vgg-fpn

vgg-fpn implemented by MXNet

Before optimization

# For GPU
python benchmark.py

# For CPU
export KMP_AFFINITY=granularity=fine,noduplicates,compact,1,0
export OMP_NUM_THREADS=CORES_PER_SOCKET
python benchmark.py --dev cpu

After optimization

1. Generate fusion and quantization format model

python ocr_gen_qsym_mkldnn.py

2. Run fusion+quantization

export KMP_AFFINITY=granularity=fine,noduplicates,compact,1,0
export OMP_NUM_THREADS=CORES_PER_SOCKET
python inference.py --symbol-file=OCR-quantized-1batches-naive-symbol.json --param-file=OCR-quantized-0000.params --batch-size=16