File size: 10,421 Bytes
85ba398 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 |
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import copy
import logging
from fairseq.models import (
FairseqEncoder,
FairseqEncoderModel,
FairseqLanguageModel,
register_model,
register_model_architecture,
)
from fairseq.models.speech_to_speech.modules.ctc_decoder import CTCDecoder
from fairseq.models.speech_to_speech.modules.stacked_embedding import StackedEmbedding
from fairseq.models.speech_to_speech.modules.transformer_decoder_aug import (
AugTransformerUnitDecoder,
)
from fairseq.models.speech_to_speech.modules.transformer_encoder import (
TransformerEncoderNoEmb,
)
from fairseq.models.speech_to_speech.s2s_conformer import S2UTConformerModel
from fairseq.models.speech_to_speech.s2s_transformer import (
TransformerUnitDecoder,
base_multitask_text_transformer_decoder_arch,
s2ut_architecture_base,
)
from fairseq.models.transformer import TransformerDecoder, TransformerModelBase
logger = logging.getLogger(__name__)
def multitask_text_transformer_decoder_arch(
args, decoder_layers, decoder_embed_dim=256, decoder_attention_heads=4
):
args.decoder_layers = decoder_layers
args.decoder_embed_dim = decoder_embed_dim
args.decoder_attention_heads = decoder_attention_heads
base_multitask_text_transformer_decoder_arch(args)
@register_model("unity_conformer")
class UnityConformerModel(S2UTConformerModel):
"""
Direct speech-to-speech translation model with Conformer encoder + MT Transformer decoder + Transformer discrete unit decoder
"""
@staticmethod
def add_args(parser):
S2UTConformerModel.add_args(parser)
parser.add_argument(
"--translation-decoder-layers",
type=int,
default=4,
metavar="N",
help="num decoder layers in the first-pass translation module",
)
parser.add_argument(
"--synthesizer",
default="transformer",
choices=["transformer"],
help="",
)
parser.add_argument(
"--synthesizer-encoder-layers",
type=int,
default=0,
metavar="N",
help="num encoder layers in the second-pass synthesizer module",
)
parser.add_argument(
"--synthesizer-augmented-cross-attention",
action="store_true",
default=False,
help="augmented cross-attention over speech encoder output",
)
@classmethod
def build_multitask_decoder(
cls,
args,
tgt_dict,
in_dim,
is_first_pass_decoder,
decoder_layers,
decoder_embed_dim,
decoder_attention_heads,
):
decoder_args = args.decoder_args
decoder_args.encoder_embed_dim = in_dim
if args.decoder_type == "transformer":
if is_first_pass_decoder:
multitask_text_transformer_decoder_arch(
decoder_args,
decoder_layers,
decoder_embed_dim,
decoder_attention_heads,
) # 4L
else:
base_multitask_text_transformer_decoder_arch(decoder_args) # 2L
task_decoder = TransformerDecoder(
decoder_args,
tgt_dict,
embed_tokens=TransformerModelBase.build_embedding(
decoder_args,
tgt_dict,
decoder_args.decoder_embed_dim,
),
)
elif args.decoder_type == "ctc":
task_decoder = CTCDecoder(
dictionary=tgt_dict,
in_dim=in_dim,
)
else:
raise NotImplementedError(
"currently only support multitask decoder_type 'transformer', 'ctc'"
)
return task_decoder
@classmethod
def build_decoder(cls, args, tgt_dict, aug_attn=False):
num_embeddings = len(tgt_dict)
padding_idx = tgt_dict.pad()
embed_tokens = StackedEmbedding(
num_embeddings,
args.decoder_embed_dim,
padding_idx,
num_stacked=args.n_frames_per_step,
)
_args = copy.deepcopy(args)
_args.encoder_embed_dim = args.decoder_embed_dim
decoder_cls = AugTransformerUnitDecoder if aug_attn else TransformerUnitDecoder
return decoder_cls(
_args,
tgt_dict,
embed_tokens,
)
@classmethod
def build_model(cls, args, task):
encoder = cls.build_encoder(args)
decoder = cls.build_decoder(
args,
task.target_dictionary,
aug_attn=getattr(args, "synthesizer_augmented_cross_attention", False),
)
base_model = cls(encoder, decoder)
base_model.t2u_augmented_cross_attn = getattr(
args, "synthesizer_augmented_cross_attention", False
)
# set up multitask decoders
base_model.mt_task_name = None
base_model.multitask_decoders = {}
has_first_pass_decoder = False
for task_name, task_obj in task.multitask_tasks.items():
if task_obj.is_first_pass_decoder:
has_first_pass_decoder = True
base_model.mt_task_name = task_name
in_dim = (
args.encoder_embed_dim
if task_obj.args.input_from == "encoder"
else args.decoder_embed_dim
)
task_decoder = cls.build_multitask_decoder(
task_obj.args,
task_obj.target_dictionary,
in_dim,
task_obj.is_first_pass_decoder,
getattr(args, "translation_decoder_layers", 4),
getattr(args, "decoder_embed_dim", 256),
getattr(args, "decoder_attention_heads", 4),
)
setattr(base_model, f"{task_name}_decoder", task_decoder)
decoder_model_cls = (
FairseqEncoderModel
if task_obj.args.decoder_type == "ctc"
else FairseqLanguageModel
)
base_model.multitask_decoders[task_name] = decoder_model_cls(
getattr(base_model, f"{task_name}_decoder")
)
assert has_first_pass_decoder, "set at least one intermediate non-CTC decoder"
# set up encoder on top of the auxiliary MT decoder
if getattr(args, "synthesizer_encoder_layers", 0) > 0:
base_model.synthesizer_encoder = cls.build_text_encoder(args)
else:
base_model.synthesizer_encoder = None
return base_model
@classmethod
def build_text_encoder(cls, args):
_args = copy.deepcopy(args)
_args.encoder_layers = args.synthesizer_encoder_layers
_args.encoder_embed_dim = args.decoder_embed_dim
_args.encoder_ffn_embed_dim = args.decoder_ffn_embed_dim
_args.encoder_attention_heads = args.decoder_attention_heads
_args.encoder_normalize_before = True
return TransformerEncoderNoEmb(_args)
def forward(
self,
src_tokens,
src_lengths,
prev_output_tokens,
prev_output_tokens_mt,
tgt_speaker=None,
return_all_hiddens=False,
):
mt_decoder = getattr(self, f"{self.mt_task_name}_decoder")
encoder_out = self.encoder(
src_tokens,
src_lengths=src_lengths,
tgt_speaker=tgt_speaker,
return_all_hiddens=return_all_hiddens,
)
# 1. MT decoder
mt_decoder_out = mt_decoder(
prev_output_tokens_mt,
encoder_out=encoder_out,
)
x = mt_decoder_out[1]["inner_states"][-1]
if mt_decoder.layer_norm is not None:
x = mt_decoder.layer_norm(x)
mt_decoder_padding_mask = None
if prev_output_tokens_mt.eq(mt_decoder.padding_idx).any():
mt_decoder_padding_mask = prev_output_tokens_mt.eq(mt_decoder.padding_idx)
# 2. T2U encoder
if self.synthesizer_encoder is not None:
t2u_encoder_out = self.synthesizer_encoder(
x,
mt_decoder_padding_mask,
return_all_hiddens=return_all_hiddens,
)
else:
t2u_encoder_out = {
"encoder_out": [x], # T x B x C
"encoder_padding_mask": [mt_decoder_padding_mask], # B x T
}
# 3. T2U decoder
if self.t2u_augmented_cross_attn:
decoder_out = self.decoder(
prev_output_tokens,
encoder_out=encoder_out,
encoder_out_aug=t2u_encoder_out,
)
else:
decoder_out = self.decoder(
prev_output_tokens,
encoder_out=t2u_encoder_out,
)
if return_all_hiddens:
decoder_out[-1]["encoder_states"] = encoder_out["encoder_states"]
decoder_out[-1]["encoder_padding_mask"] = encoder_out[
"encoder_padding_mask"
]
decoder_out[-1]["mt_decoder_out"] = mt_decoder_out
return decoder_out
@register_model_architecture(model_name="unity_conformer", arch_name="unity_conformer")
def unity_conformer_architecture_base(args):
args.conv_version = getattr(args, "conv_version", "convtransformer")
args.attn_type = getattr(args, "attn_type", None)
args.pos_enc_type = getattr(args, "pos_enc_type", "abs")
args.max_source_positions = getattr(args, "max_source_positions", 6000)
args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 256)
args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048)
args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 4)
args.dropout = getattr(args, "dropout", 0.1)
args.encoder_layers = getattr(args, "encoder_layers", 16)
args.depthwise_conv_kernel_size = getattr(args, "depthwise_conv_kernel_size", 31)
s2ut_architecture_base(args)
# for old naming
@register_model_architecture(
model_name="unity_conformer", arch_name="s2ut_conformer_translatotron2"
)
def unity_conformer_architecture_base_legacy(args):
unity_conformer_architecture_base(args)
|