could not reproduce paper results (resolved. thanks)
rptxk01 opened this issue · comments
Thank you for sharing your outstanding work.
I followed the training order in the README file.
After pre-training with L1 loss using DIV2K dataset (200 epochs), I fine-tuned on the pre-trained model with GAN (200 epochs).
All arguments were set to the default values of your code.
The losses of my train are here.
< Pretrain >
Model check_point/my_model. Epoch [200/200]. Learning rate: 5e-05
Finish train [200/200]. Loss: 6.02
Validating...
Finish valid [200/200]. Best PSNR: 27.5275dB. Cur PSNR: 27.4693dB
< Train >
Model check_point/my_model/train. Epoch [200/200]. Learning rate: 2.5e-05
Finish train [200/200]. L1: 0.00. VGG: 101.57. G: 10.80. TV: 5.45. Total G: 117.82. D: 0.05
Validating...
Finish valid [200/200]. PSNR: 24.6043dB
The test results of newly trained model is much blurred compared to your paper results and well not preserving edge component.
How should I train to reproduce your paper results?
PS. Followings are my command.
< Pretrain >
python train.py --phase pretrain --learning_rate 1e-4
YOUR SETTINGS
scale: 4
train_dataset: DIV2K
valid_dataset: PIRM
num_valids: 10
num_channels: 256
num_blocks: 32
res_scale: 0.1
phase: pretrain
pretrained_model:
batch_size: 16
learning_rate: 0.0001
lr_step: 120
num_epochs: 200
num_repeats: 20
patch_size: 24
check_point: check_point/my_model
snapshot_every: 10
gan_type: RSGAN
GP: False
spectral_norm: False
focal_loss: True
fl_gamma: 1
alpha_vgg: 50
alpha_gan: 1
alpha_tv: 1e-06
alpha_l1: 0
< Train >
python train.py --pretrained_model check_point/my_model/pretrain/best_model.pt
YOUR SETTINGS
scale: 4
train_dataset: DIV2K
valid_dataset: PIRM
num_valids: 10
num_channels: 256
num_blocks: 32
res_scale: 0.1
phase: train
pretrained_model: check_point/my_model/pretrain/best_model.pt
batch_size: 16
learning_rate: 5e-05
lr_step: 120
num_epochs: 200
num_repeats: 20
patch_size: 24
check_point: check_point/my_model
snapshot_every: 10
gan_type: RSGAN
GP: False
spectral_norm: False
focal_loss: True
fl_gamma: 1
alpha_vgg: 50
alpha_gan: 1
alpha_tv: 1e-06
alpha_l1: 0
Hi. Can you share the image comparison shown on tensorboard
Thank you for the reply :)
Following is your paper result. (DIV2K 0804.png)
I was very impressed to see that the line detail of the third man's shirt from the left is well preserved.
Despite using the GAN-based method, the lines of the shirt are clearly restored as if they were reconstructed using PSNR-based methods.
However, when I applied my newly trained model using your code, I got the following results.
The details of the sand are satisfactory, but the line of the shirt was blurred.
I think there was a mistake in my training process.
Could you give me some advice to reproduce the paper results?