-
Notifications
You must be signed in to change notification settings - Fork 2.2k
/
lm_tutorial.py
178 lines (159 loc) · 6.26 KB
/
lm_tutorial.py
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
# Copyright (c) 2019 NVIDIA Corporation
import nemo
import nemo_nlp
import math
import logging
from nemo.utils.lr_policies import CosineAnnealing
from nemo_nlp.callbacks.language_modeling import eval_iter_callback, \
eval_epochs_done_callback
console = logging.StreamHandler()
console.setLevel(logging.INFO)
# add the handler to the root logger
logging.getLogger('').addHandler(console)
parser = nemo.utils.NemoArgParser(description='Transformer for WT-103 LM')
parser.set_defaults(
train_dataset="train.txt",
eval_datasets=["valid.txt"],
checkpoint_dir="transformer_lm",
optimizer="novograd",
num_epochs=1000,
batch_size=32,
eval_batch_size=32,
lr=0.002,
beta1=0.95,
beta2=0.25,
weight_decay=0,
warmup_steps=1000,
max_steps=50000,
iter_per_step=1,
checkpoint_save_freq=10000,
eval_freq=1000
)
parser.add_argument("--data_root", default="../../tests/data/wikitext-2/",
type=str)
parser.add_argument("--d_model", default=384, type=int)
parser.add_argument("--d_inner", default=1536, type=int)
parser.add_argument("--num_layers", default=12, type=int)
parser.add_argument("--num_attn_heads", default=6, type=int)
parser.add_argument("--embedding_dropout", default=0.2, type=float)
parser.add_argument("--ffn_dropout", default=0.2, type=float)
parser.add_argument("--attn_score_dropout", default=0.2, type=float)
parser.add_argument("--attn_layer_dropout", default=0.2, type=float)
parser.add_argument("--max_sequence_length", default=256, type=int)
parser.add_argument("--label_smoothing", default=0.1, type=float)
parser.add_argument("--beam_size", default=4, type=int)
parser.add_argument("--tokenizer_model", default="vocab.txt", type=str)
parser.add_argument("--predict_last_k", default=16, type=int)
parser.add_argument("--interactive", action="store_true")
args = parser.parse_args()
# create TensorboardX logger to log training statistics
name = f"transformer-lm-lr_{args.lr}-optim_{args.optimizer}-" \
f"warmup_{args.warmup_steps}-bs_{args.batch_size}"
tb_writer = None # SummaryWriter(name)
# instantiate Neural Factory with supported backend
device = nemo.core.DeviceType.AllGpu if args.local_rank is not None \
else nemo.core.DeviceType.GPU
neural_factory = nemo.core.NeuralModuleFactory(
backend=nemo.core.Backend.PyTorch,
local_rank=args.local_rank,
optimization_level=nemo.core.Optimization.mxprO2,
placement=device)
# define tokenizer, in this example we use word-level tokenizer
# we also adjust the vocabulary size to make it multiple of 8 to accelerate
# training in fp16 mode with the use of Tensor Cores
tokenizer = nemo_nlp.WordTokenizer(f"{args.data_root}/{args.tokenizer_model}")
vocab_size = 8 * math.ceil(tokenizer.vocab_size / 8)
# instantiate necessary modules for the whole translation pipeline, namely
# data layers, encoder, decoder, output log_softmax, beam_search_translator
# and loss function
train_data_layer = nemo_nlp.LanguageModelingDataLayer(
factory=neural_factory,
tokenizer=tokenizer,
dataset=f"{args.data_root}/{args.train_dataset}",
max_seq_length=args.max_sequence_length,
batch_size=args.batch_size,
batch_step=args.max_sequence_length)
eval_data_layer = nemo_nlp.LanguageModelingDataLayer(
factory=neural_factory,
tokenizer=tokenizer,
dataset=f"{args.data_root}/{args.eval_datasets[0]}",
max_seq_length=args.max_sequence_length,
batch_size=args.batch_size,
batch_step=args.predict_last_k)
encoder = nemo_nlp.TransformerEncoderNM(
factory=neural_factory,
d_model=args.d_model,
d_inner=args.d_inner,
num_layers=args.num_layers,
embedding_dropout=args.embedding_dropout,
num_attn_heads=args.num_attn_heads,
ffn_dropout=args.ffn_dropout,
vocab_size=vocab_size,
mask_future=True,
attn_score_dropout=args.attn_score_dropout,
attn_layer_dropout=args.attn_layer_dropout,
max_seq_length=args.max_sequence_length)
log_softmax = nemo_nlp.TransformerLogSoftmaxNM(
factory=neural_factory,
vocab_size=vocab_size,
d_model=args.d_model)
loss = nemo_nlp.PaddedSmoothedCrossEntropyLossNM(
factory=neural_factory,
pad_id=tokenizer.pad_id(),
label_smoothing=args.label_smoothing)
# tie weight of embedding and log_softmax layers
log_softmax.log_softmax.dense.weight = \
encoder.embedding_layer.token_embedding.weight
# training pipeline
src, src_mask, labels = train_data_layer()
src_hiddens = encoder(input_ids=src, input_mask_src=src_mask)
log_probs = log_softmax(hidden_states=src_hiddens)
train_loss = loss(log_probs=log_probs, target_ids=labels)
# evaluation pipeline
src_, src_mask_, labels_ = eval_data_layer()
src_hiddens_ = encoder(input_ids=src_, input_mask_src=src_mask_)
log_probs_ = log_softmax(hidden_states=src_hiddens_)
eval_loss = loss(log_probs=log_probs_, target_ids=labels_)
def print_loss(x):
loss = str(x[0].item())
print(f"Training loss: {loss}")
# callback which prints training loss once in a while
callback_train = nemo.core.SimpleLossLoggerCallback(
tensors=[train_loss],
step_freq=100,
print_func=print_loss,
get_tb_values=lambda x: [["loss", x[0]]],
tb_writer=tb_writer)
# callback which calculates evaluation loss
callback_eval = nemo.core.EvaluatorCallback(
eval_tensors=[eval_loss],
user_iter_callback=eval_iter_callback,
user_epochs_done_callback=eval_epochs_done_callback,
eval_step=args.eval_freq,
tb_writer=tb_writer)
# callback which saves checkpoints once in a while
callback_ckpt = nemo.core.CheckpointCallback(
folder=args.checkpoint_dir,
step_freq=args.checkpoint_save_freq,
checkpoints_to_keep=-1)
# define learning rate decay policy
lr_policy = CosineAnnealing(args.max_steps, warmup_steps=args.warmup_steps)
# define and launch training algorithm (optimizer)
max_num_epochs = 0 if args.interactive else args.num_epochs
optimizer = neural_factory.get_trainer()
callbacks = [callback_ckpt]
if not args.interactive:
callbacks.extend([callback_train, callback_eval])
optimizer.train(
tensors_to_optimize=[train_loss],
callbacks=callbacks,
optimizer=args.optimizer,
lr_policy=lr_policy,
optimization_params={
"num_epochs": max_num_epochs,
"lr": args.lr,
"weight_decay": args.weight_decay,
"betas": (args.beta1, args.beta2)
},
batches_per_step=args.iter_per_step
)