So far(Oct, 25, 2020), it seems official ACL-GAN's trained models are not available.
I wanted to try a selfie2anime model, so I trained it from scratch.
A model is uploaded at models/gen_00570000.7z
. I compressed a model to avoid github limit size.
My models wasn't trained along original way.
I used @xunings's (unofficial) config file. thx ;-)
Original paper says, models were trained for 350K iter and batch_size is 3
. However, I reduce batch size parameter as below and trained for 350K * 3
iterations, due to my poor GPU resources. Sorry for a little complicated way.
- Until 90,000 iter, batch_size is
1
- Until 570,000 iter, batch_size is
2
Notice I FORGOT changing step_size
. So the latter part of train may not work.
For test:
First, you need to extract models/gen_00570000.7z
.
And then,
$ python test.py --config configs/selfie2anime.yaml --input <YOUR IMAGE_FILE PATH> --output_folder <YOUR OUTPUT_DIR PATH> --checkpoint ./models/test.pth
e.g. $ mkdir results; python test.py --config configs/selfie2anime.yaml --input inputs/test_male.jpg --output_folder results --checkpoint ./models/gen_00570000.pt