-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathinference.iclr.kg.py
181 lines (162 loc) · 7.39 KB
/
inference.iclr.kg.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
179
180
181
import os
import math
from pprint import PrettyPrinter
import random
import numpy as np
import torch # Torch must be imported before sklearn and tf
import sklearn
import tensorflow as tf
import better_exceptions
from tqdm import tqdm, trange
import colorlog
import colorful
from utils.etc_utils import set_logger, set_tcmalloc, set_gpus, check_none_gradients
from utils import config_utils, custom_argparsers
from models import MODELS
from modules.checkpoint_tracker import CheckpointTracker
from modules.trainer import run_wow_evaluation, Trainer
from modules.from_parlai import download_from_google_drive, unzip
from data.wizard_of_wikipedia import WowDatasetReader
from data.holle import HolleDatasetReader
from data.pchatkg import PchatKGDatasetReader
better_exceptions.hook()
_command_args = config_utils.CommandArgs()
pprint = PrettyPrinter().pprint
pformat = PrettyPrinter().pformat
BEST_N_CHECKPOINTS = 5
def main():
# Argument passing/parsing
args, model_args = config_utils.initialize_argparser(
MODELS, _command_args, custom_argparsers.DialogArgumentParser)
hparams, hparams_dict = config_utils.create_or_load_hparams(
args, model_args, args.cfg)
pprint(hparams_dict)
ckpt_fname = os.path.join('./', hparams.test_ckpt)
# if hparams.test_mode == 'wow':
# os.makedirs('./tmp', exist_ok=True)
# if not os.path.exists('tmp/wow_pretrained'):
# fname = 'wow_pretrained.zip'
# gd_id = '1lkF1QENr45j0vl-Oja3wEiqkxoNTxkXT'
# colorlog.info(f"Download pretrained checkpoint {fname}")
# download_from_google_drive(gd_id, os.path.join('tmp', fname))
# unzip('tmp', fname)
# ckpt_fname = os.path.join('tmp/wow_pretrained', 'ckpt-46070')
# elif hparams.test_mode == "holle_1":
# os.makedirs('./tmp', exist_ok=True)
# if not os.path.exists('tmp/holle_pretrained_1'):
# fname = 'holle_pretrained_1.zip'
# gd_id = '1o1-Gv5PScxlSzxW6DyZnSp3gDI5zXOhh'
# colorlog.info(f"Download pretrained checkpoint {fname}")
# download_from_google_drive(gd_id, os.path.join('tmp', fname))
# unzip('tmp', fname)
# ckpt_fname = os.path.join('tmp/holle_pretrained_1', 'ckpt-1th-best')
# elif hparams.test_mode == "holle_2":
# os.makedirs('./tmp', exist_ok=True)
# if not os.path.exists('tmp/holle_pretrained_2'):
# fname = 'holle_pretrained_2.zip'
# gd_id = '13FkCjuC0aBEenlSf-NAAgOfoWVPhqFSc'
# colorlog.info(f"Download pretrained checkpoint {fname}")
# download_from_google_drive(gd_id, os.path.join('tmp', fname))
# unzip('tmp', fname)
# ckpt_fname = os.path.join('tmp/holle_pretrained_2', 'ckpt-1th-best')
# else:
# raise ValueError("'wow' and 'holle' is currently supported")
# Set environment variables & gpus
set_logger()
set_gpus(hparams.gpus)
set_tcmalloc()
gpus = tf.config.experimental.list_physical_devices('GPU')
tf.config.experimental.set_visible_devices(gpus, 'GPU')
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
# Set random seed
tf.random.set_seed(hparams.random_seed)
np.random.seed(hparams.random_seed)
random.seed(hparams.random_seed)
# For multi-gpu
if hparams.num_gpus > 1:
mirrored_strategy = tf.distribute.MirroredStrategy() # NCCL will be used as default
else:
mirrored_strategy = None
# Download BERT pretrained model
if not os.path.exists(hparams.bert_dir):
os.makedirs(hparams.bert_dir)
fname = 'uncased_L-12_H-768_A-12.zip'
gd_id = '17rfV9CleFBwwfS7m5Yd72vvxdPLWBHl6'
download_from_google_drive(gd_id, os.path.join(hparams.bert_dir, fname))
unzip(hparams.bert_dir, fname)
# Make dataset reader
os.makedirs(hparams.cache_dir, exist_ok=True)
if hparams.data_name == 'wizard_of_wikipedia':
reader_cls = WowDatasetReader
elif hparams.data_name == 'holle':
reader_cls = HolleDatasetReader
elif hparams.data_name == "pchat":
reader_cls = PchatKGDatasetReader
else:
raise ValueError("data_name must be one of 'wizard_of_wikipedia' and 'holle'")
reader = reader_cls(
hparams.batch_size, hparams.num_epochs,
buffer_size=hparams.buffer_size,
bucket_width=hparams.bucket_width,
max_length=hparams.max_length,
max_episode_length=hparams.max_episode_length,
max_knowledge=hparams.max_knowledge,
knowledge_truncate=hparams.knowledge_truncate,
cache_dir=hparams.cache_dir,
# bert_dir=hparams.bert_dir,
bert_dir='bert_pretrained/chinese_L-12_H-768_A-12',
)
train_dataset, iters_in_train = reader.read('train', mirrored_strategy)
test_dataset, iters_in_test = reader.read('test', mirrored_strategy)
adj, unsupervise_train_dataset, iters_in_uns_train = reader.build_graph_dataset() # get adj_list and dataset
if hparams.data_name == 'wizard_of_wikipedia':
unseen_dataset, iters_in_unseen = reader.read('test_unseen', mirrored_strategy)
vocabulary = reader.vocabulary
# Build model & optimizer & trainer
if mirrored_strategy:
with mirrored_strategy.scope():
model = MODELS[hparams.model](hparams, vocabulary)
optimizer = tf.keras.optimizers.Adam(learning_rate=hparams.init_lr,
clipnorm=hparams.clipnorm)
else:
if hparams.model == 'SKT_KG':
model = MODELS[hparams.model](hparams, vocabulary, adj)
else:
model = MODELS[hparams.model](hparams, vocabulary)
optimizer = tf.keras.optimizers.Adam(learning_rate=hparams.init_lr,
clipnorm=hparams.clipnorm)
trainer = Trainer(model, optimizer, mirrored_strategy,
hparams.enable_function,
reader_cls.remove_pad)
# Setup checkpoint
global_step = tf.compat.v1.train.get_or_create_global_step()
checkpoint = tf.train.Checkpoint(optimizer=optimizer,
model=model,
optimizer_step=global_step)
# Load
train_example = next(iter(train_dataset))
_ = trainer.train_step(train_example)
#checkpoint.restore(ckpt_fname).assert_consumed()
#checkpoint.restore(ckpt_fname).expect_partial()
# a = model.layers[10].get_weights()
checkpoint.restore(ckpt_fname)
# b = model.layers[10].get_weights()
# Test
test_loop_outputs = trainer.test_loop(test_dataset, iters_in_test, 0, 'seen')
if hparams.data_name == 'wizard_of_wikipedia':
unseen_loop_outputs = trainer.test_loop(unseen_dataset, iters_in_unseen, 0, 'unseen')
test_summaries, log_dict = run_wow_evaluation(
test_loop_outputs, hparams.test_ckpt, 'test')
# test_loop_outputs, hparams.checkpoint_dir, 'seen')
if hparams.data_name == 'wizard_of_wikipedia':
unseen_summaries, unseen_log_dict = run_wow_evaluation(
unseen_loop_outputs, hparams.checkpoint_dir, 'unseen')
# Logging
tqdm.write(colorful.bold_green("seen").styled_string)
tqdm.write(colorful.bold_red(pformat(log_dict)).styled_string)
if hparams.data_name == 'wizard_of_wikipedia':
tqdm.write(colorful.bold_green("unseen").styled_string)
tqdm.write(colorful.bold_red(pformat(unseen_log_dict)).styled_string)
if __name__ == '__main__':
main()