Open c1a1o1 opened 5 years ago
It'd be helpful if you can share the dataset used, the script for training, and the results here. Thanks!
@jerryli27
For celeba to cat,
My train parameter:
--program_name=twingan --dataset_name="ren" --dataset_dir="datasets/ren2cat/ren/tfrecord/" --unpaired_target_dataset_name="cat" --unpaired_target_dataset_dir="datasets/ren2cat/cat/tfrecord/" --train_dir="./checkpoints/rencat/" --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}"
My test parameter:
--model_path="../checkpoint/rencat/256/" --image_hw=256 --input_tensor_name="sources_ph" --output_tensor_name="custom_generated_t_style_source:0" --input_image_path="../demo/face/var256/" --output_image_path="../demo/face/cat256/"
Is there some problem?
Thank you very much!
Please try to add --do_pixel_norm=True
to the training script. That stabilizes things a bit. Another tip is that the image output during training should look well at 32x32 or 64x64. If not, you don't need to train it all the way to 256. That'll be a waste of time.
And if you want, sharing the image and the tensorboard output would help a lot for debugging Let me know how it goes.
Thank you very much!
The result of my own training, using the image_translation_infer.py test, looks very bad.