xw-hu / mask-shadowgan Goto Github PK

May I have the evaluation code? Thank you！My email is [email protected].

sorry ,i can't access the USR dataset‘s link,could you share the link with Baidu cloud ？Thank you !

Hi, I've try continue training from a checkpoint, and after several epoches the loss become very large(larger than 1w). I find that the learning rate of resume training is a value less than 0, which causes the above problem. I think the solution is to set the offset parameter of lambdaLR to 0.
the relative code is:
train_Mask-ShadowGAN.py: line 97-101:

lr_scheduler_G = torch.optim.lr_scheduler.LambdaLR(optimizer_G,lr_lambda=LambdaLR(opt.n_epochs, 0, opt.decay_epoch).step)
lr_scheduler_D_A = torch.optim.lr_scheduler.LambdaLR(optimizer_D_A,lr_lambda=LambdaLR(opt.n_epochs, 0, opt.decay_epoch).step)
lr_scheduler_D_B = torch.optim.lr_scheduler.LambdaLR(optimizer_D_B,lr_lambda=LambdaLR(opt.n_epochs, 0,opt.decay_epoch).step)

For instance, let's say I take a photo of a real object, crop to the object, and add it to a fake, all-white background. Could I then use this program to transfer the shadow from the original photo to the one with the all-white background?

Hi, I have trained the Mask shadow Gan on ISTD dataset following the given setting too. But I can’t reproduce the result on the paper either. My result is 9.52 (rmse value on LAB color space). I would like to perform some academic experiments comparing with my own results. I wonder whether I make any mistakes during my evaluation. Would you mind sharing your evaluation code or the parameter setting on ISTD dataset? Thanks a lot!!

I use GTX 1080 TI to train the model ,but it's too slow,i wonder how long did you train the code for 100 epochs?

Is there a pretrained model on your mentioned 'USR' dataset?

And I'm wonderring the generality of this model.

Could it be applied to other occasion directly?

Hi, thanks for the sharing of this project. I've trained the MaskShadowGAN on ISTD dataset. Just following the setting, I evaluate the result using RMSE metric on LAB color space(using my own code). The result is 8.87, which is far from 7.61 as reported in the paper. Could you please upload the code that you used for evaluation? I want to figure out what is wrong on my experiment. Thanks!

I am using GPU NVIDIA 2070 Super, when i run "python train_Mask-ShadowGAN" command its gives the following runtime error. Please guide me how to resolve.

RuntimeError: CUDA out of memory. Tried to allocate 39.12 MiB (GPU 0; 7.79 GiB total capacity; 6.27 GiB already allocated; 38.25 MiB free; 89.02 MiB cached)

if I don't do --n_cpu=1 it gives the error ValueError: signal number 32 out of range it

I reduce the batch size from 32 to 1 (use the different values) but the error is same.
I also tries to clear the cache but it didn't work as
import gc
gc.collect()
torch.cuda.empty_cache()

Please suggest me the solution. Thanks

I run train_Mask-ShadowGAN.py and have errors as following:
Traceback (most recent call last):
File "train_Mask-ShadowGAN.py", line 140, in
batch_size=opt.batchSize, shuffle=True, num_workers=opt.n_cpu)
File "F:\python3.5\lib\site-packages\torch\utils\data\dataloader.py", line 802, in init
sampler = RandomSampler(dataset)
File "F:\python3.5\lib\site-packages\torch\utils\data\sampler.py", line 64, in init
"value, but got num_samples={}".format(self.num_samples))
ValueError: num_samples should be a positive integeral value, but got num_samples=0
I don't know where to modify the dataset path or why len(self.data_ info)= =0.Can you give me some advice?
Thanks.

when I ran the test.py .There it is “FileNotFoundError: [Errno 2] No such file or directory: 'output/netG_A2B.pth' ”
So I went to run the train_Mask-ShadowGAN.py,and I also used the dataset of shadow_USR. However it happened ValueError: num_samples should be a positive integer value, but got num_samples=0.I've racked my brains. Is there any other way? Could you help me?

Hi xiaowei,
Thanks for sharing your code! I trained the model using your code and got many unpleasant results. I wonder if it is due to the algorithm itself or some other reasons like training tricks?

Examples(shadow to shadow-free):

(shadow-free to shadow):

I've trained the model on the given USR dataset.

None parameters were changed except the dataroot.

However, the generator was unable to convergence.

Meanwhile, the discriminator convergence better than the generator.

Hi xiaowei,
would you provide the trained model files? I've trained on USR_dataset, but since I'm not so familar with tuning the net, the result seems should be better.
Thanks a lot!

Hi xiaowei,
Thanks for sharing your code! I`m a student.I trained the model by running 'train Mask-ShadowGAN.py', but got unpleasant result. May I have your trained model file ? Thank you！My email is [email protected]

In order to construct unpaired data, for unpaired training on the ISTD train dataset, the shadow and shadow free images were randomly sampled? But such random sampling may yield paired data. ISTD have different shadow images for the same shadow-free image.
Was I wrong about the way to sample the unpaired data.