Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[transformer] remove concat after to simplify the code flow #1762

Merged
merged 1 commit into from
Mar 20, 2023
Merged

[transformer] remove concat after to simplify the code flow #1762

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 2 additions & 11 deletions wenet/transformer/decoder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,9 +42,6 @@ class TransformerDecoder(torch.nn.Module):
normalize_before:
True: use layer_norm before each sub-block of a layer.
False: use layer_norm after each sub-block of a layer.
concat_after: whether to concat attention layer's input and output
True: x -> x + linear(concat(x, att(x)))
False: x -> x + att(x)
"""
def __init__(
self,
Expand All @@ -60,7 +57,6 @@ def __init__(
input_layer: str = "embed",
use_output_layer: bool = True,
normalize_before: bool = True,
concat_after: bool = False,
):
assert check_argument_types()
super().__init__()
Expand Down Expand Up @@ -90,7 +86,6 @@ def __init__(
dropout_rate),
dropout_rate,
normalize_before,
concat_after,
) for _ in range(self.num_blocks)
])

Expand Down Expand Up @@ -202,9 +197,6 @@ class BiTransformerDecoder(torch.nn.Module):
normalize_before:
True: use layer_norm before each sub-block of a layer.
False: use layer_norm after each sub-block of a layer.
concat_after: whether to concat attention layer's input and output
True: x -> x + linear(concat(x, att(x)))
False: x -> x + att(x)
"""
def __init__(
self,
Expand All @@ -221,7 +213,6 @@ def __init__(
input_layer: str = "embed",
use_output_layer: bool = True,
normalize_before: bool = True,
concat_after: bool = False,
):

assert check_argument_types()
Expand All @@ -230,13 +221,13 @@ def __init__(
vocab_size, encoder_output_size, attention_heads, linear_units,
num_blocks, dropout_rate, positional_dropout_rate,
self_attention_dropout_rate, src_attention_dropout_rate,
input_layer, use_output_layer, normalize_before, concat_after)
input_layer, use_output_layer, normalize_before)

self.right_decoder = TransformerDecoder(
vocab_size, encoder_output_size, attention_heads, linear_units,
r_num_blocks, dropout_rate, positional_dropout_rate,
self_attention_dropout_rate, src_attention_dropout_rate,
input_layer, use_output_layer, normalize_before, concat_after)
input_layer, use_output_layer, normalize_before)

def forward(
self,
Expand Down
30 changes: 4 additions & 26 deletions wenet/transformer/decoder_layer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,10 +35,6 @@ class DecoderLayer(nn.Module):
normalize_before (bool):
True: use layer_norm before each sub-block.
False: to use layer_norm after each sub-block.
concat_after (bool): Whether to concat attention layer's inpu
and output.
True: x -> x + linear(concat(x, att(x)))
False: x -> x + att(x)
"""
def __init__(
self,
Expand All @@ -48,7 +44,6 @@ def __init__(
feed_forward: nn.Module,
dropout_rate: float,
normalize_before: bool = True,
concat_after: bool = False,
):
"""Construct an DecoderLayer object."""
super().__init__()
Expand All @@ -61,13 +56,6 @@ def __init__(
self.norm3 = nn.LayerNorm(size, eps=1e-5)
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)
else:
self.concat_linear1 = nn.Identity()
self.concat_linear2 = nn.Identity()

def forward(
self,
Expand Down Expand Up @@ -115,26 +103,16 @@ def forward(
residual = residual[:, -1:, :]
tgt_q_mask = tgt_mask[:, -1:, :]

if self.concat_after:
tgt_concat = torch.cat(
(tgt_q, self.self_attn(tgt_q, tgt, tgt, tgt_q_mask)[0]), dim=-1)
x = residual + self.concat_linear1(tgt_concat)
else:
x = residual + self.dropout(
self.self_attn(tgt_q, tgt, tgt, tgt_q_mask)[0])
x = residual + self.dropout(
self.self_attn(tgt_q, tgt, tgt, tgt_q_mask)[0])
if not self.normalize_before:
x = self.norm1(x)

residual = x
if self.normalize_before:
x = self.norm2(x)
if self.concat_after:
x_concat = torch.cat(
(x, self.src_attn(x, memory, memory, memory_mask)[0]), dim=-1)
x = residual + self.concat_linear2(x_concat)
else:
x = residual + self.dropout(
self.src_attn(x, memory, memory, memory_mask)[0])
x = residual + self.dropout(
self.src_attn(x, memory, memory, memory_mask)[0])
if not self.normalize_before:
x = self.norm2(x)

Expand Down
14 changes: 3 additions & 11 deletions wenet/transformer/encoder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,7 +52,6 @@ def __init__(
input_layer: str = "conv2d",
pos_enc_layer_type: str = "abs_pos",
normalize_before: bool = True,
concat_after: bool = False,
static_chunk_size: int = 0,
use_dynamic_chunk: bool = False,
global_cmvn: torch.nn.Module = None,
Expand All @@ -77,10 +76,6 @@ def __init__(
normalize_before (bool):
True: use layer_norm before each sub-block of a layer.
False: use layer_norm after each sub-block of a layer.
concat_after (bool): whether to concat attention layer's input
and output.
True: x -> x + linear(concat(x, att(x)))
False: x -> x + att(x)
static_chunk_size (int): chunk size for static chunk training and
decoding
use_dynamic_chunk (bool): whether use dynamic chunk size for
Expand Down Expand Up @@ -341,7 +336,6 @@ def __init__(
input_layer: str = "conv2d",
pos_enc_layer_type: str = "abs_pos",
normalize_before: bool = True,
concat_after: bool = False,
static_chunk_size: int = 0,
use_dynamic_chunk: bool = False,
global_cmvn: torch.nn.Module = None,
Expand All @@ -356,7 +350,7 @@ def __init__(
linear_units, num_blocks, dropout_rate,
positional_dropout_rate, attention_dropout_rate,
input_layer, pos_enc_layer_type, normalize_before,
concat_after, static_chunk_size, use_dynamic_chunk,
static_chunk_size, use_dynamic_chunk,
global_cmvn, use_dynamic_left_chunk)
self.encoders = torch.nn.ModuleList([
TransformerEncoderLayer(
Expand All @@ -365,7 +359,7 @@ def __init__(
attention_dropout_rate),
PositionwiseFeedForward(output_size, linear_units,
dropout_rate), dropout_rate,
normalize_before, concat_after) for _ in range(num_blocks)
normalize_before) for _ in range(num_blocks)
])


Expand All @@ -384,7 +378,6 @@ def __init__(
input_layer: str = "conv2d",
pos_enc_layer_type: str = "rel_pos",
normalize_before: bool = True,
concat_after: bool = False,
static_chunk_size: int = 0,
use_dynamic_chunk: bool = False,
global_cmvn: torch.nn.Module = None,
Expand Down Expand Up @@ -419,7 +412,7 @@ def __init__(
linear_units, num_blocks, dropout_rate,
positional_dropout_rate, attention_dropout_rate,
input_layer, pos_enc_layer_type, normalize_before,
concat_after, static_chunk_size, use_dynamic_chunk,
static_chunk_size, use_dynamic_chunk,
global_cmvn, use_dynamic_left_chunk)
activation = get_activation(activation_type)

Expand Down Expand Up @@ -457,6 +450,5 @@ def __init__(
*convolution_layer_args) if use_cnn_module else None,
dropout_rate,
normalize_before,
concat_after,
) for _ in range(num_blocks)
])
35 changes: 2 additions & 33 deletions wenet/transformer/encoder_layer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,11 +36,6 @@ class TransformerEncoderLayer(nn.Module):
normalize_before (bool):
True: use layer_norm before each sub-block.
False: to use layer_norm after each sub-block.
concat_after (bool): Whether to concat attention layer's input and
output.
True: x -> x + linear(concat(x, att(x)))
False: x -> x + att(x)

"""
def __init__(
self,
Expand All @@ -49,7 +44,6 @@ def __init__(
feed_forward: torch.nn.Module,
dropout_rate: float,
normalize_before: bool = True,
concat_after: bool = False,
):
"""Construct an EncoderLayer object."""
super().__init__()
Expand All @@ -60,11 +54,6 @@ def __init__(
self.dropout = nn.Dropout(dropout_rate)
self.size = size
self.normalize_before = normalize_before
self.concat_after = concat_after
if concat_after:
self.concat_linear = nn.Linear(size + size, size)
else:
self.concat_linear = nn.Identity()

def forward(
self,
Expand Down Expand Up @@ -101,14 +90,9 @@ def forward(
residual = x
if self.normalize_before:
x = self.norm1(x)

x_att, new_att_cache = self.self_attn(
x, x, x, mask, cache=att_cache)
if self.concat_after:
x_concat = torch.cat((x, x_att), dim=-1)
x = residual + self.concat_linear(x_concat)
else:
x = residual + self.dropout(x_att)
x = residual + self.dropout(x_att)
if not self.normalize_before:
x = self.norm1(x)

Expand Down Expand Up @@ -141,10 +125,6 @@ class ConformerEncoderLayer(nn.Module):
normalize_before (bool):
True: use layer_norm before each sub-block.
False: use layer_norm after each sub-block.
concat_after (bool): Whether to concat attention layer's input and
output.
True: x -> x + linear(concat(x, att(x)))
False: x -> x + att(x)
"""
def __init__(
self,
Expand All @@ -155,7 +135,6 @@ def __init__(
conv_module: Optional[nn.Module] = None,
dropout_rate: float = 0.1,
normalize_before: bool = True,
concat_after: bool = False,
):
"""Construct an EncoderLayer object."""
super().__init__()
Expand All @@ -178,11 +157,6 @@ def __init__(
self.dropout = nn.Dropout(dropout_rate)
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)
else:
self.concat_linear = nn.Identity()


def forward(
Expand Down Expand Up @@ -230,14 +204,9 @@ def forward(
residual = x
if self.normalize_before:
x = self.norm_mha(x)

x_att, new_att_cache = self.self_attn(
x, x, x, mask, pos_emb, att_cache)
if self.concat_after:
x_concat = torch.cat((x, x_att), dim=-1)
x = residual + self.concat_linear(x_concat)
else:
x = residual + self.dropout(x_att)
x = residual + self.dropout(x_att)
if not self.normalize_before:
x = self.norm_mha(x)

Expand Down