Name | Date | Size | #Lines | LOC | ||
---|---|---|---|---|---|---|
.. | 03-May-2022 | - | ||||
README.md | H A D | 21-May-2021 | 2 KiB | 66 | 44 | |
darknet2ncnn.cpp | H A D | 21-May-2021 | 31.1 KiB | 926 | 801 |
README.md
1# Darknet To NCNN Conversion Tools 2 3This is a standalone darknet2ncnn converter without additional dependency. 4 5Support yolov4, yolov4-tiny, yolov3, yolov3-tiny and enet-coco.cfg (EfficientNetB0-Yolov3). 6 7Another conversion tool based on darknet can be found at: [darknet2ncnn](https://github.com/xiangweizeng/darknet2ncnn) 8 9## Usage 10 11``` 12Usage: darknet2ncnn [darknetcfg] [darknetweights] [ncnnparam] [ncnnbin] [merge_output] 13 [darknetcfg] .cfg file of input darknet model. 14 [darknetweights] .weights file of input darknet model. 15 [cnnparam] .param file of output ncnn model. 16 [ncnnbin] .bin file of output ncnn model. 17 [merge_output] merge all output yolo layers into one, enabled by default. 18``` 19 20## Example 21 22### 1. Convert yolov4-tiny cfg and weights 23 24Download pre-trained [yolov4-tiny.cfg](https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov4-tiny.cfg) and [yolov4-tiny.weights](https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.weights) or with your own trained weight. 25 26Convert cfg and weights: 27``` 28./darknet2ncnn yolov4-tiny.cfg yolov4-tiny.weights yolov4-tiny.param yolov4-tiny.bin 1 29``` 30 31If succeeded, the output would be: 32``` 33Loading cfg... 34WARNING: The ignore_thresh=0.700000 of yolo0 is too high. An alternative value 0.25 is written instead. 35WARNING: The ignore_thresh=0.700000 of yolo1 is too high. An alternative value 0.25 is written instead. 36Loading weights... 37Converting model... 3883 layers, 91 blobs generated. 39NOTE: The input of darknet uses: mean_vals=0 and norm_vals=1/255.f. 40NOTE: Remeber to use ncnnoptimize for better performance. 41``` 42 43### 2. Optimize graphic 44 45``` 46./ncnnoptimize yolov4-tiny.param yolov4-tiny.bin yolov4-tiny-opt.param yolov4-tiny-opt.bin 0 47``` 48 49### 3. Test 50 51build examples/yolov4.cpp and test with: 52 53``` 54./yolov4 dog.jpg 55``` 56 57The result will be: 58 59![](https://github.com/Tencent/ncnn/blob/master/tools/darknet/output.jpg) 60 61 62## How to run with benchncnn 63 64Set 2=0.3 for Yolov3DetectionOutput layer. 65 66