TensorFlow Lite now supports converting weights to 16-bit floating point values during model conversion from TensorFlow to TensorFlow Lite's flat buffer format. This results in a 2x reduction in model size.
However, this will insert new dequantize for ops (like conv2d) used for dequantize fp16 weight to fp32. Like this:
TVM doesn't support this behavior. List the things we mainly should to do:
- Support float16 type inside tflite parser
- Extend
dequantize to support fp16 to fp32
Related issue:#5774
TensorFlow Lite now supports converting weights to 16-bit floating point values during model conversion from TensorFlow to TensorFlow Lite's flat buffer format. This results in a 2x reduction in model size.
However, this will insert new
dequantizefor ops (likeconv2d) used fordequantizefp16 weight to fp32. Like this:TVM doesn't support this behavior. List the things we mainly should to do:
dequantizeto support fp16 to fp32Related issue:#5774