To populate the runtime lookup-table (LUT) inference runtime model, generate .tflite models with different MBConv types (different kernel and expansion ratio values). The scripts below automate this process.
- To generate tflite models for all different MBConvs, run:
bash gen_tflite_models.sh
which executes
python main_tflite.py --tpu=$TPU_NAME --data_dir=$DATA_DIR --model_dir=${STORAGE_BUCKET}/model-runtime-model/model-tflite-$d-$k-$e --export_dir=$(pwd)/tflite-models/model-$d-$k-$e --depth_multiplier=$d --kernel=$k --expratio=$e --mode=train --post_quantize=True
which uses the '--export_dir' flag to generate the TFLite floal and quantized models, following the MNasNet+TFLite documentation.