search

LTH14 / rcg

PyTorch implementation of RCG https://arxiv.org/abs/2312.03701
MIT License
785 stars 36 forks source link

custom train #14

Open swayampragnya-malla opened 8 months ago

swayampragnya-malla commented 8 months ago

Anyone has code for custom training on own dataset???

LTH14 commented 8 months ago

You could consider changing the default ImageNet dataset to your own dataset by changing the data_path, if your dataset follows a similar structure as ImageNet. You can also replace the original dataset with your own dataset implementation here.

LTH14 commented 8 months ago

I saw your comment in email, but I cannot find it here. Your question:

Considered the datasets size, I use the downsample version of ImageNet32(32*32). And I change the input_size parameter and so on, but still can not run main_rdm.py. I checked config/rdm/mocov3vitb_simplemlp_l12_w1536.yaml, is there any parameter in this file shall I change before training? Thank u so much.

The current codebase only supports 256x256 input size. For 32x32 input size you will need a different VQGAN and a different pre-trained Moco v3.