Repository files navigation

This convert tools is base on TensorRT 2.0 Int8 calibration tools, which use the KL algorithm to find the suitable threshold to quantize the activions from Float32 to Int8(-127 - 127).

We provide the Classification(SqueezeNet_v1.1) and Detection(MobileNet_v1 SSD 300) demo based onncnn(a high-performance neural network inference framework optimized for the mobile platform) and the community ready to support this implementation.

The pull request in ncnn

Using this newncnn-quantization-tools, you can convert your ncnn model to ncnn int8 model directly. If you just want to deploy your model with ncnn,I suggest you use it.

For details, please read the following PDF:

8-bit Inference with TensorRT

MXNet quantization implementation:

Quantization module for generating quantized (INT8) models from FP32 models

An introduction to the principles of a Chinese blog written by my friend(bruce.zhang):

The implement of Int8 quantize base on TensorRT

The purpose of this tool(caffe-int8-convert-tool-dev.py) is to test new features, such as mulit-channels quantization depend on group num.

This format is already supported in thencnn latest version. I will do my best to transform some common network models intoclassification-dev

python caffe-int8-convert-tool-dev-weight.py -husage: caffe-int8-convert-tool-dev-weight.py [-h] [--proto PROTO] [--model MODEL]                                  [--mean MEAN MEAN MEAN] [--norm NORM]                                  [--images IMAGES] [--output OUTPUT]                                  [--group GROUP] [--gpu GPU]find the pretrained caffemodel int8 quantize scale valueoptional arguments:  -h, --help            show this help message and exit  --proto PROTO         path to deploy prototxt.  --model MODEL         path to pretrained caffemodel  --mean MEAN           value of mean  --norm NORM           value of normalize(scale value or std value)  --images IMAGES       path to calibration images  --output OUTPUT       path to output calibration table file  --group GROUP         enable the group scale(0:disable,1:enable,default:1)  --gpu GPU             use gpu to forward(0:disable,1:enable,default:0)python caffe-int8-convert-tool-dev-weight.py --proto=test/models/mobilenet_v1.prototxt --model=test/models/mobilenet_v1.caffemodel --mean 103.94 116.78 123.68 --norm=0.017 --images=test/images/ output=mobilenet_v1.table --group=1 --gpu=1

For example inMobileNet_v1_dev.table

conv1_param_0 0.0 3779.48337933 482.140562772 1696.53814502conv2_1/dw_param_0 0 72.129143 149.919382 // the convdw layer's weight scale every group is 0.0 72.129 149.919 ............conv1 49.466518conv2_1/dw 123.720796 // the convdw layer's bottom blobchannel scale is 123.720......

Three steps to implement theconv1 layer int8 convolution:

1. Quantize the bottom_blob and weight:

   bottom_blob_int8 = bottom_blob_float32 * data_scale(49.466518)weight_int8 = weight_float32 * weight_scale(156.639840)

2. Convolution_Int8:

   top_blob_int32 = bottom_blob_int8 * weight_int8

3. Dequantize the TopBlob_Int32 and add the bias:

   top_blob_float32 = top_blob_int32 / [data_scale(49.466518) * weight_scale(156.639840)] + bias_float32

quantized int8 inference

The following table show the Top1 and Top5 different between Float32 and Int8 inference.

The following table show the speedup between Float32 and Int8 inference. It should be noted that the winograd algorithm is enable in the Float32 and Int8 inference. The Hardware Platform is Hisi3519(Cortex-A17@880MHz)

Runtime Memory : mbytes

Storage Memory : mbytes

Thanks to NVIDIA for providing the principle of correlation entropy and ncnn's authornihui sharing his neural network inference framework.

Thanks to the help from the following friends:

Optimization Instructor :Fugangping,bruce.zhang

Algorithm :xupengfeixupf,JansonZhu,wangxinwei,lengmm

Python :daquexian

BSD 3 Clause