forked from dheerajrajagopal/SelfExplain
-
Notifications
You must be signed in to change notification settings - Fork 1
/
run_baseline.py
executable file
·187 lines (152 loc) · 7.25 KB
/
run_baseline.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
182
183
184
185
186
187
import pytorch_lightning
from pytorch_lightning.callbacks import ModelCheckpoint, LearningRateMonitor
import random
import numpy as np
import pytorch_lightning as pl
import logging
import os
from argparse import ArgumentParser
import resource
from model.data import ClassificationData
from pytorch_lightning.loggers import TensorBoardLogger
from interpret_bangor import SwitchLMForEval
from model.SwitchLM import SwitchLM
#from model.SwitchLMSpeaker import SwitchLMSpeaker
def get_train_steps(dm):
total_devices = args.num_gpus * args.num_nodes
train_batches = len(dm.train_dataloader()) // total_devices
return (args.max_epochs * train_batches) // args.accumulate_grad_batches
rlimit = resource.getrlimit(resource.RLIMIT_NOFILE)
resource.setrlimit(resource.RLIMIT_NOFILE, (4096, rlimit[1]))
# init: important to make sure every node initializes the same weights
# argparser
parser = ArgumentParser()
parser.add_argument('--num_gpus', type=int)
parser.add_argument('--num_workers', type=int, default=4)
parser.add_argument('--batch_size', type=int, default=16)
parser.add_argument('--clip_grad', type=float, default=1.0)
parser.add_argument("--dataset_basedir", help="Base directory where the dataset is located.", type=str)
parser.add_argument("--model_name", default='xlm-roberta-base', help="Model to use.")
parser.add_argument("--use_speaker_tokens", action='store_true', help="prepend [SPK1], etc to each utterance")
parser.add_argument("--use_full_context", action='store_true', help="early fusion of speaker context")
parser.add_argument("--context_size", default=1,type=int, help="number of sentences prior for context")
parser.add_argument("--monitor", default="acc",type=str, help="monitor loss, acc, or f1")
parser.add_argument("--seed", default=18,type=int, help="seed for model run")
parser.add_argument("--control", action='store_true', help="load control data for finetuning")
parser.add_argument("--ckpt", default="",type=str, help="ckpt from which to load and finetune a model")
parser.add_argument("--tensorboard_dir", default="tb_logs_baseline",type=str, help="tensorboard directory for logs")
parser.add_argument("--use_speaker_descriptions", action='store_true', help='prepend a sentence of speaker context to the previous and current utterances')
parser = pl.Trainer.add_argparse_args(parser)
parser = SwitchLM.add_model_specific_args(parser)
args = parser.parse_args()
SEED = args.seed
np.random.seed(SEED)
random.seed(SEED)
pl.utilities.seed.seed_everything(SEED)
pytorch_lightning.seed_everything(SEED)
args.num_gpus = len(str(args.gpus).split(","))
logging.basicConfig(level=logging.INFO)
balanced = args.balanced
use_speaker_descriptions = args.use_speaker_descriptions
eval_sep_languages = args.eval_per_language
# Step 1: Init Data
logging.info("Loading the data module for Bangor")
# dm = ClassificationData(basedir=args.dataset_basedir, tokenizer_name=args.model_name,\
# batch_size=args.batch_size, codeswitch=True, num_workers=args.num_workers, balanced=balanced,\
# use_speaker_descriptions=use_speaker_descriptions,\
# get_lang_feats=eval_sep_languages, use_full_context=args.use_full_context)
dm = ClassificationData(basedir=args.dataset_basedir, tokenizer_name=args.model_name, context_size=args.context_size,load_description_data=False, \
batch_size=args.batch_size, codeswitch=True, num_workers=args.num_workers, balanced=balanced, use_speaker_tokens=args.use_speaker_tokens, \
load_control_data=args.control, load_full_control=False, do_social_predictions=args.speaker_trait_predictions, full_mtl_setup=args.full_mtl_setup)
# Step 2: Init Model
#import pdb; pdb.set_trace():q
if not args.control:
logging.info("Initializing the model")
model = SwitchLM(hparams=args, vocab_size=len(dm.tokenizer))
else:
logging.info("loading from previous checkpoint")
model = SwitchLM.load_from_checkpoint(args.ckpt)
#args.vocab_size = len(dm.tokenizer)
model.hparams.warmup_steps = int(get_train_steps(dm) * model.hparams.warmup_prop)
#lr_monitor = LearningRateMonitor(logging_interval='step')
print(model.model.vocab_size)
monitor = 'val_{}'.format(args.monitor)
print("Monitor, ", monitor)
fname = '{epoch}-{step}-{val_' + args.monitor+'_epoch:.4f}'
# Step 3: Start
# if 'non_overlapped' in args.dataset_basedir:
# print("using non overlapped nt-idx")
# desc = 'non_overlapped'
# elif 'overlapped' in args.dataset_basedir:
# print("using overlapped nt-idx")
# desc = 'overlapped'
# if args.use_full_context:
# desc += '_full_context'
# args.use_speaker_descriptions=False
desc = ''
if 'random' in args.dataset_basedir:
desc += '_random'
if args.self_explain_ngram:
desc += '_self_explain'
if args.use_speaker_descriptions:
desc += '_speaker_descriptions'
if args.phrase_embeddings:
desc += '_with_phrase'
if args.concatenate_speaker:
desc += '_cat_speaker'
if args.multihead_pool:
desc+= '_multihead_pool'
if args.multihead_pool_over_input:
desc+= '_multihead_pool_over_input'
if args.concat_speaker_logits:
desc += '_concat_speaker_logits'
# if args.threshold != 0.5:
# desc += "threshold_{}".format(args.threshold)
if args.use_speaker_tokens:
desc += '_spk_tokens'
else:
desc += '_eot_eou'
if args.no_adapter:
desc += '_no_adapter'
desc += '_ctx_size_'+str(args.context_size)
if args.speaker_trait_predictions:
desc += '_spk_trait_aux'
if args.full_mtl_setup:
desc+= "_full_mtl"
desc += "_"+monitor
mode='max'
if args.monitor == 'loss':
mode = 'min'
desc += str(args.seed)
logging.info("Starting the training")
checkpoint_callback = ModelCheckpoint(
os.path.join(os.getcwd(), 'chkpts{}'.format(desc)),
save_top_k=3,
verbose=True,
monitor='{}_epoch'.format(monitor), save_weights_only=True,
mode=mode
)
logger = TensorBoardLogger(args.tensorboard_dir, name=desc)
accelerator = None
if args.gpus > 1:
accelerator='dp'
trainer = pl.Trainer.from_argparse_args(args, callbacks=[checkpoint_callback], val_check_interval=0.5, accelerator=accelerator, gradient_clip_val=args.clip_grad, track_grad_norm=2, logger=logger)
trainer.fit(model, dm)
if not args.control:
trainer.test(ckpt_path='best')
print("evaluating/interpreting results... ")
rand='random' in args.dataset_basedir
best= SwitchLMForEval(checkpoint_callback.best_model_path, control=False, threshold=-1, do_only_eval=False, prepend_description=False, seed=SEED, random=rand)
dm2 = ClassificationData(basedir=args.dataset_basedir, tokenizer_name=args.model_name, context_size=args.context_size,load_description_data=False, \
batch_size=1, codeswitch=True, num_workers=args.num_workers, balanced=False, use_speaker_tokens=args.use_speaker_tokens, \
load_control_data=args.control, load_full_control=False, do_social_predictions=args.speaker_trait_predictions, full_mtl_setup=args.full_mtl_setup)
trainer2 = pl.Trainer(gpus=1, accelerator=None)
test_file_list = [False, True]
# if args.control_data:
# test_file_list = [True]
for testing in test_file_list:
best.testing_file=testing
if not testing:
trainer2.test(test_dataloaders=dm2.val_dataloader(), model=best)
else:
trainer2.test(test_dataloaders=dm2.test_dataloader(), model=best)