-
Notifications
You must be signed in to change notification settings - Fork 0
/
pretrain_data.py
192 lines (167 loc) · 8.34 KB
/
pretrain_data.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
188
189
190
import argparse
from collections import defaultdict
import json
import math
import os
from os.path import exists, join
from time import time, sleep
import torch
from torch.utils.data import DataLoader, ConcatDataset, ChainDataset
from torch.nn import functional as F
from torch.nn.utils import clip_grad_norm_
from data.webvid_data import WebvidFrameDataset
from tqdm import tqdm
import random
from utils.distributed import all_gather_list
from utils.logger import LOGGER, TB_LOGGER, RunningMeter, add_log_to_file
from data import PrefetchLoader
from data.data import TxtVideoDataset, txtvideo_collate, VideoMapper, TxtMapper
import ipdb
from utils.distributed import DistributedSampler_wopadding
import torch.distributed as dist
from data.data import worker_init_fn
from data.dist_data import LaionDataset, dist_data_collate
from utils.distributed import get_rank, get_local_rank
import torch.distributed as dist
from data.cc4m_data import CC4MDataset
from data.webvid10m import Webvid10M_IterableDataset
from data.coyo_webdata import CoyoWebDataset
import ipdb
def build_dataloader(dataset, collate_fn, is_train, opts, batch_size, n_workers,sampler="dist"):
if sampler=="dist":
if isinstance(dataset, torch.utils.data.IterableDataset):
sampler = None
else:
sampler = DistributedSampler_wopadding(dataset)
# sampler = DistributedSampler_wopadding(dataset)
loader = DataLoader(dataset, sampler = sampler, batch_size = batch_size,
num_workers=n_workers, pin_memory=True,
collate_fn=collate_fn, drop_last=is_train, worker_init_fn=worker_init_fn)
else:
loader = DataLoader(dataset, batch_size = batch_size,
num_workers=n_workers, pin_memory=True,
collate_fn=collate_fn, drop_last=is_train)
return loader
def compute_steps(epoch, batch_size, data_len):
gpus = int(os.getenv('WORLD_SIZE', 1))
total_batch_size = batch_size*gpus
step = int((data_len // total_batch_size) * epoch)
return step
def create_train_dataloaders(data_cfg, opts):
data_cfg = data_cfg['train']
dataloaders = {}
total_step =0
for d_cfg in data_cfg:
concate_name = ''
dataset_ls = []
dataset=None
# if 'sample_num' in d_
videogen = 'videogen' in d_cfg['task']
if not isinstance(d_cfg["datasets"], list):
dset = d_cfg["datasets"]
if dset["name"] in ["laion2b", "distcc4m", "tuchong", 'distcc12m', "coyo700m","coyo700m_web"]:
name=dset["name"]
concate_name=name
data_path = dset["data_path"]
ids_path = dset.get("ids_path","")
tokenizer = dset["tokenizer"]
rank = get_rank()
world_size = dist.get_world_size()
if name=="laion2b":
dataset = LaionDataset(data_path, ids_path, opts.video_cfg, tokenizer= tokenizer, caption="caption", is_train=True, rank=rank, world_size=world_size)
elif name=="tuchong":
dataset = LaionDataset(data_path, ids_path, opts.video_cfg, tokenizer= tokenizer, caption="desc", is_train=True, rank=rank, world_size=world_size)
elif name == "distcc4m":
dataset = CC4MDataset(data_path, opts.video_cfg, tokenizer= tokenizer, is_train=True, rank=rank, world_size=world_size)
elif name == "distcc12m":
dataset = CC4MDataset(data_path, opts.video_cfg, tokenizer= tokenizer, is_train=True, rank=rank, world_size=world_size)
elif name == "coyo700m":
dataset = LaionDataset(data_path, ids_path, opts.video_cfg, tokenizer= tokenizer, caption="caption", is_train=True, rank=rank, world_size=world_size)
elif name =="coyo700m_web":
dataset = CoyoWebDataset(data_path,opts.video_cfg, tokenizer= tokenizer, is_train=True)
# elif name == "webvid10m":
# dataset = Webvid10MDataset(data_path, opts.video_cfg, tokenizer= tokenizer, is_train=True, rank=rank, world_size=world_size)
collate_fn = dist_data_collate
LOGGER.info("Create Dataset {} success".format(name))
else:
for dset in d_cfg['datasets']:
name = dset['name']
concate_name = concate_name + name if concate_name == '' else concate_name + '_' + name
data_type = dset['datatype'] + '_' + name
video_path = dset['video']
txt_path = dset['txt']
key_path = dset['ids_path']
tokenizer = dset['tokenizer']
if name == 'webvid_frame':
dataset = WebvidFrameDataset(video_path, opts.video_cfg, tokenizer, is_train=True)
elif name == 'webvid10m':
dataset = Webvid10M_IterableDataset(
video_path,
opts.video_cfg,
rank=get_rank(),
world_size=dist.get_world_size(),
tokenizer=tokenizer,
is_train=True,
repeat=True)
else:
dataset = TxtVideoDataset(video_path,
txt_path,
key_path,
opts.video_cfg,
tokenizer=tokenizer,
is_train=True,
data_type=data_type)
collate_fn = txtvideo_collate
LOGGER.info("Create Dataset {} Success".format(name))
dataset_ls.append(dataset)
if isinstance(dataset_ls[0], torch.utils.data.IterableDataset):
dataset = ChainDataset(dataset_ls)
else:
dataset = ConcatDataset(dataset_ls)
# dataset = ConcatDataset(dataset_ls)
# dataset = dataset_ls[0]
LOGGER.info("Create Dataset {} Success".format(concate_name))
ratio = d_cfg['mix_ratio']
task = d_cfg['task']
batch_size = d_cfg['batch_size']
n_workers = d_cfg['n_workers']
epoch = d_cfg["epoch"]
data_len = d_cfg["data_len"]
steps = compute_steps(epoch, batch_size, data_len)
if concate_name in ["laion2b", "distcc4m", "tuchong", "distcc12m", "coyo700m", "coyo700m_web"]:
loader=build_dataloader(dataset, collate_fn, True, opts, batch_size, n_workers, sampler="None")
else:
loader = build_dataloader(dataset, collate_fn, True, opts, batch_size, n_workers)
LOGGER.info("Create dataloader {} Success".format(concate_name))
task_name = f'{task}--{concate_name}'
dataloaders[task_name] = [loader, ratio, steps]
total_step +=steps
for k, v in dataloaders.items():
prob = float(v[-1])/total_step
v.append(prob)
return dataloaders
def create_val_dataloaders(data_cfg, opts):
data_cfg = data_cfg['val']
dataloaders = {}
for d_cfg in data_cfg:
name = d_cfg['name']
data_type = d_cfg['datatype']
ids_path = d_cfg['ids_path']
videogen = 'videogen' in d_cfg['task']
video_token_path = d_cfg.get('video_token','')
audio_token_path = d_cfg.get('audio_token','')
video_mapper = VideoMapper(d_cfg['video'], opts.video_cfg, data_type, is_training=False,videogen=videogen, token_path= video_token_path)
txt_mapper = TxtMapper(d_cfg['txt'], opts.max_txt_len, data_type)
audio_path = d_cfg.get('audio','')
audio_mapper = AudioMapper(audio_path, opts.audio_cfg, data_type, videogen=videogen, token_path=audio_token_path)
dataset = TxtVideoAudioDataset(ids_path, txt_mapper, video_mapper, audio_mapper)
collate_fn = txtvideo_collate
raw_img_dir_for_FID = d_cfg.get('raw_img_dir_for_FID',None)
dataset.raw_img_dir_for_FID = raw_img_dir_for_FID
LOGGER.info("Create Dataset {} Success".format(name))
task = d_cfg['task']
batch_size = d_cfg['batch_size']
loader = build_dataloader(dataset, collate_fn, False, opts, batch_size)
task_name = f'{task}--{name}'
dataloaders[task_name] = PrefetchLoader(loader)
return dataloaders