天池竞赛-津南数字制造算法挑战赛赛场二解决方案源码+项目说明.zip

# Group Normalization for Mask R-CNN <div align="center"> <img src="gn.jpg" width="700px" /> </div> ## Introduction This file provides Mask R-CNN baseline results and models trained with [Group Normalization](https://arxiv.org/abs/1803.08494): ``` @article{GroupNorm2018, title={Group Normalization}, author={Yuxin Wu and Kaiming He}, journal={arXiv:1803.08494}, year={2018} } ``` **Note:** This code uses the GroupNorm op implemented in CUDA, included in the Caffe2 repo. When writing this document, Caffe2 is being merged into PyTorch, and the GroupNorm op is located [here](https://github.com/pytorch/pytorch/blob/master/caffe2/operators/group_norm_op.cu). Make sure your Caffe2 is up to date. ## Pretrained Models with GN These models are trained in Caffe2 on the standard ImageNet-1k dataset, using GroupNorm with 32 groups (G=32). - [R-50-GN.pkl](https://dl.fbaipublicfiles.com/detectron/ImageNetPretrained/47261647/R-50-GN.pkl): ResNet-50 with GN, 24.0\% top-1 error (center-crop). - [R-101-GN.pkl](https://dl.fbaipublicfiles.com/detectron/ImageNetPretrained/47592356/R-101-GN.pkl): ResNet-101 with GN, 22.6\% top-1 error (center-crop). ## Results ### Baselines with BN <table><tbody> <!-- START E2E MASK RCNN BN TABLE --> <!-- TABLE HEADER --> <!-- Info: we use wrap text in <sup><sup> to make is small --> <th valign="bottom">^{_{&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;case&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;}}</th> <th valign="bottom">^_type</th> <th valign="bottom">^_lr<br/>schd</th> <th valign="bottom">^_im/<br/>gpu</th> <th valign="bottom">^{_{train<br/>mem<br/>(GB)}}</th> <th valign="bottom">^{_{train<br/>time<br/>(s/iter)}}</th> <th valign="bottom">^{_{train<br/>time<br/>total<br/>(hr)}}</th> <th valign="bottom">^{_{inference<br/>time<br/>(s/im)}}</th> <th valign="bottom">^_box<br/>AP</th> <th valign="bottom">^_mask<br/>AP</th> <th valign="bottom">^{_{model id}}</th> <tr> <td align="left">^{_{R-50-FPN, BN*}}</td> <td align="left">^{_{Mask R-CNN}}</td> <td align="left">^_2x</td> <td align="right">^₂</td> <td align="right">^_8.6</td> <td align="right">^_0.897</td> <td align="right">^_44.9</td> <td align="right">^{_{0.099&nbsp;+&nbsp;0.018}}</td> <td align="right">^_38.6</td> <td align="right">^_34.5</td> <td align="right">^_35859007</td> </tr> <tr> <td align="left">^{_{R-101-FPN, BN*}}</td> <td align="left">^{_{Mask R-CNN}}</td> <td align="left">^_2x</td> <td align="right">^₂</td> <td align="right">^_10.2</td> <td align="right">^_0.993</td> <td align="right">^_49.7</td> <td align="right">^{_{0.126&nbsp;+&nbsp;0.017}}</td> <td align="right">^_40.9</td> <td align="right">^_36.4</td> <td align="right">^_35861858</td> </tr> <!-- END E2E MASK RCNN BN TABLE --> </tbody></table> **Notes:** - This table is copied from [Detectron Model Zoo](https://github.com/facebookresearch/Detectron/blob/master/MODEL_ZOO.md#end-to-end-faster--mask-r-cnn-baselines). - BN^* means that BatchNorm (BN) is used for pre-training and is frozen and turned into a per-channel linear layer when fine-tuning. This is the default of Faster/Mask R-CNN and Detectron. ### Mask R-CNN with GN #### Standard Mask R-CNN recipe <table><tbody> <!-- START E2E MASK RCNN GN TABLE --> <!-- TABLE HEADER --> <!-- Info: we use wrap text in <sup><sup> to make is small --> <th valign="bottom">^{_{&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;case&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;}}</th> <th valign="bottom">^_type</th> <th valign="bottom">^_lr<br/>schd</th> <th valign="bottom">^_im/<br/>gpu</th> <th valign="bottom">^{_{train<br/>mem<br/>(GB)}}</th> <th valign="bottom">^{_{train<br/>time<br/>(s/iter)}}</th> <th valign="bottom">^{_{train<br/>time<br/>total<br/>(hr)}}</th> <th valign="bottom">^{_{inference<br/>time<br/>(s/im)}}</th> <th valign="bottom">^_box<br/>AP</th> <th valign="bottom">^_mask<br/>AP</th> <th valign="bottom">^{_{model id}}</th> <th valign="bottom">^{_{download<br/>links}}</th> <!-- TABLE BODY --> <tr> <td align="left">^{_{R-50-FPN, GN}}</td> <td align="left">^{_{Mask R-CNN}}</td> <td align="left">^_2x</td> <td align="right">^₂</td> <td align="right">^_10.5</td> <td align="right">^_1.017</td> <td align="right">^_50.8</td> <td align="right">^{_{0.146&nbsp;+&nbsp;0.017}}</td> <td align="right">^_40.3</td> <td align="right">^_35.7</td> <td align="right">^_48616381</td> <td align="left">^{_{<a href="https://dl.fbaipublicfiles.com/detectron/GN/48616381/04_2018_gn_baselines/e2e_mask_rcnn_R-50-FPN_2x_gn_0416.13_23_38.bTlTI97Q/output/train/coco_2014_train%3Acoco_2014_valminusminival/generalized_rcnn/model_final.pkl">model</a> &nbsp;|&nbsp; <a href="https://dl.fbaipublicfiles.com/detectron/GN/48616381/04_2018_gn_baselines/e2e_mask_rcnn_R-50-FPN_2x_gn_0416.13_23_38.bTlTI97Q/output/test/coco_2014_minival/generalized_rcnn/bbox_coco_2014_minival_results.json">boxes</a> &nbsp;|&nbsp; <a href="https://dl.fbaipublicfiles.com/detectron/GN/48616381/04_2018_gn_baselines/e2e_mask_rcnn_R-50-FPN_2x_gn_0416.13_23_38.bTlTI97Q/output/test/coco_2014_minival/generalized_rcnn/segmentations_coco_2014_minival_results.json">masks</a>}}</td> </tr> <tr> <td align="left">^{_{R-101-FPN, GN}}</td> <td align="left">^{_{Mask R-CNN}}</td> <td align="left">^_2x</td> <td align="right">^₂</td> <td align="right">^_12.4</td> <td align="right">^_1.151</td> <td align="right">^_57.5</td> <td align="right">^{_{0.180&nbsp;+&nbsp;0.015}}</td> <td align="right">^_41.8</td> <td align="right">^_36.8</td> <td align="right">^_48616724</td> <td align="left">^{_{<a href="https://dl.fbaipublicfiles.com/detectron/GN/48616724/04_2018_gn_baselines/e2e_mask_rcnn_R-101-FPN_2x_gn_0416.13_26_34.GLnri4GR/output/train/coco_2014_train%3Acoco_2014_valminusminival/generalized_rcnn/model_final.pkl">model</a> &nbsp;|&nbsp; <a href="https://dl.fbaipublicfiles.com/detectron/GN/48616724/04_2018_gn_baselines/e2e_mask_rcnn_R-101-FPN_2x_gn_0416.13_26_34.GLnri4GR/output/test/coco_2014_minival/generalized_rcnn/bbox_coco_2014_minival_results.json">boxes</a> &nbsp;|&nbsp; <a href="https://dl.fbaipublicfiles.com/detectron/GN/48616724/04_2018_gn_baselines/e2e_mask_rcnn_R-101-FPN_2x_gn_0416.13_26_34.GLnri4GR/output/test/coco_2014_minival/generalized_rcnn/segmentations_coco_2014_minival_results.json">masks</a>}}</td> </tr> <!-- END E2E MASK RCNN GN TABLE --> </tbody></table> **Notes:** - GN is applied on: (i) ResNet layers inherited from pre-training, (ii) the FPN-specific layers, (iii) the RoI bbox head, and (iv) the RoI mask head. - These GN models use a 4conv+1fc RoI box head. The BN^* counterpart with this head performs similarly with the default 2fc head: using this co