How to train on my own dataset?
sunshineatnoon opened this issue · comments
Hi, Thanks for open-sourcing this awesome work. I would like to train the model on my own dataset. So far, I have pre-processed all images to size 256x256
by using the scripts/dataset_tool.py. Here are the issues I met when trying to train on my own images:
- How to generate image list? I used the following command to generate a list but not sure if this is correct, I actually didn't see the
datasets/ffhq/ffhq_256.txt
file when training on the FFHQ dataset.python3 -m tl2.tools.get_data_list --source_dir datasets/my_images/downsample_ffhq_256x256/ --outfile datasets/my_images/ffhq_256.txt --ext *.png
- How to change the yaml file ffhq_exp.yaml to point to my own dataset directory?
- How to pass hyperparameters to the model? I tried to use the training command in the old readme below:
But I'm not sure how to train on 32x32 images (I'd like a quick tryout), or changing the batch_size, etc. I looked into the
export CUDA_HOME=/usr/local/cuda-10.2/ export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 export PYTHONPATH=. python exp/dev/nerf_inr/scripts/train_v16.py \ --port 8888 \ --tl_config_file configs/train_ffhq.yaml \ --tl_command train_ffhq \ --tl_outdir results/train_ffhq \ --tl_opts curriculum.new_attrs.image_list_file datasets/ffhq/images256x256_image_list.txt \ D_first_layer_warmup True
tl2
library but failed to find any documentation.
Thanks for your time and any help would be appreciated!