pytorch-caffe-darknet-convert
Convert between pytorch, caffe and darknet models. Caffe darknet models can be load directly by pytorch.
- darknet2pytorch : use darknet.py to load darknet model directly
- caffe2pytorch : use caffenet.py to load caffe model directly
- darknet2caffe
- caffe2darknet
- pytorch2caffe
- pytorch2darknet : pytorch2caffe then caffe2darknet
Convert pytorch to darknet
convert resnet50 from pytorch to darknet
1. python pytorch2darknet.py
2. python main.py -a resnet50-darknet --pretrained -e /home/xiaohang/ImageNet/
=> using pre-trained model 'resnet50'
load weights from resnet50.weights
Test: [0/196] Time 16.132 (16.132) Loss 6.1005 (6.1005) Prec@1 87.109 (87.109) Prec@5 97.266 (97.266)
Test: [10/196] Time 0.387 (1.803) Loss 6.2117 (6.1357) Prec@1 77.734 (82.670) Prec@5 91.797 (95.455)
Test: [20/196] Time 0.275 (1.261) Loss 6.1050 (6.1402) Prec@1 84.375 (82.236) Prec@5 92.969 (95.424)
Test: [30/196] Time 0.162 (1.123) Loss 6.1675 (6.1257) Prec@1 80.469 (83.543) Prec@5 95.312 (95.817)
Test: [40/196] Time 0.889 (1.024) Loss 6.1888 (6.1483) Prec@1 81.250 (82.012) Prec@5 96.875 (95.770)
Test: [50/196] Time 0.164 (0.970) Loss 6.0900 (6.1520) Prec@1 88.281 (81.794) Prec@5 97.656 (95.956)
Test: [60/196] Time 0.380 (0.933) Loss 6.1949 (6.1566) Prec@1 76.172 (81.416) Prec@5 93.359 (95.946)
Test: [70/196] Time 0.427 (0.916) Loss 6.2009 (6.1525) Prec@1 78.516 (81.679) Prec@5 96.484 (96.099)
Test: [80/196] Time 0.910 (0.900) Loss 6.3763 (6.1584) Prec@1 60.938 (81.134) Prec@5 88.672 (95.751)
Test: [90/196] Time 1.035 (0.896) Loss 6.4143 (6.1703) Prec@1 55.078 (80.039) Prec@5 86.328 (95.175)
Test: [100/196] Time 0.162 (0.871) Loss 6.3073 (6.1831) Prec@1 68.359 (78.968) Prec@5 91.406 (94.609)
Test: [110/196] Time 0.658 (0.867) Loss 6.1750 (6.1885) Prec@1 76.953 (78.442) Prec@5 94.922 (94.327)
Test: [120/196] Time 0.165 (0.861) Loss 6.2904 (6.1919) Prec@1 71.094 (78.141) Prec@5 88.281 (94.034)
Test: [130/196] Time 1.751 (0.858) Loss 6.1702 (6.2008) Prec@1 81.250 (77.290) Prec@5 94.141 (93.702)
Test: [140/196] Time 0.163 (0.849) Loss 6.2442 (6.2045) Prec@1 75.000 (76.948) Prec@5 90.625 (93.448)
Test: [150/196] Time 2.031 (0.846) Loss 6.1890 (6.2082) Prec@1 76.953 (76.604) Prec@5 91.016 (93.189)
Test: [160/196] Time 0.161 (0.837) Loss 6.1399 (6.2109) Prec@1 86.719 (76.366) Prec@5 94.531 (92.998)
Test: [170/196] Time 1.687 (0.835) Loss 6.1388 (6.2159) Prec@1 82.812 (75.959) Prec@5 97.656 (92.848)
Test: [180/196] Time 1.185 (0.828) Loss 6.3454 (6.2198) Prec@1 68.750 (75.650) Prec@5 91.797 (92.705)
Test: [190/196] Time 1.367 (0.826) Loss 6.3394 (6.2196) Prec@1 67.188 (75.685) Prec@5 95.703 (92.752)
* Prec@1 75.794 Prec@5 92.798
imagenet data is processed as described here
Convert darknet to caffe
convert tiny-yolo from darknet to caffe
1. download tiny-yolo-voc.weights : https://pjreddie.com/media/files/tiny-yolo-voc.weights
https://github.com/pjreddie/darknet/blob/master/cfg/tiny-yolo-voc.cfg
2. python darknet2caffe.py tiny-yolo-voc.cfg tiny-yolo-voc.weights tiny-yolo-voc.prototxt tiny-yolo-voc.caffemodel
3. download voc data and process according to https://github.com/marvis/pytorch-yolo2
python valid.py cfg/voc.data tiny-yolo-voc.prototxt tiny-yolo-voc.caffemodel
4. python scripts/voc_eval.py results/comp4_det_test_
VOC07 metric? Yes
AP for aeroplane = 0.6094
AP for bicycle = 0.6781
AP for bird = 0.4573
AP for boat = 0.3786
AP for bottle = 0.2081
AP for bus = 0.6645
AP for car = 0.6587
AP for cat = 0.6720
AP for chair = 0.3245
AP for cow = 0.4902
AP for diningtable = 0.5549
AP for dog = 0.5905
AP for horse = 0.6871
AP for motorbike = 0.6695
AP for person = 0.5833
AP for pottedplant = 0.2535
AP for sheep = 0.5374
AP for sofa = 0.4878
AP for train = 0.7004
AP for tvmonitor = 0.5754
Mean AP = 0.5391
5. python detect.py tiny-yolo-voc.prototxt tiny-yolo-voc.caffemodel data/dog.jpg
convert tiny-yolo from darknet to caffe without bn
1. python darknet.py tiny-yolo-voc.cfg tiny-yolo-voc.weights tiny-yolo-voc-nobn.cfg tiny-yolo-voc-nobn.weights
2. python darknet2caffe.py tiny-yolo-voc-nobn.cfg tiny-yolo-voc-nobn.weights tiny-yolo-voc-nobn.prototxt tiny-yolo-voc-nobn.caffemodel
3. python valid.py cfg/voc.data tiny-yolo-voc-nobn.prototxt tiny-yolo-voc-nobn.caffemodel
4. python scripts/voc_eval.py results/comp4_det_test_
VOC07 metric? Yes
AP for aeroplane = 0.6094
AP for bicycle = 0.6781
AP for bird = 0.4573
AP for boat = 0.3786
AP for bottle = 0.2081
AP for bus = 0.6645
AP for car = 0.6587
AP for cat = 0.6720
AP for chair = 0.3245
AP for cow = 0.4902
AP for diningtable = 0.5549
AP for dog = 0.5905
AP for horse = 0.6871
AP for motorbike = 0.6695
AP for person = 0.5833
AP for pottedplant = 0.2535
AP for sheep = 0.5374
AP for sofa = 0.4878
AP for train = 0.7004
AP for tvmonitor = 0.5754
Mean AP = 0.5391
5. python detect.py tiny-yolo-voc-nobn.prototxt tiny-yolo-voc-nobn.caffemodel data/dog.jpg
Convert caffe to darknet
convert kaiming's resnet50 from caffe to darknet
1. download resnet50 from https://github.com/KaimingHe/deep-residual-networks
ResNet-50-deploy.prototxt: https://github.com/KaimingHe/deep-residual-networks/blob/master/prototxt/ResNet-50-deploy.prototxt
ResNet-50-model.caffemodel and ResNet_mean.binaryproto : https://onedrive.live.com/?authkey=%21AAFW2-FVoxeVRck&id=4006CBB8476FF777%2117887&cid=4006CBB8476FF777
python main.py -a resnet50-kaiming --pretrained -e /home/xiaohang/ImageNet/
=> using pre-trained model 'resnet50-kaiming'
load weights from ResNet-50-model.caffemodel
Loading caffemodel: ResNet-50-model.caffemodel
convlution conv1 has bias
Test: [0/196] Time 15.196 (15.196) Loss 0.5485 (0.5485) Prec@1 87.500 (87.500) Prec@5 96.875 (96.875)
Test: [10/196] Time 0.603 (1.851) Loss 1.0412 (0.7676) Prec@1 73.828 (80.717) Prec@5 92.578 (94.567)
Test: [20/196] Time 0.359 (1.361) Loss 0.7243 (0.7656) Prec@1 85.547 (80.804) Prec@5 92.969 (94.494)
Test: [30/196] Time 0.883 (1.240) Loss 0.8022 (0.7273) Prec@1 82.031 (81.981) Prec@5 95.703 (94.796)
Test: [40/196] Time 1.166 (1.159) Loss 0.8016 (0.7690) Prec@1 79.297 (80.516) Prec@5 94.922 (94.779)
Test: [50/196] Time 1.490 (1.108) Loss 0.4365 (0.7552) Prec@1 89.062 (80.668) Prec@5 98.047 (95.044)
Test: [60/196] Time 0.295 (1.072) Loss 1.0453 (0.7689) Prec@1 72.656 (80.277) Prec@5 93.750 (95.210)
Test: [70/196] Time 0.728 (1.064) Loss 0.7959 (0.7542) Prec@1 77.344 (80.573) Prec@5 94.922 (95.384)
Test: [80/196] Time 1.314 (1.047) Loss 1.5740 (0.7775) Prec@1 62.109 (80.179) Prec@5 85.938 (95.100)
Test: [90/196] Time 1.702 (1.040) Loss 2.2488 (0.8350) Prec@1 51.953 (78.979) Prec@5 82.422 (94.463)
Test: [100/196] Time 0.300 (1.016) Loss 1.2809 (0.8886) Prec@1 69.141 (77.827) Prec@5 89.844 (93.862)
Test: [110/196] Time 1.121 (1.011) Loss 0.9445 (0.9139) Prec@1 75.000 (77.404) Prec@5 92.188 (93.567)
Test: [120/196] Time 0.667 (1.007) Loss 1.4142 (0.9327) Prec@1 66.406 (77.079) Prec@5 86.328 (93.337)
Test: [130/196] Time 0.915 (0.995) Loss 0.6773 (0.9680) Prec@1 81.641 (76.273) Prec@5 95.312 (92.981)
Test: [140/196] Time 0.315 (0.989) Loss 1.1367 (0.9884) Prec@1 74.219 (75.923) Prec@5 91.016 (92.758)
Test: [150/196] Time 0.840 (0.979) Loss 1.2445 (1.0101) Prec@1 76.953 (75.492) Prec@5 88.672 (92.454)
Test: [160/196] Time 0.324 (0.978) Loss 0.9249 (1.0276) Prec@1 80.078 (75.182) Prec@5 90.234 (92.185)
Test: [170/196] Time 0.534 (0.967) Loss 0.6927 (1.0442) Prec@1 80.859 (74.737) Prec@5 96.875 (91.954)
Test: [180/196] Time 0.298 (0.965) Loss 1.3764 (1.0596) Prec@1 66.406 (74.350) Prec@5 91.797 (91.786)
Test: [190/196] Time 0.414 (0.962) Loss 1.1433 (1.0589) Prec@1 71.094 (74.317) Prec@5 94.531 (91.823)
* Prec@1 74.448 Prec@5 91.884
Kaiming: ResNet-50 24.7% 7.8% (shorter side=256)
2. python caffe2darknet.py ResNet-50-deploy.prototxt ResNet-50-model.caffemodel ResNet-50-model.cfg ResNet-50-model.weights
3. python main.py -a resnet50-kaiming-dk --pretrained -e /home/xiaohang/ImageNet/
=> using pre-trained model 'resnet50-kaiming-dk'
load weights from ResNet-50-model.weights
Test: [0/196] Time 14.963 (14.963) Loss 0.5485 (0.5485) Prec@1 87.500 (87.500) Prec@5 96.875 (96.875)
Test: [10/196] Time 0.939 (1.876) Loss 1.0412 (0.7676) Prec@1 73.828 (80.717) Prec@5 92.578 (94.567)
Test: [20/196] Time 0.331 (1.392) Loss 0.7243 (0.7656) Prec@1 85.547 (80.804) Prec@5 92.969 (94.494)
Test: [30/196] Time 1.910 (1.267) Loss 0.8022 (0.7273) Prec@1 82.031 (81.981) Prec@5 95.703 (94.796)
Test: [40/196] Time 0.352 (1.154) Loss 0.8016 (0.7690) Prec@1 79.297 (80.516) Prec@5 94.922 (94.779)
Test: [50/196] Time 1.606 (1.111) Loss 0.4365 (0.7552) Prec@1 89.062 (80.668) Prec@5 98.047 (95.044)
Test: [60/196] Time 0.714 (1.077) Loss 1.0453 (0.7689) Prec@1 72.656 (80.277) Prec@5 93.750 (95.210)
Test: [70/196] Time 0.332 (1.055) Loss 0.7959 (0.7542) Prec@1 77.344 (80.573) Prec@5 94.922 (95.384)
Test: [80/196] Time 1.654 (1.054) Loss 1.5740 (0.7775) Prec@1 62.109 (80.179) Prec@5 85.938 (95.100)
Test: [90/196] Time 0.344 (1.030) Loss 2.2488 (0.8350) Prec@1 51.953 (78.979) Prec@5 82.422 (94.463)
Test: [100/196] Time 1.332 (1.016) Loss 1.2809 (0.8886) Prec@1 69.141 (77.827) Prec@5 89.844 (93.862)
Test: [110/196] Time 0.336 (1.005) Loss 0.9445 (0.9139) Prec@1 75.000 (77.404) Prec@5 92.188 (93.567)
Test: [120/196] Time 1.411 (1.000) Loss 1.4142 (0.9327) Prec@1 66.406 (77.079) Prec@5 86.328 (93.337)
Test: [130/196] Time 1.784 (0.997) Loss 0.6773 (0.9680) Prec@1 81.641 (76.273) Prec@5 95.312 (92.981)
Test: [140/196] Time 0.374 (0.986) Loss 1.1367 (0.9884) Prec@1 74.219 (75.923) Prec@5 91.016 (92.758)
Test: [150/196] Time 1.725 (0.983) Loss 1.2445 (1.0101) Prec@1 76.953 (75.492) Prec@5 88.672 (92.454)
Test: [160/196] Time 0.345 (0.974) Loss 0.9249 (1.0276) Prec@1 80.078 (75.182) Prec@5 90.234 (92.185)
Test: [170/196] Time 1.802 (0.972) Loss 0.6927 (1.0442) Prec@1 80.859 (74.737) Prec@5 96.875 (91.954)
Test: [180/196] Time 1.748 (0.967) Loss 1.3764 (1.0596) Prec@1 66.406 (74.350) Prec@5 91.797 (91.786)
Test: [190/196] Time 1.099 (0.960) Loss 1.1433 (1.0589) Prec@1 71.094 (74.317) Prec@5 94.531 (91.823)
* Prec@1 74.448 Prec@5 91.884
convert resnet18 from caffe to darknet
1. Download resnet18 from https://github.com/HolmesShuan/ResNet-18-Caffemodel-on-ImageNet.git and save as resnet-18.caffemodel
python main.py -a resnet18-caffe --pretrained -e /home/xiaohang/ImageNet/
=> using pre-trained model 'resnet18-caffe'
load weights from resnet-18.caffemodel
Loading caffemodel: resnet-18.caffemodel
Test: [0/196] Time 14.473 (14.473) Loss 0.6839 (0.6839) Prec@1 83.594 (83.594) Prec@5 95.703 (95.703)
Test: [10/196] Time 0.313 (1.738) Loss 1.3643 (1.0104) Prec@1 62.500 (74.183) Prec@5 89.844 (91.868)
Test: [20/196] Time 0.198 (1.274) Loss 1.1714 (1.0130) Prec@1 75.391 (74.126) Prec@5 87.891 (91.853)
Test: [30/196] Time 0.205 (1.199) Loss 1.0284 (0.9888) Prec@1 74.609 (74.849) Prec@5 92.969 (92.036)
Test: [40/196] Time 0.354 (1.101) Loss 0.9944 (1.0499) Prec@1 73.047 (72.933) Prec@5 95.703 (92.073)
Test: [50/196] Time 0.451 (1.065) Loss 0.6899 (1.0433) Prec@1 86.719 (73.032) Prec@5 95.312 (92.310)
Test: [60/196] Time 0.430 (1.033) Loss 1.1791 (1.0446) Prec@1 66.406 (72.688) Prec@5 92.578 (92.508)
Test: [70/196] Time 0.487 (1.030) Loss 1.0826 (1.0279) Prec@1 71.875 (73.234) Prec@5 90.625 (92.567)
Test: [80/196] Time 0.832 (1.015) Loss 1.8822 (1.0599) Prec@1 57.422 (72.632) Prec@5 82.812 (92.110)
Test: [90/196] Time 2.098 (1.017) Loss 2.2423 (1.1228) Prec@1 48.047 (71.321) Prec@5 78.906 (91.312)
Test: [100/196] Time 0.199 (0.999) Loss 1.6156 (1.1888) Prec@1 60.938 (70.073) Prec@5 84.375 (90.447)
Test: [110/196] Time 1.784 (1.003) Loss 1.2417 (1.2139) Prec@1 69.922 (69.640) Prec@5 88.281 (90.072)
Test: [120/196] Time 0.264 (1.010) Loss 1.9448 (1.2463) Prec@1 58.984 (69.183) Prec@5 79.297 (89.550)
Test: [130/196] Time 2.169 (1.016) Loss 1.1295 (1.2835) Prec@1 73.047 (68.350) Prec@5 90.234 (89.057)
Test: [140/196] Time 0.302 (1.015) Loss 1.5492 (1.3093) Prec@1 63.672 (67.855) Prec@5 84.766 (88.683)
Test: [150/196] Time 2.651 (1.020) Loss 1.5608 (1.3354) Prec@1 69.141 (67.459) Prec@5 82.031 (88.271)
Test: [160/196] Time 0.260 (1.020) Loss 1.4529 (1.3561) Prec@1 69.922 (67.151) Prec@5 84.766 (87.976)
Test: [170/196] Time 2.067 (1.023) Loss 1.1040 (1.3789) Prec@1 69.922 (66.642) Prec@5 91.016 (87.653)
Test: [180/196] Time 1.118 (1.021) Loss 1.5395 (1.3954) Prec@1 60.547 (66.313) Prec@5 87.891 (87.412)
Test: [190/196] Time 1.031 (1.024) Loss 1.6755 (1.3916) Prec@1 56.641 (66.371) Prec@5 84.766 (87.484)
* Prec@1 66.562 Prec@5 87.562
2. python caffe2darknet.py cfg/resnet-18.prototxt resnet-18.caffemodel resnet-18.cfg resnet-18.weights
3. python main.py -a resnet18-darknet --pretrained -e /home/xiaohang/ImageNet/
=> using pre-trained model 'resnet18-darknet'
load weights from resnet-18.weights
Test: [0/196] Time 15.171 (15.171) Loss 0.6839 (0.6839) Prec@1 83.594 (83.594) Prec@5 95.703 (95.703)
Test: [10/196] Time 0.560 (1.835) Loss 1.3643 (1.0104) Prec@1 62.500 (74.183) Prec@5 89.844 (91.868)
Test: [20/196] Time 0.290 (1.345) Loss 1.1714 (1.0130) Prec@1 75.391 (74.126) Prec@5 87.891 (91.853)
Test: [30/196] Time 1.594 (1.237) Loss 1.0284 (0.9888) Prec@1 74.609 (74.849) Prec@5 92.969 (92.036)
Test: [40/196] Time 0.820 (1.151) Loss 0.9944 (1.0499) Prec@1 73.047 (72.933) Prec@5 95.703 (92.073)
Test: [50/196] Time 0.928 (1.114) Loss 0.6899 (1.0433) Prec@1 86.719 (73.032) Prec@5 95.312 (92.310)
Test: [60/196] Time 0.264 (1.081) Loss 1.1791 (1.0446) Prec@1 66.406 (72.688) Prec@5 92.578 (92.508)
Test: [70/196] Time 0.694 (1.080) Loss 1.0826 (1.0279) Prec@1 71.875 (73.234) Prec@5 90.625 (92.567)
Test: [80/196] Time 0.742 (1.059) Loss 1.8822 (1.0599) Prec@1 57.422 (72.632) Prec@5 82.812 (92.110)
Test: [90/196] Time 0.920 (1.057) Loss 2.2423 (1.1228) Prec@1 48.047 (71.321) Prec@5 78.906 (91.312)
Test: [100/196] Time 0.251 (1.035) Loss 1.6156 (1.1888) Prec@1 60.938 (70.073) Prec@5 84.375 (90.447)
Test: [110/196] Time 0.857 (1.034) Loss 1.2417 (1.2139) Prec@1 69.922 (69.640) Prec@5 88.281 (90.072)
Test: [120/196] Time 0.857 (1.029) Loss 1.9448 (1.2463) Prec@1 58.984 (69.183) Prec@5 79.297 (89.550)
Test: [130/196] Time 1.816 (1.024) Loss 1.1295 (1.2835) Prec@1 73.047 (68.350) Prec@5 90.234 (89.057)
Test: [140/196] Time 0.475 (1.010) Loss 1.5492 (1.3093) Prec@1 63.672 (67.855) Prec@5 84.766 (88.683)
Test: [150/196] Time 1.124 (1.003) Loss 1.5608 (1.3354) Prec@1 69.141 (67.459) Prec@5 82.031 (88.271)
Test: [160/196] Time 0.812 (1.000) Loss 1.4529 (1.3561) Prec@1 69.922 (67.151) Prec@5 84.766 (87.976)
Test: [170/196] Time 0.328 (0.989) Loss 1.1040 (1.3789) Prec@1 69.922 (66.642) Prec@5 91.016 (87.653)
Test: [180/196] Time 1.271 (0.988) Loss 1.5395 (1.3954) Prec@1 60.547 (66.313) Prec@5 87.891 (87.412)
Test: [190/196] Time 0.292 (0.980) Loss 1.6755 (1.3916) Prec@1 56.641 (66.371) Prec@5 84.766 (87.484)
* Prec@1 66.562 Prec@5 87.562