Caffe models (include classification, detection and segmentation) and deploy prototxt for resnet, resnext, inception_v3, inception_v4, inception_resnet, wider_resnet, densenet, aligned-inception-resne(x)t, DPNs and other networks.
We recommend using these caffe models with py-RFCN-priv
Please install py-RFCN-priv for evaluating and finetuning.
Most of the pre-train models are converted from other projects, the main contribution belongs to the original authors.
Project links:
mxnet-model-gallery、 tensorflow slim、 craftGBD、 ResNeXt、 DenseNet、 wide-residual-networks、 keras deep-learning-models、 ademxapp、 DPNs
CLS (Classification, more details are in cls)
Top-1/5 error of pre-train models in this repository (Pre-train models download urls).
| Network | 224/299 (single-crop) |
224/299 (12-crop) |
320/395 (single-crop) |
320/395 (12-crop) |
|---|---|---|---|---|
| resnet18-priv | 29.11/10.07 | 26.69/8.64 | 27.54/8.98 | 26.23/8.21 |
| resnext26-32x4d-priv | 24.93/7.75 | 23.54/6.89 | 24.20/7.21 | 23.19/6.60 |
| resnet101-v2 | 21.95/6.12 | 19.99/5.04 | 20.37/5.16 | 19.29/4.57 |
| resnet152-v2 | 20.85/5.42 | 19.24/4.68 | 19.66/4.73 | 18.84/4.32 |
| resnet269-v2 | 19.71/5.00 | 18.25/4.20 | 18.70/4.33 | 17.87/3.85 |
| resnet38a | 20.66/5.27 | ../.. | 19.25/4.66 | ../.. |
| inception-v3 | 21.67/5.75 | 19.60/4.73 | 20.10/4.82 | 19.25/4.24 |
| xception | 20.90/5.49 | 19.68/4.90 | 19.58/4.77 | 18.91/4.39 |
| inception-v4 | 20.03/5.09 | 18.60/4.30 | 18.68/4.32 | 18.12/3.92 |
| inception-resnet-v2 | 19.86/4.83 | 18.46/4.08 | 18.75/4.02 | 18.15/3.71 |
| resnext50-32x4d | 22.37/6.31 | 20.53/5.35 | 21.10/5.53 | 20.37/5.03 |
| resnext101-32x4d | 21.30/5.79 | 19.47/4.89 | 19.91/4.97 | 19.19/4.59 |
| resnext101-64x4d | 20.60/5.41 | 18.88/4.59 | 19.26/4.63 | 18.48/4.31 |
| wrn50-2 (resnet50-1x128d) |
22.13/6.13 | 20.09/5.06 | 20.68/5.28 | 19.83/4.87 |
| airx50-24x4d | 22.39/6.23 | 20.36/5.19 | 20.88/5.33 | 19.97/4.92 |
| air101 | 21.32/5.76 | ../.. | 19.92/4.75 | ../.. |
| airx101-32x4d | 21.15/5.74 | ../.. | 19.61/4.93 | ../.. |
| airx152-32x4d | 20.77/5.49 | 19.00/4.53 | ../.. | ../.. |
| dpn-92 | 20.81/5.47 | 18.99/4.59 | 19.23/4.64 | ../.. |
| dpn-98 | 20.27/5.28 | ../.. | 18.87/4.43 | ../.. |
| dpn-131 | 20.00/5.24 | ../.. | 18.63/4.31 | ../.. |
| dpn-107 | 19.70/5.06 | ../.. | 18.41/4.25 | ../.. |
- The resnet18-priv, resnext26-32x4d-priv are trained under pytorch by bupt-priv.
- The pre-train models are tested on original caffe by evaluation_cls.py, but ceil_mode:false（pooling_layer） is used for the models converted from torch, the detail in https://github.com/BVLC/caffe/pull/3057/files. If you remove ceil_mode:false, the performance will decline about 1% top1.
- 224x224(base_size=256) and 320x320(base_size=320) crop size for resnet-v2/resnext/wrn, 299x299(base_size=320) and 395x395(base_size=395) crop size for inception.
DET (Detection, more details are in det)
Original faster rcnn train on VOC 2007+2012 trainval and test on VOC 2007 test.
| Network | mAP@50 | train speed | train memory | test speed | test memory |
|---|---|---|---|---|---|
| resnet18 | 70.02 | 9.5 img/s | 1,235MB | 17.5 img/s | 989MB |
| resnet101 | -- | -- | -- | -- | -- |
| resnet101-v2 | 79.6 | 3.1 img/s | 6,495MB | 7.1 img/s | 4,573MB |
| resnet152-v2 | 80.72 | 2.8 img/s | 9,315MB | 6.2 img/s | 6,021MB |
| wrn50-2 | 78.59 | 2.1 img/s | 4,895MB | 4.9 img/s | 3,499MB |
| resnext50-32x4d | 77.99 | 3.6 img/s | 5,315MB | 7.4 img/s | 4,305MB |
| resnext101-32x4d | 79.98 | 2.7 img/s | 7,836MB | 6.3 img/s | 5,705MB |
| resnext101-64x4d | 80.71 | 2.0 img/s (batch=96) |
11,277MB | 3.7 img/s | 9,461MB |
| inception-v3 | 78.6 | 4.1 img/s | 4,325MB | 7.3 img/s | 3,445MB |
| inception-v4 | 81.49 | 2.6 img/s | 6,759MB | 5.4 img/s | 4,683MB |
| inception-resnet-v2 | 80.0 | 2.0 img/s (batch=112) |
11,497MB | 3.2 img/s | 8,409MB |
| densenet-161 | -- | -- | -- | -- | -- |
| densenet-201 | 77.53 | 3.9 img/s (batch=72) |
10,073MB | 5.5 img/s | 9,955MB |
| resnet38a | 80.1 | 1.4 img/s | 8,723MB | 3.4 img/s | 5,501MB |
- To reduce memory usage, we merge all the models batchnorm layer parameters into scale layer, more details please refer to faster-rcnn-resnet or pva-faster-rcnn;
- We also split the deploy file to rpn deploy file and rcnn deploy file for adopting more testing tricks.
- Performanc, speed and memory are calculated on py-RFCN-priv with Nvidia Titan pascal, we do not guarantee that the results can be reproduced under any other conditions;
- All the models are trained on a single scale (600*1000) with image flipping and train-batch=128 for 80,000 iterations, tested on the same single scale with test-batch=300 and nms=0.3;
caffe-model is released under the MIT License (refer to the LICENSE file for details).
I greatly thank Yangqing Jia and BVLC group for developing Caffe.
And I would like to thank all the authors of every network.