- 作者:
- 分类:知识&开发->AI->edge_AI
- 阅读:6582
- 点赞:51
- 版权:CC BY-SA 4.0
- 创建:2021-04-10
- 更新:2021-06-15
原文链接(持续更新):https://neucrack.com/p/358
直接使用 pytorch hub 的与训练模型
这里省略了模型定义和训练过程, 直接使用 pytorch hub 的 resnet18 预训练模型进行简单介绍:
https://pytorch.org/hub/pytorch_vision_resnet/
在 PC 端测试模型推理
根据上面链接的使用说明, 使用如下代码可以运行模型
其中, label 下载: https://raw.githubusercontent.com/pytorch/hub/master/imagenet_classes.txt
import osimport torchfrom torchsummary import summary## modelmodel = 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 imagefilename = "out/dog.jpg"if not os.path.exists(filename):if not os.path.exists("out"):os.makedirs("out")import urlliburl, 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 datafrom PIL import Imageimport numpy as npfrom torchvision import transformsinput_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 availableif 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]))
运行后 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] 9,408BatchNorm2d-2 [-1, 64, 112, 112] 128ReLU-3 [-1, 64, 112, 112] 0MaxPool2d-4 [-1, 64, 56, 56] 0Conv2d-5 [-1, 64, 56, 56] 36,864BatchNorm2d-6 [-1, 64, 56, 56] 128ReLU-7 [-1, 64, 56, 56] 0Conv2d-8 [-1, 64, 56, 56] 36,864BatchNorm2d-9 [-1, 64, 56, 56] 128ReLU-10 [-1, 64, 56, 56] 0BasicBlock-11 [-1, 64, 56, 56] 0Conv2d-12 [-1, 64, 56, 56] 36,864BatchNorm2d-13 [-1, 64, 56, 56] 128ReLU-14 [-1, 64, 56, 56] 0Conv2d-15 [-1, 64, 56, 56] 36,864BatchNorm2d-16 [-1, 64, 56, 56] 128ReLU-17 [-1, 64, 56, 56] 0BasicBlock-18 [-1, 64, 56, 56] 0Conv2d-19 [-1, 128, 28, 28] 73,728BatchNorm2d-20 [-1, 128, 28, 28] 256ReLU-21 [-1, 128, 28, 28] 0Conv2d-22 [-1, 128, 28, 28] 147,456BatchNorm2d-23 [-1, 128, 28, 28] 256Conv2d-24 [-1, 128, 28, 28] 8,192BatchNorm2d-25 [-1, 128, 28, 28] 256ReLU-26 [-1, 128, 28, 28] 0BasicBlock-27 [-1, 128, 28, 28] 0Conv2d-28 [-1, 128, 28, 28] 147,456BatchNorm2d-29 [-1, 128, 28, 28] 256ReLU-30 [-1, 128, 28, 28] 0Conv2d-31 [-1, 128, 28, 28] 147,456BatchNorm2d-32 [-1, 128, 28, 28] 256ReLU-33 [-1, 128, 28, 28] 0BasicBlock-34 [-1, 128, 28, 28] 0Conv2d-35 [-1, 256, 14, 14] 294,912BatchNorm2d-36 [-1, 256, 14, 14] 512ReLU-37 [-1, 256, 14, 14] 0Conv2d-38 [-1, 256, 14, 14] 589,824BatchNorm2d-39 [-1, 256, 14, 14] 512Conv2d-40 [-1, 256, 14, 14] 32,768BatchNorm2d-41 [-1, 256, 14, 14] 512ReLU-42 [-1, 256, 14, 14] 0BasicBlock-43 [-1, 256, 14, 14] 0Conv2d-44 [-1, 256, 14, 14] 589,824BatchNorm2d-45 [-1, 256, 14, 14] 512ReLU-46 [-1, 256, 14, 14] 0Conv2d-47 [-1, 256, 14, 14] 589,824BatchNorm2d-48 [-1, 256, 14, 14] 512ReLU-49 [-1, 256, 14, 14] 0BasicBlock-50 [-1, 256, 14, 14] 0Conv2d-51 [-1, 512, 7, 7] 1,179,648BatchNorm2d-52 [-1, 512, 7, 7] 1,024ReLU-53 [-1, 512, 7, 7] 0Conv2d-54 [-1, 512, 7, 7] 2,359,296BatchNorm2d-55 [-1, 512, 7, 7] 1,024Conv2d-56 [-1, 512, 7, 7] 131,072BatchNorm2d-57 [-1, 512, 7, 7] 1,024ReLU-58 [-1, 512, 7, 7] 0BasicBlock-59 [-1, 512, 7, 7] 0Conv2d-60 [-1, 512, 7, 7] 2,359,296BatchNorm2d-61 [-1, 512, 7, 7] 1,024ReLU-62 [-1, 512, 7, 7] 0Conv2d-63 [-1, 512, 7, 7] 2,359,296BatchNorm2d-64 [-1, 512, 7, 7] 1,024ReLU-65 [-1, 512, 7, 7] 0BasicBlock-66 [-1, 512, 7, 7] 0AdaptiveAvgPool2d-67 [-1, 512, 1, 1] 0Linear-68 [-1, 1000] 513,000================================================================Total params: 11,689,512Trainable params: 11,689,512Non-trainable params: 0----------------------------------------------------------------Input size (MB): 0.57Forward/backward pass size (MB): 62.79Params size (MB): 44.59Estimated Total Size (MB): 107.96----------------------------------------------------------------out/dog.jpgtensor(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
将模型转换为 V831 能使用的模型文件
转换过程如下:
使用 Pytorch 将模型导出为 onnx模型, 得到onnx文件
def torch_to_onnx(net, input_shape, out_name="out/model.onnx", input_names=["input0"], output_names=["output0"], device="cpu"):batch_size = 1if 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 = filenametorch_to_onnx(model, input_shape, onnx_out, device="cuda:0")
如果你不是使用 pytorch 转换的, 而是使用了现成的 ncnn 模型, 不知道输出层的名字, 可以在 https://netron.app/ 打开模型查看输出层的名字
使用 onnx2ncnn 工具将onnx转成ncnn模型,得到一个.param文件和一个.bin文件
按照ncnn项目的编译说明编译,在
build/tools/onnx目录下得到onnx2ncnn可执行文件
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 dircmd = 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)
使用全志提供的awnn工具将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
使用模型,在v831上推理
可以使用 python 或者 C 写代码,以下两种方式
MaixPy3
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/blob/main/ext_modules/_maix_nn/example/load_forward_camera.py
label 在这里: https://github.com/sipeed/MaixPy3/blob/main/ext_modules/_maix_nn/example/classes_label.py
from maix import nnfrom PIL import Image, ImageDrawfrom maix import camera, displaytest_jpg = "/root/test_input/input.jpg"model = {"param": "/root/models/resnet_awnn.param","bin": "/root/models/resnet_awnn.bin"}camera.config(size=(224, 224))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:", model)m = nn.load(model, opt=options)print("-- load ok")print("-- read image")img = Image.open(test_jpg)print("-- read image ok")print("-- forward model with image as input")out = m.forward(img, quantize=True)print("-- read image ok")print("-- out:", out.shape)out = nn.F.softmax(out)print(out.max(), out.argmax())from classes_label import labelswhile 1:img = camera.capture()if not img:time.sleep(0.02)continueout = m.forward(img, quantize=True)out = nn.F.softmax(out)msg = "{:.2f}: {}".format(out.max(), labels[out.argmax()])print(msg)draw = ImageDraw.Draw(img)draw.text((0, 0), msg, fill=(255, 0, 0))display.show(img)
C语言 SDK, libmaix
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
