Error when train the model

[hanchaoyuan]

During training, this error occurs！Do you know how to solve?

After I declared the variable I got another error

[captain-pool]

Facing the same problem. For the first one, I set imfd_initial to '' and trained the MTM model. Next when I try to train the DRM model, it fails since input['person_fdepth'] is not a torch tensor, instead it's a list which looks like ['', '', '']. How can I fix this? CC: @fyviezhao

[captain-pool]

This is most likely happening because, the data loader is setting imfd to '' in case of DRM or TFM training and returning it.

M3D-VTON/data/aligned_MPV3dDataset.py

Lines 184 to 187 in 3c9d16c

	elif self.model == 'DRM' or self.model == 'TFM':
	imfd = ''
	imfd_initial = np.load(os.path.join(self.warproot, 'initial-depth', im_name.replace('whole_front.png', 'initial_front_depth.npy')))
	imfd_initial = torch.from_numpy(imfd_initial).unsqueeze(0)

This data dictionary is being sent to DRM model's set_input() function, where it is being loaded as a torch tensor and being used to compute gradients.

M3D-VTON/models/DRM_model.py

Lines 142 to 152 in 3c9d16c

	if self.isTrain:
	self.imfd = input['person_fdepth'].to(self.device) # for ground truth
	self.imbd = input['person_bdepth'].to(self.device) # for ground truth
	if self.use_grad_loss:
	self.fgrad = self.compute_grad(self.imfd) # for ground truth
	self.bgrad = self.compute_grad(self.imbd) # for ground truth
	if self.use_normal_loss or self.use_gan_loss:
	self.im_mask = input['person_mask'].to(self.device) # for input
	self.imfn = util.depth2normal_ortho(self.imfd).to(self.device) # for ground truth
	self.imbn = util.depth2normal_ortho(self.imbd).to(self.device) # for ground truth

[hanchaoyuan]

This is most likely happening because, the data loader is setting imfd to '' in case of DRM or TFM training and returning it.

M3D-VTON/data/aligned_MPV3dDataset.py

Lines 184 to 187 in 3c9d16c

	elif self.model == 'DRM' or self.model == 'TFM':
	imfd = ''
	imfd_initial = np.load(os.path.join(self.warproot, 'initial-depth', im_name.replace('whole_front.png', 'initial_front_depth.npy')))
	imfd_initial = torch.from_numpy(imfd_initial).unsqueeze(0)

This data dictionary is being sent to DRM model's set_input() function, where it is being loaded as a torch tensor and being used to compute gradients.

M3D-VTON/models/DRM_model.py

Lines 142 to 152 in 3c9d16c

	if self.isTrain:
	self.imfd = input['person_fdepth'].to(self.device) # for ground truth
	self.imbd = input['person_bdepth'].to(self.device) # for ground truth
	if self.use_grad_loss:
	self.fgrad = self.compute_grad(self.imfd) # for ground truth
	self.bgrad = self.compute_grad(self.imbd) # for ground truth
	if self.use_normal_loss or self.use_gan_loss:
	self.im_mask = input['person_mask'].to(self.device) # for input
	self.imfn = util.depth2normal_ortho(self.imfd).to(self.device) # for ground truth
	self.imbn = util.depth2normal_ortho(self.imbd).to(self.device) # for ground truth

I think so, but I have not found a suitable solution, do you have a good solution?

[jasmine-97]

I face the same problem, so could you tell me how you solve the problem in MTM model?

[fyviezhao]

Sorry for the late reply. Please check the latest code for training, which should now work well.