Training details

[Akila-Ayanthi]

Hi,
This is exactly what I was looking for. Thank you.
But, it is not clear to me how the training needs to be done. Could you please help me with that?

[moono]

I used followings to train on FFHQ dataset,

• Refer to FFHQ dataset repo to create *.tfrecord for resolution of your interest.
• Please use docker for training envrionment.
  • Recommended docker image : nvcr.io/nvidia/tensorflow:21.03-tf2-py3
• Train with train.py
  • example 256x256 resolution

# adjust shuffle_buffer_size and batch_size_per_replica to fit your hardware
~$ python train.py \
--allow_memory_growth false \
--use_tf_function true \ 
--use_custom_cuda true \
--model_base_dir <directory-of-output-trained-model> \
--tfrecord_dir <FFHQ-tfrecord-dir> \
--train_res 256 \
--shuffle_buffer_size 5000 \
--batch_size_per_replica 16

[Akila-Ayanthi]

Thank you.