def train(self, train_loader):
args = self.args
logger = self.logger
steps_per_epoch = len(train_loader)
device = self.device
self.aanet.train()
if args.freeze_bn:
def set_bn_eval(m):
classname = m.__class__.__name__
if classname.find('BatchNorm') != -1:
m.eval()
self.aanet.apply(set_bn_eval)
# Learning rate summary
base_lr = self.optimizer.param_groups[0]['lr']
offset_lr = self.optimizer.param_groups[1]['lr']
self.train_writer.add_scalar('base_lr', base_lr, self.epoch + 1)
self.train_writer.add_scalar('offset_lr', offset_lr, self.epoch + 1)
last_print_time = time.time()
for i, sample in enumerate(train_loader):
left = sample['left'].to(device) # [B, 3, H, W]
right = sample['right'].to(device)
gt_disp = sample['disp'].to(device) # [B, H, W]
mask = (gt_disp > 0) & (gt_disp < args.max_disp)
if args.load_pseudo_gt:
pseudo_gt_disp = sample['pseudo_disp'].to(device)
pseudo_mask = (pseudo_gt_disp > 0) & (
pseudo_gt_disp < args.max_disp) & (~mask) # inverse mask
if not mask.any():
continue
pred_disp_pyramid = self.aanet(
left, right) # list of H/12, H/6, H/3, H/2, H
if args.highest_loss_only:
pred_disp_pyramid = [
pred_disp_pyramid[-1]
] # only the last highest resolution output
disp_loss = 0
pseudo_disp_loss = 0
pyramid_loss = []
pseudo_pyramid_loss = []
# Loss weights
if len(pred_disp_pyramid) == 5:
pyramid_weight = [1 / 3, 2 / 3, 1.0, 1.0, 1.0] # AANet
elif len(pred_disp_pyramid) == 4:
pyramid_weight = [1 / 3, 2 / 3, 1.0, 1.0] # AANet+
elif len(pred_disp_pyramid) == 3:
pyramid_weight = [1.0, 1.0, 1.0] # 1 scale only
elif len(pred_disp_pyramid) == 1:
pyramid_weight = [1.0] # highest loss only
else:
raise NotImplementedError
assert len(pyramid_weight) == len(pred_disp_pyramid)
for k in range(len(pred_disp_pyramid)):
pred_disp = pred_disp_pyramid[k]
weight = pyramid_weight[k]
if pred_disp.size(-1) != gt_disp.size(-1):
pred_disp = pred_disp.unsqueeze(1) # [B, 1, H, W]
pred_disp = F.interpolate(
pred_disp,
size=(gt_disp.size(-2), gt_disp.size(-1)),
mode='bilinear') * (gt_disp.size(-1) /
pred_disp.size(-1))
pred_disp = pred_disp.squeeze(1) # [B, H, W]
curr_loss = F.smooth_l1_loss(pred_disp[mask],
gt_disp[mask],
reduction='mean')
disp_loss += weight * curr_loss
pyramid_loss.append(curr_loss)
# Pseudo gt loss
if args.load_pseudo_gt:
pseudo_curr_loss = F.smooth_l1_loss(
pred_disp[pseudo_mask],
pseudo_gt_disp[pseudo_mask],
reduction='mean')
pseudo_disp_loss += weight * pseudo_curr_loss
pseudo_pyramid_loss.append(pseudo_curr_loss)
total_loss = disp_loss + pseudo_disp_loss
self.optimizer.zero_grad()
total_loss.backward()
self.optimizer.step()
self.num_iter += 1
if self.num_iter % args.print_freq == 0:
this_cycle = time.time() - last_print_time
last_print_time += this_cycle
logger.info(
'Epoch: [%3d/%3d] [%5d/%5d] time: %4.2fs disp_loss: %.3f' %
(self.epoch + 1, args.max_epoch, i + 1, steps_per_epoch,
this_cycle, disp_loss.item()))
if self.num_iter % args.summary_freq == 0:
img_summary = dict()
img_summary['left'] = left
img_summary['right'] = right
img_summary['gt_disp'] = gt_disp
if args.load_pseudo_gt:
img_summary['pseudo_gt_disp'] = pseudo_gt_disp
# Save pyramid disparity prediction
for s in range(len(pred_disp_pyramid)):
# Scale from low to high, reverse
save_name = 'pred_disp' + str(
len(pred_disp_pyramid) - s - 1)
save_value = pred_disp_pyramid[s]
img_summary[save_name] = save_value
pred_disp = pred_disp_pyramid[-1]
if pred_disp.size(-1) != gt_disp.size(-1):
pred_disp = pred_disp.unsqueeze(1) # [B, 1, H, W]
pred_disp = F.interpolate(
pred_disp,
size=(gt_disp.size(-2), gt_disp.size(-1)),
mode='bilinear') * (gt_disp.size(-1) /
pred_disp.size(-1))
pred_disp = pred_disp.squeeze(1) # [B, H, W]
img_summary['disp_error'] = disp_error_img(pred_disp, gt_disp)
save_images(self.train_writer, 'train', img_summary,
self.num_iter)
epe = F.l1_loss(gt_disp[mask],
pred_disp[mask],
reduction='mean')
self.train_writer.add_scalar('train/epe', epe.item(),
self.num_iter)
self.train_writer.add_scalar('train/disp_loss',
disp_loss.item(), self.num_iter)
self.train_writer.add_scalar('train/total_loss',
total_loss.item(), self.num_iter)
# Save loss of different scale
for s in range(len(pyramid_loss)):
save_name = 'train/loss' + str(len(pyramid_loss) - s - 1)
save_value = pyramid_loss[s]
self.train_writer.add_scalar(save_name, save_value,
self.num_iter)
d1 = d1_metric(pred_disp, gt_disp, mask)
self.train_writer.add_scalar('train/d1', d1.item(),
self.num_iter)
thres1 = thres_metric(pred_disp, gt_disp, mask, 1.0)
thres2 = thres_metric(pred_disp, gt_disp, mask, 2.0)
thres3 = thres_metric(pred_disp, gt_disp, mask, 3.0)
self.train_writer.add_scalar('train/thres1', thres1.item(),
self.num_iter)
self.train_writer.add_scalar('train/thres2', thres2.item(),
self.num_iter)
self.train_writer.add_scalar('train/thres3', thres3.item(),
self.num_iter)
self.epoch += 1
# Always save the latest model for resuming training
if args.no_validate:
utils.save_checkpoint(args.checkpoint_dir,
self.optimizer,
self.aanet,
epoch=self.epoch,
num_iter=self.num_iter,
epe=-1,
best_epe=self.best_epe,
best_epoch=self.best_epoch,
filename='aanet_latest.pth')
# Save checkpoint of specific epoch
if self.epoch % args.save_ckpt_freq == 0:
model_dir = os.path.join(args.checkpoint_dir, 'models')
utils.check_path(model_dir)
utils.save_checkpoint(model_dir,
self.optimizer,
self.aanet,
epoch=self.epoch,
num_iter=self.num_iter,
epe=-1,
best_epe=self.best_epe,
best_epoch=self.best_epoch,
save_optimizer=False)
def validate(self, val_loader):
args = self.args
logger = self.logger
logger.info('=> Start validation...')
if args.evaluate_only is True:
if args.pretrained_aanet is not None:
pretrained_aanet = args.pretrained_aanet
else:
model_name = 'aanet_best.pth'
pretrained_aanet = os.path.join(args.checkpoint_dir,
model_name)
if not os.path.exists(
pretrained_aanet): # KITTI without validation
pretrained_aanet = pretrained_aanet.replace(
model_name, 'aanet_latest.pth')
logger.info('=> loading pretrained aanet: %s' % pretrained_aanet)
utils.load_pretrained_net(self.aanet,
pretrained_aanet,
no_strict=True)
self.aanet.eval()
num_samples = len(val_loader)
logger.info('=> %d samples found in the validation set' % num_samples)
val_epe = 0
val_d1 = 0
val_thres1 = 0
val_thres2 = 0
val_thres3 = 0
val_count = 0
val_file = os.path.join(args.checkpoint_dir, 'val_results.txt')
num_imgs = 0
valid_samples = 0
for i, sample in enumerate(val_loader):
if i % 100 == 0:
logger.info('=> Validating %d/%d' % (i, num_samples))
left = sample['left'].to(self.device) # [B, 3, H, W]
right = sample['right'].to(self.device)
gt_disp = sample['disp'].to(self.device) # [B, H, W]
mask = (gt_disp > 0) & (gt_disp < args.max_disp)
if not mask.any():
continue
valid_samples += 1
num_imgs += gt_disp.size(0)
with torch.no_grad():
pred_disp = self.aanet(left, right)[-1] # [B, H, W]
if pred_disp.size(-1) < gt_disp.size(-1):
pred_disp = pred_disp.unsqueeze(1) # [B, 1, H, W]
pred_disp = F.interpolate(
pred_disp, (gt_disp.size(-2), gt_disp.size(-1)),
mode='bilinear') * (gt_disp.size(-1) / pred_disp.size(-1))
pred_disp = pred_disp.squeeze(1) # [B, H, W]
epe = F.l1_loss(gt_disp[mask], pred_disp[mask], reduction='mean')
d1 = d1_metric(pred_disp, gt_disp, mask)
thres1 = thres_metric(pred_disp, gt_disp, mask, 1.0)
thres2 = thres_metric(pred_disp, gt_disp, mask, 2.0)
thres3 = thres_metric(pred_disp, gt_disp, mask, 3.0)
val_epe += epe.item()
val_d1 += d1.item()
val_thres1 += thres1.item()
val_thres2 += thres2.item()
val_thres3 += thres3.item()
# Save 3 images for visualization
if not args.evaluate_only:
if i in [
num_samples // 4, num_samples // 2,
num_samples // 4 * 3
]:
img_summary = dict()
img_summary['disp_error'] = disp_error_img(
pred_disp, gt_disp)
img_summary['left'] = left
img_summary['right'] = right
img_summary['gt_disp'] = gt_disp
img_summary['pred_disp'] = pred_disp
save_images(self.train_writer, 'val' + str(val_count),
img_summary, self.epoch)
val_count += 1
logger.info('=> Validation done!')
mean_epe = val_epe / valid_samples
mean_d1 = val_d1 / valid_samples
mean_thres1 = val_thres1 / valid_samples
mean_thres2 = val_thres2 / valid_samples
mean_thres3 = val_thres3 / valid_samples
# Save validation results
with open(val_file, 'a') as f:
f.write('epoch: %03d\t' % self.epoch)
f.write('epe: %.3f\t' % mean_epe)
f.write('d1: %.4f\t' % mean_d1)
f.write('thres1: %.4f\t' % mean_thres1)
f.write('thres2: %.4f\t' % mean_thres2)
f.write('thres3: %.4f\n' % mean_thres3)
logger.info('=> Mean validation epe of epoch %d: %.3f' %
(self.epoch, mean_epe))
if not args.evaluate_only:
self.train_writer.add_scalar('val/epe', mean_epe, self.epoch)
self.train_writer.add_scalar('val/d1', mean_d1, self.epoch)
self.train_writer.add_scalar('val/thres1', mean_thres1, self.epoch)
self.train_writer.add_scalar('val/thres2', mean_thres2, self.epoch)
self.train_writer.add_scalar('val/thres3', mean_thres3, self.epoch)
if not args.evaluate_only:
if args.val_metric == 'd1':
if mean_d1 < self.best_epe:
# Actually best_epe here is d1
self.best_epe = mean_d1
self.best_epoch = self.epoch
utils.save_checkpoint(args.checkpoint_dir,
self.optimizer,
self.aanet,
epoch=self.epoch,
num_iter=self.num_iter,
epe=mean_d1,
best_epe=self.best_epe,
best_epoch=self.best_epoch,
filename='aanet_best.pth')
elif args.val_metric == 'epe':
if mean_epe < self.best_epe:
self.best_epe = mean_epe
self.best_epoch = self.epoch
utils.save_checkpoint(args.checkpoint_dir,
self.optimizer,
self.aanet,
epoch=self.epoch,
num_iter=self.num_iter,
epe=mean_epe,
best_epe=self.best_epe,
best_epoch=self.best_epoch,
filename='aanet_best.pth')
else:
raise NotImplementedError
if self.epoch == args.max_epoch:
# Save best validation results
with open(val_file, 'a') as f:
f.write('\nbest epoch: %03d \t best %s: %.3f\n\n' %
(self.best_epoch, args.val_metric, self.best_epe))
logger.info('=> best epoch: %03d \t best %s: %.3f\n' %
(self.best_epoch, args.val_metric, self.best_epe))
# Always save the latest model for resuming training
if not args.evaluate_only:
utils.save_checkpoint(args.checkpoint_dir,
self.optimizer,
self.aanet,
epoch=self.epoch,
num_iter=self.num_iter,
epe=mean_epe,
best_epe=self.best_epe,
best_epoch=self.best_epoch,
filename='aanet_latest.pth')
# Save checkpoint of specific epochs
if self.epoch % args.save_ckpt_freq == 0:
model_dir = os.path.join(args.checkpoint_dir, 'models')
utils.check_path(model_dir)
utils.save_checkpoint(model_dir,
self.optimizer,
self.aanet,
epoch=self.epoch,
num_iter=self.num_iter,
epe=mean_epe,
best_epe=self.best_epe,
best_epoch=self.best_epoch,
save_optimizer=False)
def validate(self, val_loader, local_master, valLossDict, valLossKey):
args = self.args
logger = self.logger
logger.info('=> Start validation...')
# 只做evaluate，则需要从文件加载训练好的模型。否则，直接使用本model类中保存的(尚未完成全部的Epoach训练的)self.aanet即可。
if args.evaluate_only is True:
if args.pretrained_aanet is not None:
pretrained_aanet = args.pretrained_aanet
else:
model_name = 'aanet_best.pth'
pretrained_aanet = os.path.join(args.checkpoint_dir, model_name)
if not os.path.exists(pretrained_aanet): # KITTI without validation
pretrained_aanet = pretrained_aanet.replace(model_name, 'aanet_latest.pth')
logger.info('=> loading pretrained aanet: %s' % pretrained_aanet)
utils.load_pretrained_net(self.aanet, pretrained_aanet, no_strict=True)
self.aanet.eval()
num_samples = len(val_loader)
logger.info('=> %d samples found in the validation set' % num_samples)
val_epe = 0
val_d1 = 0
val_thres1 = 0
val_thres2 = 0
val_thres3 = 0
val_thres10 = 0
val_thres20 = 0
val_count = 0
val_file = os.path.join(args.checkpoint_dir, 'val_results.txt')
num_imgs = 0
valid_samples = 0
# 遍历验证样本或测试样本
for i, sample in enumerate(val_loader):
if (i + 1) % 100 == 0:
logger.info('=> Validating %d/%d' % (i, num_samples))
left = sample['left'].to(self.device) # [B, 3, H, W]
right = sample['right'].to(self.device)
gt_disp = sample['disp'].to(self.device) # [B, H, W]
mask = (gt_disp > 0) & (gt_disp < args.max_disp)
if not mask.any():
continue
valid_samples += 1
num_imgs += gt_disp.size(0)
with torch.no_grad():
disparity_pyramid = self.aanet(left, right) # [B, H, W]
pred_disp = disparity_pyramid[-1]
if pred_disp.size(-1) < gt_disp.size(-1):
pred_disp = pred_disp.unsqueeze(1) # [B, 1, H, W]
pred_disp = F.interpolate(pred_disp, (gt_disp.size(-2), gt_disp.size(-1)),
mode='bilinear', align_corners=False) * (
gt_disp.size(-1) / pred_disp.size(-1))
pred_disp = pred_disp.squeeze(1) # [B, H, W]
epe = F.l1_loss(gt_disp[mask], pred_disp[mask], reduction='mean')
d1 = d1_metric(pred_disp, gt_disp, mask)
thres1 = thres_metric(pred_disp, gt_disp, mask, 1.0)
thres2 = thres_metric(pred_disp, gt_disp, mask, 2.0)
thres3 = thres_metric(pred_disp, gt_disp, mask, 3.0)
thres10 = thres_metric(pred_disp, gt_disp, mask, 10.0)
thres20 = thres_metric(pred_disp, gt_disp, mask, 20.0)
val_epe += epe.item()
val_d1 += d1.item()
val_thres1 += thres1.item()
val_thres2 += thres2.item()
val_thres3 += thres3.item()
val_thres10 += thres10.item()
val_thres20 += thres20.item()
# save Image For Error Analysis
# saveForErrorAnalysis(index, img_name, dstPath, dstName, left, right, gt_disp, disparity_pyramid):
with torch.no_grad():
saveImgErrorAnalysis(i, sample['left_name'], './myDataAnalysis', 'SceneFlow_valIdx_{}'.format(i),
left, right, gt_disp, disparity_pyramid, disp_error_img(pred_disp, gt_disp))
# Save 3 images for visualization
if not args.evaluate_only or args.mode == 'test':
# if i in [num_samples // 4, num_samples // 2, num_samples // 4 * 3]:
if i in [num_samples // 6, num_samples // 6 * 2, num_samples // 6 * 3, num_samples // 6 * 4,
num_samples // 6 * 5]:
img_summary = dict()
img_summary['disp_error'] = disp_error_img(pred_disp, gt_disp)
img_summary['left'] = left
img_summary['right'] = right
img_summary['gt_disp'] = gt_disp
img_summary['pred_disp'] = pred_disp
save_images(self.train_writer, 'val' + str(val_count), img_summary, self.epoch)
disp_error = disp_error_hist(pred_disp, gt_disp, args.max_disp)
save_hist(self.train_writer, '{}/{}'.format('val' + str(val_count), 'hist'), disp_error, self.epoch)
val_count += 1
# 遍历验证样本或测试样本完成
logger.info('=> Validation done!')
mean_epe = val_epe / valid_samples
mean_d1 = val_d1 / valid_samples
mean_thres1 = val_thres1 / valid_samples
mean_thres2 = val_thres2 / valid_samples
mean_thres3 = val_thres3 / valid_samples
mean_thres10 = val_thres10 / valid_samples
mean_thres20 = val_thres20 / valid_samples
# 记录数据为matlab的mat文件，用于分析和对比
valLossDict[valLossKey]["epochs"].append(self.epoch)
valLossDict[valLossKey]["avgEPE"].append(mean_epe)
valLossDict[valLossKey]["avg_d1"].append(mean_d1)
valLossDict[valLossKey]["avg_thres1"].append(mean_thres1)
valLossDict[valLossKey]["avg_thres2"].append(mean_thres2)
valLossDict[valLossKey]["avg_thres3"].append(mean_thres3)
valLossDict[valLossKey]["avg_thres10"].append(mean_thres10)
valLossDict[valLossKey]["avg_thres20"].append(mean_thres20)
# Save validation results
with open(val_file, 'a') as f:
f.write('epoch: %03d\t' % self.epoch)
f.write('epe: %.3f\t' % mean_epe)
f.write('d1: %.4f\t' % mean_d1)
f.write('thres1: %.4f\t' % mean_thres1)
f.write('thres2: %.4f\t' % mean_thres2)
f.write('thres3: %.4f\t' % mean_thres3)
f.write('thres10: %.4f\t' % mean_thres10)
f.write('thres20: %.4f\n' % mean_thres20)
f.write('dataset_name= %s\t mode=%s\n' % (args.dataset_name, args.mode))
logger.info('=> Mean validation epe of epoch %d: %.3f' % (self.epoch, mean_epe))
if not args.evaluate_only:
self.train_writer.add_scalar('val/epe', mean_epe, self.epoch)
self.train_writer.add_scalar('val/d1', mean_d1, self.epoch)
self.train_writer.add_scalar('val/thres1', mean_thres1, self.epoch)
self.train_writer.add_scalar('val/thres2', mean_thres2, self.epoch)
self.train_writer.add_scalar('val/thres3', mean_thres3, self.epoch)
self.train_writer.add_scalar('val/thres10', mean_thres10, self.epoch)
self.train_writer.add_scalar('val/thres20', mean_thres20, self.epoch)
if not args.evaluate_only:
if args.val_metric == 'd1':
if mean_d1 < self.best_epe:
# Actually best_epe here is d1
self.best_epe = mean_d1
self.best_epoch = self.epoch
utils.save_checkpoint(args.checkpoint_dir, self.optimizer, self.aanet,
epoch=self.epoch, num_iter=self.num_iter,
epe=mean_d1, best_epe=self.best_epe,
best_epoch=self.best_epoch,
filename='aanet_best.pth') if local_master else None
elif args.val_metric == 'epe':
if mean_epe < self.best_epe:
self.best_epe = mean_epe
self.best_epoch = self.epoch
utils.save_checkpoint(args.checkpoint_dir, self.optimizer, self.aanet,
epoch=self.epoch, num_iter=self.num_iter,
epe=mean_epe, best_epe=self.best_epe,
best_epoch=self.best_epoch,
filename='aanet_best.pth') if local_master else None
else:
raise NotImplementedError
if self.epoch == args.max_epoch:
# Save best validation results
with open(val_file, 'a') as f:
f.write('\nbest epoch: %03d \t best %s: %.3f\n\n' % (self.best_epoch,
args.val_metric,
self.best_epe))
logger.info('=> best epoch: %03d \t best %s: %.3f\n' % (self.best_epoch,
args.val_metric,
self.best_epe))
# Always save the latest model for resuming training
if not args.evaluate_only:
utils.save_checkpoint(args.checkpoint_dir, self.optimizer, self.aanet,
epoch=self.epoch, num_iter=self.num_iter,
epe=mean_epe, best_epe=self.best_epe,
best_epoch=self.best_epoch,
filename='aanet_latest.pth') if local_master else None
# Save checkpoint of specific epochs
if self.epoch % args.save_ckpt_freq == 0:
model_dir = os.path.join(args.checkpoint_dir, 'models')
utils.check_path(model_dir)
utils.save_checkpoint(model_dir, self.optimizer, self.aanet,
epoch=self.epoch, num_iter=self.num_iter,
epe=mean_epe, best_epe=self.best_epe,
best_epoch=self.best_epoch,
save_optimizer=False) if local_master else None
def train(self, train_loader, local_master, trainLoss_dict, trainLossKey):
args = self.args
logger = self.logger
steps_per_epoch = len(train_loader) / args.accumulation_steps # len(train_loader)返回的是Batch的个数
device = self.device
self.aanet.train() # 设置模型为训练模式！
if args.freeze_bn:
def set_bn_eval(m):
classname = m.__class__.__name__ # 实例调用__class__属性时会指向该实例对应的类。.__class__将实例变量指向类，然后再去调用__name__类属性
if classname.find('BatchNorm') != -1:
m.eval()
self.aanet.apply(
set_bn_eval) # apply(fn: Callable[Module, None])：Applies fn recursively to every submodule (as returned by .children()) as well as self. Typical use includes initializing the parameters of a model (see also torch.nn.init).
# Learning rate summary
base_lr = self.optimizer.param_groups[0]['lr']
offset_lr = self.optimizer.param_groups[1]['lr']
self.train_writer.add_scalar('lr/base_lr', base_lr, self.epoch + 1)
self.train_writer.add_scalar('lr/offset_lr', offset_lr, self.epoch + 1)
last_print_time = time.time()
validate_count = 0
total_epe = 0
total_d1 = 0
total_thres1 = 0
total_thres2 = 0
total_thres3 = 0
total_thres10 = 0
total_thres20 = 0
loss_acum = 0
for i, sample in enumerate(train_loader):
left = sample['left'].to(device) # [B, 3, H, W]
right = sample['right'].to(device)
gt_disp = sample['disp'].to(device) # [B, H, W]
mask = (gt_disp > 0) & (gt_disp < args.max_disp) # KITTI数据集约定：视差为0，表示无效视差。
if args.load_pseudo_gt:
pseudo_gt_disp = sample['pseudo_disp'].to(device)
pseudo_mask = (pseudo_gt_disp > 0) & (pseudo_gt_disp < args.max_disp) & (
~mask) # inverse mask # 需要修补的像素位置的mask
if not mask.any(): # np.array.any()是或操作，任意一个元素为True，输出为True。
continue
# 尝试分布式训练
# 只在DDP模式下，轮数不是args.accumulation_steps整数倍的时候使用no_sync。
# 博客：https://blog.csdn.net/a40850273/article/details/111829836
my_context = self.aanet.no_sync if args.distributed and (
i + 1) % args.accumulation_steps != 0 else nullcontext
with my_context():
pred_disp_pyramid = self.aanet(left, right) # list of H/12, H/6, H/3, H/2, H
if args.highest_loss_only:
pred_disp_pyramid = [pred_disp_pyramid[-1]] # only the last highest resolution output
disp_loss = 0
pseudo_disp_loss = 0
pyramid_loss = []
pseudo_pyramid_loss = []
# Loss weights
if len(pred_disp_pyramid) == 5:
pyramid_weight = [1 / 3, 2 / 3, 1.0, 1.0, 1.0] # AANet and AANet+
elif len(pred_disp_pyramid) == 4:
pyramid_weight = [1 / 3, 2 / 3, 1.0, 1.0]
elif len(pred_disp_pyramid) == 3:
pyramid_weight = [1.0, 1.0, 1.0] # 1 scale only
elif len(pred_disp_pyramid) == 1:
pyramid_weight = [1.0] # highest loss only
else:
raise NotImplementedError
assert len(pyramid_weight) == len(pred_disp_pyramid)
for k in range(len(pred_disp_pyramid)):
pred_disp = pred_disp_pyramid[k]
weight = pyramid_weight[k]
if pred_disp.size(-1) != gt_disp.size(-1):
pred_disp = pred_disp.unsqueeze(1) # [B, 1, H, W]
pred_disp = F.interpolate(pred_disp, size=(gt_disp.size(-2), gt_disp.size(-1)),
mode='bilinear', align_corners=False) * (
gt_disp.size(-1) / pred_disp.size(-1)) # 最后乘上这一项是必须的。因为图像放大，视差要相应增大。
pred_disp = pred_disp.squeeze(1) # [B, H, W]
curr_loss = F.smooth_l1_loss(pred_disp[mask], gt_disp[mask],
reduction='mean')
disp_loss += weight * curr_loss
pyramid_loss.append(curr_loss)
# Pseudo gt loss
if args.load_pseudo_gt:
pseudo_curr_loss = F.smooth_l1_loss(pred_disp[pseudo_mask], pseudo_gt_disp[pseudo_mask],
reduction='mean')
pseudo_disp_loss += weight * pseudo_curr_loss
pseudo_pyramid_loss.append(pseudo_curr_loss)
total_loss = disp_loss + pseudo_disp_loss
total_loss /= args.accumulation_steps
total_loss.backward()
# 仅用于记录和分析数据
with torch.no_grad():
validate_count += 1
total_epe += F.l1_loss(gt_disp[mask], pred_disp_pyramid[-1][mask], reduction='mean').detach().cpu().numpy()
total_d1 += d1_metric(pred_disp_pyramid[-1], gt_disp, mask).detach().cpu().numpy()
total_thres1 += thres_metric(pred_disp_pyramid[-1], gt_disp, mask, 1.0).detach().cpu().numpy() # mask.shape=[B, H, W]
total_thres2 += thres_metric(pred_disp_pyramid[-1], gt_disp, mask, 2.0).detach().cpu().numpy()
total_thres3 += thres_metric(pred_disp_pyramid[-1], gt_disp, mask, 3.0).detach().cpu().numpy()
total_thres10 += thres_metric(pred_disp_pyramid[-1], gt_disp, mask, 10.0).detach().cpu().numpy()
total_thres20 += thres_metric(pred_disp_pyramid[-1], gt_disp, mask, 20.0).detach().cpu().numpy()
loss_acum += total_loss.detach().cpu().numpy()
if (i + 1) % args.accumulation_steps == 0:
self.optimizer.step()
self.optimizer.zero_grad()
self.num_iter += 1
if self.num_iter % args.print_freq == 0:
this_cycle = time.time() - last_print_time
last_print_time += this_cycle
time_to_finish = (args.max_epoch - self.epoch) * (1.0 * steps_per_epoch / args.print_freq) * \
this_cycle / 3600.0 # 还有多久才能完成训练。单位：小时
logger.info('Epoch: [%3d/%3d] [%5d/%5d] time: %4.2fs remainT: %4.2fh disp_loss: %.3f' %
(self.epoch + 1, args.max_epoch, self.num_iter, steps_per_epoch, this_cycle,
time_to_finish,
disp_loss.item()))
# self.num_iter：表示当前一共进行了多少次迭代，一次参数更新表示一次迭代。
# steps_per_epoch：表示一个epoch中有多少次迭代，一次参数更新表示一次迭代。
if self.num_iter % args.summary_freq == 0:
img_summary = dict()
img_summary['left'] = left # [B, C=3, H, W]
img_summary['right'] = right # [B, C=3, H, W]
img_summary['gt_disp'] = gt_disp # [B, H, W]
if args.load_pseudo_gt:
img_summary['pseudo_gt_disp'] = pseudo_gt_disp
# Save pyramid disparity prediction
for s in range(len(pred_disp_pyramid)):
# Scale from low to high, reverse
save_name = 'pred_disp' + str(
len(pred_disp_pyramid) - s - 1) # pred_disp0-->pred_disp4：高分辨率->低分辨率
save_value = pred_disp_pyramid[s]
img_summary[save_name] = save_value
pred_disp = pred_disp_pyramid[-1]
if pred_disp.size(-1) != gt_disp.size(-1):
pred_disp = pred_disp.unsqueeze(1) # [B, 1, H, W]
pred_disp = F.interpolate(pred_disp, size=(gt_disp.size(-2), gt_disp.size(-1)),
mode='bilinear', align_corners=False) * (
gt_disp.size(-1) / pred_disp.size(-1))
pred_disp = pred_disp.squeeze(1) # [B, H, W]
img_summary['disp_error'] = disp_error_img(pred_disp, gt_disp) # [B, C=3, H, W]
save_images(self.train_writer, 'train', img_summary, self.num_iter)
epe = F.l1_loss(gt_disp[mask], pred_disp[mask], reduction='mean')
self.train_writer.add_scalar('train/epe', epe.item(), self.num_iter)
self.train_writer.add_scalar('train/disp_loss', disp_loss.item(), self.num_iter)
self.train_writer.add_scalar('train/total_loss', total_loss.item(), self.num_iter)
# Save loss of different scale
for s in range(len(pyramid_loss)):
save_name = 'train/loss' + str(len(pyramid_loss) - s - 1) # loss0-->loss4：低分辨率~高分辨率
save_value = pyramid_loss[s]
self.train_writer.add_scalar(save_name, save_value, self.num_iter)
d1 = d1_metric(pred_disp, gt_disp, mask) # pred_disp.shape=[B, H, W], gt_disp.shape=[B, H, W]
self.train_writer.add_scalar('train/d1', d1.item(), self.num_iter)
thres1 = thres_metric(pred_disp, gt_disp, mask, 1.0) # mask.shape=[B, H, W]
thres2 = thres_metric(pred_disp, gt_disp, mask, 2.0)
thres3 = thres_metric(pred_disp, gt_disp, mask, 3.0)
thres10 = thres_metric(pred_disp, gt_disp, mask, 10.0)
thres20 = thres_metric(pred_disp, gt_disp, mask, 20.0)
self.train_writer.add_scalar('train/thres1', thres1.item(), self.num_iter)
self.train_writer.add_scalar('train/thres2', thres2.item(), self.num_iter)
self.train_writer.add_scalar('train/thres3', thres3.item(), self.num_iter)
self.train_writer.add_scalar('train/thres10', thres10.item(), self.num_iter)
self.train_writer.add_scalar('train/thres20', thres20.item(), self.num_iter)
self.epoch += 1
# 记录数据为matlab的mat文件，用于分析和对比
trainLoss_dict[trainLossKey]['epochs'].append(self.epoch)
trainLoss_dict[trainLossKey]['avgEPE'].append(total_epe / validate_count)
trainLoss_dict[trainLossKey]['avg_d1'].append(total_d1 / validate_count)
trainLoss_dict[trainLossKey]['avg_thres1'].append(total_thres1 / validate_count)
trainLoss_dict[trainLossKey]['avg_thres2'].append(total_thres2 / validate_count)
trainLoss_dict[trainLossKey]['avg_thres3'].append(total_thres3 / validate_count)
trainLoss_dict[trainLossKey]['avg_thres10'].append(total_thres10 / validate_count)
trainLoss_dict[trainLossKey]['avg_thres20'].append(total_thres20 / validate_count)
trainLoss_dict[trainLossKey]['avg_loss'].append(loss_acum / validate_count)
# 一个epoch结束：
# args.no_validate=False,则后面不会做self.validate()，故需要在此处记录如下信息。
# args.no_validate=True,则后面会做self.validate()，会在elf.validate()中记录如下信息，此处不必记录。
# 需记录的信息包括：
# 1.最新的训练的模型和状态(写入aanet_latest.pth、optimizer_latest.pth文件) for resuming training;
# 2.Save checkpoint of specific epoch.
# Always save the latest model for resuming training
if args.no_validate:
utils.save_checkpoint(args.checkpoint_dir, self.optimizer, self.aanet,
epoch=self.epoch, num_iter=self.num_iter,
epe=-1, best_epe=self.best_epe,
best_epoch=self.best_epoch,
filename='aanet_latest.pth') if local_master else None
# Save checkpoint of specific epoch
if self.epoch % args.save_ckpt_freq == 0:
model_dir = os.path.join(args.checkpoint_dir, 'models')
utils.check_path(model_dir)
utils.save_checkpoint(model_dir, self.optimizer, self.aanet,
epoch=self.epoch, num_iter=self.num_iter,
epe=-1, best_epe=self.best_epe,
best_epoch=self.best_epoch,
save_optimizer=False) if local_master else None
def validate(self, val_loader):
args = self.args
logger = self.logger
logger.info('=> Start validation...')
if args.evaluate_only is True:
if args.pretrained_aanet is not None:
pretrained_aanet = args.pretrained_aanet
else:
model_name = 'aanet_best.pth'
pretrained_aanet = os.path.join(args.checkpoint_dir,
model_name)
if not os.path.exists(
pretrained_aanet): # KITTI without validation
pretrained_aanet = pretrained_aanet.replace(
model_name, 'aanet_latest.pth')
logger.info('=> loading pretrained aanet: %s' % pretrained_aanet)
utils.load_pretrained_net(self.aanet,
pretrained_aanet,
no_strict=True)
self.aanet.train()
num_samples = len(val_loader)
logger.info('=> %d samples found in the validation set' % num_samples)
val_epe = 0
val_d1 = 0
# val_thres1 = 0
# val_thres2 = 0
# val_thres3 = 0
val_bad1 = 0
val_bad2 = 0
val_abs = 0
val_mm2 = 0
val_mm4 = 0
val_mm8 = 0
val_count = 0
val_file = os.path.join(args.checkpoint_dir, 'val_results.txt')
num_imgs = 0
valid_samples = 0
baseline = 0.055
intrinsic = [[1387.095, 0.0, 960.0], [0.0, 1387.095, 540.0],
[0.0, 0.0, 1.0]]
for i, sample in enumerate(val_loader):
if i % 100 == 0:
logger.info('=> Validating %d/%d' % (i, num_samples))
left = sample['left'].to(self.device) # [B, 3, H, W]
right = sample['right'].to(self.device)
gt_disp = sample['disp'].to(self.device) # [B, H, W]
gt_depth = []
pred_disp = []
if args.dataset_name == 'custom_dataset': # going to be depthL_fromR_down if from custom_dataset
gt_disp_1 = (baseline * 1000 * intrinsic[0][0] /
2) / (gt_disp * 256.)
gt_disp_1[gt_disp_1 == inf] = 0
gt_depth = gt_disp * 256.
gt_disp = gt_disp_1
if (args.dataset_name == 'custom_dataset_full'
or args.dataset_name == 'custom_dataset_obj'):
# convert to disparity then apply warp ops
temp = gt_disp * 256.
for x in range(left.shape[0]):
baseline = sample['baseline'][x].to(self.device)
intrinsic = sample['intrinsic'][x].to(self.device)
temp[x] = (baseline * 1000 * intrinsic[0][0] /
2) / (temp[x])
temp[x][temp[x] == inf] = 0
# gt_disp = torch.clone(temp)
gt_disp = apply_disparity_cu(temp.unsqueeze(1),
temp.type(torch.int))
gt_disp = torch.squeeze(gt_disp)
gt_depth = temp
# convert to gt_depth
for x in range(left.shape[0]):
baseline = sample['baseline'][x].to(self.device)
intrinsic = sample['intrinsic'][x].to(self.device)
gt_depth[x] = (baseline * 1000 * intrinsic[0][0] /
2) / (gt_disp[x])
gt_depth[x][gt_depth[x] == inf] = 0
gt_depth = gt_depth.to(self.device)
if (args.dataset_name == 'custom_dataset_sim'
or args.dataset_name == 'custom_dataset_real'):
temp = gt_disp * 256.
for x in range(left.shape[0]):
baseline = sample['baseline'][x].to(self.device)
intrinsic = sample['intrinsic'][x].to(self.device)
temp[x] = (baseline * 1000 * intrinsic[0][0] /
2) / (temp[x])
temp[x][temp[x] == inf] = 0
# gt_disp = torch.clone(temp)
gt_disp = apply_disparity_cu(temp.unsqueeze(1),
temp.type(torch.int))
gt_disp = torch.squeeze(gt_disp)
gt_disp = torch.unsqueeze(gt_disp, 0)
gt_depth = temp
# convert to gt_depth
for x in range(left.shape[0]):
baseline = sample['baseline'][x].to(self.device)
intrinsic = sample['intrinsic'][x].to(self.device)
gt_depth[x] = (baseline * 1000 * intrinsic[0][0] /
2) / (gt_disp[x])
gt_depth[x][gt_depth[x] == inf] = 0
gt_depth = gt_depth.to(self.device)
mask_disp = (gt_disp > 0.) & (gt_disp < args.max_disp)
if not mask_disp.any():
continue
valid_samples += 1
num_imgs += gt_disp.size(0)
with torch.no_grad():
pred_disp = self.aanet(left, right)[-1] # [B, H, W]
if pred_disp.size(-1) < gt_disp.size(-1):
pred_disp = pred_disp.unsqueeze(1) # [B, 1, H, W]
pred_disp = F.interpolate(
pred_disp, (gt_disp.size(-2), gt_disp.size(-1)),
mode='bilinear',
align_corners=False) * (gt_disp.size(-1) /
pred_disp.size(-1))
pred_disp = pred_disp.squeeze(1) # [B, H, W]
if (onlyObj):
gt_disp[sample['label'] >= 17] = 0
mask_disp = (gt_disp > 0.) & (gt_disp < args.max_disp)
epe = F.l1_loss(gt_disp[mask_disp],
pred_disp[mask_disp],
reduction='mean')
d1 = d1_metric(pred_disp, gt_disp, mask_disp)
bad1 = bad(pred_disp, gt_disp, mask_disp)
bad2 = bad(pred_disp, gt_disp, mask_disp, threshold=2)
pred_depth = []
if (args.dataset_name == 'custom_dataset_full'
or args.dataset_name == 'custom_dataset_sim'
or args.dataset_name == 'custom_dataset_real'
or args.dataset_name == 'custom_dataset_obj'):
temp = torch.zeros((pred_disp.shape)).to(self.device)
for x in range(left.shape[0]):
baseline = sample['baseline'][x].to(self.device)
intrinsic = sample['intrinsic'][x].to(self.device)
temp[x] = (baseline * 1000 * intrinsic[0][0] /
2) / (pred_disp[x])
temp[x][temp[x] == inf] = 0
pred_depth = temp
else:
pred_depth = (baseline * 1000 * intrinsic[0][0] /
2) / (pred_disp)
pred_depth[pred_depth == inf] = 0
mask_depth = (gt_depth > 0.) & (gt_depth < 2000)
if (onlyObj):
gt_depth[sample['label'] >= 17] = 0
mask_depth = (gt_depth > 0.) & (gt_disp < args.max_disp)
abs = F.l1_loss(gt_depth[mask_depth],
pred_depth[mask_depth],
reduction='mean')
mm2 = mm_error(pred_depth, gt_depth, mask_depth)
mm4 = mm_error(pred_depth, gt_depth, mask_depth, threshold=4)
mm8 = mm_error(pred_depth, gt_depth, mask_depth, threshold=8)
pred_depth[pred_depth > 2000] = 0
if (perObject):
for x in range(left.shape[0]):
labels = sample['label'][x].detach().numpy().astype(
np.uint8)
for obj in np.unique(labels):
gtObjectDepth = gt_depth[x].detach().clone()
gtObjectDepth[labels != obj] = 0
predObjectDepth = pred_depth[x].detach().clone()
predObjectDepth[labels != obj] = 0
gtObjectDisp = gt_disp[x].detach().clone()
gtObjectDisp[labels != obj] = 0
predObjectDisp = pred_disp[x].detach().clone()
predObjectDisp[labels != obj] = 0
mask_depth = (gtObjectDepth > 0.)
mask_disp = (gtObjectDisp > 0.)
objectCount[obj] += 1
perObjectDisp[obj] += F.l1_loss(
gtObjectDisp[mask_disp],
predObjectDisp[mask_disp],
reduction='mean')
perObjectDepth[obj] += F.l1_loss(
gtObjectDepth[mask_depth],
predObjectDepth[mask_depth],
reduction='mean')
# thres1 = thres_metric(pred_disp, gt_disp, mask, 1.0)
# thres2 = thres_metric(pred_disp, gt_disp, mask, 2.0)
# thres3 = thres_metric(pred_disp, gt_disp, mask, 3.0)
val_epe += epe.item()
val_d1 += d1.item()
val_bad1 += bad1.item()
val_bad2 += bad2.item()
val_abs += abs.item()
val_mm2 += mm2.item()
val_mm4 += mm4.item()
val_mm8 += mm8.item()
# val_thres1 += thres1.item()
# val_thres2 += thres2.item()
# val_thres3 += thres3.item()
# Save 3 images for visualization
if i in [num_samples // 4, num_samples // 2, num_samples // 4 * 3]:
if args.evaluate_only:
im = (pred_depth[0]).detach().cpu().numpy().astype(
np.uint16)
if not os.path.isdir('/cephfs/edward/depths'):
os.mkdir('/cephfs/edward/depths')
imageio.imwrite('/cephfs/edward/depths/' + str(i) + ".png",
im)
im = (gt_depth[0]).detach().cpu().numpy().astype(np.uint16)
imageio.imwrite(
'/cephfs/edward/depths/' + str(i) + "gt.png", im)
imageio.imwrite(
'/cephfs/edward/depths/' + str(i) + "label.png",
sample['label'][x].detach().numpy().astype(np.uint8))
info = {
'baseline': sample['baseline'][x],
'intrinsic': sample['intrinsic'][x],
'object_ids': sample['object_ids'][x],
'extrinsic': sample['extrinsic'][x]
}
filename = '/cephfs/edward/depths/meta' + str(i) + '.pkl'
with open(filename, 'wb') as f:
pickle.dump(info, f)
img_summary = {}
img_summary['left'] = left
img_summary['right'] = right
img_summary['gt_depth'] = gt_depth
img_summary['gt_disp'] = gt_disp
if (onlyObj):
pred_disp[sample['label'] >= 17] = 0
pred_depth[sample['label'] >= 17] = 0
img_summary['disp_error'] = disp_error_img(pred_disp, gt_disp)
img_summary['depth_error'] = depth_error_img(
pred_depth, gt_depth)
img_summary['pred_disp'] = pred_disp
img_summary['pred_depth'] = pred_depth
save_images(self.train_writer, 'val' + str(val_count),
img_summary, self.epoch)
val_count += 1
logger.info('=> Validation done!')
if (perObject):
for key, value in objectCount.items():
perObjectDisp[key] = float(perObjectDisp[key]) / value
perObjectDepth[key] = float(perObjectDepth[key]) / value
print(perObjectDisp, perObjectDepth, objectCount)
mean_epe = val_epe / valid_samples
mean_d1 = val_d1 / valid_samples
mean_bad1 = val_bad1 / valid_samples
mean_bad2 = val_bad2 / valid_samples
mean_abs = val_abs / valid_samples
mean_mm2 = val_mm2 / valid_samples
mean_mm4 = val_mm4 / valid_samples
mean_mm8 = val_mm8 / valid_samples
# mean_thres1 = val_thres1 / valid_samples
# mean_thres2 = val_thres2 / valid_samples
# mean_thres3 = val_thres3 / valid_samples
# Save validation results
with open(val_file, 'a') as f:
f.write('epoch: %03d\t' % self.epoch)
f.write('epe: %.4f\t' % mean_epe)
f.write('d1: %.4f\t' % mean_d1)
f.write('bad1: %.4f\t' % mean_bad1)
f.write('bad2: %.4f\t' % mean_bad2)
f.write('abs: %.4f\t' % mean_abs)
f.write('mm2: %.4f\t' % mean_mm2)
f.write('mm4: %.4f\t' % mean_mm4)
f.write('mm8: %.4f\t' % mean_mm8)
# f.write('thres1: %.4f\t' % mean_thres1)
# f.write('thres2: %.4f\t' % mean_thres2)
# f.write('thres3: %.4f\n' % mean_thres3)
logger.info('=> Mean validation epe of epoch %d: %.3f' %
(self.epoch, mean_epe))
self.train_writer.add_scalar('val/epe', mean_epe, self.epoch)
self.train_writer.add_scalar('val/d1', mean_d1, self.epoch)
self.train_writer.add_scalar('val/bad1', mean_bad1, self.epoch)
self.train_writer.add_scalar('val/bad2', mean_bad2, self.epoch)
self.train_writer.add_scalar('val/abs', mean_abs, self.epoch)
self.train_writer.add_scalar('val/mm2', mean_mm2, self.epoch)
self.train_writer.add_scalar('val/mm4', mean_mm4, self.epoch)
self.train_writer.add_scalar('val/mm8', mean_mm8, self.epoch)
# self.train_writer.add_scalar('val/thres1', mean_thres1, self.epoch)
# self.train_writer.add_scalar('val/thres2', mean_thres2, self.epoch)
# self.train_writer.add_scalar('val/thres3', mean_thres3, self.epoch)
if not args.evaluate_only:
if args.val_metric == 'd1':
if mean_d1 < self.best_epe:
# Actually best_epe here is d1
self.best_epe = mean_d1
self.best_epoch = self.epoch
utils.save_checkpoint(args.checkpoint_dir,
self.optimizer,
self.aanet,
epoch=self.epoch,
num_iter=self.num_iter,
epe=mean_d1,
best_epe=self.best_epe,
best_epoch=self.best_epoch,
filename='aanet_best.pth')
elif args.val_metric == 'epe':
if mean_epe < self.best_epe:
self.best_epe = mean_epe
self.best_epoch = self.epoch
utils.save_checkpoint(args.checkpoint_dir,
self.optimizer,
self.aanet,
epoch=self.epoch,
num_iter=self.num_iter,
epe=mean_epe,
best_epe=self.best_epe,
best_epoch=self.best_epoch,
filename='aanet_best.pth')
else:
raise NotImplementedError
if self.epoch == args.max_epoch:
# Save best validation results
with open(val_file, 'a') as f:
f.write('\nbest epoch: %03d \t best %s: %.3f\n\n' %
(self.best_epoch, args.val_metric, self.best_epe))
logger.info('=> best epoch: %03d \t best %s: %.3f\n' %
(self.best_epoch, args.val_metric, self.best_epe))
# Always save the latest model for resuming training
if not args.evaluate_only:
utils.save_checkpoint(args.checkpoint_dir,
self.optimizer,
self.aanet,
epoch=self.epoch,
num_iter=self.num_iter,
epe=mean_epe,
best_epe=self.best_epe,
best_epoch=self.best_epoch,
filename='aanet_latest.pth')
# Save checkpoint of specific epochs
if self.epoch % args.save_ckpt_freq == 0:
model_dir = os.path.join(args.checkpoint_dir, 'models')
utils.check_path(model_dir)
utils.save_checkpoint(model_dir,
self.optimizer,
self.aanet,
epoch=self.epoch,
num_iter=self.num_iter,
epe=mean_epe,
best_epe=self.best_epe,
best_epoch=self.best_epoch,
save_optimizer=False)
