def main(cfg):
# build model, optimizer and scheduler
model = make_model(cfg)
model = model.to(cfg.DEVICE)
if os.path.isfile(cfg.CKPT_DIR):
model.load_state_dict(torch.load(cfg.CKPT_DIR))
print(colored('Loaded checkpoint:{}'.format(cfg.CKPT_DIR), 'blue', 'on_green'))
else:
print(colored('The cfg.CKPT_DIR id not a file: {}'.format(cfg.CKPT_DIR), 'green', 'on_red'))
if cfg.USE_WANDB:
logger = Logger("MPED_RNN",
cfg,
project = cfg.PROJECT,
viz_backend="wandb"
)
else:
logger = logging.Logger("MPED_RNN")
# get dataloaders
test_dataloader = make_dataloader(cfg, 'test')
if hasattr(logger, 'run_id'):
run_id = logger.run_id
else:
run_id = 'no_wandb'
_, _, inference = build_engine(cfg)
inference(cfg, 0, model, test_dataloader, cfg.DEVICE, logger=logger, eval_kde_nll=True, test_mode=True)
def either_dataloader_dataset_to_both(data: Union[DataLoader, Dataset],
*,
batch_size=None,
eval=False,
**kwargs):
if isinstance(data, DataLoader):
dataloader = data
dataset = data.dataset
elif isinstance(data, Dataset):
dataset = data
dl_kwargs = {}
if eval:
dl_kwargs.update(
dict(batch_size=1000, shuffle=False, drop_last=False))
else:
dl_kwargs.update(dict(batch_size=128, shuffle=True))
if batch_size is not None:
dl_kwargs["batch_size"] = batch_size
dl_kwargs.update(kwargs)
dataloader = datasets.make_dataloader(data, **dl_kwargs)
else:
raise NotImplementedError()
return dataloader, dataset
def train():
logger.info('Initializing....')
cudnn.enable = True
cudnn.benchmark = True
# torch.manual_seed(1)
# torch.cuda.manual_seed(1)
write_config_into_log(cfg)
logger.info('Building model......')
if cfg.pretrained:
model = make_model(cfg)
model.load_param(cfg)
logger.info('Loaded pretrained model from {0}'.format(cfg.pretrained))
else:
model = make_model(cfg)
model.cuda()
model = torch.nn.DataParallel(model)
optimizer = torch.optim.Adam(
[{
'params': model.module.base.parameters(),
'lr': cfg.get_lr(0)[0]
}, {
'params': model.module.classifiers.parameters(),
'lr': cfg.get_lr(0)[1]
}],
weight_decay=cfg.weight_decay)
celoss = nn.CrossEntropyLoss().cuda()
softloss = SoftLoss()
sp_kd_loss = SP_KD_Loss()
criterions = [celoss, softloss, sp_kd_loss]
train_loader, val_loader = make_dataloader(cfg)
logger.info('Begin training......')
for epoch in range(cfg.start_epoch, cfg.max_epoch):
train_one_epoch(train_loader, val_loader, model, criterions, optimizer,
epoch, cfg)
total_acc = test(cfg, val_loader, model, epoch)
with open(cfg.test_log, 'a+') as f:
f.write('Epoch {0}: Acc is {1:.4f}\n'.format(epoch, total_acc))
torch.save(obj=model.state_dict(),
f=os.path.join(
cfg.snapshot_dir,
'ep{}_acc{:.4f}.pth'.format(epoch, total_acc)))
logger.info('Model saved')
def __init__(self, cfg):
self.cfg = cfg
model = RRNet(cfg).cuda(cfg.Distributed.gpu_id)
model = nn.SyncBatchNorm.convert_sync_batchnorm(model)
self.optimizer = optim.Adam(model.parameters(), lr=cfg.Train.lr)
self.lr_sch = optim.lr_scheduler.MultiStepLR(
self.optimizer, milestones=cfg.Train.lr_milestones, gamma=0.1)
self.training_loader, self.validation_loader = make_dataloader(
cfg, collate_fn='rrnet')
super(RRNetOperator, self).__init__(cfg=self.cfg,
model=model,
lr_sch=self.lr_sch)
# TODO: change it to our class
self.hm_focal_loss = FocalLossHM()
self.l1_loss = RegL1Loss()
self.main_proc_flag = cfg.Distributed.gpu_id == 0
def __init__(self, cfg):
self.cfg = cfg
#print(self.cfg.Val.threshold)
model = CenterNet(cfg).cuda(cfg.Distributed.gpu_id)
self.optimizer = optim.Adam(model.parameters(), lr=cfg.Train.lr)
self.lr_sch = optim.lr_scheduler.MultiStepLR(
self.optimizer, milestones=cfg.Train.lr_milestones, gamma=0.1)
self.training_loader, self.validation_loader = make_dataloader(
cfg, collate_fn='ctnet')
super(CenterNetOperator, self).__init__(cfg=self.cfg,
model=model,
lr_sch=self.lr_sch)
# TODO: change it to our class
self.focal_loss = FocalLossHM()
self.l1_loss = RegL1Loss()
self.main_proc_flag = cfg.Distributed.gpu_id == 0
def __init__(self, cfg):
self.cfg = cfg
model = RetinaNet(cfg).cuda(cfg.Distributed.gpu_id)
self.optimizer = optim.Adam(model.parameters(), lr=cfg.Train.lr)
self.lr_sch = optim.lr_scheduler.MultiStepLR(
self.optimizer, milestones=cfg.Train.lr_milestones, gamma=0.1)
self.training_loader, self.validation_loader = make_dataloader(cfg)
super(RetinaNetOperator, self).__init__(cfg=self.cfg,
model=model,
lr_sch=self.lr_sch)
self.anchor_maker = Anchors(sizes=(16, 64, 128))
self.anchors, self.anchors_widths, self.anchors_heights, self.anchors_ctr_x, self.anchors_ctr_y = \
self.make_anchor(cfg.Train.crop_size)
self.focal_loss = FocalLoss()
self.main_proc_flag = cfg.Distributed.gpu_id == 0
import os
import sys
from config.cfg import Cfg
from torch.backends import cudnn
sys.path.append('.')
from datasets import make_dataloader
from processor import do_inference
from model import make_model
from utils.logger import setup_logger
if __name__ == "__main__":
# with open('cfg_test.json') as f:
# cfg = json.load(f)
Cfg.freeze()
log_dir = Cfg.DATALOADER.LOG_DIR
logger = setup_logger('Extract Feats', log_dir)
logger.info("Running with config:\n{}".format(Cfg))
os.environ['CUDA_VISIBLE_DEVICES'] = Cfg.MODEL.DEVICE_ID
cudnn.benchmark = True
val_loader = make_dataloader(Cfg)
model = make_model(Cfg,255)
model.load_param(Cfg.TEST.WEIGHT)
do_inference(Cfg, model, val_loader)
if output_dir and not os.path.exists(output_dir):
os.makedirs(output_dir)
logger = setup_logger("reid_baseline", output_dir, if_train=True)
logger.info("Saving model in the path :{}".format(cfg.OUTPUT_DIR))
logger.info(args)
if args.config_file != "":
logger.info("Loaded configuration file {}".format(args.config_file))
with open(args.config_file, 'r') as cf:
config_str = "\n" + cf.read()
logger.info(config_str)
logger.info("Running with config:\n{}".format(cfg))
os.environ['CUDA_VISIBLE_DEVICES'] = cfg.MODEL.DEVICE_ID
train_loader, val_loader, num_query, num_classes = make_dataloader(cfg)
if cfg.MODEL.PRETRAIN_CHOICE == 'finetune':
model = make_model(cfg, num_class=num_classes)
model.load_param_finetune(cfg.MODEL.PRETRAIN_PATH)
print('Loading pretrained model for finetuning......')
else:
model = make_model(cfg, num_class=num_classes)
loss_func, center_criterion = make_loss(cfg, num_classes=num_classes)
optimizer, optimizer_center = make_optimizer(cfg, model, center_criterion)
scheduler = WarmupMultiStepLR(optimizer, cfg.SOLVER.STEPS,
cfg.SOLVER.GAMMA, cfg.SOLVER.WARMUP_FACTOR,
cfg.SOLVER.WARMUP_EPOCHS,
cfg.SOLVER.WARMUP_METHOD)
def main():
parser = argparse.ArgumentParser(
description="PyTorch Object Detection Training")
parser.add_argument('--gpu', default='0', type=str)
parser.add_argument(
"--config_file",
default="",
metavar="FILE",
help="path to config file",
type=str,
)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
# build model, optimizer and scheduler
model = make_model(cfg)
model = model.to(cfg.DEVICE)
optimizer = build_optimizer(cfg, model)
print('optimizer built!')
# NOTE: add separate optimizers to train single object predictor and interaction predictor
if cfg.USE_WANDB:
logger = Logger("FOL", cfg, project=cfg.PROJECT, viz_backend="wandb")
else:
logger = logging.Logger("FOL")
dataloader_params = {
"batch_size": cfg.SOLVER.BATCH_SIZE,
"shuffle": True,
"num_workers": cfg.DATALOADER.NUM_WORKERS
}
# get dataloaders
train_dataloader = make_dataloader(cfg, 'train')
val_dataloader = make_dataloader(cfg, 'val')
test_dataloader = make_dataloader(cfg, 'test')
print('Dataloader built!')
# get train_val_test engines
do_train, do_val, inference = build_engine(cfg)
print('Training engine built!')
if hasattr(logger, 'run_id'):
run_id = logger.run_id
else:
run_id = 'no_wandb'
save_checkpoint_dir = os.path.join(cfg.CKPT_DIR, run_id)
if not os.path.exists(save_checkpoint_dir):
os.makedirs(save_checkpoint_dir)
# NOTE: hyperparameter scheduler
model.param_scheduler = ParamScheduler()
model.param_scheduler.create_new_scheduler(name='kld_weight',
annealer=sigmoid_anneal,
annealer_kws={
'device': cfg.DEVICE,
'start': 0,
'finish': 100.0,
'center_step': 400.0,
'steps_lo_to_hi': 100.0,
})
model.param_scheduler.create_new_scheduler(name='z_logit_clip',
annealer=sigmoid_anneal,
annealer_kws={
'device': cfg.DEVICE,
'start': 0.05,
'finish': 5.0,
'center_step': 300.0,
'steps_lo_to_hi': 300.0 / 5.
})
if cfg.SOLVER.scheduler == 'exp':
# exponential schedule
lr_scheduler = optim.lr_scheduler.ExponentialLR(optimizer,
gamma=cfg.SOLVER.GAMMA)
elif cfg.SOLVER.scheduler == 'plateau':
# Plateau scheduler
lr_scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
factor=0.2,
patience=5,
min_lr=1e-07,
verbose=1)
else:
lr_scheduler = optim.lr_scheduler.MultiStepLR(optimizer,
milestones=[25, 40],
gamma=0.2)
print('Schedulers built!')
for epoch in range(cfg.SOLVER.MAX_EPOCH):
logger.info("Epoch:{}".format(epoch))
do_train(cfg,
epoch,
model,
optimizer,
train_dataloader,
cfg.DEVICE,
logger=logger,
lr_scheduler=lr_scheduler)
val_loss = do_val(cfg,
epoch,
model,
val_dataloader,
cfg.DEVICE,
logger=logger)
if (epoch + 1) % 1 == 0:
inference(cfg,
epoch,
model,
test_dataloader,
cfg.DEVICE,
logger=logger,
eval_kde_nll=False)
torch.save(
model.state_dict(),
os.path.join(save_checkpoint_dir,
'Epoch_{}.pth'.format(str(epoch).zfill(3))))
# update LR
if cfg.SOLVER.scheduler != 'exp':
lr_scheduler.step(val_loss)
if output_dir and not os.path.exists(output_dir):
os.makedirs(output_dir)
logger = setup_logger("reid_baseline", output_dir, if_train=True)
logger.info("Saving model in the path :{}".format(cfg.OUTPUT_DIR))
logger.info(args)
if args.config_file != "":
logger.info("Loaded configuration file {}".format(args.config_file))
with open(args.config_file, 'r') as cf:
config_str = "\n" + cf.read()
logger.info(config_str)
logger.info("Running with config:\n{}".format(cfg))
os.environ['CUDA_VISIBLE_DEVICES'] = cfg.MODEL.DEVICE_ID
train_loader, val_loader_green, val_loader_normal, num_query_green, num_query_normal, num_classes = make_dataloader(
cfg)
if cfg.MODEL.PRETRAIN_CHOICE == 'finetune':
model = make_model(cfg, num_class=num_classes)
model.load_param_finetune(cfg.MODEL.PRETRAIN_PATH)
print('Loading pretrained model for finetuning......')
else:
model = make_model(cfg, num_class=num_classes)
loss_func, center_criterion = make_loss(cfg, num_classes=num_classes)
optimizer, optimizer_center = make_optimizer(cfg, model, center_criterion)
if cfg.SOLVER.TYPE == 'warmup':
scheduler = WarmupMultiStepLR(optimizer, cfg.SOLVER.STEPS,
cfg.SOLVER.GAMMA,
cfg.SOLVER.WARMUP_FACTOR,
from loss import make_loss
from processor import do_train
if __name__ == '__main__':
Cfg.freeze()
log_dir = Cfg.DATALOADER.LOG_DIR
logger = setup_logger('pose-transfer-gan.train', log_dir)
logger.info("Running with config:\n{}".format(Cfg))
os.environ['CUDA_VISIBLE_DEVICES'] = Cfg.MODEL.DEVICE_ID
cudnn.benchmark = True
# This flag allows you to enable the inbuilt cudnn auto-tuner to find the best algorithm to use for your hardware.
train_loader, val_loader = make_dataloader(Cfg)
model_G, model_Dip, model_Dii, model_D_reid = make_model(Cfg)
optimizerG = make_optimizer(Cfg, model_G)
optimizerDip = make_optimizer(Cfg, model_Dip)
optimizerDii = make_optimizer(Cfg, model_Dii)
schedulerG = WarmupMultiStepLR(optimizerG, Cfg.SOLVER.STEPS,
Cfg.SOLVER.GAMMA, Cfg.SOLVER.WARMUP_FACTOR,
Cfg.SOLVER.WARMUP_EPOCHS,
Cfg.SOLVER.WARMUP_METHOD)
schedulerDip = WarmupMultiStepLR(optimizerDip, Cfg.SOLVER.STEPS,
Cfg.SOLVER.GAMMA,
Cfg.SOLVER.WARMUP_FACTOR,
Cfg.SOLVER.WARMUP_EPOCHS,
Cfg.SOLVER.WARMUP_METHOD)
def do_train(cfg, model, center_criterion, train_loader, val_loader, optimizer,
optimizer_center, scheduler, loss_fn, num_query):
log_period = cfg.SOLVER.LOG_PERIOD
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
eval_period = cfg.SOLVER.EVAL_PERIOD
device = "cuda"
epochs = cfg.SOLVER.MAX_EPOCHS
logger = logging.getLogger("reid_baseline.train")
logger.info('start training')
print("torch.cuda.device_count()", torch.cuda.device_count())
if device:
model.to(device)
if torch.cuda.device_count() > 1:
print('Using {} GPUs for training'.format(
torch.cuda.device_count()))
print("多卡训练")
# model = DDP(model, delay_allreduce=True) # 必须在initialze之后
# model = nn.DataParallel(model)
# model, optimizer = amp.initialize(model, optimizer, opt_level="O1") # 字母小写o,不是零。
torch.distributed.init_process_group(
'gloo',
init_method='file:///tmp/somefile',
rank=0,
world_size=1)
# model = convert_syncbn_model(model)
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
# model = DistributedDataParallel(model, delay_allreduce=True)
# model = torch.nn.parallel.DistributedDataParallel(model, find_unused_parameters=True)
model = nn.DataParallel(model)
# model = convert_syncbn_model(model)
else:
print("单卡训练")
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
model.to(device=0)
loss_meter = AverageMeter()
acc_meter = AverageMeter()
# evaluator = R1_mAP_eval(num_query, max_rank=50, feat_norm=cfg.TEST.FEAT_NORM)
# model.base._freeze_stages()
logger.info('Freezing the stages number:{}'.format(cfg.MODEL.FROZEN))
# model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
for epoch in range(1, epochs + 1):
if epoch == 5:
print("balance 数据训练")
# cfg.DATASETS.ROOT_DIR = '/home/lab3/bi/0716/Veri/ai_city/tools/mix_train_balance_flip.pkl'
cfg.DATASETS.ROOT_DIR = 'datasets/mix_train_balance.pkl'
train_loader, val_loader, num_query, num_classes = make_dataloader(
cfg)
# model.base._freeze_stages()
start_time = time.time()
loss_meter.reset()
acc_meter.reset()
# evaluator.reset()
scheduler.step()
model.train()
# print(scheduler.get_lr()[0])
for n_iter, (img, vid) in enumerate(tqdm(train_loader)):
optimizer.zero_grad()
optimizer_center.zero_grad()
img = img.to(device)
target = vid.to(device)
#grid mask
# img = grid(img)
# score, feat,score_f1,score_f2,score_f3,f4,f4_score = model(img, target)
# score, feat,score_f1,score_f2,feat1,score_layer2 = model(img, target)
score, feat, score_f1, score_f2, feat1 = model(img, target)
# print(feat.shape)
loss = loss_fn(score, feat, target, score_f1, score_f2, feat1)
# loss = loss_fn(score, feat, target,score_f1,score_f2,feat1,score_layer2)
if cfg.SOLVER.FP16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
# scaled_loss.backward(retain_graph=True)
else:
loss.backward()
# loss.backward()
optimizer.step()
if 'center' in cfg.MODEL.METRIC_LOSS_TYPE:
for param in center_criterion.parameters():
param.grad.data *= (1. / cfg.SOLVER.CENTER_LOSS_WEIGHT)
optimizer_center.step()
acc = (score.max(1)[1] == target).float().mean()
loss_meter.update(loss.item(), img.shape[0])
acc_meter.update(acc, 1)
# print(loss_meter.val)
if (n_iter + 1) % log_period == 0:
logger.info(
"Epoch[{}] Iteration[{}/{}] Loss: {:.3f}, Acc: {:.3f}, Base Lr: {:.2e}"
.format(epoch, (n_iter + 1), len(train_loader),
loss_meter.avg, acc_meter.avg,
scheduler.get_lr()[0]))
end_time = time.time()
time_per_batch = (end_time - start_time) / (n_iter + 1)
logger.info(
"Epoch {} done. Time per batch: {:.3f}[s] Speed: {:.1f}[samples/s]"
.format(epoch, time_per_batch,
train_loader.batch_size / time_per_batch))
if epoch % checkpoint_period == 0:
torch.save(
model.state_dict(),
os.path.join(cfg.OUTPUT_DIR,
cfg.MODEL.NAME + '_epoch{}.pth'.format(epoch)))
if epoch == 10:
reduce_model_dict = model.half().state_dict()
del_keys = []
for key in reduce_model_dict.keys():
if 'class' in key or 'sub1' in key or 'sub2' in key or 'base.fc' in key:
del_keys.append(key)
for key in del_keys:
del reduce_model_dict[key]
torch.save(
reduce_model_dict,
os.path.join(
cfg.OUTPUT_DIR, cfg.MODEL.NAME +
str(cfg.INPUT.SIZE_TRAIN[0]) + 'half.pth'))
logger = setup_logger("reid_baseline", output_dir, if_train=False)
logger.info(args)
if args.config_file != "":
logger.info("Loaded configuration file {}".format(args.config_file))
with open(args.config_file, 'r') as cf:
config_str = "\n" + cf.read()
logger.info(config_str)
logger.info("Running with config:\n{}".format(cfg))
os.environ['CUDA_VISIBLE_DEVICES'] = cfg.MODEL.DEVICE_ID
if cfg.TEST.FLIP_FEATS != 'on':
train_loader, val_loader, num_query_normal, num_classes,val_loader_center,val_loader_lb,val_loader_rb,val_loader_rt,val_loader_lt = make_dataloader(cfg)
val_loader_normal = [val_loader,val_loader_center,val_loader_lt,val_loader_rt,val_loader_lb,val_loader_rb]
model = make_model(cfg, num_class=num_classes)
model.load_param(cfg.TEST.WEIGHT)
do_inference_multi(cfg,
model,
val_loader_normal,
num_query_normal)
else:
train_loader, val_loader_normal, num_query_normal, num_classes = make_dataloader(cfg)
model = make_model(cfg, num_class=num_classes)
model.load_param(cfg.TEST.WEIGHT)
do_inference(cfg,
# model_name = 'MixNet'
# model_path = ' '
model = Baseline(model='train', model_name=model_name, model_path=model_path)
#model.load_param('models/model_1_180000.pth')
model = model.cuda()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
#exp_lr_scheduler = lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.1)
# kd_id = 0
# kd_num = 7
# batch_size = 48
# instance_num = 1
train_data, val_data, trains, vals = make_dataloader(kd_id, kd_num)
train_loader = DataLoader(dataset=train_data,
batch_size=batch_size,
sampler=RandomSampler(trains, batch_size,
instance_num),
shuffle=False,
num_workers=2,
collate_fn=train_collate)
#train_loader = DataLoader(dataset=train_data, batch_size=48, shuffle=False, num_workers=2, collate_fn=train_collate)
val_loader = DataLoader(dataset=val_data,
batch_size=64,
shuffle=False,
num_workers=2,
collate_fn=train_collate)
train_length = len(train_loader)
val_length = len(val_loader)
from datasets import make_dataloader
from configs.kmeans_config import Config
from tqdm import tqdm
import torch
from ext.kmeans.kmeans import lloyd
train_loader, val_loader = make_dataloader(Config)
all_w = []
all_h = []
all_d = []
with torch.no_grad():
for i, batch in enumerate(tqdm(train_loader)):
annos = batch[1][0]
# d = (annos[:, 2].pow(2) + annos[:, 3].pow(2)).sqrt()
all_w.append(annos[:, 3].clone())
all_h.append(annos[:, 2].clone())
all_w = torch.cat(all_w).unsqueeze(1)
all_h = torch.cat(all_h).unsqueeze(1)
h_results = lloyd(all_h, 3)
print(h_results) # 20.3807, 73.2261, 182.68274
w_results = lloyd(all_w, 3)
print(w_results) # 21.9839, 63.8345, 155.8799
os.makedirs(output_dir)
logger = setup_logger("reid_baseline", output_dir, if_train=True)
logger.info("Saving model in the path :{}".format(cfg.OUTPUT_DIR))
logger.info(args)
if args.config_file != "":
logger.info("Loaded configuration file {}".format(args.config_file))
with open(args.config_file, 'r') as cf:
config_str = "\n" + cf.read()
logger.info(config_str)
logger.info("Running with config:\n{}".format(cfg))
os.environ['CUDA_VISIBLE_DEVICES'] = cfg.MODEL.DEVICE_ID
#train_loader, val_loader_green, val_loader_normal, num_query_green, num_query_normal, num_classes = make_dataloader(cfg)
train_loader, num_classes = make_dataloader(cfg)
val_loader_green = None
num_query_green = None
if cfg.MODEL.PRETRAIN_CHOICE == 'finetune':
model = make_model(cfg, num_class=num_classes)
model.load_param_finetune(cfg.MODEL.PRETRAIN_PATH)
print('Loading pretrained model for finetuning......')
#print(model.base.state_dict()['conv1.weight'])
for param in model.parameters(): # 冻结参数
param.requires_grad = True
print("参数解冻")
else:
model = make_model(cfg, num_class=num_classes)
loss_func, center_criterion = make_loss(cfg, num_classes=num_classes)
def train():
if args.local_rank == 0:
logger.info('Initializing....')
cudnn.enable = True
cudnn.benchmark = True
# torch.manual_seed(1)
# torch.cuda.manual_seed(1)
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
args.gpu = 0
args.world_size = 1
if args.distributed:
args.gpu = args.local_rank
torch.cuda.set_device(args.gpu)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
args.world_size = torch.distributed.get_world_size()
if args.local_rank == 0:
write_config_into_log(cfg)
if args.local_rank == 0:
logger.info('Building model......')
if cfg.pretrained:
model = make_model(cfg)
model.load_param(cfg)
if args.local_rank == 0:
logger.info('Loaded pretrained model from {0}'.format(
cfg.pretrained))
else:
model = make_model(cfg)
if args.sync_bn:
if args.local_rank == 0: logging.info("using apex synced BN")
model = convert_syncbn_model(model)
model.cuda()
if args.distributed:
# By default, apex.parallel.DistributedDataParallel overlaps communication with
# computation in the backward pass.
# delay_allreduce delays all communication to the end of the backward pass.
model = DistributedDataParallel(model, delay_allreduce=True)
else:
model = torch.nn.DataParallel(model)
optimizer = torch.optim.Adam(
[{
'params': model.module.base.parameters(),
'lr': cfg.get_lr(0)[0]
}, {
'params': model.module.classifiers.parameters(),
'lr': cfg.get_lr(0)[1]
}],
weight_decay=cfg.weight_decay)
celoss = nn.CrossEntropyLoss().cuda()
softloss = SoftLoss()
sp_kd_loss = SP_KD_Loss()
criterions = [celoss, softloss, sp_kd_loss]
cfg.batch_size = cfg.batch_size // args.world_size
cfg.num_workers = cfg.num_workers // args.world_size
train_loader, val_loader = make_dataloader(cfg)
if args.local_rank == 0:
logger.info('Begin training......')
for epoch in range(cfg.start_epoch, cfg.max_epoch):
train_one_epoch(train_loader, val_loader, model, criterions, optimizer,
epoch, cfg)
total_acc = test(cfg, val_loader, model, epoch)
if args.local_rank == 0:
with open(cfg.test_log, 'a+') as f:
f.write('Epoch {0}: Acc is {1:.4f}\n'.format(epoch, total_acc))
torch.save(obj=model.state_dict(),
f=os.path.join(
cfg.snapshot_dir,
'ep{}_acc{:.4f}.pth'.format(epoch, total_acc)))
logger.info('Model saved')
if output_dir and not os.path.exists(output_dir):
os.makedirs(output_dir)
logger = setup_logger("transreid", output_dir, if_train=False)
logger.info(args)
if args.config_file != "":
logger.info("Loaded configuration file {}".format(args.config_file))
with open(args.config_file, 'r') as cf:
config_str = "\n" + cf.read()
logger.info(config_str)
logger.info("Running with config:\n{}".format(cfg))
os.environ['CUDA_VISIBLE_DEVICES'] = cfg.MODEL.DEVICE_ID
train_loader, train_loader_normal, val_loader, num_query, num_classes, camera_num, view_num = make_dataloader(
cfg)
model = make_model(cfg,
num_class=num_classes,
camera_num=camera_num,
view_num=view_num)
model.load_param(cfg.TEST.WEIGHT)
if cfg.DATASETS.NAMES == 'VehicleID':
for trial in range(10):
train_loader, train_loader_normal, val_loader, num_query, num_classes, camera_num, view_num = make_dataloader(
cfg)
rank_1, rank5 = do_inference(cfg, model, val_loader, num_query)
if trial == 0:
all_rank_1 = rank_1
all_rank_5 = rank5
import os
import sys
from config.config import Configuration
import torch
from torch.backends import cudnn
sys.path.append('.')
from datasets import make_dataloader
from processor import do_inference
from model import make_model
from utils.logger import setup_logger
if __name__ == "__main__":
Cfg = Configuration()
log_dir = Cfg.LOG_DIR
logger = setup_logger('{}.test'.format(Cfg.PROJECT_NAME), log_dir)
os.environ['CUDA_VISIBLE_DEVICES'] = Cfg.DEVICE_ID
cudnn.benchmark = True
# This flag allows you to enable the inbuilt cudnn auto-tuner to find the best algorithm to use for your hardware.
train_loader, test_loader = make_dataloader(Cfg)
model = make_model(Cfg)
model.load_state_dict(torch.load(Cfg.TEST_WEIGHT))
do_inference(Cfg, model, test_loader)
