With the help of MobileNet-Caffe and pytorch-caffe-darknet-convert, the MobileNet v2 pretrained model is transferred to darknet
Only the src/image.c and examples/classifier.c are modified, search for "mobilenet" to see what are changed
If you want to evaluate with single image, use
./darknet classifier predict mobilenet/imagenet1k.data mobilenet/test.cfg mobilenet/test.weights mobilenet/cat.jpg
Expected outputs (GPU=1, OPENCV=0):
19.65%: Egyptian cat
16.53%: tiger cat
10.64%: tabby
7.38%: red fox
6.09%: kit fox
If you want to validate with ILSVRC2012, please refer to pjreddie.com to make a valid list
The result is:
| Network | Top-1 | Top-5 | Ops |
|---|---|---|---|
| MobileNet v2 | 71.44 | 90.33 | 0.86Bn |
- With the parameter of
predict, the outputs will be different according to options combination in Makefile - The network structure is not exactly the same to the paper, therefore the Ops is larger