nihui / ncnn-assets Goto Github PK

can you write the steps on how to convert yolov5s to the ncnn format?
I converted the torch model to onnx then using ncnn tools converted to bin/param but it did not work

hello,，nihui大佬，我用github.com/eric612/MobileNet-YOLO/blob/master/models/mobilenetv2_voc/yolo_lite/yolov3_m2.prototxt的caffe模型，通过caffe2ncnn做转换，转换生成了param和bin，load_model的时候报错: layer load_model 3 failed。

请问您在转化这个caffemodel的时候，做了一些什么预处理吗，我需要做什么改动才能转换成功呢？请指教，多谢～

(卷卷卷qwq)

用官方模型转换为onnx，然后再转换成param和bin，模型也能转换成功，但是报错。想知道你们是如何转换的，报错如下：
Shape not supported yet!
Shape not supported yet!
Shape not supported yet!
Unsupported unsqueeze axes !
Unsupported unsqueeze axes !
Unsupported unsqueeze axes !
Unknown data type 0
Shape not supported yet!
Shape not supported yet!
Shape not supported yet!
Unsupported unsqueeze axes !
Unsupported unsqueeze axes !
Unsupported unsqueeze axes !
Unknown data type 0
Shape not supported yet!
Shape not supported yet!
Shape not supported yet!
Unsupported unsqueeze axes !
Unsupported unsqueeze axes !
Unsupported unsqueeze axes !
Unknown data type 0

在转模型的时候感觉不太能确定输入输出，在调试的时候，参考了nihui转的mobilenetssd的内容，查看了下输出。突然就想到有没有中间生成的onnx文件，也能检验一下自己是不是onnx本身转对了。

I trained yolov7-tiny on a custom dataset, converted it to onnx using this example and then converted that onnx to ncnn using this.

pt to onnx conversion

python export.py --weights MY_MODEL.pt --simplify --topk-all 100 --iou-thres 0.65 --conf-thres 0.35 --img-size 416 416 --max-wh 416

onnx to ncnn conversion

onnx2ncnn MY_MODEL.onnx MY_MODEL.param MY_MODEL.bin

When I open up the MY_MODEL.param with netron, I see that the value of w parameter for the reshape layer of all the outputs is some positive integer (e.g. 676). And the detection does not work (does not show any of the class).

When I compared MY_MODEL.param with your yolov7-tiny.param , I see that the value of w parameter for the reshape layer of all the outputs is -1. If I manually change this value to -1 in my own MY_MODEL.param, my model works with proper detection.

Can you please explain how you created your yolov7-tiny ncnn? What are the exact steps you did?

loongson@ls2k:~/Desktop/ncnn/build/examples$ ./yolov5 01.jpg
0 = 0.88773 at 613.13 309.97 129.40 x 293.63
0 = 0.85717 at 717.60 414.13 116.84 x 182.27
0 = 0.83715 at 310.91 331.21 109.08 x 272.36
0 = 0.80530 at 818.37 296.96 71.07 x 209.47
0 = 0.79004 at 874.96 290.36 127.99 x 308.92
2 = 0.77431 at 993.14 334.77 85.86 x 87.92
0 = 0.77314 at 461.77 333.91 74.34 x 262.11
2 = 0.77174 at 91.87 364.67 105.94 x 101.68
0 = 0.70869 at 0.40 366.16 75.08 x 122.26
2 = 0.67867 at 146.39 357.82 287.35 x 151.01
0 = 0.65766 at 739.59 315.20 67.53 x 137.53
2 = 0.63094 at 353.93 347.06 256.41 x 129.11
2 = 0.62434 at 48.99 362.04 66.74 x 61.99
3 = 0.38121 at 5.74 421.51 67.24 x 91.54
0 = 0.34597 at 795.92 336.61 31.88 x 90.78
imshow save image to image.png
waitKey stub

loongson@ls2k:~/Desktop/ncnn/build/examples$ ./yolox 01.jpg
浮点数例外

同样的tiny版本，为什么我生成的bin 比官方提供的文件大一倍

@nihui 你好， 我在用ncnn编译生成的caffe2ncnn工具转caffe模型至ncnn模型时， 生成的.param文件没有这两行：
Input data 0 1 data 0=300 1=300 2=3
Split splitncnn_0 1 7 data data_splitncnn_0 data_splitncnn_1 data_splitncnn_2 data_splitncnn_3 data_splitncnn_4 data_splitncnn_5 data_splitncnn_6
这样也导致.param文件的第2行s数据”148 181“的splitncnn_blob_count数为148， 因为没有Input，所以少1层。
我手动修改了148-->149，文件能正常加载。
看了下caffe2ncnn.cpp源代码，还是不太清楚为啥？ 麻烦有空解答一下， 谢谢！

Hi @nihui ,
Thanks for your assets. Could you please guide me on how to convert yolov5s_6.0 to the ncnn format? I have seen the converted output in your repo.
Thank you so much!