-
Notifications
You must be signed in to change notification settings - Fork 3
/
train.py
317 lines (294 loc) · 15.8 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
from system_utils import get_gpt_id
dev = get_gpt_id()
import os
os.environ["CUDA_VISIBLE_DEVICES"] = dev
import signal
import time
import csv
import sys
import warnings
import random
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
import torch.multiprocessing as mp
import numpy as np
import time
import pprint
from loguru import logger
import smplx
from torch.utils.tensorboard import SummaryWriter
import wandb
import matplotlib.pyplot as plt
from utils import config, logger_tools, other_tools, metric
from dataloaders import data_tools
from dataloaders.build_vocab import Vocab
from optimizers.optim_factory import create_optimizer
from optimizers.scheduler_factory import create_scheduler
from optimizers.loss_factory import get_loss_func
class BaseTrainer(object):
def __init__(self, args):
self.args = args
self.rank = 0
self.checkpoint_path = args.out_path + "custom/" + args.name + args.notes + "/" #wandb.run.dir #args.cache_path+args.out_path+"/"+args.name
if self.rank==0:
if self.args.stat == "ts":
self.writer = SummaryWriter(log_dir=args.out_path + "custom/" + args.name + args.notes + "/")
else:
wandb.init(project=args.project, entity="liu1997", dir=args.out_path, name=args.name[12:] + args.notes)
wandb.config.update(args)
self.writer = None
#self.test_demo = args.data_path + args.test_data_path + "bvh_full/"
self.train_data_beatx = __import__(f"dataloaders.{args.dataset}", fromlist=["something"]).CustomDataset(args, "train")
self.train_data = self.train_data_beatx
if args.use_amass:
self.train_data_amass = __import__(f"dataloaders.amass_sep_lower", fromlist=["something"]).CustomDataset(args, "train")
beatx_len = len(self.train_data_beatx)
smplx_len = len(self.train_data_amass)
self.train_data = torch.utils.data.ConcatDataset([self.train_data_beatx,self.train_data_amass,])
self.train_loader = torch.utils.data.DataLoader(
self.train_data,
batch_size=args.batch_size,
shuffle=False if args.ddp else True,
num_workers=args.loader_workers,
drop_last=True,
sampler=torch.utils.data.distributed.DistributedSampler(self.train_data) if args.ddp else None,
)
self.train_length = len(self.train_loader)
logger.info(f"Init train dataloader success")
# self.val_data = __import__(f"dataloaders.{args.dataset}", fromlist=["something"]).CustomDataset(args, "val")
# self.val_loader = torch.utils.data.DataLoader(
# self.val_data,
# batch_size=args.batch_size,
# shuffle=False,
# num_workers=args.loader_workers,
# drop_last=False,
# sampler=torch.utils.data.distributed.DistributedSampler(self.val_data) if args.ddp else None,
# )
logger.info(f"Init val dataloader success")
if self.rank == 0:
self.test_data = __import__(f"dataloaders.{args.dataset}", fromlist=["something"]).CustomDataset(args, "test")
self.test_loader = torch.utils.data.DataLoader(
self.test_data,
batch_size=1,
shuffle=False,
num_workers=args.loader_workers,
drop_last=False,
)
logger.info(f"Init test dataloader success")
model_module = __import__(f"models.{args.model}", fromlist=["something"])
if args.ddp:
self.model = getattr(model_module, args.g_name)(args).to(self.rank)
process_group = torch.distributed.new_group()
self.model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(self.model, process_group)
self.model = DDP(self.model, device_ids=[self.rank], output_device=self.rank,
broadcast_buffers=False, find_unused_parameters=False)
else:
self.model = torch.nn.DataParallel(getattr(model_module, args.g_name)(args), args.gpus).cuda()
if self.rank == 0:
logger.info(self.model)
logger.info(f"init {args.g_name} success")
if args.stat == "wandb":
wandb.watch(self.model)
if args.d_name is not None:
if args.ddp:
self.d_model = getattr(model_module, args.d_name)(args).to(self.rank)
self.d_model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(self.d_model, process_group)
self.d_model = DDP(self.d_model, device_ids=[self.rank], output_device=self.rank,
broadcast_buffers=False, find_unused_parameters=False)
else:
self.d_model = torch.nn.DataParallel(getattr(model_module, args.d_name)(args), args.gpus).cuda()
if self.rank == 0:
logger.info(self.d_model)
logger.info(f"init {args.d_name} success")
if args.stat == "wandb":
wandb.watch(self.d_model)
self.opt_d = create_optimizer(args, self.d_model, lr_weight=args.d_lr_weight)
self.opt_d_s = create_scheduler(args, self.opt_d)
if args.e_name is not None:
"""
bugs on DDP training using eval_model, using additional eval_copy for evaluation
"""
eval_model_module = __import__(f"models.{args.eval_model}", fromlist=["something"])
# eval copy is for single card evaluation
if self.args.ddp:
self.eval_model = getattr(eval_model_module, args.e_name)(args).to(self.rank)
self.eval_copy = getattr(eval_model_module, args.e_name)(args).to(self.rank)
else:
self.eval_model = getattr(eval_model_module, args.e_name)(args)
self.eval_copy = getattr(eval_model_module, args.e_name)(args).to(self.rank)
#if self.rank == 0:
other_tools.load_checkpoints(self.eval_copy, args.data_path+args.e_path, args.e_name)
other_tools.load_checkpoints(self.eval_model, args.data_path+args.e_path, args.e_name)
if self.args.ddp:
self.eval_model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(self.eval_model, process_group)
self.eval_model = DDP(self.eval_model, device_ids=[self.rank], output_device=self.rank,
broadcast_buffers=False, find_unused_parameters=False)
self.eval_model.eval()
self.eval_copy.eval()
if self.rank == 0:
logger.info(self.eval_model)
logger.info(f"init {args.e_name} success")
if args.stat == "wandb":
wandb.watch(self.eval_model)
self.opt = create_optimizer(args, self.model)
self.opt_s = create_scheduler(args, self.opt)
self.smplx = smplx.create(
self.args.data_path_1+"smplx_models/",
model_type='smplx',
gender='NEUTRAL_2020',
use_face_contour=False,
num_betas=300,
num_expression_coeffs=100,
ext='npz',
use_pca=False,
).to(self.rank).eval()
self.alignmenter = metric.alignment(0.3, 7, self.train_data.avg_vel, upper_body=[3,6,9,12,13,14,15,16,17,18,19,20,21]) if self.rank == 0 else None
self.align_mask = 60
self.l1_calculator = metric.L1div() if self.rank == 0 else None
def inverse_selection(self, filtered_t, selection_array, n):
original_shape_t = np.zeros((n, selection_array.size))
selected_indices = np.where(selection_array == 1)[0]
for i in range(n):
original_shape_t[i, selected_indices] = filtered_t[i]
return original_shape_t
# def inverse_selection_6d(self, filtered_t, selection_array, n):
# original_shape_t = np.zeros((n, selection_array.size))
# selected_indices = np.where(selection_array == 1)[0]
# new_selected_indices = np.zeros((n, selected_indices.size*2))
# new_selected_indices[:, ::2] = selected_indices
# new_selected_indices[:, 1::2] = selected_indices
# selected_indices = new_selected_indices.astype(np.bool)
# for i in range(n):
# original_shape_t[i, selected_indices] = filtered_t[i]
# return original_shape_t
def inverse_selection_tensor(self, filtered_t, selection_array, n):
selection_array = torch.from_numpy(selection_array).cuda()
selected_indices = torch.where(selection_array == 1)[0]
if len(filtered_t.shape) == 2:
original_shape_t = torch.zeros((n, 165)).cuda()
for i in range(n):
original_shape_t[i, selected_indices] = filtered_t[i]
elif len(filtered_t.shape) == 3:
bs, n, _ = filtered_t.shape
original_shape_t = torch.zeros((bs, n, 165), device='cuda')
expanded_indices = selected_indices.unsqueeze(0).unsqueeze(0).expand(bs, n, -1)
original_shape_t.scatter_(2, expanded_indices, filtered_t)
return original_shape_t
def inverse_selection_tensor_6d(self, filtered_t, selection_array, n):
new_selected_array = np.zeros((330))
new_selected_array[::2] = selection_array
new_selected_array[1::2] = selection_array
selection_array = new_selected_array
selection_array = torch.from_numpy(selection_array).cuda()
selected_indices = torch.where(selection_array == 1)[0]
if len(filtered_t.shape) == 2:
original_shape_t = torch.zeros((n, 330)).cuda()
for i in range(n):
original_shape_t[i, selected_indices] = filtered_t[i]
elif len(filtered_t.shape) == 3:
bs, n, _ = filtered_t.shape
original_shape_t = torch.zeros((bs, n, 330), device='cuda')
expanded_indices = selected_indices.unsqueeze(0).unsqueeze(0).expand(bs, n, -1)
original_shape_t.scatter_(2, expanded_indices, filtered_t)
return original_shape_t
def train_recording(self, epoch, its, t_data, t_train, mem_cost, lr_g, lr_d=None):
pstr = "[%03d][%03d/%03d] "%(epoch, its, self.train_length)
for name, states in self.tracker.loss_meters.items():
metric = states['train']
if metric.count > 0:
pstr += "{}: {:.3f}\t".format(name, metric.avg)
self.writer.add_scalar(f"train/{name}", metric.avg, epoch*self.train_length+its) if self.args.stat == "ts" else wandb.log({name: metric.avg}, step=epoch*self.train_length+its)
pstr += "glr: {:.1e}\t".format(lr_g)
self.writer.add_scalar("lr/glr", lr_g, epoch*self.train_length+its) if self.args.stat == "ts" else wandb.log({'glr': lr_g}, step=epoch*self.train_length+its)
if lr_d is not None:
pstr += "dlr: {:.1e}\t".format(lr_d)
self.writer.add_scalar("lr/dlr", lr_d, epoch*self.train_length+its) if self.args.stat == "ts" else wandb.log({'dlr': lr_d}, step=epoch*self.train_length+its)
pstr += "dtime: %04d\t"%(t_data*1000)
pstr += "ntime: %04d\t"%(t_train*1000)
pstr += "mem: {:.2f} ".format(mem_cost*len(self.args.gpus))
logger.info(pstr)
def val_recording(self, epoch):
pstr_curr = "Curr info >>>> "
pstr_best = "Best info >>>> "
for name, states in self.tracker.loss_meters.items():
metric = states['val']
if metric.count > 0:
pstr_curr += "{}: {:.3f} \t".format(name, metric.avg)
if epoch != 0:
if self.args.stat == "ts":
self.writer.add_scalars(f"val/{name}", {name+"_val":metric.avg, name+"_train":states['train'].avg}, epoch*self.train_length)
else:
wandb.log({name+"_val": metric.avg, name+"_train":states['train'].avg}, step=epoch*self.train_length)
new_best_train, new_best_val = self.tracker.update_and_plot(name, epoch, self.checkpoint_path+f"{name}_{self.args.name+self.args.notes}.png")
if new_best_val:
other_tools.save_checkpoints(os.path.join(self.checkpoint_path, f"{name}.bin"), self.model, opt=None, epoch=None, lrs=None)
for k, v in self.tracker.values.items():
metric = v['val']['best']
if self.tracker.loss_meters[k]['val'].count > 0:
pstr_best += "{}: {:.3f}({:03d})\t".format(k, metric['value'], metric['epoch'])
logger.info(pstr_curr)
logger.info(pstr_best)
def test_recording(self, dict_name, value, epoch):
self.tracker.update_meter(dict_name, "test", value)
_ = self.tracker.update_values(dict_name, 'test', epoch)
@logger.catch
def main_worker(rank, world_size, args):
#os.environ['TRANSFORMERS_CACHE'] = args.data_path_1 + "hub/"
if not sys.warnoptions:
warnings.simplefilter("ignore")
dist.init_process_group(backend="nccl", rank=rank, world_size=world_size)
logger_tools.set_args_and_logger(args, rank)
other_tools.set_random_seed(args)
other_tools.print_exp_info(args)
# return one intance of trainer
trainer = __import__(f"{args.trainer}_trainer", fromlist=["something"]).CustomTrainer(args) if args.trainer != "base" else BaseTrainer(args)
logger.info("Training from scratch ...")
start_time = time.time()
for epoch in range(args.epochs+1):
if args.ddp: trainer.val_loader.sampler.set_epoch(epoch)
# trainer.val(epoch)
# if (epoch) % args.test_period == 1: trainer.val(epoch)
epoch_time = time.time()-start_time
if trainer.rank == 0: logger.info("Time info >>>> elapsed: %.2f mins\t"%(epoch_time/60)+"remain: %.2f mins"%((args.epochs/(epoch+1e-7)-1)*epoch_time/60))
if epoch != args.epochs:
if args.ddp: trainer.train_loader.sampler.set_epoch(epoch)
trainer.tracker.reset()
trainer.train(epoch)
if args.debug:
other_tools.save_checkpoints(os.path.join(trainer.checkpoint_path, f"last_{epoch}.bin"), trainer.model, opt=None, epoch=None, lrs=None)
other_tools.load_checkpoints(trainer.model, os.path.join(trainer.checkpoint_path, f"last_{epoch}.bin"), args.g_name)
#other_tools.load_checkpoints(trainer.model, "/home/s24273/datasets/hub/pretrained_vq/last_140.bin", args.g_name)
trainer.test(epoch)
if (epoch) % args.test_period == 0 and epoch !=0:
if rank == 0:
other_tools.save_checkpoints(os.path.join(trainer.checkpoint_path, f"last_{epoch}.bin"), trainer.model, opt=None, epoch=None, lrs=None)
#trainer.test(epoch)
if rank == 0:
for k, v in trainer.tracker.values.items():
if trainer.tracker.loss_meters[k]['val'].count > 0:
other_tools.load_checkpoints(trainer.model, os.path.join(trainer.checkpoint_path, f"{k}.bin"), args.g_name)
logger.info(f"inference on ckpt {k}_val_{v['val']['best']['epoch']}:")
trainer.test(v['valb']['est']['epoch'])
other_tools.record_trial(args, trainer.tracker)
wandb.log({"fid_test": trainer.tracker["fid"]["test"]["best"]})
if args.stat == "ts":
trainer.writer.close()
else:
wandb.finish()
if __name__ == "__main__":
os.environ["MASTER_ADDR"]='127.0.0.1'
os.environ["MASTER_PORT"]=f'9{dev}75'
#os.environ["TORCH_DISTRIBUTED_DEBUG"] = "DETAIL"
args = config.parse_args()
if args.ddp:
mp.set_start_method("spawn", force=True)
mp.spawn(
main_worker,
args=(len(args.gpus), args,),
nprocs=len(args.gpus),
)
else:
main_worker(0, 1, args)