File size: 7,821 Bytes
d596074 |
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 |
# Copyright (c) 2022 Xiaomi Corporation (authors: Xiaoyu Yang)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
# 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.
import argparse
import logging
import torch
from icefall.checkpoint import average_checkpoints, load_checkpoint
from icefall.rnn_lm.model import RnnLmModel
from icefall.transformer_lm.model import TransformerLM
from icefall.utils import AttributeDict, str2bool
class LmScorer(torch.nn.Module):
"""This is a wrapper for NN LMs
The language models supported include:
RNN,
Transformer
"""
def __init__(
self,
lm_type: str,
params: AttributeDict,
device,
lm_scale: float = 0.3,
):
super(LmScorer, self).__init__()
assert lm_type in ["rnn", "transformer"], f"{lm_type} is not supported"
self.lm_type = lm_type
self.lm = self.get_lm(lm_type, device, params)
self.lm_scale = lm_scale
self.params = params
@classmethod
def add_arguments(cls, parser):
# LM general arguments
parser.add_argument(
"--lm-vocab-size",
type=int,
default=500,
)
parser.add_argument(
"--lm-epoch",
type=int,
default=7,
help="""Which epoch to be used
""",
)
parser.add_argument(
"--lm-avg",
type=int,
default=1,
help="""Number of checkpoints to be averaged
""",
)
parser.add_argument("--lm-exp-dir", type=str, help="Path to LM experiments")
# Now RNNLM related arguments
parser.add_argument(
"--rnn-lm-embedding-dim",
type=int,
default=2048,
help="Embedding dim of the model",
)
parser.add_argument(
"--rnn-lm-hidden-dim",
type=int,
default=2048,
help="Hidden dim of the model",
)
parser.add_argument(
"--rnn-lm-num-layers",
type=int,
default=3,
help="Number of RNN layers the model",
)
parser.add_argument(
"--rnn-lm-tie-weights",
type=str2bool,
default=True,
help="""True to share the weights between the input embedding layer and the
last output linear layer
""",
)
# Now transformers
parser.add_argument(
"--transformer-lm-exp-dir", type=str, help="Directory of transformer LM exp"
)
parser.add_argument(
"--transformer-lm-dim-feedforward",
type=int,
default=2048,
help="Dimension of FFW module in transformer",
)
parser.add_argument(
"--transformer-lm-encoder-dim",
type=int,
default=768,
help="Encoder dimension of transformer",
)
parser.add_argument(
"--transformer-lm-embedding-dim",
type=int,
default=768,
help="Input embedding dimension of transformer",
)
parser.add_argument(
"--transformer-lm-nhead",
type=int,
default=8,
help="Number of attention heads in transformer",
)
parser.add_argument(
"--transformer-lm-num-layers",
type=int,
default=16,
help="Number of encoder layers in transformer",
)
parser.add_argument(
"--transformer-lm-tie-weights",
type=str2bool,
default=True,
help="If tie weights in transformer LM",
)
def get_lm(self, lm_type: str, device, params: AttributeDict) -> torch.nn.Module:
"""Return the neural network LM
Args:
lm_type (str): Type name of NN LM
"""
if lm_type == "rnn":
model = RnnLmModel(
vocab_size=params.lm_vocab_size,
embedding_dim=params.rnn_lm_embedding_dim,
hidden_dim=params.rnn_lm_hidden_dim,
num_layers=params.rnn_lm_num_layers,
tie_weights=params.rnn_lm_tie_weights,
)
if params.lm_avg == 1:
load_checkpoint(
f"{params.lm_exp_dir}/epoch-{params.lm_epoch}.pt", model
)
model.to(device)
else:
start = params.lm_epoch - params.lm_avg + 1
filenames = []
for i in range(start, params.lm_epoch + 1):
if start >= 0:
filenames.append(f"{params.lm_exp_dir}/epoch-{i}.pt")
logging.info(f"averaging {filenames}")
model.to(device)
model.load_state_dict(average_checkpoints(filenames, device=device))
elif lm_type == "transformer":
model = TransformerLM(
vocab_size=params.lm_vocab_size,
d_model=params.transformer_lm_encoder_dim,
embedding_dim=params.transformer_lm_embedding_dim,
dim_feedforward=params.transformer_lm_dim_feedforward,
nhead=params.transformer_lm_nhead,
num_layers=params.transformer_lm_num_layers,
tie_weights=params.transformer_lm_tie_weights,
params=params,
)
if params.lm_avg == 1:
load_checkpoint(
f"{params.lm_exp_dir}/epoch-{params.lm_epoch}.pt", model
)
model.to(device)
else:
start = params.lm_epoch - params.lm_avg + 1
filenames = []
for i in range(start, params.lm_epoch + 1):
if start >= 0:
filenames.append(f"{params.lm_exp_dir}/epoch-{i}.pt")
logging.info(f"averaging {filenames}")
model.to(device)
model.load_state_dict(average_checkpoints(filenames, device=device))
else:
raise NotImplementedError()
return model
def score_token(self, x: torch.Tensor, x_lens: torch.Tensor, state=None):
"""Score the input and return the prediction
This requires the lm to have the method `score_token`
Args:
x (torch.Tensor): Input tokens
x_lens (torch.Tensor): Length of the input tokens
state (optional): LM states
"""
return self.lm.score_token(x, x_lens, state)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
LmScorer.add_arguments(parser)
args = parser.parse_args()
params = AttributeDict()
params.update(vars(args))
device = torch.device("cpu")
if torch.cuda.is_available():
device = torch.device("cuda", 0)
Scorer = LmScorer(params=params, device=device)
Scorer.eval()
x = (
torch.tensor([[1, 4, 19, 256, 77], [1, 4, 19, 256, 77]])
.to(device)
.to(torch.int64)
)
x_lens = torch.tensor([5, 5]).to(device)
state = None
score, state = Scorer.score(x, x_lens)
print(score.shape)
print(score[0])
print(score[1])
|