search

cysmith / neural-style-tf

TensorFlow (Python API) implementation of Neural Style
GNU General Public License v3.0
3.11k stars 828 forks source link

TF2 Migration, issue with L-BFGS-B #123

Open splicergroup opened 2 years ago

splicergroup commented 2 years ago

I was able to get this to work with Tensorflow 2 on Windows 10 by converting a couple things for compatibility. However I'm stumped on one problem regarding the optimizer portion of the code. The Adam optimizer works fine with optimizer = tf.compat.v1.train.AdamOptimizer(args.learning_rate)

The L-BFGS-B optimizer does not seem to work with Tensorflow 2 as it's been removed. This is the part that needs to be migrated to work with Tensorflow 2

optimizer = tf.contrib.opt.ScipyOptimizerInterface(
    loss, method='L-BFGS-B',
    options={'maxiter': args.max_iterations,
        'disp': print_iterations})

I looked into tfp.optimizer.lbfgs_minimize but I have no idea how to get it working in the existing code. Do you have any ideas? I am using Nvidia Ampere architecture cards (RTX 3090) that do not support Tensorflow 1.x due to CUDA compute version. Would really like to use L-BFGS instead of Adam for better results.

sbetzin commented 2 years ago

I had the same problem and stopped there with the Migration. Has anyone an migrated script to TF2 ?