TensorFlow benchmark scripts for single and multi Nodes with Multi GPUs
git clone https://github.com/reger-men/tensorflow_benchmark.git
pip3 install -r requirements.txt
Train with custom loop:
python3 train.py -train-mode loop
Train with Keras fit:
python3 train.py -train-mode fit
Mirrored strategy will be used as defalt with num_gpus>1
python3 train.py -train_mode fit -num_gpus 2
Loop still under development.
Experimental: launch Multi-Nodes training from the chief Worker Node
python3 train.py --train_mode=fit --workers="localhost:122,localhost:123" --w_type="worker" --w_index=0 --distribution_strategy=MultiWorker
Loop still under development.