train error： RuntimeError: the derivative for 'weight' is not implemented
Marshall-yao opened this issue · comments
Hi, thanks for your outstanding work.
Problem:
I met an error when i finetune with pretrained model. RuntimeError: the derivative for 'weight' is not implemented. The details are as follows.
After pre-training with L1 loss using DIV2K dataset (200 epochs), I plan to finetune on the pre-trained model with GAN (200 epochs).
The loss of pretraining with L1 loss is here.
< Pretrain >
Model check_point/my_model. Epoch [200/200]. Learning rate: 5e-05
100%|██████████████████████████████████████████████████████████████████████████████████| 1000/1000 [02:08<00:00, 7.77it/s]
Finish train [200/200]. Loss: 5.98
Validating...
100%|██████████████████████████████████████████████████████████████████████████████████████| 10/10 [00:06<00:00, 1.62it/s]
Finish valid [200/200]. Best PSNR: 27.5345dB. Cur PSNR: 27.4683dB
Followings are my command of pretrain .
< Pretrain >
python train.py --phase pretrain --learning_rate 1e-4
YOUR SETTINGS
fl_gamma: 1
valid_dataset: PIRM
num_epochs: 200
gan_type: RSGAN
check_point: check_point/my_model
spectral_norm: False
batch_size: 16
alpha_vgg: 50
res_scale: 0.1
focal_loss: True
lr_step: 120
snapshot_every: 10
GP: False
scale: 4
train_dataset: DIV2K
alpha_tv: 1e-06
learning_rate: 0.0001
alpha_gan: 1
pretrained_model:
num_valids: 10
num_channels: 256
num_repeats: 20
patch_size: 24
phase: pretrain
num_blocks: 32
alpha_l1: 0
Then i use pretrained model of best_model.pt saved in check_point/my_modedl/pretrain to finetune the model with GAN . It gave the error of RuntimeError: the derivative for 'weight' is not implemented.
Command of finetune .
python train.py --pretrained_model check_point/my_model/pretrain/best_model.pt
YOUR SETTINGS
num_repeats: 20
GP: False
spectral_norm: False
snapshot_every: 10
num_epochs: 200
gan_type: RSGAN
num_channels: 256
lr_step: 120
alpha_vgg: 50
num_blocks: 32
alpha_l1: 0
phase: train
num_valids: 10
batch_size: 16
focal_loss: True
valid_dataset: PIRM
pretrained_model: check_point/my_model/pretrain/best_model.pt
scale: 4
res_scale: 0.1
alpha_gan: 1
train_dataset: DIV2K
check_point: check_point/my_model
alpha_tv: 1e-06
learning_rate: 5e-05
patch_size: 24
fl_gamma: 1
Loading dataset...
Loading model using 1 GPU(s)
Fetching pretrained model check_point/my_model/pretrain/best_model.pt
Model check_point/my_model/train. Epoch [1/200]. Learning rate: 5e-05
0%| | 0/1000 [00:02<?, ?it/s]
Traceback (most recent call last):
File "train.py", line 323, in
main()
File "train.py", line 257, in main
G_loss = f_loss_fn(pred_fake - pred_real, target_real) #Focal loss
File "/home/anaconda3/lib/python3.5/site-packages/torch/nn/modules/module.py", line 489, in call
result = self.forward(*input, **kwargs)
File "/home/PESR-master/model/focal_loss.py", line 13, in forward
return F.binary_cross_entropy_with_logits(x, t, w)
File "/home/anaconda3/lib/python3.5/site-packages/torch/nn/functional.py", line 2077, in binary_cross_entropy_with_logits
return torch.binary_cross_entropy_with_logits(input, target, weight, pos_weight, reduction_enum)
RuntimeError: the derivative for 'weight' is not implemented
Does anyone have some idea about this problem?
I guess that the derivative for weight is not supported in binary_cross_entropy. You can follow this post.
OK. Thanks so much.
I will have a try.
Best regards.
After reading the post you recommended, i think it needs the derivative, but i dont know how to implement binary_cross_entropy_with_logits.
Could you give some suggestions ?
To my knowledge, you dont need the derivative for focal loss weight. So just detach it.
@thangvubk
Thanks so much. I misunderstood the post you commented.
I have changed the w with w.detach() in focal_loss.py .Then it runs well.
Then i will test the model to check if it reproduces well.