记录下我对fp32转int8或者fp16的理解，有不对地方欢迎指正。
tensort支持通过输入大量图像的办法获得int8或者fp16的权重。获得低bit权重就能获得更快的计算速度。 最近在做FPGA，以我对FPGA的理解，xilinx的zynq系列soc中的FPGA内包含的DSP模块，这个DSP不是DSP芯片，是xilinx设计的DSP电路，可以完成乘法累加操作。
乘累加有个高大上的英文缩写，MACC。众所周知，卷积都是转换成矩阵做乘法，而矩阵乘法的基本操作就是乘累加。
xilinx的DSP可以在一个时钟周期内完成两次int8的乘累加，因为DSP输入位宽固定，输入数的位宽越小，一个时钟周期内算出来的结果也就越短，并行度也就越大。
再如以ARM为例， 这是我做的ARM neon指令集的3×3卷积程序，https://github.com/canteen-man/arm_neon_conv_3-3 int类型的整数最高可以以SIMD的计算方式达到8个寄存器同时计算，而float32则没有这么大的并行度。
所以如果cnn的权重都用int8，会根据不同的硬件设备加快计算速度。
但需要注意的是，这不能降低读写的速度，具体为什么，就要研究下float32是怎么转到int8的了。
参考了网上其他人的总结和分享，我大致理解float32转int8的流程如下:
首先对已有的float32做直方图统计，既然要做直方图统计就要分成每一个小柱形的统计区间，float32分布密集，就要自己设计统计小区间，按照nvidia的设计，是2048个bin。不过这2048个bin貌似是一侧的，以0位中心分为左右两侧。
不能直接把最大的两个值映射到int8的最大值127，因为饱和映射没有不饱和映射好，这个好理解，不能为了一个极端的值而压缩了另外许多值。
既然是非饱和映射，就要有阈值T。
下一步，利用KL散度优化，开始不断的取阈值T，求出KL散度优化最小的T，即为要做的阈值。这里其实只用从127到2048这个区间试就行了，要是127以下的那就好了，就直接一一映射了。
最近发现channel pruning，BNN和int8量化的核心都是优化方法。
想想channel pruning，刚读论文的时候还很是惊讶，什么索套回归，之前都没听过。后来发现巧妙之处，太巧妙了，这索套回归的优化目的和cnn删除冗余卷积核后保证特征图重构误差最小的优化目的完全相同。
研究BNN时发现也是优化方法，研究转int8的时候又发现KL散度优化还是优化方法。
如果我要是博士的话，我就先把优化方法彻底的好好学一遍，然后没准哪个最优化的公式就能套进cnn，没准就能发个CCF A类。
得出阈值T后就能映射到int8了，所以就会有个scale。
但这个scale是从权重读进后再做scale，因为每一层的scale不同，每次做完卷积要再从int8回到float32，再从float32转到下一次int8，所以读取速度就没变。
按照ARM+FPGA的理解，就是AXI总线的传输效率没变，传输的还是float32权重，每次从DDR读图或者读权重的时间没变，读进BRAM后在折腾成int8，再输入DSP做计算，加速是加速在计算这步时间了。
貌似在哪听说tensorrt转换完的trt模型不能打印出权重，所以想拿tensort转换int8后用到其他地方的思路貌似是走不通了。
tensorrt转int8效果好的另一个原因是他是输入标定图像的，这个标定图像不是标定相机的那意思，就是把一堆图像输入，这样统计激活的分布就会更准，然后做KL散度优化的时候就更有代表性。
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(-128 - 127).
We provide the Classification(SqueezeNet_v1.1) and Detection(MobileNet_v1 SSD 300) demos based on ncnn(It is a high-performance neural network inference framework optimized for the mobile platform),and the community ready to support this implment.
For details, please read the following PDF:
$ python caffe-int8-convert-tool.py --help
usage: caffe-int8-convert-tool.py [-h] [--proto PROTO] [--model MODEL]
[--mean MEAN MEAN MEAN] [--norm NORM]
[--images IMAGES] [--output OUTPUT]
[--gpu GPU]
find the pretrained caffe models int8 quantize scale value
optional arguments:
-h, --help show this help message and exit
--proto PROTO path to deploy prototxt.
--model MODEL path to pretrained weights
--mean MEAN value of mean
--norm NORM value of normalize
--images IMAGES path to calibration images
--output OUTPUT path to output calibration table file
--gpu GPU use gpu to forward
$ python caffe-int8-convert-tool.py --proto=squeezenet_v1.1.prototxt --model=squeezenet.caffemodel --mean 104 117 123 --images=ILSVRC2012_1k --output=squeezenet_v1.1.table --gpu=1
Pay attention to the type of images,it is just the original image format,such as jpg or jpeg,do not us the type of caffe dataset(lmdb). It is recommended to provide representative calibration dataset for the given model use case,such as the validation or test dataset.
For example in squeezenet_v1_1.table
conv1_param_0 138.066410
fire2/squeeze1x1_param_0 92.028103 // the conv layer's weight scale is 92.028
......
data 0.841264
conv1 0.295743
pool1 0.161700 // the pool layer's top blob scale is 0.1617
fire2/squeeze1x1 0.089383 // the conv layer's top blob scale is 0.0893
......
Three steps to implement the fire2/squeeze1x1 layer int8 convolution:
-
Quantize the bottom_blob and weight:
bottom_blob_int8 = bottom_blob_float32 * data_scale(0.1617) weight_int8 = weight_float32 * weight_scale(92.028) -
Convolution_Int8:
top_blob_int32 = bottom_blob_int8 * weight_int8 -
Dequantize the TopBlob_Int32 and add the bias:
top_blob_float32 = top_blob_int32 / [data_scale(0.1617) * weight_scale(92.028)] + bias_float32
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,sparse calculation and so on.
This format is already supported in the ncnn latest version.I will do my best to transform some common network models into classification-dev
python caffe-int8-convert-tool-dev.py -h
usage: caffe-int8-convert-tool.py [-h] [--proto PROTO] [--model MODEL]
[--mean MEAN MEAN MEAN] [--norm NORM]
[--images IMAGES] [--output OUTPUT]
[--group GROUP] [--gpu GPU]
find the pretrained caffe models int8 quantize scale value
optional arguments:
-h, --help show this help message and exit
--proto PROTO path to deploy prototxt.
--model MODEL path to pretrained weights
--mean MEAN value of mean
--norm NORM value of normalize
--images IMAGES path to calibration images
--output OUTPUT path to output calibration table file
--group GROUP enable the group scale
--gpu GPU use gpu to forward
python caffe-int8-convert-tool-dev.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/ --group=1
Although it's done,but the speed of group quanization is very slow......The difference from the release tool is that we try to get the int8_scale of bottom blob not the top blob.
For example in MobileNet_v1_dev.table
conv1_param_0 156.639840
conv2_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 49.466518
conv2_1/dw 0 123.720796 48.705349 ...... // the convdw layer's bottom blob every group channel scale is 0.0 123.720 48.705 ......
......
We used ImageNet2012 Dataset to complete some experiments.
| Type | Detail |
|---|---|
| Calibration Dataset | ILSVRC2012_img_test 1k |
| Test Dataset | ILSVRC2012_img_val 5k |
| Framework | ncnn-int8 |
| Support Layer | Convolution3x3,Convolution1x1,ConvolutionDepthwise3x3 |
The following table show the Top1 and Top5 different between Float32 and Int8 inference.
| FP32 | INT8 | |||
|---|---|---|---|---|
| NETWORK | Top1 | Top5 | Top1 | Top5 |
| SqueezeNet v1.1 | 57.86% | 79.86% | 57.36% | 79.84% |
| MobileNet v1 | 67.78% | 87.62% | 64.92% | 85.22% |
| MobileNet v2 | 70.20% | 89.20% | 69.00% | 88.04% |
| GoogleNet v1 | 67.70% | 88.32% | 67.64% | 88.26% |
| ResNet-18 | 65.50% | 86.46% | 65.48% | 86.44% |
| ResNet-50 | 71.68% | 89.94% | 71.38% | 89.52% |
| NETWORK | Top1 | Top5 | Diff Top1 | Diff Top5 |
| SqueezeNet v1.1 | 57.86% | 79.86% | 0.50% | 0.02% |
| MobileNet v1 | 67.78% | 87.62% | 2.86% | 2.40% |
| MobileNet v2 | 70.20% | 89.20% | 1.06% | 1.16% |
| GoogleNet v1 | 67.70% | 88.32% | 0.06% | 0.06% |
| ResNet-18 | 65.50% | 86.46% | 0.02% | 0.02% |
| ResNet-50 | 71.68% | 89.94% | 0.30% | 0.32% |
The following table show the speedup between Float32 and Int8 inference.It should be noted that the winograd algorithm is not used in the Float32 inference.The Hardware Platform is Hisi3519([email protected])
| Uint(ms) | SqueezeNet v1.1 | MobileNet v1 | MobileNet v2 | GoogleNet | ResNet18 | MobileNetv1 SSD |
|---|---|---|---|---|---|---|
| Float32 | 382 | 568 | 392 | 1662 | 1869 | 1120 |
| Int8 | 242 | 369 | 311 | 1159 | 1159 | 701 |
| Ratio | x1.30 | x1.41 | x1.28 | x1.43 | x1.61 | x1.47 |
Thanks to our company SenseNets to support the open source project,and NVIDIA for providing the principle of correlation entropy,and ncnn's author nihui sharing his neural network inference framework.
Thanks to the help from the following friends:
Algorithm : xupengfeixupf, JansonZhu, wangxinwei, lengmm
Python : daquexian
BSD 3 Clause