MobileNetV3 TPU Estimator implementation

MobileNetV3 tf.keras implementation with pre-trained weights using the TPU Estimator API.

Requirements

I have tested this implementation using tensorflow 1.13

ImageNet pre-trained weights

	Top-1 Acc.	Top-5 Acc.	Path to weights:
V3Large	75.0%	92.0%	Google drive
V3Small	67.3%	87.4%	Google drive

Train on ImageNet

Please see main.py for detailed information on available flags.

export TPU_NAME=<your TPU name>
export MODEL_NAME=MobileNetV3Small 
export STORAGE_BUCKET=<your imagenet bucket location>
export DATA_DIR=${STORAGE_BUCKET}
export OUTPUT_DIR=${STORAGE_BUCKET}/mobilenet-test
python3 main.py --tpu=$TPU_NAME --data_dir=$DATA_DIR --model_name=$MODEL_NAME \
                --model_dir=$OUTPUT_DIR --export_dir=$OUTPUT_DIR/export 

Acknowledgement

This repo's implementation is built upon the codebase of MnasNet's TPU implementation as well as that of Single-path NAS.