Tensorflow2 with jit (xla) compiling on multi-gpu training.

• CIFAR and ILSVRC training code with jit compiling and distributed learning on the multi-GPU system.
• I highly recommend using Jit compiling because most of the algorithm is static and can be compiled, which gives memory usage reduction and training speed improvement.
• This repository is built by custom layers and custom training loop for my project, but if you only want to check how to use jit compiling with distributed learning, check 'train.py' and 'op_util.py'.

Requirement

• Tensorflow >= 2.5
• Pillow

How to run

• ILSVRC

python train.py --compile --gpu_id {} --dataset ILSVRC --data_path /path/to/your/ILSVRC/home --train_path /path/to/log

• CIFAR{10,100}

python train.py --compile --gpu_id {} --dataset CIFAR{10,100} --train_path /path/to/log

Experimental results

• I used four 1080ti.
• Jit compiling gives a 40% speedup for training time.

	Accuracy	Training time
Distributed only	75.83	94.61
Distributed with Jit	75.57	56.98