[ACM MM 2022] [GT-MUST] GT-MUST: Gated Try-on by Learning the Mannequin-Specific Transformation
There are several arguments that can be used, which are
--data_root +str #where to get the images
--data_mpv #whether to load MPV dataset (employ dataset_mpv.py for DataLoader)
--data_mode +str #determine to get the training data or testing data
--stage +str #determine the traning stage, ILM, MWM, or GTM
--model_save_path +str #where to save the model during training
--result_save_path +str #where to save the inpainting results during testing
--num_iters +int #the max training iterations
--model_path +str #the pretrained generator to use during training/testing
--batch_size +int #the size of mini-batch for training
--n_threads +int
--test #test the model
--gpu_id +int #which gpu to use
--load_iter +int #for loading pretrained modules
To fully exploit the performance of the network, we suggest to use the following training procedure, in specific
- Train the first module ILM,
python run.py --data_root ../datasets/viton_resize --data_mode train --model_save_path checkpoint --model_path checkpoint --stage ILM --gpu_id 0 --num_iters 200000
Use the --data_mpv for MPV dataset, i.e.,
python run.py --data_root ../datasets/MPV --data_mpv --data_mode train --model_save_path checkpoint --model_path checkpoint --stage ILM --gpu_id 0 --num_iters 200000
- Train the second module MWM,
python run.py --data_root ../datasets/viton_resize --data_mode train --model_save_path checkpoint --model_path checkpoint --stage MWM --gpu_id 1 --load_iter 200000 --num_iters 200000
- Train the last module GTM,
python run.py --data_root ../datasets/viton_resize --data_mode train --model_save_path checkpoint --model_path checkpoint --stage GTM --gpu_id 2 --load_iter 200000 --num_iters 800000
- Test the model
python run.py --test --data_root ../datasets/viton_resize --data_mode test --gpu_id 1 --model_path checkpoint --result_save_path results --load_iter 200000
All the descriptions below are under the assumption that the size of mini-batch is 16,
| Module | ILM | MWM | GTM |
|---|---|---|---|
| Iters | 200K | 200K | 400K |
If you find the article or code useful for your project, please refer to
@inproceedings{wang2022gt,
title={GT-MUST: Gated Try-on by Learning the Mannequin-Specific Transformation},
author={Wang, Ning and Zhang, Jing and Zhang, Lefei and Tao, Dacheng},
booktitle={Proceedings of the 30th ACM International Conference on Multimedia},
pages={2182--2190},
year={2022},
doi = {10.1145/3503161.3547775},
url = {https://doi.org/10.1145/3503161.3547775}
}
See the Paper folder