-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
148 lines (124 loc) · 4.59 KB
/
train.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
import os
import sys
import random
sys.path.append("./model/")
sys.path.append("./preprocess/")
from transformers import Trainer, TrainingArguments, EarlyStoppingCallback
import numpy as np
import evaluate
import torch
from preprocess.roundataset import roundataset
from model.modeling_deberta_visual import DebertaForPhotobookListener
from model.configuration_deberta_visual import DebertaWithVisualConfig
from model.variables import (
EPOCHS, CKPT_DIR, RND_SEED,
BATCH_SIZE, PEAK_LR, WARMUP_STEPS, WEIGHT_DECAY,
PRETRAINED_MODEL_NAME, DLS
)
metric = evaluate.load("accuracy")
config_json = sys.argv[1]
CKPT_DIR = sys.argv[2] if len(sys.argv) > 2 else CKPT_DIR
RND_SEED = int(sys.argv[3]) if len(sys.argv) > 3 else RND_SEED
# NOTE (Shih-Lun): borrowed from
# https://wandb.ai/sauravmaheshkar/RSNA-MICCAI/reports/How-to-Set-Random-Seeds-in-PyTorch-and-Tensorflow--VmlldzoxMDA2MDQy
def set_rnd_seed():
np.random.seed(RND_SEED)
random.seed(RND_SEED)
torch.manual_seed(RND_SEED)
torch.cuda.manual_seed(RND_SEED)
# When running on the CuDNN backend, two further options must be set
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# Set a fixed value for the hash seed
os.environ["PYTHONHASHSEED"] = str(RND_SEED)
print(f"[info] Random seed set as {RND_SEED}")
def compute_metrics(eval_pairs):
predictions, labels = eval_pairs
# print (labels[0])
predictions = np.argmax(predictions[..., 1:], axis=-1) + 1
true_predictions = []
true_labels = []
# fetch last timestep outputs only
bsize, seqlen = predictions.shape[0], predictions.shape[1]
for b in range(bsize):
for pos in range(seqlen - 1, -1, -1):
if labels[b, pos, 0] != -100:
true_predictions.extend(predictions[b, pos])
true_labels.extend(labels[b, pos])
break
# print (true_predictions, true_labels)
results = metric.compute(predictions=true_predictions, references=true_labels)
# print (results)
return results
if __name__ == '__main__':
set_rnd_seed()
train_dset = roundataset(
'data/train_clean_sections.pickle',
'data/image_feats.pickle',
separate_images=False,
dense_learning_signals=DLS,
)
print ("[info] train set loaded, len =", len(train_dset))
val_dset = roundataset(
'data/valid_clean_sections.pickle',
'data/image_feats.pickle',
separate_images=False,
dense_learning_signals=DLS,
)
print ("[info] valid set loaded, len =", len(val_dset))
test_dset = roundataset(
'data/test_clean_sections.pickle',
'data/image_feats.pickle',
separate_images=False,
dense_learning_signals=DLS,
)
print ("[info] test dset loaded, len =", len(test_dset))
print ("[info] using config:", config_json)
model = DebertaForPhotobookListener(
DebertaWithVisualConfig.from_json_file(
config_json
)
)
model.deberta.load_hf_pretrained_deberta(PRETRAINED_MODEL_NAME)
trainer = Trainer(
model,
TrainingArguments(
output_dir=CKPT_DIR,
do_train=True,
do_eval=True,
per_device_eval_batch_size=BATCH_SIZE,
per_device_train_batch_size=BATCH_SIZE,
learning_rate=PEAK_LR,
weight_decay=WEIGHT_DECAY,
warmup_steps=WARMUP_STEPS,
num_train_epochs=EPOCHS,
evaluation_strategy='epoch',
save_strategy='epoch',
metric_for_best_model="eval_accuracy",
save_total_limit=3,
dataloader_num_workers=8,
logging_steps=50,
load_best_model_at_end=True,
gradient_accumulation_steps=1
),
train_dataset=train_dset,
eval_dataset=val_dset,
compute_metrics=compute_metrics,
callbacks=[EarlyStoppingCallback(early_stopping_patience=10)]
)
train_result = trainer.train()
metrics = train_result.metrics
trainer.save_model() # Saves the tokenizer too for easy upload
metrics["train_samples"] = len(train_dset)
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# # Evaluation
metrics = trainer.evaluate()
metrics["eval_samples"] = len(val_dset)
trainer.log_metrics("val", metrics)
trainer.save_metrics("val", metrics)
metrics = trainer.evaluate(eval_dataset=test_dset)
metrics["test_samples"] = len(test_dset)
trainer.log_metrics("test", metrics)
trainer.save_metrics("test", metrics)