance(self.embed, StreamSinusoidalPositionEncoder):
|
||||
self.embed = None
|
||||
self.model = model
|
||||
self.feats_dim = feats_dim
|
||||
self._output_size = model._output_size
|
||||
|
||||
from funasr.utils.torch_function import sequence_mask
|
||||
|
||||
self.make_pad_mask = sequence_mask(max_seq_len, flip=False)
|
||||
|
||||
from funasr.models.sanm.attention import MultiHeadedAttentionSANMExport
|
||||
|
||||
if hasattr(model, "encoders0"):
|
||||
for i, d in enumerate(self.model.encoders0):
|
||||
if isinstance(d.self_attn, MultiHeadedAttentionSANM):
|
||||
d.self_attn = MultiHeadedAttentionSANMExport(d.self_attn)
|
||||
self.model.encoders0[i] = EncoderLayerSANMExport(d)
|
||||
|
||||
for i, d in enumerate(self.model.encoders):
|
||||
if isinstance(d.self_attn, MultiHeadedAttentionSANM):
|
||||
d.self_attn = MultiHeadedAttentionSANMExport(d.self_attn)
|
||||
self.model.encoders[i] = EncoderLayerSANMExport(d)
|
||||
|
||||
self.model_name = model_name
|
||||
self.num_heads = model.encoders[0].self_attn.h
|
||||
self.hidden_size = model.encoders[0].self_attn.linear_out.out_features
|
||||
|
||||
self.ctc_linear = ctc_linear
|
||||
|
||||
def prepare_mask(self, mask):
|
||||
"""Prepare mask.
|
||||
|
||||
Args:
|
||||
mask: TODO.
|
||||
"""
|
||||
mask_3d_btd = mask[:, :, None]
|
||||
if len(mask.shape) == 2:
|
||||
mask_4d_bhlt = 1 - mask[:, None, None, :]
|
||||
elif len(mask.shape) == 3:
|
||||
mask_4d_bhlt = 1 - mask[:, None, :]
|
||||
mask_4d_bhlt = mask_4d_bhlt * -10000.0
|
||||
|
||||
return mask_3d_btd, mask_4d_bhlt
|
||||
|
||||
def forward(self, speech: torch.Tensor, speech_lengths: torch.Tensor, online: bool = False):
|
||||
"""Forward pass for training.
|
||||
|
||||
Args:
|
||||
speech: Speech audio tensor, shape (batch, time).
|
||||
speech_lengths: Length of each speech sample.
|
||||
online: TODO.
|
||||
"""
|
||||
if not online:
|
||||
speech = speech * self._output_size**0.5
|
||||
|
||||
mask = self.make_pad_mask(speech_lengths)
|
||||
mask = self.prepare_mask(mask)
|
||||
if self.embed is None:
|
||||
xs_pad = speech
|
||||
else:
|
||||
xs_pad = self.embed(speech)
|
||||
|
||||
encoder_outs = self.model.encoders0(xs_pad, mask)
|
||||
xs_pad, masks = encoder_outs[0], encoder_outs[1]
|
||||
|
||||
encoder_outs = self.model.encoders(xs_pad, mask)
|
||||
xs_pad, masks = encoder_outs[0], encoder_outs[1]
|
||||
|
||||
xs_pad = self.model.after_norm(xs_pad)
|
||||
|
||||
if self.ctc_linear is not None:
|
||||
xs_pad = self.ctc_linear(xs_pad)
|
||||
xs_pad = F.softmax(xs_pad, dim=2)
|
||||
|
||||
return xs_pad, speech_lengths
|
||||
|
||||
def get_output_size(self):
|
||||
"""Get output size."""
|
||||
return self.model.encoders[0].size
|
||||
|
||||
def get_dummy_inputs(self):
|
||||
"""Get dummy inputs."""
|
||||
feats = torch.randn(1, 100, self.feats_dim)
|
||||
return feats
|
||||
|
||||
def get_input_names(self):
|
||||
"""Get input names."""
|
||||
return ["feats"]
|
||||
|
||||
def get_output_names(self):
|
||||
"""Get output names."""
|
||||
return ["encoder_out", "encoder_out_lens", "predictor_weight"]
|
||||
|
||||
def get_dynamic_axes(self):
|
||||
"""Get dynamic axes."""
|
||||
return {
|
||||
"feats": {1: "feats_length"},
|
||||
"encoder_out": {1: "enc_out_length"},
|
||||
"predictor_weight": {1: "pre_out_length"},
|
||||
}
|
||||
@@ -0,0 +1,34 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- encoding: utf-8 -*-
|
||||
# Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
|
||||
# MIT License (https://opensource.org/licenses/MIT)
|
||||
|
||||
import logging
|
||||
|
||||
import torch
|
||||
|
||||
from funasr.models.transformer.model import Transformer
|
||||
from funasr.register import tables
|
||||
|
||||
|
||||
@tables.register("model_classes", "SANM")
|
||||
class SANM(Transformer):
|
||||
"""
|
||||
Author: Zhifu Gao, Shiliang Zhang, Ming Lei, Ian McLoughlin
|
||||
San-m: Memory equipped self-attention for end-to-end speech recognition
|
||||
https://arxiv.org/abs/2006.01713
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
*args,
|
||||
**kwargs,
|
||||
):
|
||||
|
||||
"""Initialize SANM.
|
||||
|
||||
Args:
|
||||
*args: Variable positional arguments.
|
||||
**kwargs: Additional keyword arguments.
|
||||
"""
|
||||
super().__init__(*args, **kwargs)
|
||||
@@ -0,0 +1,384 @@
|
||||
import os
|
||||
import math
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
|
||||
class MultiHeadedAttentionSANMExport(nn.Module):
|
||||
def __init__(self, model):
|
||||
"""Initialize MultiHeadedAttentionSANMExport.
|
||||
|
||||
Args:
|
||||
model: Model instance or model name.
|
||||
"""
|
||||
super().__init__()
|
||||
self.d_k = model.d_k
|
||||
self.h = model.h
|
||||
self.linear_out = model.linear_out
|
||||
self.linear_q_k_v = model.linear_q_k_v
|
||||
self.fsmn_block = model.fsmn_block
|
||||
self.pad_fn = model.pad_fn
|
||||
|
||||
self.attn = None
|
||||
self.all_head_size = self.h * self.d_k
|
||||
|
||||
def forward(self, x, mask):
|
||||
"""Forward pass for training.
|
||||
|
||||
Args:
|
||||
x: TODO.
|
||||
mask: TODO.
|
||||
"""
|
||||
mask_3d_btd, mask_4d_bhlt = mask
|
||||
q_h, k_h, v_h, v = self.forward_qkv(x)
|
||||
fsmn_memory = self.forward_fsmn(v, mask_3d_btd)
|
||||
q_h = q_h * self.d_k ** (-0.5)
|
||||
scores = torch.matmul(q_h, k_h.transpose(-2, -1))
|
||||
att_outs = self.forward_attention(v_h, scores, mask_4d_bhlt)
|
||||
return att_outs + fsmn_memory
|
||||
|
||||
def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
|
||||
"""Transpose for scores.
|
||||
|
||||
Args:
|
||||
x: TODO.
|
||||
"""
|
||||
new_x_shape = x.size()[:-1] + (self.h, self.d_k)
|
||||
x = x.view(new_x_shape)
|
||||
return x.permute(0, 2, 1, 3)
|
||||
|
||||
def forward_qkv(self, x):
|
||||
"""Forward qkv.
|
||||
|
||||
Args:
|
||||
x: TODO.
|
||||
"""
|
||||
q_k_v = self.linear_q_k_v(x)
|
||||
q, k, v = torch.split(q_k_v, int(self.h * self.d_k), dim=-1)
|
||||
q_h = self.transpose_for_scores(q)
|
||||
k_h = self.transpose_for_scores(k)
|
||||
v_h = self.transpose_for_scores(v)
|
||||
return q_h, k_h, v_h, v
|
||||
|
||||
def forward_fsmn(self, inputs, mask):
|
||||
# b, t, d = inputs.size()
|
||||
# mask = torch.reshape(mask, (b, -1, 1))
|
||||
"""Forward fsmn.
|
||||
|
||||
Args:
|
||||
inputs: TODO.
|
||||
mask: TODO.
|
||||
"""
|
||||
inputs = inputs * mask
|
||||
x = inputs.transpose(1, 2)
|
||||
x = self.pad_fn(x)
|
||||
x = self.fsmn_block(x)
|
||||
x = x.transpose(1, 2)
|
||||
x = x + inputs
|
||||
x = x * mask
|
||||
return x
|
||||
|
||||
def forward_attention(self, value, scores, mask):
|
||||
"""Forward attention.
|
||||
|
||||
Args:
|
||||
value: TODO.
|
||||
scores: TODO.
|
||||
mask: TODO.
|
||||
"""
|
||||
scores = scores + mask
|
||||
|
||||
attn = torch.softmax(scores, dim=-1)
|
||||
context_layer = torch.matmul(attn, value) # (batch, head, time1, d_k)
|
||||
|
||||
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
|
||||
new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
|
||||
context_layer = context_layer.view(new_context_layer_shape)
|
||||
return self.linear_out(context_layer) # (batch, time1, d_model)
|
||||
|
||||
|
||||
def preprocess_for_attn(x, mask, cache, pad_fn, kernel_size):
|
||||
"""Preprocess for attn.
|
||||
|
||||
Args:
|
||||
x: TODO.
|
||||
mask: TODO.
|
||||
cache: State cache dict for streaming inference.
|
||||
pad_fn: TODO.
|
||||
kernel_size: Size/dimension parameter.
|
||||
"""
|
||||
x = x * mask
|
||||
x = x.transpose(1, 2)
|
||||
if cache is None:
|
||||
x = pad_fn(x)
|
||||
else:
|
||||
x = torch.cat((cache, x), dim=2)
|
||||
cache = x[:, :, -(kernel_size - 1) :]
|
||||
return x, cache
|
||||
|
||||
|
||||
torch_version = tuple([int(i) for i in torch.__version__.split(".")[:2]])
|
||||
if torch_version >= (1, 8):
|
||||
import torch.fx
|
||||
|
||||
torch.fx.wrap("preprocess_for_attn")
|
||||
|
||||
|
||||
class MultiHeadedAttentionSANMDecoderExport(nn.Module):
|
||||
def __init__(self, model):
|
||||
"""Initialize MultiHeadedAttentionSANMDecoderExport.
|
||||
|
||||
Args:
|
||||
model: Model instance or model name.
|
||||
"""
|
||||
super().__init__()
|
||||
self.fsmn_block = model.fsmn_block
|
||||
self.pad_fn = model.pad_fn
|
||||
self.kernel_size = model.kernel_size
|
||||
self.attn = None
|
||||
|
||||
def forward(self, inputs, mask, cache=None):
|
||||
"""Forward pass for training.
|
||||
|
||||
Args:
|
||||
inputs: TODO.
|
||||
mask: TODO.
|
||||
cache: State cache dict for streaming inference.
|
||||
"""
|
||||
x, cache = preprocess_for_attn(inputs, mask, cache, self.pad_fn, self.kernel_size)
|
||||
x = self.fsmn_block(x)
|
||||
x = x.transpose(1, 2)
|
||||
|
||||
x = x + inputs
|
||||
x = x * mask
|
||||
return x, cache
|
||||
|
||||
|
||||
class MultiHeadedAttentionCrossAttExport(nn.Module):
|
||||
def __init__(self, model):
|
||||
"""Initialize MultiHeadedAttentionCrossAttExport.
|
||||
|
||||
Args:
|
||||
model: Model instance or model name.
|
||||
"""
|
||||
super().__init__()
|
||||
self.d_k = model.d_k
|
||||
self.h = model.h
|
||||
self.linear_q = model.linear_q
|
||||
self.linear_k_v = model.linear_k_v
|
||||
self.linear_out = model.linear_out
|
||||
self.attn = None
|
||||
self.all_head_size = self.h * self.d_k
|
||||
|
||||
def forward(self, x, memory, memory_mask):
|
||||
"""Forward pass for training.
|
||||
|
||||
Args:
|
||||
x: TODO.
|
||||
memory: TODO.
|
||||
memory_mask: TODO.
|
||||
"""
|
||||
q, k, v = self.forward_qkv(x, memory)
|
||||
scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k)
|
||||
return self.forward_attention(v, scores, memory_mask)
|
||||
|
||||
def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
|
||||
"""Transpose for scores.
|
||||
|
||||
Args:
|
||||
x: TODO.
|
||||
"""
|
||||
new_x_shape = x.size()[:-1] + (self.h, self.d_k)
|
||||
x = x.view(new_x_shape)
|
||||
return x.permute(0, 2, 1, 3)
|
||||
|
||||
def forward_qkv(self, x, memory):
|
||||
"""Forward qkv.
|
||||
|
||||
Args:
|
||||
x: TODO.
|
||||
memory: TODO.
|
||||
"""
|
||||
q = self.linear_q(x)
|
||||
|
||||
k_v = self.linear_k_v(memory)
|
||||
k, v = torch.split(k_v, int(self.h * self.d_k), dim=-1)
|
||||
q = self.transpose_for_scores(q)
|
||||
k = self.transpose_for_scores(k)
|
||||
v = self.transpose_for_scores(v)
|
||||
return q, k, v
|
||||
|
||||
def forward_attention(self, value, scores, mask):
|
||||
"""Forward attention.
|
||||
|
||||
Args:
|
||||
value: TODO.
|
||||
scores: TODO.
|
||||
mask: TODO.
|
||||
"""
|
||||
scores = scores + mask
|
||||
|
||||
attn = torch.softmax(scores, dim=-1)
|
||||
context_layer = torch.matmul(attn, value) # (batch, head, time1, d_k)
|
||||
|
||||
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
|
||||
new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
|
||||
context_layer = context_layer.view(new_context_layer_shape)
|
||||
return self.linear_out(context_layer) # (batch, time1, d_model)
|
||||
|
||||
|
||||
class OnnxMultiHeadedAttention(nn.Module):
|
||||
def __init__(self, model):
|
||||
"""Initialize OnnxMultiHeadedAttention.
|
||||
|
||||
Args:
|
||||
model: Model instance or model name.
|
||||
"""
|
||||
super().__init__()
|
||||
self.d_k = model.d_k
|
||||
self.h = model.h
|
||||
self.linear_q = model.linear_q
|
||||
self.linear_k = model.linear_k
|
||||
self.linear_v = model.linear_v
|
||||
self.linear_out = model.linear_out
|
||||
self.attn = None
|
||||
self.all_head_size = self.h * self.d_k
|
||||
|
||||
def forward(self, query, key, value, mask):
|
||||
"""Forward pass for training.
|
||||
|
||||
Args:
|
||||
query: TODO.
|
||||
key: Sample identifiers.
|
||||
value: TODO.
|
||||
mask: TODO.
|
||||
"""
|
||||
q, k, v = self.forward_qkv(query, key, value)
|
||||
scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k)
|
||||
return self.forward_attention(v, scores, mask)
|
||||
|
||||
def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
|
||||
"""Transpose for scores.
|
||||
|
||||
Args:
|
||||
x: TODO.
|
||||
"""
|
||||
new_x_shape = x.size()[:-1] + (self.h, self.d_k)
|
||||
x = x.view(new_x_shape)
|
||||
return x.permute(0, 2, 1, 3)
|
||||
|
||||
def forward_qkv(self, query, key, value):
|
||||
"""Forward qkv.
|
||||
|
||||
Args:
|
||||
query: TODO.
|
||||
key: Sample identifiers.
|
||||
value: TODO.
|
||||
"""
|
||||
q = self.linear_q(query)
|
||||
k = self.linear_k(key)
|
||||
v = self.linear_v(value)
|
||||
q = self.transpose_for_scores(q)
|
||||
k = self.transpose_for_scores(k)
|
||||
v = self.transpose_for_scores(v)
|
||||
return q, k, v
|
||||
|
||||
def forward_attention(self, value, scores, mask):
|
||||
"""Forward attention.
|
||||
|
||||
Args:
|
||||
value: TODO.
|
||||
scores: TODO.
|
||||
mask: TODO.
|
||||
"""
|
||||
scores = scores + mask
|
||||
|
||||
attn = torch.softmax(scores, dim=-1)
|
||||
context_layer = torch.matmul(attn, value) # (batch, head, time1, d_k)
|
||||
|
||||
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
|
||||
new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
|
||||
context_layer = context_layer.view(new_context_layer_shape)
|
||||
return self.linear_out(context_layer) # (batch, time1, d_model)
|
||||
|
||||
|
||||
class OnnxRelPosMultiHeadedAttention(OnnxMultiHeadedAttention):
|
||||
def __init__(self, model):
|
||||
"""Initialize OnnxRelPosMultiHeadedAttention.
|
||||
|
||||
Args:
|
||||
model: Model instance or model name.
|
||||
"""
|
||||
super().__init__(model)
|
||||
self.linear_pos = model.linear_pos
|
||||
self.pos_bias_u = model.pos_bias_u
|
||||
self.pos_bias_v = model.pos_bias_v
|
||||
|
||||
def forward(self, query, key, value, pos_emb, mask):
|
||||
"""Forward pass for training.
|
||||
|
||||
Args:
|
||||
query: TODO.
|
||||
key: Sample identifiers.
|
||||
value: TODO.
|
||||
pos_emb: TODO.
|
||||
mask: TODO.
|
||||
"""
|
||||
q, k, v = self.forward_qkv(query, key, value)
|
||||
q = q.transpose(1, 2) # (batch, time1, head, d_k)
|
||||
|
||||
p = self.transpose_for_scores(self.linear_pos(pos_emb)) # (batch, head, time1, d_k)
|
||||
|
||||
# (batch, head, time1, d_k)
|
||||
q_with_bias_u = (q + self.pos_bias_u).transpose(1, 2)
|
||||
# (batch, head, time1, d_k)
|
||||
q_with_bias_v = (q + self.pos_bias_v).transpose(1, 2)
|
||||
|
||||
# compute attention score
|
||||
# first compute matrix a and matrix c
|
||||
# as described in https://arxiv.org/abs/1901.02860 Section 3.3
|
||||
# (batch, head, time1, time2)
|
||||
matrix_ac = torch.matmul(q_with_bias_u, k.transpose(-2, -1))
|
||||
|
||||
# compute matrix b and matrix d
|
||||
# (batch, head, time1, time1)
|
||||
matrix_bd = torch.matmul(q_with_bias_v, p.transpose(-2, -1))
|
||||
matrix_bd = self.rel_shift(matrix_bd)
|
||||
|
||||
scores = (matrix_ac + matrix_bd) / math.sqrt(self.d_k) # (batch, head, time1, time2)
|
||||
|
||||
return self.forward_attention(v, scores, mask)
|
||||
|
||||
def rel_shift(self, x):
|
||||
"""Rel shift.
|
||||
|
||||
Args:
|
||||
x: TODO.
|
||||
"""
|
||||
zero_pad = torch.zeros((*x.size()[:3], 1), device=x.device, dtype=x.dtype)
|
||||
x_padded = torch.cat([zero_pad, x], dim=-1)
|
||||
|
||||
x_padded = x_padded.view(*x.size()[:2], x.size(3) + 1, x.size(2))
|
||||
x = x_padded[:, :, 1:].view_as(x)[
|
||||
:, :, :, : x.size(-1) // 2 + 1
|
||||
] # only keep the positions from 0 to time2
|
||||
return x
|
||||
|
||||
def forward_attention(self, value, scores, mask):
|
||||
"""Forward attention.
|
||||
|
||||
Args:
|
||||
value: TODO.
|
||||
scores: TODO.
|
||||
mask: TODO.
|
||||
"""
|
||||
scores = scores + mask
|
||||
|
||||
attn = torch.softmax(scores, dim=-1)
|
||||
context_layer = torch.matmul(attn, value) # (batch, head, time1, d_k)
|
||||
|
||||
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
|
||||
new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
|
||||
context_layer = context_layer.view(new_context_layer_shape)
|
||||
return self.linear_out(context_layer) # (batch, time1, d_model)
|
||||
@@ -0,0 +1,33 @@
|
||||
#!/usr/bin/env python3
|
||||
# -*- coding: utf-8 -*-
|
||||
|
||||
|
||||
"""Positionwise feed forward layer definition."""
|
||||
|
||||
import torch
|
||||
|
||||
from funasr.models.transformer.layer_norm import LayerNorm
|
||||
|
||||
|
||||
class PositionwiseFeedForwardDecoderSANM(torch.nn.Module):
|
||||
"""Positionwise feed forward layer.
|
||||
|
||||
Args:
|
||||
idim (int): Input dimenstion.
|
||||
hidden_units (int): The number of hidden units.
|
||||
dropout_rate (float): Dropout rate.
|
||||
|
||||
"""
|
||||
|
||||
def __init__(self, idim, hidden_units, dropout_rate, adim=None, activation=torch.nn.ReLU()):
|
||||
"""Construct an PositionwiseFeedForward object."""
|
||||
super(PositionwiseFeedForwardDecoderSANM, self).__init__()
|
||||
self.w_1 = torch.nn.Linear(idim, hidden_units)
|
||||
self.w_2 = torch.nn.Linear(hidden_units, idim if adim is None else adim, bias=False)
|
||||
self.dropout = torch.nn.Dropout(dropout_rate)
|
||||
self.activation = activation
|
||||
self.norm = LayerNorm(hidden_units)
|
||||
|
||||
def forward(self, x):
|
||||
"""Forward function."""
|
||||
return self.w_2(self.norm(self.dropout(self.activation(self.w_1(x)))))
|
||||
@@ -0,0 +1,121 @@
|
||||
# This is an example that demonstrates how to configure a model file.
|
||||
# You can modify the configuration according to your own requirements.
|
||||
|
||||
# to print the register_table:
|
||||
# from funasr.register import tables
|
||||
# tables.print()
|
||||
|
||||
# network architecture
|
||||
model: SANM
|
||||
model_conf:
|
||||
ctc_weight: 0.0
|
||||
lsm_weight: 0.1
|
||||
length_normalized_loss: true
|
||||
|
||||
# encoder
|
||||
encoder: SANMEncoder
|
||||
encoder_conf:
|
||||
output_size: 512
|
||||
attention_heads: 4
|
||||
linear_units: 2048
|
||||
num_blocks: 50
|
||||
dropout_rate: 0.1
|
||||
positional_dropout_rate: 0.1
|
||||
attention_dropout_rate: 0.1
|
||||
input_layer: pe
|
||||
pos_enc_class: SinusoidalPositionEncoder
|
||||
normalize_before: true
|
||||
kernel_size: 11
|
||||
sanm_shfit: 0
|
||||
selfattention_layer_type: sanm
|
||||
|
||||
# decoder
|
||||
decoder: FsmnDecoder
|
||||
decoder_conf:
|
||||
attention_heads: 4
|
||||
linear_units: 2048
|
||||
num_blocks: 16
|
||||
dropout_rate: 0.1
|
||||
positional_dropout_rate: 0.1
|
||||
self_attention_dropout_rate: 0.1
|
||||
src_attention_dropout_rate: 0.1
|
||||
att_layer_num: 16
|
||||
kernel_size: 11
|
||||
sanm_shfit: 0
|
||||
|
||||
|
||||
|
||||
# frontend related
|
||||
frontend: WavFrontend
|
||||
frontend_conf:
|
||||
fs: 16000
|
||||
window: hamming
|
||||
n_mels: 80
|
||||
frame_length: 25
|
||||
frame_shift: 10
|
||||
lfr_m: 7
|
||||
lfr_n: 6
|
||||
|
||||
specaug: SpecAugLFR
|
||||
specaug_conf:
|
||||
apply_time_warp: false
|
||||
time_warp_window: 5
|
||||
time_warp_mode: bicubic
|
||||
apply_freq_mask: true
|
||||
freq_mask_width_range:
|
||||
- 0
|
||||
- 30
|
||||
lfr_rate: 6
|
||||
num_freq_mask: 1
|
||||
apply_time_mask: true
|
||||
time_mask_width_range:
|
||||
- 0
|
||||
- 12
|
||||
num_time_mask: 1
|
||||
|
||||
train_conf:
|
||||
accum_grad: 1
|
||||
grad_clip: 5
|
||||
max_epoch: 150
|
||||
val_scheduler_criterion:
|
||||
- valid
|
||||
- acc
|
||||
best_model_criterion:
|
||||
- - valid
|
||||
- acc
|
||||
- max
|
||||
keep_nbest_models: 10
|
||||
avg_nbest_model: 10
|
||||
log_interval: 50
|
||||
|
||||
optim: adam
|
||||
optim_conf:
|
||||
lr: 0.0005
|
||||
scheduler: warmuplr
|
||||
scheduler_conf:
|
||||
warmup_steps: 30000
|
||||
|
||||
dataset: AudioDataset
|
||||
dataset_conf:
|
||||
index_ds: IndexDSJsonl
|
||||
batch_sampler: BatchSampler
|
||||
batch_type: example # example or length
|
||||
batch_size: 1 # if batch_type is example, batch_size is the numbers of samples; if length, batch_size is source_token_len+target_token_len;
|
||||
max_token_length: 2048 # filter samples if source_token_len+target_token_len > max_token_length,
|
||||
buffer_size: 500
|
||||
shuffle: True
|
||||
num_workers: 0
|
||||
|
||||
tokenizer: CharTokenizer
|
||||
tokenizer_conf:
|
||||
unk_symbol: <unk>
|
||||
split_with_space: true
|
||||
|
||||
|
||||
ctc_conf:
|
||||
dropout_rate: 0.0
|
||||
ctc_type: builtin
|
||||
reduce: true
|
||||
ignore_nan_grad: true
|
||||
|
||||
normalize: null
|
||||
Reference in New Issue
Block a user