- train mobilenetv3-ssd use pytorch(provide *.pth model)
- convert to ncnn model(provide *.param, *.bin)
Reference paper: MobileNetv3 https://arxiv.org/pdf/1905.02244.pdf
We mainly train mobilenetv3-ssd detection network rather than classification network, for convenience, we use trained mobiletnetv3-large network from https://github.com/xiaolai-sqlai/mobilenetv3 (We are also trying to use https://github.com/rwightman/gen-efficientnet-pytorch provided mobilenetv3-large classification network)
open-source mobilenetv3-large classification network
| mobilenetv3-large | top-1 accuracy | params(million) | flops/Madds(million) |
|---|---|---|---|
| https://github.com/xiaolai-sqlai/mobilenetv3 | 75.5 | 3.96 | 272 |
| https://github.com/d-li14/mobilenetv3.pytorch | 73.2 | 5.15 | 246 |
| https://github.com/Randl/MobileNetV3-pytorch | 73.5 | 5.48 | 220 |
| https://github.com/rwightman/gen-efficientnet-pytorch | 75.6 | 5.5 | 219 |
| official mobilenetv3 | 75.2 | 5.4 | 219 |
| official mobilenetv2 | 72.0 | 3.4 | 300 |
| official efficient B0 | 76.3 | 5.3 | 390 |
For extra-body, we use 1x1 conv + 3x3 dw conv + 1x1 conv block follow mobilenetv2-ssd setting(official tensorflow version), details below:
1x1 256 conv -> 3x3 256 s=2 conv -> 1x1 512 conv
1x1 128 conv -> 3x3 128 s=2 conv -> 1x1 256 conv
1x1 128 conv -> 3x3 128 s=2 conv -> 1x1 256 conv
1x1 64 conv -> 3x3 64 s=2 conv -> 1x1 128 conv
For head, we use 3x3 dw conv + 1x1 conv block follow mobilenetv2-ssd-lite setting(official tensorflow version)
We choose 6 feature maps to predict box coordinates and label, their dimenstions are 19x19, 10x10, 5x5, 3x3, 2x2, 1x1. their anchor numbers are 4, 6, 6, 6, 4, 4.
We train mobilenetv3-ssd use mmdetection framework(based on pytorch), we use PASCAL VOC0712 trainval dataset to train, it reaches 71.7mAP on VOC2007 test dataset.
img test:
- convert .pth model to onnx(not included priorbox layer, detection_output layer) -> I provide origin pytorch model
- use onnx-simplifier to simplify onnx model
- convert simplified .onnx model to ncnn
- modify .param manually(add priorbox layer, detection_output layer, etc.) -> I provide converted ncnn model
you can refer to https://github.com/Tencent/ncnn/blob/master/examples/mobilenetv3ssdlite.cpp
mobilenetv3-ssd pytorch model 百度网盘链接: https://pan.baidu.com/s/1sTGrTHxpv4yZJUpTJD8BNw 提取码: sid9 mobilenetv3-ssd ncnn model 百度网盘链接: https://pan.baidu.com/s/1zBqGnp4utJGi6-IzYs7lTg 提取码: phdx google drive link: https://drive.google.com/file/d/11_C_ko-arXnzM60udcXOMM5_PDNXuCcs/view?usp=sharing