FoxPandaNet

The 3rd place winner of the 2020 On-device Visual Intelligence Competition (OVIC) of Low-Power Computer Vision Challenge (LPCVC), subtrack of Real-time Image Classification on LG G8.

Model

Tested with tflite-runtime 2.1.0 on Raspberry Pi 4 (single core)

Model	Input Image Size	Accuracy (UINT8)	Latency	Download	SHA256 Checksum
fpnet_pixel4 (uint8)	192x192	71.93%	60.1ms	Download Link	6e927d7af8da1eb9297017ebe92a67632ce73f612ff32cbfa3917f88d761a5f9
fpnet_dsp (uint8)	224x224	74.18%	94.9ms	Download Link	b77445326ef3f64fc8d3236b213e121aba5004dee4449deceb13f246477add4e
fpnet_fpga (uint8)	160x160	70.28%	48ms	Download Link	7b15050f0f2f723b13cfc001026a36133f78103049c94ae8fe023807e355fc20

Compared with MobileNetV3/MobileNetV2 on Raspberry Pi 4 CPU (single core)

Training and Evaluation

$ python3 train.py --model_name fpnet_pixel4 \
  --batch_size 1024 \
  --epochs 250 \
  --warmup_epochs 5 \
  --base_lr 0.4 \
  --init_lr 0.1 \
  --image_size 192 \
  --use_cache \
  --imagenet_path $IMAGENET_PATH \
  --checkpoint_path $CHECKPOINT_PATH
$ python3 convert_quant.py --keras_model_file $KERAS_FILE \
  --output_file $TFLITE_FILE \
  --image_size 192 \
  --imagenet_path $IMAGENET_PATH
$ python3 val_quant.py --tflite_model_file $TFLITE_FILE \
  --image_size 192 \
  --imagenet_path $IMAGENET_PATH

Methodology

• Neural architecture search with multivariate regression
• Once-for-All supernet
• MobileNet V3 backbone
• Replacing Hard-swish with ReLU6 for better quantization performace