csqiangwen / deepfillv2_pytorch Goto Github PK

could you provide a script to export an onnx model?

is this some wrong with me? the masked region is overlaped with some red strange patch.
python = 3.6.
torch = 1.5.0.

Hi there,

I have noticed that you have set the ratio between l1 loss, perceptual loss and adversarial loss as 100: 10: 1, which is not the same as the official implementation. The tensorflow official implementation set l1 loss: adversarial loss = 1: 1. Are you sure this ensure the reproducibility? I really doubt that.

By the way, mmediting version code also set l1 loss: adversarial loss = 1: 0.1. I have tried the ratio of 1: 1. But the model easily collapse. Does anyone know what cause this problem? Is it lead by the difference between PyTorch and TensorFlow?

I am a novice. Can you tell me some tips for training with other data sets

Hi!

Thanks for the awesome repo.
I have used the repo to create a google colab link for inferencing. The link to my repo is here

Please let me know if any other citations are required.

The train set and value set are placed under the 'dataset_path'? the train set is good pictures and the value is poor pictures.

Used the pretrained model against a test image of my own, and the results are not good at all. Curious if there are any thoughts on what might be causing this. Everything seems to run correctly.

Hello,

after running !bash ./run_test.sh

I get :

Traceback (most recent call last): File "/content/DeepFillv2_Pytorch/test.py", line 41, in <module> tester.WGAN_tester(opt) File "/content/DeepFillv2_Pytorch/tester.py", line 30, in WGAN_tester if not os.path.exists(results_path): NameError: name 'results_path' is not defined

Hi, thank you for your code.
I have some questions.

1. Is there a reason why sn is set to True in the class TransposeGatedConv2d of network_module.py file? It is used directly on the generator.
2. Do you think excluding perceptual loss in this implementation will not affect the result?

Thank you!

Hi, thanks for your great job! Could you share the discriminator pretrained model for training? Thanks again!

几个epoch之后就爆显存，把batch_size设置为1或2时不会出现这个问题，但是在test时直接把输入图像输出了，mask部分变成了黑色

Hi,
Firstly, Thanks for your sharing the code.
When I read your code, I found the last layer in generator is nn.Tanh, but the inputs of images with range [0, 1]
Could you please to explain the reason why you training like that ?

Hello, I'm interested in your DeepFillv2 Model for 2D Image Inpainting. And your suggested images show awesome at 2D Image Inpainting. I'd like to get more test examples that looks good on 2D Image Inpainting. Could you provide me more??
Why I left this issue is that your test images are different comparing with Original DeepFillv2 Examples but Inpainting performance is perfect, I think.
Thank you for your attention

Hi. Thanks for your re-implementation. I got several questions when I was training other settings according to the paper.
The author of DeepFillv2 said that this model can be trained with input channel of 5, which consists of masked image, mask and sketch. I want to train a model to inpaint image with user-guided sketch. But the model cannot converge all the time and the output result was a totally black color image.

Recently I had checked out another repo which is the pytorch re-implementation of the paper Contextual Residual Aggregation by Huawei. The first-stage model is just the same as DeepFillv2. But I notice that the hyper-parameters are not the same. Besides, I noticed that the loss function of the discriminator was hinge loss. But it wasn't implemented in your code. I was quite confused about this, could you please explain why the above issues happen? Thanks a lot!

Hello thanks for sharing your work.
I am trying training the model with CelebA datasets but the loss function when training is too large. Do you think this loss function is problematic for me?

when i tried the test, the result made me not satisfied

can you help me ? Thank you