Hi,

I trained a model with unet_learner but I can't convert it to jit.

I run the following code: torch.jit.save(torch.jit.script(learn.model), 'jit.pt')

Here is the error:

UnsupportedNodeError: GeneratorExp aren't supported: File "/usr/local/lib/python3.7/dist-packages/fastai/callbacks/hooks.py", line 21 "Applieshook_functomodule,input,output." if self.detach: input = (o.detach() for o in input ) if is_listy(input ) else input.detach() ~ <--- HERE output = (o.detach() for o in output) if is_listy(output) else output.detach() self.stored = self.hook_func(module, input, output)

May I know how you convert it to a jit model? Thanks

Добрый вечер.В ArcaneGAN на colab for videos,выдаёт ошибку:

RuntimeError: CUDA out of memory. Tried to allocate 2.80 GiB (GPU 0; 11.17 GiB total capacity; 5.74 GiB already allocated; 2.21 GiB free; 8.44 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

Помогите пожалуйста!

Nice work! My only confusion is how you change the style of the whole image instead of just the face. Usually, StyleGAN generates aligned face images by fine-tuning the FFHQ checkpoint. How does the pix2pix model trained with these face image pairs work with the full image or frame.

Hi, what does the architecture look like? Is it similar to Pix2Pix? And for processing of the video, are you doing anything extra to make sure the frames are consistent?

Hello, I try your model and it's amazing, but I find in some pictures if the nose is too big, there will be eyes in the nose. I try to lower the 'target_face' and it can work. But the details like the light of the eyes and background will also lose when I lower the 'target_face'. So I wonder is there a way to prevent the eyes occurs in the nose and keep the details in the meantime?

Great work!

The unet prediction result will be cropped to be the same size as the training input, e.g. 256 or 512. For arbitrary image size (e.g. 1280*720), how to config or set the model to output the same size of the input image as your colab did? Thank you.

Добрый вечер.Извините,это опять я.Снова эта ошибка появляется.Можно ли,самому эту ошибку решать?Или исправлять можете только вы?Обьясните пожалуйста подробно.

how do you make sure the background and expression similarity between the generated input(face) and target(style face) ? I find that the style is too weak when less finetune and the similarity is too weak when more finetune, how do you solve it ? Would you like to share the paired datasets generated code with me ? thanks a lot ~

Interesting topic... I wonder how you trained the model, especially the augmentation part. Fixed crop limitation is a well-known problem and would like to know how you handle it. :)

Was trying the ArcaneGan video colab but I am having a tuple issue can you please help, i am really excited to try the Arcane video can you please help out

Hi,

I want to train the Fastai u-net model. However, when I try to train the critic (learn_critic.fit_one_cycle(6, 1e-3)), I get the following error:

CUDA out of memory. Tried to allocate 4.00 GiB (GPU 0; 14.76 GiB total capacity; 9.78 GiB already allocated; 891.75 MiB free; 12.57 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

The GPU is a Tesla T4 with 16 GB of VRAM. My batch size is 4 and the training images size is 512*512. I also tried with lower numbers, but I'm still getting the same error.

The following are input image, my training output from pair label supervision, and the output from your test model。 I trained my model (Super-Resolution model) on the images from your model outputs, I find it difficult to change the facial features。 Like the eyes and face texture are changed, how to do it ? I use L1Loss (weight is 1) + PerceptualLoss (weight is 1)+ GANLoss (weight is 0.1),