Prepare quantized tflite models available on GPU/DSP/TPU

[kdh0102]

For more interesting experiments, models available on DSP, GPU are required.

As DSP needs INT8 quantized models (in case of GPU, INT8/FLOAT16), we need quantized tflite models:

• Current available
  • MobileNet V2
  • Inception V2/V3
  • YOLO, YOLO-tiny V3/V4
• Need quantization
  • Models in mediapipe repo
  • Object Detection (SSD)

Note

~Succeeded in making INT8 quantized YOLOv4-tiny.tflite model.~ Quantization fails due to the minimum filter size in a conv. layer.

But,

• 3 SPLIT ops fall back to CPU, creating 4 GPU delegate ops.
• DSP Delegate op is not generated. (needs more investigation.)

Even though quantization is done, the lack of op coverage can be a big obstacle.

[kdh0102]

Refer to our Google Drive for available models. Maybe we can create a .tflite model zoo Github repo later.