Repository navigation
Expand file tree
/
Copy pathtrain.py
More file actions
166 lines (132 loc) · 6.2 KB
/
Copy pathtrain.py
File metadata and controls
166 lines (132 loc) · 6.2 KB
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
from argparse import ArgumentParser
import pytorch_lightning as pl
import torch
import torch.nn as nn
import torch.optim as optim
from pytorch_lightning.callbacks import LearningRateMonitor
from torch.nn.utils.rnn import pack_padded_sequence
from torchtext.data.metrics import bleu_score
from torchvision import transforms
from dataset import CaptionDataset
from model import CaptionModel
from preprocess import Tokenizer
from utils import *
class LitModel(pl.LightningModule):
def __init__(self, epochs, root_dir, tokenizer_path, fine_tune):
super().__init__()
train_tfms = transforms.Compose([
# radom crop of image resized to 256
transforms.RandomResizedCrop(256, scale=(0.9, 1)),
# horizontally flip image with probability=0.5
transforms.RandomHorizontalFlip(),
# convert the PIL Image to a tensor
transforms.ToTensor(),
# normalize image for pre-trained model
transforms.Normalize((0.485, 0.456, 0.406),
(0.229, 0.224, 0.225))
])
val_tfms = transforms.Compose([
# smaller edge of image resized to 256
transforms.Resize((256, 256)),
transforms.ToTensor(),
# normalize image for pre-trained model
transforms.Normalize((0.485, 0.456, 0.406),
(0.229, 0.224, 0.225))
])
self.tokenizer = Tokenizer(root_dir)
self.epochs = epochs
if tokenizer_path is not None:
self.tokenizer.load_tokenizer(tokenizer_path)
else:
self.tokenizer.tokenize('Flicker8k_text/Flickr_8k.trainImages.txt')
self.fine_tune = fine_tune
self.model = CaptionModel(self.tokenizer)
self.model.encoder.fine_tune(fine_tune=self.fine_tune)
self.train_ds = CaptionDataset(
root_dir, 'Flicker8k_text/Flickr_8k.trainImages.txt', self.tokenizer, transform=train_tfms)
self.val_ds = CaptionDataset(
root_dir, 'Flicker8k_text/Flickr_8k.devImages.txt', self.tokenizer, mode='val', transform=val_tfms)
self.loss_func = nn.CrossEntropyLoss()
self.val_bleu_meter = AverageMeter()
def forward(self, x):
pass
def configure_optimizers(self):
optimizers = []
optimizer1 = optim.Adam(self.model.decoder.parameters(), lr=4e-4)
optimizers.append(optimizer1)
if self.fine_tune:
optimizer2 = optim.Adam(self.model.encoder.parameters(), lr=1e-4)
optimizers.append(optimizer2)
"""steps_per_epoch = math.ceil(len(self.train_ds)/128)
sched = optim.lr_scheduler.OneCycleLR(optimizer,
max_lr=max_lr, epochs=self.epochs,
steps_per_epoch=steps_per_epoch)
scheduler = {
"scheduler": sched,
"interval": "step",
"frequency": 1,
}"""
return optimizers#, [scheduler]
def training_step(self, batch, batch_idx):
# training_step defines the train loop.
x, captions, caplens = batch
predictions, captions, caplens, sort_idx = self.model(
x, captions, caplens, self.device)
# Since we decoded starting with <start>, the targets are all words after <start>, up to <end>
targets = captions[:, 1:]
# Remove timesteps that we didn't decode at, or are pads
# pack_padded_sequence is an easy trick to do this
scores = pack_padded_sequence(predictions, caplens, batch_first=True)
targets = pack_padded_sequence(targets, caplens, batch_first=True)
loss = self.loss_func(scores.data, targets.data)
load, mem_util = get_gpu_usage()
self.log('train_loss', loss)
self.log('gpu_load', load)
self.log('gpu_mem_usage', mem_util)
return loss
def validation_step(self, batch, batch_idx):
# this is the validation loop
x, y, z = batch
captions = self.model.predict(x, self.device)
for i in range(len(captions)):
captions[i] = [self.tokenizer.idx2val[j] for j in captions[i]]
bleu4 = bleu_score(captions, y)
self.val_bleu_meter.update(bleu4)
def predict_step(self, batch, batch_idx):
# this is the prediction loop
x, y = batch
out = self.model.encoder(x)
captions = self.model(out, self.device)
return captions
def val_dataloader(self):
loader = torch.utils.data.DataLoader(
self.val_ds, batch_size=128, num_workers=2, pin_memory=True, collate_fn=val_collate)
return loader
def train_dataloader(self):
loader = torch.utils.data.DataLoader(
self.train_ds, batch_size=128, shuffle=True, pin_memory=True, num_workers=2, collate_fn=train_collate)
return loader
if __name__ == '__main__':
parser = ArgumentParser()
parser.add_argument("--root_dir", default="./",
type=str, help="root directory path")
parser.add_argument("--epochs", default=10, type=int,
help="num of epochs to train")
parser.add_argument("--tokenizer_path", default=None, type=str,
help="path to saved tokenizer")
parser.add_argument("--fine_tune", action='store_true',
help="fine tuning switch for training")
parser.add_argument("--ckpt_path", default=None, type=str,
help="path to load checkpoint stored")
args = parser.parse_args()
model = LitModel(args.epochs, args.root_dir,
args.tokenizer_path, args.fine_tune)
"""if args.ckpt_path is not None:
if os.path.exists(args.ckpt_path):
model = model.load_from_checkpoint(
args.ckpt_path, epochs=args.epochs, root_dir=args.root_dir, tokenizer_path=args.tokenizer_path, fine_tune=args.fine_tune)
"""
lr_monitor = LearningRateMonitor(logging_interval='step')
cust_metric_logger = CustomMetricLogger()
trainer = pl.Trainer(accelerator='gpu', devices=1, log_every_n_steps=10, max_epochs=args.epochs, callbacks=[lr_monitor, cust_metric_logger])
trainer.fit(model, ckpt_path=args.ckpt_path)