def train(train_loader, model, optimizer, lr_scheduler, epoch, cfg):
global tb_index, best_acc, cur_lr, logger
cur_lr = lr_scheduler.get_cur_lr()
logger = logging.getLogger('global')
avg = AverageMeter()
model.train()
model = model.cuda()
end = time.time()
def is_valid_number(x):
return not (math.isnan(x) or math.isinf(x) or x > 1e4)
num_per_epoch = len(train_loader.dataset) // args.epochs // args.batch
start_epoch = epoch
epoch = epoch
for iter, input in enumerate(train_loader):
if epoch != iter // num_per_epoch + start_epoch: # next epoch
epoch = iter // num_per_epoch + start_epoch
if not os.path.exists(args.save_dir): # makedir/save model
os.makedirs(args.save_dir)
save_checkpoint(
{
'epoch': epoch,
'arch': args.arch,
'state_dict': model.module.state_dict(),
'best_acc': best_acc,
'optimizer': optimizer.state_dict(),
'anchor_cfg': cfg['anchors']
}, False,
os.path.join(args.save_dir, 'checkpoint_e%d.pth' % (epoch)),
os.path.join(args.save_dir, 'best.pth'))
if epoch == args.epochs:
return
if model.module.features.unfix(epoch / args.epochs):
logger.info('unfix part model.')
optimizer, lr_scheduler = build_opt_lr(model.module, cfg, args,
epoch)
lr_scheduler.step(epoch)
cur_lr = lr_scheduler.get_cur_lr()
logger.info('epoch:{}'.format(epoch))
tb_index = iter
if iter % num_per_epoch == 0 and iter != 0:
for idx, pg in enumerate(optimizer.param_groups):
logger.info("epoch {} lr {}".format(epoch, pg['lr']))
tb_writer.add_scalar('lr/group%d' % (idx + 1), pg['lr'],
tb_index)
data_time = time.time() - end
avg.update(data_time=data_time)
x = {
'cfg': cfg,
'template': torch.autograd.Variable(input[0]).cuda(),
'search': torch.autograd.Variable(input[1]).cuda(),
'label_cls': torch.autograd.Variable(input[2]).cuda(),
'label_loc': torch.autograd.Variable(input[3]).cuda(),
'label_loc_weight': torch.autograd.Variable(input[4]).cuda(),
'label_mask': torch.autograd.Variable(input[6]).cuda(),
'label_mask_weight': torch.autograd.Variable(input[7]).cuda(),
}
outputs = model(x)
rpn_cls_loss, rpn_loc_loss, rpn_mask_loss = torch.mean(
outputs['losses'][0]), torch.mean(
outputs['losses'][1]), torch.mean(outputs['losses'][2])
mask_iou_mean, mask_iou_at_5, mask_iou_at_7 = torch.mean(
outputs['accuracy'][0]), torch.mean(
outputs['accuracy'][1]), torch.mean(outputs['accuracy'][2])
cls_weight, reg_weight, mask_weight = cfg['loss']['weight']
loss = rpn_cls_loss * cls_weight + rpn_loc_loss * reg_weight + rpn_mask_loss * mask_weight
optimizer.zero_grad()
loss.backward()
if cfg['clip']['split']:
torch.nn.utils.clip_grad_norm_(model.module.features.parameters(),
cfg['clip']['feature'])
torch.nn.utils.clip_grad_norm_(model.module.rpn_model.parameters(),
cfg['clip']['rpn'])
torch.nn.utils.clip_grad_norm_(
model.module.mask_model.parameters(), cfg['clip']['mask'])
else:
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.clip) # gradient clip
if is_valid_number(loss.item()):
optimizer.step()
siammask_loss = loss.item()
batch_time = time.time() - end
avg.update(batch_time=batch_time,
rpn_cls_loss=rpn_cls_loss,
rpn_loc_loss=rpn_loc_loss,
rpn_mask_loss=rpn_mask_loss,
siammask_loss=siammask_loss,
mask_iou_mean=mask_iou_mean,
mask_iou_at_5=mask_iou_at_5,
mask_iou_at_7=mask_iou_at_7)
tb_writer.add_scalar('loss/cls', rpn_cls_loss, tb_index)
tb_writer.add_scalar('loss/loc', rpn_loc_loss, tb_index)
tb_writer.add_scalar('loss/mask', rpn_mask_loss, tb_index)
tb_writer.add_scalar('mask/mIoU', mask_iou_mean, tb_index)
tb_writer.add_scalar('mask/[email protected]', mask_iou_at_5, tb_index)
tb_writer.add_scalar('mask/[email protected]', mask_iou_at_7, tb_index)
end = time.time()
if (iter + 1) % args.print_freq == 0:
logger.info(
'Epoch: [{0}][{1}/{2}] lr: {lr:.6f}\t{batch_time:s}\t{data_time:s}'
'\t{rpn_cls_loss:s}\t{rpn_loc_loss:s}\t{rpn_mask_loss:s}\t{siammask_loss:s}'
'\t{mask_iou_mean:s}\t{mask_iou_at_5:s}\t{mask_iou_at_7:s}'.
format(epoch + 1, (iter + 1) % num_per_epoch,
num_per_epoch,
lr=cur_lr,
batch_time=avg.batch_time,
data_time=avg.data_time,
rpn_cls_loss=avg.rpn_cls_loss,
rpn_loc_loss=avg.rpn_loc_loss,
rpn_mask_loss=avg.rpn_mask_loss,
siammask_loss=avg.siammask_loss,
mask_iou_mean=avg.mask_iou_mean,
mask_iou_at_5=avg.mask_iou_at_5,
mask_iou_at_7=avg.mask_iou_at_7))
print_speed(iter + 1, avg.batch_time.avg,
args.epochs * num_per_epoch)
def train(train_loader, model, optimizer, lr_scheduler, epoch, cfg):
global tb_index, best_acc, cur_lr, logger
cur_lr = lr_scheduler.get_cur_lr()
logger = logging.getLogger('global')
avg = AverageMeter()
model.train()
# model.module.features.eval()
# model.module.rpn_model.eval()
# model.module.features.apply(BNtoFixed)
# model.module.rpn_model.apply(BNtoFixed)
#
# model.module.mask_model.train()
# model.module.refine_model.train()
model = model.cuda()
end = time.time()
def is_valid_number(x):
return not (math.isnan(x) or math.isinf(x) or x > 1e4)
num_per_epoch = len(train_loader.dataset) // args.epochs // args.batch
start_epoch = epoch
epoch = epoch
with torch.no_grad():
for iter, input in enumerate(train_loader):
if iter > 100:
break
if epoch != iter // num_per_epoch + start_epoch: # next epoch
epoch = iter // num_per_epoch + start_epoch
if epoch == args.epochs:
return
if model.module.features.unfix(epoch / args.epochs):
logger.info('unfix part model.')
optimizer, lr_scheduler = build_opt_lr(
model.module, cfg, args, epoch)
lr_scheduler.step(epoch)
cur_lr = lr_scheduler.get_cur_lr()
logger.info('epoch:{}'.format(epoch))
tb_index = iter
if iter % num_per_epoch == 0 and iter != 0:
for idx, pg in enumerate(optimizer.param_groups):
logger.info("epoch {} lr {}".format(epoch, pg['lr']))
tb_writer.add_scalar('lr/group%d' % (idx + 1), pg['lr'],
tb_index)
data_time = time.time() - end
avg.update(data_time=data_time)
x_rpn = {
'cfg': cfg,
'template': torch.autograd.Variable(input[0]).cuda(),
'search': torch.autograd.Variable(input[1]).cuda(),
'label_cls': torch.autograd.Variable(input[2]).cuda(),
'label_loc': torch.autograd.Variable(input[3]).cuda(),
'label_loc_weight': torch.autograd.Variable(input[4]).cuda(),
'label_mask': torch.autograd.Variable(input[6]).cuda()
}
x_kp = input[7]
x_kp = {
x: torch.autograd.Variable(y).cuda()
for x, y in x_kp.items()
}
x_rpn['anchors'] = train_loader.dataset.anchors.all_anchors[0]
outputs = model(x_rpn, x_kp)
roi_box = outputs['predict'][-1]
pred_kp = outputs['predict'][2]['hm_hp']
batch_img = x_rpn['search'].expand(x_kp['hm_hp'].size(0), -1, -1,
-1)
gt_img, pred_img = save_gt_pred_heatmaps(
batch_img, x_kp['hm_hp'], pred_kp,
'test_imgs/test_{}.jpg'.format(iter))
# rpn_pred_cls, rpn_pred_loc = outputs['predict'][:2]
# rpn_pred_cls = outputs['predict'][-1]
# anchors = train_loader.dataset.anchors.all_anchors[0]
#
# normalized_boxes = proposal_layer([rpn_pred_cls, rpn_pred_loc], anchors, config=cfg)
# print('rpn_pred_cls: ', rpn_pred_cls.shape)
rpn_cls_loss, rpn_loc_loss, kp_losses = torch.mean(outputs['losses'][0]),\
torch.mean(outputs['losses'][1]),\
outputs['losses'][3]
kp_loss = torch.mean(kp_losses['loss'])
kp_hp_loss = torch.mean(kp_losses['hp_loss'])
kp_hm_hp_loss = torch.mean(kp_losses['hm_hp_loss'])
kp_hp_offset_loss = torch.mean(kp_losses['hp_offset_loss'])
# mask_iou_mean, mask_iou_at_5, mask_iou_at_7 = torch.mean(outputs['accuracy'][0]), torch.mean(outputs['accuracy'][1]), torch.mean(outputs['accuracy'][2])
cls_weight, reg_weight, kp_weight = cfg['loss']['weight']
loss = rpn_cls_loss * cls_weight + rpn_loc_loss * reg_weight + kp_loss * kp_weight
optimizer.zero_grad()
loss.backward()
if cfg['clip']['split']:
torch.nn.utils.clip_grad_norm_(
model.module.features.parameters(), cfg['clip']['feature'])
torch.nn.utils.clip_grad_norm_(
model.module.rpn_model.parameters(), cfg['clip']['rpn'])
torch.nn.utils.clip_grad_norm_(
model.module.mask_model.parameters(), cfg['clip']['mask'])
else:
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.clip) # gradient clip
if is_valid_number(loss.item()):
optimizer.step()
siammask_loss = loss.item()
batch_time = time.time() - end
avg.update(batch_time=batch_time,
rpn_cls_loss=rpn_cls_loss,
rpn_loc_loss=rpn_loc_loss,
kp_hp_loss=kp_hp_loss,
kp_hm_hp_loss=kp_hm_hp_loss,
kp_hp_offset_loss=kp_hp_offset_loss,
kp_loss=kp_loss,
siammask_loss=siammask_loss)
# mask_iou_mean=mask_iou_mean, mask_iou_at_5=mask_iou_at_5, mask_iou_at_7=mask_iou_at_7)
tb_writer.add_scalar('loss/cls', rpn_cls_loss, tb_index)
tb_writer.add_scalar('loss/loc', rpn_loc_loss, tb_index)
tb_writer.add_scalar('loss/kp_hp_loss', kp_hp_loss, tb_index)
tb_writer.add_scalar('loss/kp_hm_hp_loss', kp_hm_hp_loss, tb_index)
tb_writer.add_scalar('loss/kp_hp_offset_loss', kp_hp_offset_loss,
tb_index)
# tb_writer.add_scalar('loss/kp', kp_loss, tb_index)
end = time.time()
if (iter + 1) % args.print_freq == 0:
logger.info(
'Epoch: [{0}][{1}/{2}] lr: {lr:.6f}\t{batch_time:s}\t{data_time:s}'
'\t{rpn_cls_loss:s}\t{rpn_loc_loss:s}'
'\t{kp_hp_loss:s}\t{kp_hm_hp_loss:s}\t{kp_hp_offset_loss:s}'
'\t{kp_loss:s}\t{siammask_loss:s}'.format(
epoch + 1,
(iter + 1) % num_per_epoch,
num_per_epoch,
lr=cur_lr,
batch_time=avg.batch_time,
data_time=avg.data_time,
rpn_cls_loss=avg.rpn_cls_loss,
rpn_loc_loss=avg.rpn_loc_loss,
kp_hp_loss=avg.kp_hp_loss,
kp_hm_hp_loss=avg.kp_hm_hp_loss,
kp_hp_offset_loss=avg.kp_hp_offset_loss,
kp_loss=avg.kp_loss,
siammask_loss=avg.siammask_loss,
))
# mask_iou_mean=avg.mask_iou_mean,
# mask_iou_at_5=avg.mask_iou_at_5,mask_iou_at_7=avg.mask_iou_at_7))
print_speed(iter + 1, avg.batch_time.avg,
args.epochs * num_per_epoch)
def validation(epoch, log_interval, test_dataloader, model, loss, writer,
device):
"""Validate on test dataset.
Current validation is only for loss, pos|neg_distance.
In future, we will add more validation like MAP5|10|50|100.
(maybe in another file.)
Args:
log_interval:
How many time will the logger log once.
test_dataloader:
It should not be none! A Triplet dataloader to validate data.
model:
The model that used to test on dataset.
loss:
Loss metric.
writer:
Tensorboard writer
device:
Device that model compute on
Return:
epoch avrage value:
triplet_loss, pos_dists, neg_dists
"""
logger.info(
"\n------------------------- Start validation -------------------------\n"
)
# epoch average meter
avg_test = AverageMeter()
# get test batch count
current_test_batch = 0
total_test_batch = len(test_dataloader)
# check dataloader is not None
assert test_dataloader is not None, "test_dataloader should not be None."
for batch_idx, batch_sample in enumerate(test_dataloader):
# Skip last iteration to avoid the problem of having different number of tensors while calculating
# averages (sizes of tensors must be the same for pairwise distance calculation)
if batch_idx + 1 == len(test_dataloader):
continue
# switch to evaluation mode.
for param in model.parameters():
param.requires_grad = False
model.eval()
# start time counting
batch_start_time_test = time.time()
# Forward pass - compute embeddings
anc_imgs = batch_sample['anchor_img']
pos_imgs = batch_sample['pos_img']
neg_imgs = batch_sample['neg_img']
pos_cls = batch_sample['pos_cls']
neg_cls = batch_sample['neg_cls']
# move to device
anc_imgs = anc_imgs.to(device)
pos_imgs = pos_imgs.to(device)
neg_imgs = neg_imgs.to(device)
pos_cls = pos_cls.to(device)
neg_cls = neg_cls.to(device)
# forward
output = model.forward_triplet(anc_imgs, pos_imgs, neg_imgs)
# get output
anc_emb = output['anchor_map']
pos_emb = output['pos_map']
neg_emb = output['neg_map']
pos_dists = torch.mean(output['dist_pos'])
neg_dists = torch.mean(output['dist_neg'])
# loss compute
loss_value = loss(anc_emb, pos_emb, neg_emb)
# batch time & batch count
current_test_batch += 1
batch_time = time.time() - batch_start_time_test
# update avg
avg_test.update(time=batch_time,
triplet_loss=loss_value,
pos_dists=pos_dists,
neg_dists=neg_dists)
if current_test_batch % log_interval == 0:
print_speed(current_test_batch, batch_time, total_test_batch,
"global")
logger.info(
"\n current global average information:\n batch_time {0:.5f} | triplet_loss: {1:.5f} | pos_dists: {2:.5f} | neg_dists: {3:.5f} \n"
.format(avg_test.time.avg, avg_test.triplet_loss.avg,
avg_test.pos_dists.avg, avg_test.neg_dists.avg))
else:
writer.add_scalar("Validate/Loss/train",
avg_test.triplet_loss.avg,
global_step=epoch)
writer.add_scalar("Validate/Other/pos_dists",
avg_test.pos_dists.avg,
global_step=epoch)
writer.add_scalar("Validate/Other/neg_dists",
avg_test.neg_dists.avg,
global_step=epoch)
return avg_test.triplet_loss.avg, avg_test.pos_dists.avg, avg_test.neg_dists.avg
epoch * train_lenth + step + 1)
writer.add_video('test_seq/gt_seq', gt_seq_test,
epoch * train_lenth + step + 1)
writer.add_video('test_seq/pred_seq', test_output,
epoch * train_lenth + step + 1)
writer.add_scalars(
'loss/merge', {
"train_loss": train_loss,
"test_loss": test_loss,
"train_metric": train_metric,
"test_metric": test_metric
}, epoch * train_lenth + step + 1)
# 更新avrager
avg.update(step_time=step_time,
train_loss=train_loss,
test_loss=test_loss,
train_metric=train_metric) # 算平均值
# 打印结果
if (step + 1) % print_freq == 0:
global_logger.info(
'Epoch: [{0}][{1}/{2}] {step_time:s}\t{train_loss:s}\t{test_loss:s}\t{train_metric:s}'
.format(epoch + 1, (step + 1) % train_lenth,
train_lenth,
step_time=avg.step_time,
train_loss=avg.train_loss,
test_loss=avg.test_loss,
train_metric=avg.train_metric))
print_speed(epoch * train_lenth + step + 1, avg.step_time.avg,
epoches * train_lenth)
def train(train_loader, model, optimizer, lr_scheduler, epoch, cfg):
global tb_index, best_acc, cur_lr, logger
cur_lr = lr_scheduler.get_cur_lr()
logger = logging.getLogger('global')
avg = AverageMeter()
model.train()
model = model.cuda()
end = time.time()
def is_valid_number(x):
return not (math.isnan(x) or math.isinf(x) or x > 1e4)
num_per_epoch = len(train_loader.dataset) // args.epochs // args.batch
start_epoch = epoch
epoch = epoch
for iter, input in enumerate(train_loader):
if epoch != iter // num_per_epoch + start_epoch: # next epoch
epoch = iter // num_per_epoch + start_epoch
if not os.path.exists(args.save_dir): # makedir/save model
os.makedirs(args.save_dir)
save_checkpoint(
{
'epoch': epoch,
'arch': args.arch,
'state_dict': model.module.state_dict(),
'best_acc': best_acc,
'optimizer': optimizer.state_dict(),
'anchor_cfg': cfg['anchors']
}, False,
os.path.join(args.save_dir, 'checkpoint_e%d.pth' % (epoch)),
os.path.join(args.save_dir, 'best.pth'))
if epoch == args.epochs:
return
if model.module.features.unfix(epoch / args.epochs):
logger.info('unfix part model.')
optimizer, lr_scheduler = build_opt_lr(model.module, cfg, args,
epoch)
lr_scheduler.step(epoch)
cur_lr = lr_scheduler.get_cur_lr()
logger.info('epoch:{}'.format(epoch))
tb_index = iter
if iter % num_per_epoch == 0 and iter != 0:
for idx, pg in enumerate(optimizer.param_groups):
logger.info("epoch {} lr {}".format(epoch, pg['lr']))
tb_writer.add_scalar('lr/group%d' % (idx + 1), pg['lr'],
tb_index)
data_time = time.time() - end
avg.update(data_time=data_time)
x = {
'cfg': cfg,
'template': torch.autograd.Variable(input[0]).cuda(),
'search': torch.autograd.Variable(input[1]).cuda(),
'label_cls': torch.autograd.Variable(input[2]).cuda(),
'label_loc': torch.autograd.Variable(input[3]).cuda(),
'label_loc_weight': torch.autograd.Variable(input[4]).cuda(),
'label_mask': torch.autograd.Variable(input[6]).cuda(),
'label_kp_weight': torch.autograd.Variable(input[7]).cuda(),
'label_mask_weight': torch.autograd.Variable(input[8]).cuda(),
}
outputs = model(x)
# print(x['search'].shape)
pred_mask = outputs['predict'][2]
pred_mask = select_pred_heatmap(
pred_mask,
x['label_mask_weight']) #is rpn_pred_mask (bs, 17, 127, 127)
true_search = select_gt_img(x['search'], x['label_mask_weight'])
if true_search.shape:
save_batch_heatmaps(true_search,
pred_mask,
vis_outpath + '{}.jpg'.format(iter),
normalize=True)
# pred_mask = pred_mask.cpu(.sh).detach().numpy()
# true_search = true_search.cpu().detach().numpy()
# print("pose_mask", pred_mask.shape)
# pose_heat = np.transpose(pred_mask[0,:,:,:],(1,2,0)) #shape (127,127,17)
# plt.figure(num='image', figsize=(128,128))
#
# plt.subplot(1, 2, 1)
# plt.title('origin image')
# plt.imshow(np.transpose(true_search[0,:,:,:], (1,2,0)))
#
# plt.subplot(1, 2, 2)
# plt.title('heatmap')
# pose_map = np.zeros((127,127), np.float32)
# for i in range(pred_mask.shape[1]):
# pose_map += pose_heat[:,:,i]
# plt.imshow(pose_map)
# plt.axis('off')
#
#
# plt.show()
# 可视化： 把17个map都投影到一张黑色图片上
rpn_cls_loss, rpn_loc_loss, rpn_mask_loss = torch.mean(outputs['losses'][0]),\
torch.mean(outputs['losses'][1]),\
torch.mean(outputs['losses'][2])
# mask_iou_mean, mask_iou_at_5, mask_iou_at_7 = torch.mean(outputs['accuracy'][0]), torch.mean(outputs['accuracy'][1]), torch.mean(outputs['accuracy'][2])
cls_weight, reg_weight, mask_weight = cfg['loss']['weight']
loss = rpn_cls_loss * cls_weight + rpn_loc_loss * reg_weight + rpn_mask_loss * mask_weight
optimizer.zero_grad()
loss.backward()
if cfg['clip']['split']:
torch.nn.utils.clip_grad_norm_(model.module.features.parameters(),
cfg['clip']['feature'])
torch.nn.utils.clip_grad_norm_(model.module.rpn_model.parameters(),
cfg['clip']['rpn'])
torch.nn.utils.clip_grad_norm_(
model.module.mask_model.parameters(), cfg['clip']['mask'])
else:
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.clip) # gradient clip
if is_valid_number(loss.item()):
optimizer.step()
siammask_loss = loss.item()
batch_time = time.time() - end
avg.update(batch_time=batch_time,
rpn_cls_loss=rpn_cls_loss,
rpn_loc_loss=rpn_loc_loss,
rpn_mask_loss=rpn_mask_loss * mask_weight,
siammask_loss=siammask_loss)
# mask_iou_mean=mask_iou_mean, mask_iou_at_5=mask_iou_at_5, mask_iou_at_7=mask_iou_at_7)
tb_writer.add_scalar('loss/cls', rpn_cls_loss, tb_index)
tb_writer.add_scalar('loss/loc', rpn_loc_loss, tb_index)
tb_writer.add_scalar('loss/mask', rpn_mask_loss * mask_weight,
tb_index)
# tb_writer.add_scalar('mask/mIoU', mask_iou_mean, tb_index)
# tb_writer.add_scalar('mask/[email protected]', mask_iou_at_5, tb_index)
# tb_writer.add_scalar('mask/[email protected]', mask_iou_at_7, tb_index)
end = time.time()
if (iter + 1) % args.print_freq == 0:
logger.info(
'Epoch: [{0}][{1}/{2}] lr: {lr:.6f}\t{batch_time:s}\t{data_time:s}'
'\t{rpn_cls_loss:s}\t{rpn_loc_loss:s}\t{rpn_mask_loss:s}\t{siammask_loss:s}'
.format(
epoch + 1,
(iter + 1) % num_per_epoch,
num_per_epoch,
lr=cur_lr,
batch_time=avg.batch_time,
data_time=avg.data_time,
rpn_cls_loss=avg.rpn_cls_loss,
rpn_loc_loss=avg.rpn_loc_loss,
rpn_mask_loss=avg.rpn_mask_loss,
siammask_loss=avg.siammask_loss,
))
# mask_iou_mean=avg.mask_iou_mean,
# mask_iou_at_5=avg.mask_iou_at_5,mask_iou_at_7=avg.mask_iou_at_7))
print_speed(iter + 1, avg.batch_time.avg,
args.epochs * num_per_epoch)
pos_dists = torch.mean(output['dist_pos'])
neg_dists = torch.mean(output['dist_neg'])
# loss compute
loss_value = loss(anc_emb, pos_emb, neg_emb)
# Backward pass
optimizer_model.zero_grad()
loss_value.backward()
optimizer_model.step()
current_batch += 1
batch_time = time.time() - batch_start_time
avg.update(time=batch_time,
triplet_loss=loss_value,
pos_dists=pos_dists,
neg_dists=neg_dists)
writer.add_scalar("Train_Batch/Loss/train_loss",
loss_value,
global_step=current_batch)
writer.add_scalar("Train_Batch/Distance/pos_dists",
pos_dists,
global_step=current_batch)
writer.add_scalar("Train_Batch/Distance/neg_dists",
neg_dists,
global_step=current_batch)
writer.add_scalar("Train_Batch_Global_AVG/loss",
avg.triplet_loss.avg,
global_step=current_batch)
writer.add_scalar("Train_Batch_Global_AVG/pos_dists",
avg.pos_dists.avg,
# 送入模型进行推断
test_output = model(seq_test, future=input_num)
# loss计算
test_loss = loss(seq_test[:, -input_num:, :, :, :], gt_seq_test[:, -input_num:, :, :, :])
step_time = time.time() - step_time
# 将有用的信息存进tensorboard中
if (step+1) % print_freq == 0:
writer.add_video('seq/train_seq', seq, epoch*train_lenth + step + 1)
writer.add_video('seq/gt_seq', seq_target, epoch*train_lenth + step + 1)
writer.add_video('seq/pred_seq', layer_output, epoch*train_lenth + step + 1)
writer.add_scalars('loss/merge', {"train_loss": loss_,"test_loss":test_loss}, epoch*train_lenth + step + 1)
# 更新avrager
avg.update(step_time=step_time, train_loss=loss_, test_loss=test_loss) # 算平均值
# 打印结果
if (step+1) % print_freq == 0:
global_logger.info('Epoch: [{0}][{1}/{2}] {step_time:s}\t{train_loss:s}\t{test_loss:s}'.format(
epoch+1, (step + 1) % train_lenth, train_lenth, step_time=avg.step_time, train_loss=avg.train_loss, test_loss=avg.test_loss))
print_speed(epoch*train_lenth + step + 1, avg.step_time.avg, epoches * train_lenth)
# scheduler更新
scheduler.step()
points_feats = model(points_graph)
global_feats = torch.cat((visual_feats, points_feats),
dim=0).contiguous()
labels = torch.cat((visual_graph.y, points_graph.y),
dim=0).contiguous()
iloss, _, _, _, _, _ = global_loss(trip_loss, global_feats,
labels)
batch_loss += iloss
batch_loss = batch_loss / len(value[0])
optimizer.zero_grad()
batch_loss.backward()
# clip_gradient(model, 10)
optimizer.step()
batch_time = time.time() - end_time
avg.update(batch_time=batch_time, loss=batch_loss.item())
if (step + 1) % args.disp_interval == 0:
vis.plot('loss', avg.avg('loss'))
log_str = '(Train) Epoch: [{0}][{1}/{2}]\t lr: {lr:.6f} \t {batch_time:s} \t {loss:s} \n'.format(
epoch,
step + 1,
len(train_dataloader),
lr=lr,
batch_time=avg.batch_time,
loss=avg.loss)
vis.log(log_str)
if args.trainval:
# validation
model.eval()
valid_avg = AverageMeter()
valid_disp_interval = int(args.disp_interval /