Some details about the training parameters. about twingan

[jerryli27]

I added two lines to the training script. It should work now.
--gradient_penalty_lambda=0.25 --use_unet=True

The whole script now looks like:

python pggan_runner.py
--program_name=twingan
--dataset_name="image_only"
# Assume you have data like 
# ./data/celeba/train-00000-of-00100.tfrecord,  
# ./data/celeba/train-00001-of-00100.tfrecord ...
--dataset_dir="./data/celeba/"
--unpaired_target_dataset_name="anime_faces"
--unpaired_target_dataset_dir="./data/anime_faces/"
--train_dir="./checkpoints/twingan_faces/"
--dataset_split_name=train
--preprocessing_name="danbooru"
--resize_mode=RESHAPE
--do_random_cropping=True
--learning_rate=0.0001
--learning_rate_decay_type=fixed
--is_training=True
--generator_network="pggan"
--use_unet=True
--num_images_per_resolution=300000
--loss_architecture=dragan
--gradient_penalty_lambda=0.25
--pggan_max_num_channels=256
--generator_norm_type=batch_renorm
--hw_to_batch_size="{4: 8, 8: 8, 16: 8, 32: 8, 64: 8, 128: 4, 256: 3, 512: 2}"

I haven't tested with the multi-gpu setting thoroughly yet due to limits in hardware, so yes there may be some bug, but you can try to add the following flags.

--sync_replicas=False
--replicas_to_aggregate=1
--num_clones=2
--worker_replicas=1

I updated the training readme with the comments above.

[jerryli27]

Hi @lionel3 I updated the training documentation. There was indeed a bug in my default parameters. After fixing that I am able to reproduce my previous results.

Please sync to the latest version and see https://github.com/jerryli27/TwinGAN/blob/master/docs/training.md .

The parameters I added are:

--do_pixel_norm=True
--l_content_weight=0.1
--l_cycle_weight=1.0

Please reopen this issue if you cannot reproduce. Thanks!

[jerryli27]

Yes you're right. Sorry for the wrong documentation. I'll push a newer version shortly.

1. The num_image_per_resolution I used was '300000'. Of course 600000 should also work, but it takes longer to train.
2. Please change to --resize_mode=RESHAPE.

FYI. The --do_random_cropping=True is in case You can try RANDOM_CROP as well if at inference time the quality is too bad because the face is not at the center of the image.

I am rerunning the exact code that I provided in the training example code. It will take a day or two for me to verify that it works.

[lionel3]

Thanks for your answer.

Besides, when training with
'hw_to_batch_size', '{4: 16, 8: 16, 16: 16, 32: 16, 64: 12, 128: 12, 256: 12, 512: 6}.
I got ResourceExhaustedError: OOM when allocating tensor with ... during fade-in phase from resolution 128 to 256. Same error when trying 2 GPUs.
I am not familiar with Tensorflow. I guess there may be some bug with Multi-GPU training.

I will try to reproduce the error and show more training details once I have idle GPU.

[lionel3]

Thanks, I will try it out asap.