louisrouss / diffusion-based-model-for-colorization Goto Github PK

It is amazing that you have done this implementation. Looking at the other issue, I could see that training is needed. I don't have a lot of experience with this or a top of the line PC, but I think I can experiment with training with the equipment I have. I would like to support the project, but I would like to know if you could add some more detailed instructions on how to install, train and test it (for beginners).
Thank you very much.

The results shown in the paper look really promising and I am interested in using it for inpainting.
What do I need to do/change to use this for inpainting?
Is it a lot, because this would be my first project of this kind?

Hi @LouisRouss any idea when will you release the pretrained model for colorization ?

I had set the batchsize is ONE, but there is still “CUDA out of memory”.

My GPU is 3060 and the memory is 12GB, so how could I train the model, thanks!

Hi Louis! We're exploring new approaches to colorization. Ours at Hotpot.ai is currently GAN-based. Could you email us. Perhaps we could offer some compute and work on a collaboration?

Hi! Thank you for uploading a good report for colorization!

I conducted training based on your explanation, but error has occurred.

Config loaded Traceback (most recent call last): File "/content/drive/MyDrive/Diffusion-Based-Model-for-Colorization-main/main.py", line 19, in <module> main() File "/content/drive/MyDrive/Diffusion-Based-Model-for-Colorization-main/main.py", line 14, in main train(config) File "/content/drive/MyDrive/Diffusion-Based-Model-for-Colorization-main/src/train.py", line 6, in train trainer = Trainer(config) File "/content/drive/MyDrive/Diffusion-Based-Model-for-Colorization-main/src/trainer.py", line 48, in __init__ dataset_train = gray_color_data(self.path_train_color,self.path_train_grey) File "/content/drive/MyDrive/Diffusion-Based-Model-for-Colorization-main/src/dataloader.py", line 13, in __init__ self.data_color = np.load(path_color) File "/usr/local/lib/python3.7/dist-packages/numpy/lib/npyio.py", line 417, in load fid = stack.enter_context(open(os_fspath(file), "rb")) FileNotFoundError: [Errno 2] No such file or directory: '/content/drive/MyDrive/pixel-guide-diffusion-for-anime-colorization-main/data/HistoricalColor-ECCV2012/train.npy'

My conf.yml is like this.
MODE : 1 # 1 Train, 2 Validation IMAGE_SIZE : [224,224] CHANNEL_X : 1 CHANNEL_Y : 3 TIMESTEPS : 2000 MODEL_CHANNELS : 128 NUM_RESBLOCKS : 4 ATTENTION_RESOLUTIONS : [2,4,8] DROPOUT : 0 CHANNEL_MULT : [1,2,4,8] CONV_RESAMPLE : 'True' USE_CHECKPOINT : 'False' USE_FP16 : 'False' NUM_HEADS : 1 NUM_HEAD_CHANNELS : 64 NUM_HEAD_UPSAMPLE : -1 USE_SCALE_SHIFT_NORM : 'False' RESBLOCK_UPDOWN : 'False' USE_NEW_ATTENTION_ORDER : 'False' PATH_COLOR : '/content/drive/MyDrive/pixel-guide-diffusion-for-anime-colorization-main/data/HistoricalColor-ECCV2012/' PATH_GREY : '/content/drive/MyDrive/pixel-guide-diffusion-for-anime-colorization-main/data/Gray_HistoricalColor-ECCV2012/' BATCH_SIZE : 1 BATCH_SIZE_VAL : 8 ITERATION_MAX : 1000000 LR : 0.0001 LOSS : 'L2' VALIDATION_EVERY : 1000 EMA_EVERY : 100 START_EMA : 2000 SAVE_MODEL_EVERY : 10000

Where should I fix to proceed training?

I'm looking forward to hearing from you soon!

Hi Louis,
Really enjoying this repo. I'm attempting to train this model but don't have the train or val.npy file. I understand that they are a numpy array file of the training and validation set but am unsure about how to construct this. Could you provide example files or else instructions about this?
Thanks,
Rory

There is no definition of partial in trainer.py line 77, can you tell me how to implement it?

Can you kindly explain the methodology applied by you for achieving this ?

Hi there! I just found your repo and think this is a super worthy project and I saw you mention all you are lacking is compute. I joined the LAOIN discord a while back and they are a great community for discussing projects like this, theyre all about opensourcing these big models and multimodality and also seem quite generous in offering compute to worthy projects, they have a bunch of A100s at their disposal. They're also the ones behind the 5 billion image/text pair dataset recently released openly.

I was wondering if you wanted to join the discord?

Hope you can pop by and introduce yourself there https://discord.gg/AagfYHQd