Inquiry into training issue when training the GLCIC-Pytorch on the Mnist dataset

[yorklyb]

Hi! Thanks for your work! I encountered some issues when training the GLCIC-Pytorch on the Mnist dataset. I tried to adjust the parameters in train.py according to Mnist dataset (image size = 28*28, num = 48000). The following are the ones I’m using. arser = argparse.ArgumentParser() parser.add_argument('data_dir') parser.add_argument('result_dir') parser.add_argument('--data_parallel', action='store_true') parser.add_argument('--recursive_search', action='store_true', default=False) parser.add_argument('--init_model_cn', type=str, default=None) parser.add_argument('--init_model_cd', type=str, default=None) parser.add_argument('--steps_1', type=int, default=10000) parser.add_argument('--steps_2', type=int, default=20000) parser.add_argument('--steps_3', type=int, default=40000) parser.add_argument('--snaperiod_1', type=int, default=10000) parser.add_argument('--snaperiod_2', type=int, default=2000) parser.add_argument('--snaperiod_3', type=int, default=10000) parser.add_argument('--max_holes', type=int, default=1) parser.add_argument('--hole_min_w', type=int, default=7) parser.add_argument('--hole_max_w', type=int, default=7) parser.add_argument('--hole_min_h', type=int, default=3) parser.add_argument('--hole_max_h', type=int, default=3) parser.add_argument('--cn_input_size', type=int, default=28) parser.add_argument('--ld_input_size', type=int, default=14) parser.add_argument('--bsize', type=int, default=16) parser.add_argument('--bdivs', type=int, default=1) parser.add_argument('--num_test_completions', type=int, default=16) parser.add_argument('--mpv', nargs=3, type=float, default=None) parser.add_argument('--alpha', type=float, default=4e-4) parser.add_argument('--arc', type=str, choices=['celeba', 'places2'], default='celeba')

However, the training collapsed at phase 2:

I’m wondering if you could help solve this issue. Thanks in advance!

[qwexacc]

Hello, I also changed the size of the global picture and the size of the local picture during training, and the error reported is the same as yours, may I ask if you have solved the problem? Thanks in advance