def train(**kwargs):
opt._parse(kwargs)
opt.model_name = 'bfe_test'
# set random seed and cudnn benchmark
torch.manual_seed(opt.seed)
os.makedirs(opt.save_dir, exist_ok=True)
use_gpu = torch.cuda.is_available()
sys.stdout = Logger(osp.join(opt.save_dir, 'log_train.txt'))
print('=========user config==========')
pprint(opt._state_dict())
print('============end===============')
if use_gpu:
print('currently using GPU')
cudnn.benchmark = True
torch.cuda.manual_seed_all(opt.seed)
else:
print('currently using cpu')
print('initializing dataset {}'.format(opt.dataset))
dataset = data_manager.init_dataset(name=opt.dataset, mode=opt.mode)
pin_memory = True if use_gpu else False
summary_writer = SummaryWriter(osp.join(opt.save_dir, 'tensorboard_log'))
trainloader = DataLoader(ImageData(dataset.train,
TrainTransform(opt.datatype)),
sampler=RandomIdentitySampler(
dataset.train, opt.num_instances),
batch_size=opt.train_batch,
num_workers=opt.workers,
pin_memory=pin_memory,
drop_last=True)
queryloader = DataLoader(ImageData(dataset.query,
TestTransform(opt.datatype)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
galleryloader = DataLoader(ImageData(dataset.gallery,
TestTransform(opt.datatype)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
queryFliploader = DataLoader(ImageData(dataset.query,
TestTransform(opt.datatype, True)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
galleryFliploader = DataLoader(ImageData(dataset.gallery,
TestTransform(opt.datatype,
True)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
print('initializing model ...')
model = BFE(dataset.num_train_pids, 1.0, 0.33)
optim_policy = model.get_optim_policy()
if opt.pretrained_model:
state_dict = torch.load(opt.pretrained_model)['state_dict']
# state_dict = {k: v for k, v in state_dict.items() \
# if not ('reduction' in k or 'softmax' in k)}
model.load_state_dict(state_dict, False)
print('load pretrained model ' + opt.pretrained_model)
print('model size: {:.5f}M'.format(
sum(p.numel() for p in model.parameters()) / 1e6))
if use_gpu:
model = nn.DataParallel(model).cuda()
reid_evaluator = ResNetEvaluator(model)
if opt.evaluate:
reid_evaluator.evaluate(queryloader,
galleryloader,
queryFliploader,
galleryFliploader,
re_ranking=opt.re_ranking,
savefig=opt.savefig)
return
# xent_criterion = nn.CrossEntropyLoss()
xent_criterion = CrossEntropyLabelSmooth(dataset.num_train_pids)
if opt.loss == 'triplet':
embedding_criterion = TripletLoss(opt.margin)
elif opt.loss == 'lifted':
embedding_criterion = LiftedStructureLoss(hard_mining=True)
elif opt.loss == 'weight':
embedding_criterion = Margin()
def criterion(triplet_y, softmax_y, labels):
losses = [embedding_criterion(output, labels)[0] for output in triplet_y] + \
[xent_criterion(output, labels) for output in softmax_y]
loss = sum(losses)
return loss
# get optimizer
if opt.optim == "sgd":
optimizer = torch.optim.SGD(optim_policy,
lr=opt.lr,
momentum=0.9,
weight_decay=opt.weight_decay)
else:
optimizer = torch.optim.Adam(optim_policy,
lr=opt.lr,
weight_decay=opt.weight_decay)
start_epoch = opt.start_epoch
# get trainer and evaluator
reid_trainer = cls_tripletTrainer(opt, model, optimizer, criterion,
summary_writer)
def adjust_lr(optimizer, ep):
if ep < 10:
lr = opt.lr * 0.1 * (ep / 10.0) # warm_up
elif ep < 50:
lr = opt.lr * (ep // 5 + 1)
elif ep < 200:
lr = opt.lr * 10.0
elif ep < 300:
lr = opt.lr
else:
lr = opt.lr * 0.1
for p in optimizer.param_groups:
p['lr'] = lr
# start training
best_rank1 = opt.best_rank
best_epoch = 0
for epoch in range(start_epoch, opt.max_epoch):
if opt.adjust_lr:
adjust_lr(optimizer, epoch + 1)
reid_trainer.train(epoch, trainloader)
# skip if not save model
if opt.eval_step > 0 and (epoch + 1) % opt.eval_step == 0 or (
epoch + 1) == opt.max_epoch:
if opt.mode == 'class':
rank1 = test(model, queryloader)
else:
rank1 = reid_evaluator.evaluate(queryloader, galleryloader,
queryFliploader,
galleryFliploader)
is_best = rank1 > best_rank1
if is_best:
best_rank1 = rank1
best_epoch = epoch + 1
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint({
'state_dict': state_dict,
'epoch': epoch + 1
},
is_best=is_best,
save_dir=opt.save_dir,
filename='checkpoint_ep' + str(epoch + 1) +
'.pth.tar')
print('Best rank-1 {:.1%}, achived at epoch {}'.format(
best_rank1, best_epoch))
def train(**kwargs):
opt._parse(kwargs)
# set random seed and cudnn benchmark
torch.manual_seed(opt.seed)
use_gpu = torch.cuda.is_available()
sys.stdout = Logger(osp.join(opt.save_dir, 'log_train.txt'))
print('=========user config==========')
pprint(opt._state_dict())
print('============end===============')
if use_gpu:
print('currently using GPU')
cudnn.benchmark = True
torch.cuda.manual_seed_all(opt.seed)
else:
print('currently using cpu')
print('initializing dataset {}'.format(opt.dataset))
dataset = data_manager.init_dataset(name=opt.dataset)
pin_memory = True if use_gpu else False
summary_writer = SummaryWriter(osp.join(opt.save_dir, 'tensorboard_log'))
if 'triplet' in opt.model_name:
trainloader = DataLoader(
ImageData(dataset.train, TrainTransform(opt.height, opt.width)),
sampler=RandomIdentitySampler(dataset.train, opt.num_instances),
batch_size=opt.train_batch,
num_workers=opt.workers,
pin_memory=pin_memory,
drop_last=True)
else:
trainloader = DataLoader(ImageData(
dataset.train, TrainTransform(opt.height, opt.width)),
batch_size=opt.train_batch,
shuffle=True,
num_workers=opt.workers,
pin_memory=pin_memory)
queryloader = DataLoader(ImageData(dataset.query,
TestTransform(opt.height, opt.width)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
galleryloader = DataLoader(ImageData(dataset.gallery,
TestTransform(opt.height, opt.width)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
print('initializing model ...')
if opt.model_name == 'softmax' or opt.model_name == 'softmax_triplet':
model, optim_policy = get_baseline_model(dataset.num_train_pids)
elif opt.model_name == 'triplet':
model, optim_policy = get_baseline_model(num_classes=None)
print('model size: {:.5f}M'.format(
sum(p.numel() for p in model.parameters()) / 1e6))
# xent_criterion = nn.CrossEntropyLoss()
xent_criterion = CrossEntropyLabelSmooth(dataset.num_train_pids)
tri_criterion = TripletLoss(opt.margin)
def cls_criterion(cls_scores, targets):
cls_loss = xent_criterion(cls_scores, targets)
return cls_loss
def triplet_criterion(feat, targets):
triplet_loss, _, _ = tri_criterion(feat, targets)
return triplet_loss
def cls_tri_criterion(cls_scores, feat, targets):
cls_loss = xent_criterion(cls_scores, targets)
triplet_loss, _, _ = tri_criterion(feat, targets)
loss = cls_loss + triplet_loss
return loss
# get optimizer
optimizer = torch.optim.Adam(optim_policy,
lr=opt.lr,
weight_decay=opt.weight_decay)
def adjust_lr(optimizer, ep):
if ep < 20:
lr = 1e-4 * (ep + 1) / 2
elif ep < 80:
lr = 1e-3 * opt.num_gpu
elif ep < 180:
lr = 1e-4 * opt.num_gpu
elif ep < 300:
lr = 1e-5 * opt.num_gpu
elif ep < 320:
lr = 1e-5 * 0.1**((ep - 320) / 80) * opt.num_gpu
elif ep < 400:
lr = 1e-6
elif ep < 480:
lr = 1e-4 * opt.num_gpu
else:
lr = 1e-5 * opt.num_gpu
for p in optimizer.param_groups:
p['lr'] = lr
start_epoch = opt.start_epoch
if use_gpu:
model = nn.DataParallel(model).cuda()
# get trainer and evaluator
if opt.model_name == 'softmax':
reid_trainer = clsTrainer(opt, model, optimizer, cls_criterion,
summary_writer)
elif opt.model_name == 'softmax_triplet':
reid_trainer = cls_tripletTrainer(opt, model, optimizer,
cls_tri_criterion, summary_writer)
elif opt.model_name == 'triplet':
reid_trainer = tripletTrainer(opt, model, optimizer, triplet_criterion,
summary_writer)
reid_evaluator = ResNetEvaluator(model)
# start training
best_rank1 = -np.inf
best_epoch = 0
for epoch in range(start_epoch, opt.max_epoch):
if opt.step_size > 0:
adjust_lr(optimizer, epoch + 1)
reid_trainer.train(epoch, trainloader)
# skip if not save model
if opt.eval_step > 0 and (epoch + 1) % opt.eval_step == 0 or (
epoch + 1) == opt.max_epoch:
rank1 = reid_evaluator.evaluate(queryloader, galleryloader)
is_best = rank1 > best_rank1
if is_best:
best_rank1 = rank1
best_epoch = epoch + 1
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint({
'state_dict': state_dict,
'epoch': epoch + 1,
},
is_best=is_best,
save_dir=opt.save_dir,
filename='checkpoint_ep' + str(epoch + 1) +
'.pth.tar')
print('Best rank-1 {:.1%}, achived at epoch {}'.format(
best_rank1, best_epoch))
def main():
logger = logging.getLogger('global')
global criterion_xent, criterion_triplet, criterion_center
if os.path.exists(cfg.TRAIN.LOG_DIR):
shutil.rmtree(cfg.TRAIN.LOG_DIR)
os.makedirs(cfg.TRAIN.LOG_DIR)
init_log('global', logging.INFO) # log
add_file_handler('global', os.path.join(cfg.TRAIN.LOG_DIR, 'logs.txt'),
logging.INFO)
summary_writer = SummaryWriter(cfg.TRAIN.LOG_DIR) # visualise
dataset, train_loader, _, _ = build_data_loader()
model = BagReID_RESNET(dataset.num_train_bags)
criterion_xent = CrossEntropyLabelSmooth(dataset.num_train_bags,
use_gpu=cfg.CUDA)
criterion_triplet = TripletLoss(margin=cfg.TRAIN.MARGIN)
criterion_center = CenterLoss(dataset.num_train_bags,
cfg.MODEL.GLOBAL_FEATS +
cfg.MODEL.PART_FEATS,
use_gpu=cfg.CUDA)
if cfg.TRAIN.OPTIM == "sgd":
optimizer = torch.optim.SGD(model.parameters(),
lr=cfg.SOLVER.LEARNING_RATE,
momentum=cfg.SOLVER.MOMENTUM,
weight_decay=cfg.SOLVER.WEIGHT_DECAY)
else:
optimizer = torch.optim.Adam(model.parameters(),
lr=cfg.SOLVER.LEARNING_RATE,
weight_decay=cfg.SOLVER.WEIGHT_DECAY)
center_optimizer = torch.optim.SGD(criterion_center.parameters(),
lr=cfg.SOLVER.LEARNING_RATE_CENTER)
optimizers = [optimizer, center_optimizer]
schedulers = build_lr_schedulers(optimizers)
if cfg.CUDA:
model.cuda()
if torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model, device_ids=cfg.DEVICES)
logger.info("model prepare done")
# start training
for epoch in range(cfg.TRAIN.NUM_EPOCHS):
train(epoch, train_loader, model, criterion, optimizers,
summary_writer)
for scheduler in schedulers:
scheduler.step()
# skip if not save model
if cfg.TRAIN.EVAL_STEP > 0 and (epoch + 1) % cfg.TRAIN.EVAL_STEP == 0 \
or (epoch + 1) == cfg.TRAIN.NUM_EPOCHS:
if cfg.CUDA and torch.cuda.device_count() > 1:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint({
'state_dict': state_dict,
'epoch': epoch + 1
},
is_best=False,
save_dir=cfg.TRAIN.SNAPSHOT_DIR,
filename='checkpoint_ep' + str(epoch + 1) + '.pth')
def train(**kwargs):
opt._parse(kwargs)
#opt.lr=0.00002
opt.model_name = 'AlignedReid'
# set random seed and cudnn benchmark
torch.manual_seed(opt.seed)
os.makedirs(opt.save_dir, exist_ok=True)
use_gpu = torch.cuda.is_available()
sys.stdout = Logger(osp.join(opt.save_dir, 'log_train.txt'))
print('=========user config==========')
pprint(opt._state_dict())
print('============end===============')
if use_gpu:
print('currently using GPU')
cudnn.benchmark = True
torch.cuda.manual_seed_all(opt.seed)
else:
print('currently using cpu')
print('initializing dataset {}'.format(opt.dataset))
dataset = data_manager.init_dataset(name=opt.dataset, mode=opt.mode)
pin_memory = True if use_gpu else False
summary_writer = SummaryWriter(osp.join(opt.save_dir, 'tensorboard_log'))
trainloader = DataLoader(ImageData(dataset.train,
TrainTransform(opt.datatype)),
sampler=RandomIdentitySampler(
dataset.train, opt.num_instances),
batch_size=opt.train_batch,
num_workers=opt.workers,
pin_memory=pin_memory,
drop_last=True)
queryloader = DataLoader(ImageData(dataset.query,
TestTransform(opt.datatype)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
galleryloader = DataLoader(ImageData(dataset.gallery,
TestTransform(opt.datatype)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
queryFliploader = queryloader
galleryFliploader = galleryloader
# queryFliploader = DataLoader(
# ImageData(dataset.query, TestTransform(opt.datatype, True)),
# batch_size=opt.test_batch, num_workers=opt.workers,
# pin_memory=pin_memory
# )
#
# galleryFliploader = DataLoader(
# ImageData(dataset.gallery, TestTransform(opt.datatype, True)),
# batch_size=opt.test_batch, num_workers=opt.workers,
# pin_memory=pin_memory
# )
print('initializing model ...')
model = AlignedResNet50(num_classes=dataset.num_train_pids,
loss={'softmax', 'metric'},
aligned=True,
use_gpu=use_gpu)
optim_policy = model.get_optim_policy()
if opt.pretrained_model:
state_dict = torch.load(opt.pretrained_model)['state_dict']
# state_dict = {k: v for k, v in state_dict.items() \
# if not ('reduction' in k or 'softmax' in k)}
model.load_state_dict(state_dict, False)
print('load pretrained model ' + opt.pretrained_model)
print('model size: {:.5f}M'.format(
sum(p.numel() for p in model.parameters()) / 1e6))
if use_gpu:
model = nn.DataParallel(model).cuda()
reid_evaluator = AlignedEvaluator(model)
if opt.evaluate:
#rank1 = test(model, queryloader, galleryloader, use_gpu)
reid_evaluator.evaluate(queryloader,
galleryloader,
queryFliploader,
galleryFliploader,
re_ranking=opt.re_ranking,
savefig=opt.savefig,
test_distance='global')
return
# xent_criterion = nn.CrossEntropyLoss()
xent_criterion = CrossEntropyLabelSmooth(dataset.num_train_pids,
use_gpu=use_gpu)
embedding_criterion = TripletLossAlignedReID(margin=opt.margin)
# def criterion(triplet_y, softmax_y, labels):
# losses = [embedding_criterion(output, labels)[0] for output in triplet_y] + \
# [xent_criterion(output, labels) for output in softmax_y]
# loss = sum(losses)
# return loss
def criterion(outputs, features, local_features, labels):
if opt.htri_only:
if isinstance(features, tuple):
global_loss, local_loss = DeepSupervision(
embedding_criterion, features, labels)
else:
global_loss, local_loss = embedding_criterion(
features, labels, local_features)
else:
if isinstance(outputs, tuple):
xent_loss = DeepSupervision(xent_criterion, outputs, labels)
else:
xent_loss = xent_criterion(outputs, labels)
if isinstance(features, tuple):
global_loss, local_loss = DeepSupervision(
embedding_criterion, features, labels)
else:
global_loss, local_loss = embedding_criterion(
features, labels, local_features)
loss = xent_loss + global_loss + local_loss
return loss
# get optimizer
if opt.optim == "sgd":
optimizer = torch.optim.SGD(optim_policy,
lr=opt.lr,
momentum=0.9,
weight_decay=opt.weight_decay)
else:
optimizer = torch.optim.Adam(optim_policy,
lr=opt.lr,
weight_decay=opt.weight_decay)
start_epoch = opt.start_epoch
# get trainer and evaluator
reid_trainer = AlignedTrainer(opt, model, optimizer, criterion,
summary_writer)
def adjust_lr(optimizer, ep):
if ep < 50:
lr = opt.lr * (ep // 5 + 1)
elif ep < 200:
lr = opt.lr * 10
elif ep < 300:
lr = opt.lr
else:
lr = opt.lr * 0.1
for p in optimizer.param_groups:
p['lr'] = lr
# start training
best_rank1 = opt.best_rank
best_epoch = 0
print('start train......')
for epoch in range(start_epoch, opt.max_epoch):
if opt.adjust_lr:
adjust_lr(optimizer, epoch + 1)
reid_trainer.train(epoch, trainloader)
# skip if not save model
if opt.eval_step > 0 and (epoch + 1) % opt.eval_step == 0 or (
epoch + 1) == opt.max_epoch:
# just avoid out of memory during eval,and can't save the model
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint({
'state_dict': state_dict,
'epoch': epoch + 1
},
is_best=0,
save_dir=opt.save_dir,
filename='checkpoint_ep' + str(epoch + 1) +
'.pth.tar')
if opt.mode == 'class':
rank1 = test(model, queryloader)
else:
rank1 = reid_evaluator.evaluate(queryloader, galleryloader,
queryFliploader,
galleryFliploader)
#rank1 = test(model, queryloader, galleryloader, use_gpu)
is_best = rank1 > best_rank1
if is_best:
best_rank1 = rank1
best_epoch = epoch + 1
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
if is_best:
save_checkpoint({
'state_dict': state_dict,
'epoch': epoch + 1
},
is_best=is_best,
save_dir=opt.save_dir,
filename='checkpoint_ep' + str(epoch + 1) +
'.pth.tar')
print('Best rank-1 {:.1%}, achived at epoch {}'.format(
best_rank1, best_epoch))
model.load_state_dict(state_dict, False)
print('load pretrained model ' + opt.pretrained_model)
print('model size: {:.5f}M'.format(sum(p.numel() for p in model.parameters()) / 1e6))
if use_gpu:
model = nn.DataParallel(model).cuda()
reid_evaluator = ResNetEvaluator(model)
if opt.evaluate:
reid_evaluator.evaluate(queryloader, galleryloader,
queryFliploader, galleryFliploader, re_ranking=opt.re_ranking, savefig=opt.savefig)
return
#xent_criterion = nn.CrossEntropyLoss()
#print(dataset.num_train_pids)
xent_criterion = CrossEntropyLabelSmooth(dataset.num_train_pids)
xent_criterion2 = CrossEntropyLabelSmooth2(dataset.num_train_pids)
if opt.loss == 'triplet':
embedding_criterion = TripletLoss(opt.margin)
elif opt.loss == 'lifted':
embedding_criterion = LiftedStructureLoss(hard_mining=True)
elif opt.loss == 'weight':
embedding_criterion = Margin()
def criterion(a ,b, c, d, e, f, triplet_y, softmax_y, labels):
losses = [embedding_criterion(output, labels)[0] for output in triplet_y] + \
[xent_criterion(output, labels) for output in softmax_y]+\
[xent_criterion2(i, j, labels) for i, j in zip(a,a)]+\
[xent_criterion2(i, j, labels) for i, j in zip(b,b)]+\
[xent_criterion2(i, j, labels) for i, j in zip(c,c)]+\
def train(**kwargs):
opt._parse(kwargs) ##设置程序的所有参数
# set random seed and cudnn benchmark
torch.manual_seed(opt.seed)
os.makedirs(opt.save_dir, exist_ok=True)
use_gpu = torch.cuda.is_available()
sys.stdout = Logger(osp.join(opt.save_dir, 'log_train.txt'))
print('=========user config==========')
pprint(opt._state_dict())
print('============end===============')
if use_gpu:
print('currently using GPU')
cudnn.benchmark = True
torch.cuda.manual_seed_all(opt.seed)
else:
print('currently using cpu')
print('initializing dataset {}'.format(opt.dataset))
dataset = data_manager.init_dataset(name=opt.dataset, mode=opt.mode)
pin_memory = True if use_gpu else False
trainloader = DataLoader(ImageData(dataset.train,
TrainTransform(opt.datatype)),
batch_size=opt.train_batch,
num_workers=opt.workers,
pin_memory=pin_memory,
drop_last=True)
queryloader = DataLoader(ImageData(dataset.query,
TestTransform(opt.datatype)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
galleryloader = DataLoader(ImageData(dataset.gallery,
TestTransform(opt.datatype)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
queryFliploader = DataLoader(ImageData(dataset.query,
TestTransform(opt.datatype, True)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
galleryFliploader = DataLoader(ImageData(dataset.gallery,
TestTransform(opt.datatype,
True)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
print('initializing model ...')
if opt.model_name == 'softmax' or opt.model_name == 'softmax_triplet':
model = ResNetBuilder(dataset.num_train_pids, 1, True)
elif opt.model_name == 'triplet':
model = ResNetBuilder(None, 1, True)
elif opt.model_name == 'bfe':
if opt.datatype == "person":
model = BFE(dataset.num_train_pids, 1.0, 0.33)
else:
model = BFE(dataset.num_train_pids, 0.5, 0.5)
elif opt.model_name == 'ide':
model = IDE(dataset.num_train_pids)
elif opt.model_name == 'resnet':
model = Resnet(dataset.num_train_pids)
elif opt.model_name == 'strongBaseline':
model = StrongBaseline(dataset.num_train_pids)
optim_policy = model.get_optim_policy()
# update model
model = resnet18(True)
model.fc.out_features = 10
print('model size: {:.5f}M'.format(
sum(p.numel() for p in model.parameters()) / 1e6))
if use_gpu:
model = nn.DataParallel(model).cuda()
# get optimizer
if opt.optim == "sgd":
optimizer = torch.optim.SGD(optim_policy,
lr=opt.lr,
momentum=0.9,
weight_decay=opt.weight_decay)
else:
optimizer = torch.optim.Adam(optim_policy,
lr=opt.lr,
weight_decay=opt.weight_decay)
#xent_criterion = nn.CrossEntropyLoss()
criterion = CrossEntropyLabelSmooth(10)
epochs = 100
best = 0.0
b_e = 0
for e in range(epochs):
model.train()
for i, inputs in enumerate(trainloader):
imgs, pid, _ = inputs
imgs, pid = imgs.cuda(), pid.cuda()
outputs = model(imgs)
loss = criterion(outputs, pid)
optimizer.zero_grad()
loss.backward()
optimizer.step()
print(('epoch=%s \t batch loss=%s') % (e, loss.item()))
