-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainer.py
165 lines (144 loc) · 7.66 KB
/
trainer.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
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import torchvision.models as models
import torchvision.utils as vutils
from torch.utils.data import DataLoader
from tqdm import trange
from torch.utils.tensorboard import SummaryWriter
from PIL import Image, ImageFile
from dataset import collate, ContentStyleDataset, loader, denorm
from net import Model
import os
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.distributed import init_process_group, destroy_process_group
import torch.multiprocessing as mp
from torch.utils.data.distributed import DistributedSampler
class Trainer():
def __init__(self, rank, model, dataloader, learning_rate, save_every, writer, progress_dir, checkpoint_path, checkpoint_save_path, optimizer = "adam", lr_decay = 5e-5, alpha = 1.0, k = 5.0):
self.device = torch.device(rank)
self.rank = rank
self.checkpoint_path = checkpoint_path
self.checkpoint_save_path = checkpoint_save_path
self.model = DDP(model, device_ids=[rank])
self.dataloader = dataloader
self.lr = learning_rate
self.save_every = save_every
self.writer = writer
self.progress_dir = progress_dir + '/'
if optimizer == "adam":
self.optimizer = torch.optim.Adam(self.model.module.decoder.parameters(), lr = self.lr)
else:
self.optimizer = torch.optim.SGD(self.model.module.decoder.parameters(), lr = self.lr)
self.lr_decay = lr_decay
self.alpha = alpha
self.k = k
self.epoch = 0
self.step = 0
def train(self, num_epochs):
with trange(self.epoch, num_epochs, desc="All epochs") as epochs:
for epoch in epochs:
self.run_epoch(epoch)
def logs(self, step, content, style, stylized, content_loss, style_loss, loss):
self.writer.add_scalar("Style Loss", style_loss, step)
self.writer.add_scalar("Content Loss", content_loss, step)
self.writer.add_scalar("Total Loss", loss, step)
if step % 10 == 0:
grid = vutils.make_grid([denorm(content[0].squeeze(), self.device), denorm(style[0].squeeze(), self.device), denorm(stylized[0].squeeze(), self.device)], nrow = 3)
self.writer.add_image("Style Transfer Result", grid, step)
vutils.save_image(grid, self.progress_dir + str(self.step)+'.png')
def learning_rate_decay(self):
'''Learning rate decay based on the torch implementation by the authors of the paper '''
lr = self.lr / (1.0 + self.lr_decay * self.step)
for param_group in self.optimizer.param_groups:
param_group['lr'] = lr
def overfit(self, iters):
'''Overfit the model on a single batch for debugging purposes, sanity check that the encoder is set up properly and backprop is doing its thing'''
data = next(iter(self.dataloader))
data = data.to(self.device)
with trange(0, iters, desc="All epochs") as progress_bar:
for i in range(iters):
self.training_step(data)
content_loss, style_loss, loss = self.training_step(data)
progress_bar.update(1)
def run_epoch(self, epoch_num):
num_batches = len(self.dataloader)
progress_bar = trange(num_batches, desc = "epoch" + str(epoch_num) + " gpu: " + str(self.device))
for batch_num, data in enumerate(self.dataloader):
self.learning_rate_decay()
data = data.to(self.device)
content_loss, style_loss, loss = self.training_step(data)
progress_bar.update(1)
progress_bar.set_postfix(content_loss = content_loss.item(), style_loss = style_loss.item(), total_loss = loss.item())
if batch_num % 200 == 0 and self.rank == 0:
self.save_model(epoch_num)
def training_step(self, data):
content, style = data
content_loss, style_loss, stylized = self.model(data)
loss = content_loss + self.k * style_loss
self.step += 1
if self.rank == 0:
self.logs(self.step, content, style, stylized, content_loss, style_loss, loss)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
return content_loss, style_loss, loss
def save_model(self, epoch):
checkpoint = {
"model" : self.model.state_dict(),
"epoch" : epoch + 1,
"optimizer": self.optimizer.state_dict(),
"step": self.step
}
torch.save(checkpoint, self.checkpoint_save_path)
def load_checkpoint(self):
checkpoint = torch.load(self.checkpoint_path, map_location = "cuda:{}".format(self.rank))
self.model.load_state_dict(checkpoint['model'])
self.optimizer.load_state_dict(checkpoint['optimizer'])
self.epoch = checkpoint['epoch']
self.step = checkpoint['step']
print("RESUMING FROM EPOCH ", self.epoch)
def ddp_setup(rank, world_size):
"""
Args:
rank: Unique identifier of each process
world_size: Total number of processes
"""
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = "12974"
init_process_group(backend='gloo', rank=rank, world_size=world_size)
def main(rank, world_size, args):
torch.cuda.set_device(rank)
device = torch.device("cuda", rank)
ddp_setup(rank, world_size)
model, dataloader, writer = load_train_objects(rank, args.writer_dir)
trainer = Trainer(rank, model, dataloader, learning_rate = 1e-4, save_every = 1, writer = writer, progress_dir = args.progress_dir, checkpoint_path = args.checkpoint_path, checkpoint_save_path = args.checkpoint_save_path)
if args.load_model:
trainer.load_checkpoint()
if args.overfit == 'overfit':
trainer.overfit(10000)
else:
trainer.train(args.total_epochs)
destroy_process_group()
def load_train_objects(rank, writer_dir):
model = Model(rank)
dataset = ContentStyleDataset('content/', 'style/', transform = loader)
dataloader = DataLoader(dataset, batch_size = 1, collate_fn = collate, shuffle = False, num_workers = 0, sampler = DistributedSampler(dataset))
writer = SummaryWriter(writer_dir)
return model, dataloader, writer
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser(description='Neural Style Transfer following the AdaIN paper by Huang, et. al.')
parser.add_argument('total_epochs', type=int, help='Total epochs to train the model', default = 1)
parser.add_argument('--load_model', type=int, help='0 to start a new model, non 0 to continue from checkpoint.pt, default 0', default = 0)
parser.add_argument('--writer_dir', type=str, help='Directory to save tensorboard logs, default is ./runs', default = './runs')
parser.add_argument('--overfit', type=str, help="overfit to overfit, any other str to run normally", default = "regular")
parser.add_argument('--progress_dir', type=str, help="Directory to save progress images, default is ./outputs", default = "./outputs")
parser.add_argument('--save', type = bool, help = 'True by default, meaning the model saves checkpoints while training.', default = True)
parser.add_argument('--checkpoint_path', type = str, help = 'path for the checkpoint to load model from, default is \'checkpoint.pt\'', default = 'checkpoint.pt')
parser.add_argument('--checkpoint_save_path', type = str, help = 'path to save model checkpoints in this run, default is \'checkpoint.pt\'', default = 'checkpoint.pt')
args = parser.parse_args()
world_size = torch.cuda.device_count()
print("Starting multi processes")
mp.spawn(main, args=(world_size, args), nprocs=world_size)