在V831上使用resnet18分类网络,我们需要在linux环境下进行。windows系统可以使用虚拟机,或者是使用WSL,具体的安装教程请自行百度,这里就不过多的进行描述
我们需要在系统中安装pytorch,通过在pytorch官网上可以知道安装pytorch需要执行
pip3 install torch==1.9.0+cpu torchvision==0.10.0+cpu torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html
或者是通过conda环境进行安装
conda install pytorch torchvision torchaudio cpuonly -c pytorch
我们还需要安装一个torchsummary
库来进行神经网络的可视化
pip3 install torchsummary
通过 git clone https://github.com/Tencent/ncnn.git
将ncnn的仓库拉取到本地,进行编译
安装编译环境的依赖
sudo apt update
sudo apt install build-essential git cmake libprotobuf-dev protobuf-compiler libvulkan-dev vulkan-utils libopencv-dev
编译ncnn需要使用到 Vulkan 后端
要使用 Vulkan 后端,请安装 Vulkan 头文件、一个 vulkan 驱动程序加载器、GLSL 到 SPIR-V 编译器和 vulkaninfo 工具。或者从https://vulkan.lunarg.com/sdk/home下载并安装完整的 Vulkan SDK(大约 200MB;它包含所有头文件、文档和预构建的加载程序,以及一些额外的工具和所有源代码)
wget https://sdk.lunarg.com/sdk/download/1.2.182.0/linux/vulkansdk-linux-x86_64-1.2.182.0.tar.gz
tar xvf vulkansdk-linux-x86_64-1.2.182.0.tar.gz
export VULKAN_SDK=$(pwd)/1.2.182.0/x86_64
拉取ncnn的子仓库
cd ncnn
git submodule update --init
开始编译ncnn
mkdir -p build
cd build
cmake -DCMAKE_BUILD_TYPE=Release -DNCNN_VULKAN=ON -DNCNN_SYSTEM_GLSLANG=ON -DNCNN_BUILD_EXAMPLES=ON ..
make -j$(nproc)
编译结束之后会在build/tools/onnx/下的到onnx2ncnn可执行文件,这个是就用ncnn的转换工具
以下代码建议在jupyter中运行
通过pytorch hub来获取resnet18的预训练模型,这里并不细说训练的过程和模型定义
label下载
使用以下代码进行模型的下载和推理
import os
import torch
from torchsummary import summary
torch.hub._validate_not_a_forked_repo=lambda a,b,c: True
## model
model = torch.hub.load('pytorch/vision:v0.6.0', 'resnet18', pretrained=True)
model.eval()
input_shape = (3, 224, 224)
summary(model, input_shape, device="cpu")
## test image
filename = "out/dog.jpg"
if not os.path.exists(filename):
if not os.path.exists("out"):
os.makedirs("out")
import urllib
url, filename = ("https://github.com/pytorch/hub/raw/master/images/dog.jpg", filename)
try: urllib.URLopener().retrieve(url, filename)
except: urllib.request.urlretrieve(url, filename)
print("test image:", filename)
## preparing input data
from PIL import Image
import numpy as np
from torchvision import transforms
input_image = Image.open(filename)
# input_image.show()
preprocess = transforms.Compose([
transforms.Resize(max(input_shape[1:3])),
transforms.CenterCrop(input_shape[1:3]),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
input_tensor = preprocess(input_image)
print("input data max value: {}, min value: {}".format(torch.max(input_tensor), torch.min(input_tensor)))
input_batch = input_tensor.unsqueeze(0) # create a mini-batch as expected by the model
## forward model
# move the input and model to GPU for speed if available
if torch.cuda.is_available():
input_batch = input_batch.to('cuda')
model.to('cuda')
with torch.no_grad():
output = model(input_batch)
## result
# Tensor of shape 1000, with confidence scores over Imagenet's 1000 classes
# print(output[0])
# The output has unnormalized scores. To get probabilities, you can run a softmax on it.
max_1000 = torch.nn.functional.softmax(output[0], dim=0)
max_idx = int(torch.argmax(max_1000))
with open("imagenet_classes.txt") as f:
labels = f.read().split("\n")
print("result: idx:{}, name:{}".format(max_idx, labels[max_idx]))
wsl下是无法使用GPU需要将cuda:0修改成cpu
运行后得到结果:
Using cache found in /home/neucrack/.cache/torch/hub/pytorch_vision_v0.6.0
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 64, 112, 112] 9,408
BatchNorm2d-2 [-1, 64, 112, 112] 128
ReLU-3 [-1, 64, 112, 112] 0
MaxPool2d-4 [-1, 64, 56, 56] 0
Conv2d-5 [-1, 64, 56, 56] 36,864
BatchNorm2d-6 [-1, 64, 56, 56] 128
ReLU-7 [-1, 64, 56, 56] 0
Conv2d-8 [-1, 64, 56, 56] 36,864
BatchNorm2d-9 [-1, 64, 56, 56] 128
ReLU-10 [-1, 64, 56, 56] 0
BasicBlock-11 [-1, 64, 56, 56] 0
Conv2d-12 [-1, 64, 56, 56] 36,864
BatchNorm2d-13 [-1, 64, 56, 56] 128
ReLU-14 [-1, 64, 56, 56] 0
Conv2d-15 [-1, 64, 56, 56] 36,864
BatchNorm2d-16 [-1, 64, 56, 56] 128
ReLU-17 [-1, 64, 56, 56] 0
BasicBlock-18 [-1, 64, 56, 56] 0
Conv2d-19 [-1, 128, 28, 28] 73,728
BatchNorm2d-20 [-1, 128, 28, 28] 256
ReLU-21 [-1, 128, 28, 28] 0
Conv2d-22 [-1, 128, 28, 28] 147,456
BatchNorm2d-23 [-1, 128, 28, 28] 256
Conv2d-24 [-1, 128, 28, 28] 8,192
BatchNorm2d-25 [-1, 128, 28, 28] 256
ReLU-26 [-1, 128, 28, 28] 0
BasicBlock-27 [-1, 128, 28, 28] 0
Conv2d-28 [-1, 128, 28, 28] 147,456
BatchNorm2d-29 [-1, 128, 28, 28] 256
ReLU-30 [-1, 128, 28, 28] 0
Conv2d-31 [-1, 128, 28, 28] 147,456
BatchNorm2d-32 [-1, 128, 28, 28] 256
ReLU-33 [-1, 128, 28, 28] 0
BasicBlock-34 [-1, 128, 28, 28] 0
Conv2d-35 [-1, 256, 14, 14] 294,912
BatchNorm2d-36 [-1, 256, 14, 14] 512
ReLU-37 [-1, 256, 14, 14] 0
Conv2d-38 [-1, 256, 14, 14] 589,824
BatchNorm2d-39 [-1, 256, 14, 14] 512
Conv2d-40 [-1, 256, 14, 14] 32,768
BatchNorm2d-41 [-1, 256, 14, 14] 512
ReLU-42 [-1, 256, 14, 14] 0
BasicBlock-43 [-1, 256, 14, 14] 0
Conv2d-44 [-1, 256, 14, 14] 589,824
BatchNorm2d-45 [-1, 256, 14, 14] 512
ReLU-46 [-1, 256, 14, 14] 0
Conv2d-47 [-1, 256, 14, 14] 589,824
BatchNorm2d-48 [-1, 256, 14, 14] 512
ReLU-49 [-1, 256, 14, 14] 0
BasicBlock-50 [-1, 256, 14, 14] 0
Conv2d-51 [-1, 512, 7, 7] 1,179,648
BatchNorm2d-52 [-1, 512, 7, 7] 1,024
ReLU-53 [-1, 512, 7, 7] 0
Conv2d-54 [-1, 512, 7, 7] 2,359,296
BatchNorm2d-55 [-1, 512, 7, 7] 1,024
Conv2d-56 [-1, 512, 7, 7] 131,072
BatchNorm2d-57 [-1, 512, 7, 7] 1,024
ReLU-58 [-1, 512, 7, 7] 0
BasicBlock-59 [-1, 512, 7, 7] 0
Conv2d-60 [-1, 512, 7, 7] 2,359,296
BatchNorm2d-61 [-1, 512, 7, 7] 1,024
ReLU-62 [-1, 512, 7, 7] 0
Conv2d-63 [-1, 512, 7, 7] 2,359,296
BatchNorm2d-64 [-1, 512, 7, 7] 1,024
ReLU-65 [-1, 512, 7, 7] 0
BasicBlock-66 [-1, 512, 7, 7] 0
AdaptiveAvgPool2d-67 [-1, 512, 1, 1] 0
Linear-68 [-1, 1000] 513,000
================================================================
Total params: 11,689,512
Trainable params: 11,689,512
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 62.79
Params size (MB): 44.59
Estimated Total Size (MB): 107.96
----------------------------------------------------------------
out/dog.jpg
tensor(2.6400) tensor(-2.1008)
idx:258, name:Samoyed, Samoyede
可以看到模型有 11,689,512的参数, 即 11MiB左右, 这个大小也就几乎是实际在 831 上运行的模型的大小了
通过pytorch hub获取到的resnet18 模型是pth格式的,需要转换成onnx格式的模型
转换代码
def torch_to_onnx(net, input_shape, out_name="out/model.onnx", input_names=["input0"], output_names=["output0"], device="cpu"):
batch_size = 1
if len(input_shape) == 3:
x = torch.randn(batch_size, input_shape[0], input_shape[1], input_shape[2], dtype=torch.float32, requires_grad=True).to(device)
elif len(input_shape) == 1:
x = torch.randn(batch_size, input_shape[0], dtype=torch.float32, requires_grad=False).to(device)
else:
raise Exception("not support input shape")
print("input shape:", x.shape)
# torch.onnx._export(net, x, "out/conv0.onnx", export_params=True)
torch.onnx.export(net, x, out_name, export_params=True, input_names = input_names, output_names=output_names)
onnx_out="out/resnet_1000.onnx"
ncnn_out_param = "out/resnet_1000.param"
ncnn_out_bin = "out/resnet_1000.bin"
input_img = filename
torch_to_onnx(model, input_shape, onnx_out, device="cuda:0")
> wsl下是无法使用GPU需要将cuda:0修改成cpu
在out文件夹中得到onnx格式模型文件
然后再利用前面编译出来的onnx2ncnn转换工具进行ncnn格式的转换
def onnx_to_ncnn(input_shape, onnx="out/model.onnx", ncnn_param="out/conv0.param", ncnn_bin = "out/conv0.bin"):
import os
# onnx2ncnn tool compiled from ncnn/tools/onnx, and in the buld dir
cmd = f"onnx2ncnn {
onnx} {
ncnn_param} {
ncnn_bin}"
os.system(cmd)
with open(ncnn_param) as f:
content = f.read().split("\n")
if len(input_shape) == 1:
content[2] += " 0={}".format(input_shape[0])
else:
content[2] += " 0={} 1={} 2={}".format(input_shape[2], input_shape[1], input_shape[0])
content = "\n".join(content)
with open(ncnn_param, "w") as f:
f.write(content)
onnx_to_ncnn(input_shape, onnx=onnx_out, ncnn_param=ncnn_out_param, ncnn_bin=ncnn_out_bin)
通过maixhub将ncnn模型进行量化到int8模型
在 maixhub 模型转换 将 ncnn 模型转换为 awnn 支持的 int8 模型 (网页在线转换很方便人为操作,另一个方面因为全志要求不开放 awnn 所以暂时只能这样做)
阅读转换说明,可以获得更多详细的转换说明
这里有几组参数:
均值 和 归一化因子: 在 pytorch 中一般是 (输入值 - mean ) / std, awnn对输入的处理是 (输入值 - mean ) * norm, 总之,让你训练的时候的输入到第一层网络的值范围和给awnn量化工具经过(输入值 - mean ) * norm 计算后的值范围一致既可。 比如 这里打印了实际数据的输入范围是[-2.1008, 2.6400], 是代码中preprocess 对象处理后得到的,即x = (x - mean) / std ==> (0-0.485)/0.229 = -2.1179, 到awnn就是x = (x - mean_2*255) * (1 / std * 255) 即 mean2 = mean * 255, norm = 1/(std * 255), 更多可以看这里。
所以我们这里可以设置 均值为 0.485 * 255 = 123.675, 设置 归一化因子为1/ (0.229 * 255) = 0.017125, 另外两个通道同理,但是目前 awnn 只能支持三个通道值一样。。。所以填123.675, 123.675, 123.675,0.017125, 0.017125, 0.017125 即可,因为这里用了pytorch hub的预训练的参数,就这样吧, 如果自己训练,可以好好设置一下
图片输入层尺寸(问不是图片怎么办?貌似 awnn 暂时之考虑到了图片。。)
RGB 格式: 如果训练输入的图片是 RGB 就选 RGB
量化图片, 选择一些和输入尺寸相同的图片,可以从测试集中拿一些,不一定要图片非常多,但尽量覆盖全场景(摊手
自己写的其它模型转换如果失败,多半是啥算子不支持,需要在 使用说明里面看支持的 算子,比如现在的版本view、 flatten、reshape 都不支持所以写模型要相当小心, 后面的版本会支持 flatten reshape 等 CPU 算子
如果不出意外, 终于得到了量化好的 awnn 能使用的模型, *.param 和 *.bin
可以使用 python 或者 C 写代码,以下两种方式
python的是需要在终端下运行的,不要使用jupyter,建议使用ssh,这样放文件什么都比较方便
python 请看MaixPy3
不想看文档的话,就是在系统开机使用的基础上, 更新 MaixPy3 就可以了:
pip install --upgrade maixpy3
然后在终端使用 python 运行脚本(可能需要根据你的文件名参数什么的改一下代码):
https://github.com/sipeed/MaixPy3_scripts/blob/main/basic/v1.0/resnet.py
label 在这里: https://github.com/sipeed/MaixPy3/blob/main/ext_modules/_maix_nn/example/classes_label.py
baars.ttf 在这里:https://github.com/sipeed/MaixPy3_scripts/blob/main/application/base/res/baars.ttf
from maix import nn
from PIL import Image, ImageFont, ImageDraw
from maix import display
from classes_label import labels
import time
from maix import camera
class funation:
model = None
options = None
m = None
font = None
def __init__(self):
camera.config(size=(224, 224))
self.model = {
"param": "/root/res/resnet.param",
"bin": "/root/res/resnet.bin"
}
self.options = {
"model_type": "awnn",
"inputs": {
"input0": (224, 224, 3)
},
"outputs": {
"output0": (1, 1, 1000)
},
"first_layer_conv_no_pad": False,
"mean": [127.5, 127.5, 127.5],
"norm": [0.00784313725490196, 0.00784313725490196, 0.00784313725490196],
}
print("-- load model:", self.model)
self.m = nn.load(self.model, opt=self.options)
print("-- load ok")
self.font = ImageFont.truetype("/root/res/baars.ttf",20, encoding="unic")
def run(self):
img = camera.capture()
t = time.time()
out = self.m.forward(img, quantize=True)
t = time.time() - t
print("-- forward time: {}s".format(t))
t = time.time()
out2 = nn.F.softmax(out)
t = time.time() - t
print("-- softmax time: {}s".format(t))
msg = "{:.2f}: {}".format(out.max(), labels[out.argmax()])
print(msg)
img = Image.new("RGBA", (240, 240), "#00000000")
draw = ImageDraw.Draw(img)
draw.text((0, 0), msg, (255, 0, 0), self.font)
display.show(img)
if __name__ == "__main__":
start = funation()
while True:
start.run()
如果运行报错了,请更新maixpy3再运行
访问这里,按照 https://github.com/sipeed/libmaix 的说明克隆仓库,并编译 https://github.com/sipeed/libmaix/tree/master/examples/nn_resnet
上传编译成功后dist目录下的所有内容到 v831, 然后执行./start_app.sh即可
以上内容出至:https://neucrack.com/p/358