-
Notifications
You must be signed in to change notification settings - Fork 25
/
Copy pathtrain.py
183 lines (160 loc) · 8.97 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
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
from tqdm import tqdm
from einops import rearrange
from PIL import Image
from copy import deepcopy
from torchvision.utils import save_image
from typing import List, Optional, Union
from torch import autocast
from torchvision import utils as vutils
from utils.util import EditingJsonDataset, EditingSingleImageDataset, plot_images
from lr_schedule import WarmupLinearLRSchedule
from torch.optim.lr_scheduler import CosineAnnealingLR
from torch.utils.tensorboard import SummaryWriter
from models.model import RGN
from models.utils import visualize_images, read_image_from_url, draw_image_with_bbox_new, Bbox
from utils.util2 import compose_text_with_templates, get_augmentations_template
from torchvision.utils import draw_bounding_boxes
from timm.data.constants import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
from torchvision import datasets, transforms
from engine import *
from vis import *
import os, jax, cv2, pdb
import numpy as np
import argparse, torch, inspect
import PIL, time, json, datetime
import random
import torch.nn as nn
import torch.nn.functional as F
import torch.backends.cudnn as cudnn
import utils.misc as misc
import torchvision.transforms as T
import torch.distributed as dist
def configure_optimizers(model, lr, betas=(0.9, 0.96), weight_decay=4.5e-2):
optimizer = torch.optim.Adam(model.module.anchor_net.parameters(), lr=lr, betas=betas, weight_decay=weight_decay)
return optimizer
def train(args, lr_schedule, model, template, len_train_dataset, data_loader_train, optim, device_id):
save_path = args.save_path
rank = dist.get_rank()
if not os.path.exists(save_path) and rank == 0:
os.mkdir(save_path)
for epoch in range(1, args.epochs+1):
data_loader_train.sampler.set_epoch(epoch)
metric_logger = misc.MetricLogger(delimiter=" ")
metric_logger.add_meter('lr', misc.SmoothedValue(window_size=1, fmt='{value:.6f}'))
if rank == 0:
print(f'Epoch {epoch}:')
for data_iter_step, (imgs, o_prompt, e_prompt) in enumerate(tqdm(data_loader_train)):
lr_schedule.step()
imgs = imgs.to(device=device_id, non_blocking=True)
o_prompt, e_prompt = o_prompt[0], e_prompt[0]
e_prompt = compose_text_with_templates(e_prompt, template)
with torch.cuda.amp.autocast():
bboxs = torch.ceil(map_cooridates(model.module.get_anchor_box(imgs)))
imgs_new, mask_imgs = get_mask_imgs(imgs, bboxs)
results = model.module.generate_result(imgs_new.to(device_id), mask_imgs.to(device_id), e_prompt).to(device_id)
loss, loss_clip, loss_cip_dir, loss_structure = model.module.get_loss(imgs_new, results, e_prompt, o_prompt)
loss.backward()
if data_iter_step % args.accum_grad == 0:
optim.step()
optim.zero_grad()
metric_logger.update(loss=loss.item())
if rank == 0:
if epoch % args.ckpt_interval == 0:
torch.save(model.state_dict(), os.path.join(save_path, f'transformer_epoch_{epoch}.pth'))
torch.save(model.state_dict(), os.path.join(save_path, 'last.pth'))
return model
def main(args):
dist.init_process_group("nccl", init_method='env://')
rank = dist.get_rank()
device_id = rank % torch.cuda.device_count()
device = torch.device(args.device)
seed = args.seed + misc.get_rank()
torch.manual_seed(seed)
np.random.seed(seed)
cudnn.benchmark= True
template = get_augmentations_template()
if not os.path.exists(args.save_path) and rank == 0:
os.mkdir(args.save_path)
model = RGN(image_size=args.image_size, device=device_id, args=args).to(device_id)
model = torch.nn.parallel.DistributedDataParallel(model, find_unused_parameters=True)
if rank == 0 and not os.path.exists(args.output_dir):
os.mkdir(args.output_dir)
dirs = os.listdir(args.output_dir)
if args.json_file is not None:
train_dataset = EditingJsonDataset(args, args.per_image_iteration)
test_dataset = EditingJsonDataset(args)
else:
train_dataset = EditingSingleImageDataset(args, args.per_image_iteration)
test_dataset = EditingSingleImageDataset(args)
len_train_dataset = len(train_dataset)
num_tasks = misc.get_world_size()
sampler = torch.utils.data.DistributedSampler(
train_dataset, num_replicas=num_tasks, rank=rank, shuffle=False, drop_last=False
)
data_loader_train = torch.utils.data.DataLoader(
train_dataset, sampler=sampler,
batch_size=args.batch_size,
num_workers=args.num_workers,
pin_memory=args.pin_mem,
drop_last=False,
shuffle=False)
data_loader_test = torch.utils.data.DataLoader(
test_dataset,
batch_size=1,
num_workers=args.num_workers,
shuffle=False,
pin_memory=args.pin_mem)
optim = configure_optimizers(model, args.lr)
total_steps = len_train_dataset / (args.batch_size*num_tasks)
lr_schedule = CosineAnnealingLR(optim, T_max=args.epochs*total_steps)
optim.zero_grad()
model = train(args, lr_schedule, model, template, len_train_dataset, data_loader_train, optim, device_id)
if rank == 0:
print('Generating edited images!')
model.eval()
predict(args, model, template, data_loader_test, device_id)
def get_args_parser():
parser = argparse.ArgumentParser(description="train models")
parser.add_argument('--run_name', type=str, default="exp")
parser.add_argument("--nodes", default=1, type=int, help='number of nodes to request')
parser.add_argument('--image_size', type=int, default=256, help='image height and width.')
parser.add_argument('--image_dir_path', type=str, default=None, help='dir path to input images.')
parser.add_argument('--image_file_path', type=str, default=None, help='path to input images.')
parser.add_argument('--image_caption', type=str, default=None, help='caption of the input image.')
parser.add_argument('--editing_prompt', type=str, default=None, help='editing prompt.')
parser.add_argument('--json_file', type=str, default=None, help='path to image-prompt file.')
parser.add_argument('--draw_box', action='store_true', help='draw boxes')
parser.add_argument('--diffusion_model_path', type=str, default='runwayml/stable-diffusion-inpainting', help='path to stable diffusion model.')
parser.add_argument('--save_path', type=str, default='./checkpoints', help='path to save checkpoint.')
parser.add_argument('--load_checkpoint_path', type=str, default=None, help='path to save checkpoint.')
parser.add_argument('--output_dir', type=str, default='./output', help='path to output dir.')
parser.add_argument('--device', type=str, default="cuda", help='device the training is on.')
parser.add_argument('--batch_size', type=int, default=192, help='batch size for training.')
parser.add_argument('--accum_grad', type=int, default=25, help='number for gradient accumulation.')
parser.add_argument('--epochs', type=int, default=10, help='number of epochs to train.')
parser.add_argument('--per_image_iteration', type=int, default=1, help='training iterations for each image.')
parser.add_argument('--loss_alpha', type=int, default=1, help='coefficient of clip loss.')
parser.add_argument('--loss_beta', type=int, default=1, help='coefficient of directional clip loss.')
parser.add_argument('--loss_gamma', type=int, default=1, help='coefficient of sturcture loss.')
parser.add_argument('--test_alpha', type=int, default=2, help='coefficient of text-to-image similarity.')
parser.add_argument('--test_beta', type=int, default=1, help='coefficient of image-to-image similarity.')
parser.add_argument('--ckpt_interval', type=int, default=10, help='number of epochs to save.')
parser.add_argument('--lr', type=float, default=1e-2, help='learning rate.')
parser.add_argument('--max_window_size', type=int, default=6, help='max window size')
parser.add_argument('--point_number', type=int, default=3, help='point sample number')
parser.add_argument('--num_workers', default=16, type=int)
parser.add_argument('--seed', default=42, type=int)
parser.add_argument('--pin_mem', action='store_true',
help='Pin CPU memory in DataLoader for more efficient (sometimes) transfer to GPU.')
# distributed training parameters
parser.add_argument('--world_size', default=1, type=int,
help='number of distributed processes')
parser.add_argument('--local_rank', default=0, type=int)
parser.add_argument('--dist_on_itp', action='store_true')
parser.add_argument('--dist_url', default='env://',
help='url used to set up distributed training')
args = parser.parse_args()
return args
if __name__ == '__main__':
args = get_args_parser()
main(args)