Update
====
Some critical bugs of ONNX importer are fixed.
Quantization mechanism is improved. 8-bit quantization works well for the included examples.
- Algorithm to determine right-shift amounts in quantization is improved. Quantization speed is enhanced.
- Bug fixes of ONNX importer for Gemm: The weight matrix of a Gemm (matmul) is transposed when its input has a different layout. A bug of Bias/Scale bit-width of Gemm is also fixed.
- A bug fix of ONNX importer for GlobalAveragePool: In the previous version, max_pool is wrongly generated. It has been replaced with avg_pool.
- Examples of VGG-11 and ResNet-18 are improved.
Test environment
====