-
Notifications
You must be signed in to change notification settings - Fork 4
/
train.py
349 lines (303 loc) · 15.2 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
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
import os
import sys
import time
import math
import torch
import logging
import argparse
from random import choice
from loss import ClipLoss
from email.policy import default
from torch.nn import functional as F
from dali import dali_dataloader
from torch import distributed, optim
from torch.utils.tensorboard import SummaryWriter
rank = int(os.environ["RANK"])
local_rank = int(os.environ["LOCAL_RANK"])
world_size = int(os.environ["WORLD_SIZE"])
distributed.init_process_group(backend="nccl")
torch.cuda.set_device(local_rank)
class AverageMeter(object):
def __init__(self):
self.reset()
def reset(self):
self.val = 0
self.avg = 0
self.sum = 0
self.count = 0
def update(self, val, n=1):
self.val = val
self.sum += val * n
self.count += n
self.avg = self.sum / self.count
class GlobalStep:
def __init__(self, step: int = 0):
self.step = int(step)
def update(self):
self.step += 1
class SpeedCallBack(object):
def __init__(self, frequent, steps_total, batch_size):
self.batch_size = batch_size
self.frequent = frequent
self.steps_total = steps_total
self.loss_metric = AverageMeter()
self.rank = int(os.environ["RANK"])
self.world_size = int(os.environ["WORLD_SIZE"])
self.time_start = time.time()
self.init = False
self.tic = 0
def __call__(
self,
lr_scheduler: optim.lr_scheduler._LRScheduler,
loss,
global_step,
scale):
assert isinstance(loss, float)
self.loss_metric.update(loss)
if global_step > 0 and global_step % self.frequent == 0:
if self.init:
try:
speed: float = (self.frequent * self.batch_size / (time.time() - self.tic))
self.tic = time.time()
speed_total = speed * self.world_size
except ZeroDivisionError:
speed = float("inf")
speed_total = float("inf")
loss_str_format = f"[{self.loss_metric.avg :.3f}]"
self.loss_metric.reset()
time_now = (time.time() - self.time_start) / 3600
time_total = time_now / ((global_step + 1) / self.steps_total)
time_for_end = time_total - time_now
lr_1 = lr_scheduler.get_last_lr()[0]
msg = f"rank:{int(speed) :d} "
msg += f"total:{int(speed_total) :d} "
msg += f"lr:[{lr_1 :.8f}] "
msg += f"step:{global_step :d} "
msg += f"amp:{int(scale) :d} "
msg += f"required:{time_for_end :.1f} hours "
msg += loss_str_format
if self.rank == 0:
logging.info(msg)
else:
self.init = True
self.tic = time.time()
def unwrap_model(model):
if hasattr(model, 'module'):
return model.module
else:
return model
def init_logging(rank, models_root):
if rank == 0:
log_root = logging.getLogger()
log_root.setLevel(logging.INFO)
formatter = logging.Formatter("Training: %(asctime)s-%(message)s")
handler_file = logging.FileHandler(os.path.join(models_root, "training.log"))
handler_stream = logging.StreamHandler(sys.stdout)
handler_file.setFormatter(formatter)
handler_stream.setFormatter(formatter)
log_root.addHandler(handler_file)
log_root.addHandler(handler_stream)
log_root.info('rank_id: %d' % rank)
def get_model_RWKV_CLIP(args):
from model import Text_RWKV, Image_RWKV, get_model
model_image_rwkv = Image_RWKV(img_size = args.input_size,
patch_size= args.image_patch_size,
embed_dims = args.image_embed_dims,
hidden_rate= args.image_hidden_rate,
depth=args.image_depth,
num_heads=args.image_num_heads,
output_cls_token=args.image_output_cls_token,
with_cls_token=args.image_with_cls_token,
with_cp=args.with_cp,
drop_path_rate=args.drop_path_rate)
model_text_rwkv = Text_RWKV(args)
model = get_model(model_image_rwkv, model_text_rwkv, image_cls_token=args.image_output_cls_token)
return model
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument("--batch-size", type=int, default=256)
parser.add_argument("--beta1", type=float, default=0.9, help="adamw")
parser.add_argument("--beta2", type=float, default=0.98, help="adamw")
parser.add_argument("--epochs", type=int, default=32)
parser.add_argument("--gradient-acc", type=int, default=1)
parser.add_argument("--lr", type=float, default=0.1, help="Learning rate.")
parser.add_argument("--lr-scheduler", default="cosine")
parser.add_argument("--warmup", type=float, default=0.1)
parser.add_argument("--is-normlize", type=int, default=1)
parser.add_argument("--local-loss",default=False,help="calculate loss w/ local features @ global (instead of realizing full global @ global matrix)")
parser.add_argument("--gather-with-grad",default=False,help="enable full distributed gradient for feature gather")
parser.add_argument("--horovod",default=False,action="store_true",help="Use horovod for distributed training.")
parser.add_argument("--optimizer", default="sgd")
parser.add_argument("--output", required=True)
parser.add_argument('--cfg', type=str, default='', metavar="FILE", help='path to config file')
parser.add_argument("--opts", help="Modify config options by adding 'KEY VALUE' pairs. ", default=None, nargs='+')
parser.add_argument('--drop-path', type=float, default=0.1, metavar='PCT',help='Drop path rate (default: 0.1)')
parser.add_argument("--train-data", required=True)
parser.add_argument("--train-num-samples", type=int, required=True)
parser.add_argument("--weight-decay", type=float, default=5e-4, help="Weight decay.")
parser.add_argument("--workers", type=int, default=10)
parser.add_argument("--precision", default="bf16", type=str)
parser.add_argument('--dropout', type=float, default=0.0, metavar='PCT', help='Dropout rate (default: 0.)')
parser.add_argument("--open-checkpoint", default="False", type=str)
################################################################################
################################# Image RWKV ####################################
################################################################################
parser.add_argument("--input-size", default=224, type=int, help="input_image_size")
parser.add_argument("--image-depth", default=12, type=int)
parser.add_argument("--image-embed-dims", default=384, type=int)
parser.add_argument("--image-patch-size", default=16, type=int)
parser.add_argument("--image-hidden-rate", default=4, type=int)
parser.add_argument("--image-num-heads", default=8, type=int)
parser.add_argument("--image-output-cls-token", default="False", type=str)
parser.add_argument("--image-with-cls-token", default="False", type=str)
parser.add_argument("--drop-path-rate", default=0.0, type=float)
################################################################################
################################# Text RWKV ####################################
################################################################################
parser.add_argument("--data-type", default="utf-8", type=str)
parser.add_argument("--ctx-len", default=77, type=int, help="")
parser.add_argument("--vocab-size", default=49408, type=int, help="Vocabular number")
parser.add_argument("--text-initialization", default="True", type=str)
parser.add_argument("--head-size", default=8, type=int)
parser.add_argument("--text-num-head", default=0, type=int)
parser.add_argument("--pos-emb", default=0, type=int)
parser.add_argument("--head-size-divisor", default=8, type=int)
parser.add_argument("--n-layer", default=12, type=int)
parser.add_argument("--n-embd", default=384, type=int)
parser.add_argument("--dim-att", default=0, type=int)
parser.add_argument("--dim-ffn", default=0, type=int)
parser.add_argument("--head-qk", default=0, type=int)
parser.add_argument("--tiny-att-dim", default=0, type=int)
parser.add_argument("--tiny-att-layer", default=-999, type=int)
args = parser.parse_args()
args.text_initialization = True if args.text_initialization == "True" else False
args.image_output_cls_token = True if args.image_output_cls_token == "True" else False
args.image_with_cls_token = True if args.image_output_cls_token == "True" else False
args.with_cp = True if args.open_checkpoint == "True" else False
assert args.data_type in ["utf-8", "utf-16le", "numpy", "binidx", "dummy", "uint16"]
assert args.precision in ["fp32", "tf32", "fp16", "bf16"]
assert args.image_embed_dims == args.n_embd, "Image embedding dimension must be the same as the Text embedding dimension"
assert args.image_output_cls_token == args.image_with_cls_token, "with_cls_token must be True if set output_cls_token to True"
if args.dim_att <= 0:
args.dim_att = args.n_embd
if args.dim_ffn <= 0:
args.dim_ffn = int((args.n_embd * 3.5) // 32 * 32) # default = 3.5x emb size
if args.text_num_head != 0:
assert args.n_embd % args.text_num_head == 0,
args.head_size = args.n_embd // args.text_num_head
os.environ["RWKV_CTXLEN"] = str(args.ctx_len)
os.environ["RWKV_HEAD_SIZE"] = str(args.head_size)
os.environ['RWKV_FLOAT_MODE'] = str(args.precision)
os.environ['Image_T_max'] = str((args.input_size / args.image_patch_size)**2)
os.environ['Text_T_max'] = str(256)
os.environ['Image_HEAD_SIE'] = str(args.image_embed_dims // args.image_num_heads)
return args
def main(args):
os.makedirs(args.output, exist_ok=True)
init_logging(rank, args.output)
if rank == 0:
summary_writer = SummaryWriter(os.path.join(args.output, "tensorboard"))
else:
summary_writer = None
start_epoch = 0
RWKV_CLIP_model = get_model_RWKV_CLIP(args)
train_loader = dali_dataloader(args)
training_precision_type = torch.bfloat16 if args.precision == "bf16" else torch.float16
RWKV_CLIP_model.train().cuda()
RWKV_CLIP_model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(RWKV_CLIP_model)
RWKV_CLIP_model = torch.nn.parallel.DistributedDataParallel(
module=RWKV_CLIP_model,
bucket_cap_mb=32,
find_unused_parameters=True,
static_graph=True)
global_step = GlobalStep()
steps_per_epoch = args.train_num_samples // world_size // args.batch_size + 1
steps_total = int(args.epochs * steps_per_epoch)
contrastive_loss = ClipLoss(
local_loss=args.local_loss,
gather_with_grad=args.gather_with_grad,
cache_labels=True,
rank=int(os.environ["RANK"]),
world_size=int(os.environ["WORLD_SIZE"]),
use_horovod=args.horovod)
opt = torch.optim.AdamW(
params=[{"params": RWKV_CLIP_model.parameters()}],
lr=args.lr, weight_decay=args.weight_decay, betas=(args.beta1, args.beta2))
if args.lr_scheduler == "cosine":
assert isinstance(args.epochs, int)
lr_scheduler = optim.lr_scheduler.OneCycleLR(
optimizer=opt,
max_lr=[args.lr],
steps_per_epoch=steps_per_epoch,
epochs=args.epochs,
pct_start=0.1,
)
elif args.lr_scheduler == "linear":
lr_scheduler = optim.lr_scheduler.LinearLR(
optimizer=opt, start_factor=1.0, end_factor=0.0,
total_iters=int(args.epochs * steps_per_epoch))
else:
raise
callback_func = SpeedCallBack(5, steps_total, args.batch_size)
auto_scaler = torch.cuda.amp.grad_scaler.GradScaler(init_scale=128, growth_interval=200)
for epoch in range(start_epoch, math.ceil(args.epochs)):
for _, (img, text_token) in enumerate(train_loader):
opt.zero_grad()
new_text_token = []
# text random augmentation
for i in range(text_token.size(0)):
choose = choice([0,1,2])
# Raw Text
if choose == 0:
new_text_token.append(text_token[i, :77].long().cuda())
# Synthetic Text
elif choose == 1:
new_text_token.append(text_token[i, 77*1:77*2].long().cuda())
# Generated Text
elif choose == 2:
new_text_token.append(text_token[i, 77*2:77*3].long().cuda())
text_token = torch.stack(new_text_token, dim=0)
img = img.cuda()
with torch.cuda.amp.autocast(True, dtype=training_precision_type):
image_embeddings, text_embeddings, logit_scale = RWKV_CLIP_model(img, text_token)
image_embedding_norm = F.normalize(image_embeddings, dim=-1)
text_embedding_norm = F.normalize(text_embeddings, dim=-1)
loss = contrastive_loss(image_embedding_norm, text_embedding_norm, logit_scale)
if args.precision == "bf16":
loss.backward()
if global_step.step % args.gradient_acc == 0:
torch.nn.utils.clip_grad_norm_(RWKV_CLIP_model.parameters(), 1)
opt.step()
opt.zero_grad()
else:
auto_scaler.scale(loss).backward()
if global_step.step % args.gradient_acc == 0:
auto_scaler.unscale_(opt)
torch.nn.utils.clip_grad_norm_(RWKV_CLIP_model.parameters(), 1)
auto_scaler.step(opt)
auto_scaler.update()
opt.zero_grad()
with torch.no_grad():
unwrap_model(RWKV_CLIP_model).logit_scale.clamp_(0, math.log(100))
lr_scheduler.step()
global_step.step += 1
with torch.no_grad():
callback_func(lr_scheduler, float(loss), global_step.step, auto_scaler.get_scale())
if summary_writer is not None:
summary_writer.add_scalar(tag="loss", scalar_value=loss.item(), global_step=global_step.step)
summary_writer.add_scalar(tag="lr_backbone",
scalar_value=lr_scheduler.get_last_lr()[0],
global_step=global_step.step)
summary_writer.add_scalar(tag="logit_scale",
scalar_value=logit_scale.item(),
global_step=global_step.step)
if global_step.step > steps_total:
break
train_loader.reset()
if rank == 0:
torch.save(obj=RWKV_CLIP_model.state_dict(), f=os.path.join(args.output, "RWKV_CLIP_model_" + str(epoch) + ".pt"))
if summary_writer is not None:
summary_writer.close()
if __name__ == "__main__":
args = get_args()
main(args)