which parameter did you set during the train ?

[vanillaHill]

Hello,
I've trained it with the places365_standard, with batch_size=16, the mask that you mentionned on the issue #1 and the other parameters by default.
But i didn't get that interesting results.
Could you tell me what parameters did you put to get the results that you show in your video?

[erhuodaosi]

Hello,
I've trained it with the places365_standard, with batch_size=16, the mask that you mentionned on the issue #1 and the other parameters by default.
But i didn't get that interesting results.
Could you tell me what parameters did you put to get the results that you show in your video?
Hi,I think the solution to the problem is that you could train more steps or you could use the model the author given,you could compare with them,maybe you could change the loss weight,they are different with the original paper,wish you could get the wonderful results as the videos.

I have a question,while training the model,have you met such similar question:
(py3) [fl@ibcu05 place2]$ python train.py
Loaded training dataset with 1434892 samples and 55116 masks
Loaded model to device...
Setup Adam optimizer...
Setup loss function...

EPOCH:0 of 3 - starting training loop from iteration:0 to iteration:89680

0%| | 0/89680 [00:00<?, ?it/s]
Traceback (most recent call last):
File "train.py", line 141, in
loss_dict = loss_func(image, mask, output, gt)
File "/home/fl/.local/lib/python3.6/site-packages/torch/nn/modules/module.py", line 489, in call
result = self.forward(*input, **kwargs)
File "/home/fl/place2/loss.py", line 94, in forward
loss_dict["tv"] = total_variation_loss(composed_output, self.l1) * LAMBDAS["tv"]
File "/home/fl/place2/loss.py", line 50, in total_variation_loss
loss = l1(image[:, :, :, :-1] - image[:, :, :, 1:]) + l1(image[:, :, :-1, :] - image[:, :, 1:, :])
File "/home/fl/.local/lib/python3.6/site-packages/torch/nn/modules/module.py", line 489, in call
result = self.forward(*input, **kwargs)
TypeError: forward() missing 1 required positional argument: 'target'

I would very appreciate if you could help me, thank you!

[erhuodaosi]

Hi,I think the solution to the problem is that you could train more steps or you could use the model the author given,you could compare with them,maybe you could change the loss weight,they are different with the original paper,wish you could get the wonderful results as the videos.

[chen849157649]

@erhuodaosi
loss function error, modify :
def total_variation_loss(image):
# shift one pixel and get difference (for both x and y direction)
loss = torch.mean(torch.abs(image[:, :, :, :-1] - image[:, :, :, 1:])) +
torch.mean(torch.abs(image[:, :, :-1, :] - image[:, :, 1:, :]))
return loss

[erhuodaosi]

@chen849157649
Forgive me for seeing the information today.And the program could run perfectly!Thank you very much!