# gamma **Repository Path**: chenx2ovo/gamma ## Basic Information - **Project Name**: gamma - **Description**: miccai 2021 bd gamma task 1-3 - **Primary Language**: Unknown - **License**: Not specified - **Default Branch**: master - **Homepage**: None - **GVP Project**: No ## Statistics - **Stars**: 0 - **Forks**: 0 - **Created**: 2021-07-27 - **Last Updated**: 2022-05-30 ## Categories & Tags **Categories**: Uncategorized **Tags**: None ## README # MICCAI 2021 GAMMA ## 比赛介绍 https://aistudio.baidu.com/aistudio/competition/detail/90 ## 脚本 ln -s /data/chenx2/checkpoints /root/.cache/torch/hub ### task1 使用 timm() 镜像 py task1 infer transform res diff python3 -m torch.distributed.launch --nproc_per_node=2 --use_env task1.py --model_name=deit_deit_selayer --batch-size=8 --epochs=600 --save_checkpoint=300 --frame_num=32 infer python3 task1.py --infer --resume=task1/20210806171021_deit_deit_selayer_SELayerFusion_bs8/checkpoint_max_kappa.pth --frame_num=32 --model_name=deit_deit_selayer ### task2 使用 timm python3 task2.py --lr=1e-4 --warmup-lr=1e-5 --min-lr=1e-5 --num_workers=8 --batch-size=16 --backbone_name=resnest26d --input-size=224 --warmup-epochs=10 --epochs=1000 --loss_alpha=0.5 --criterion_name=AEDWithClsLossV4 python3 task2.py --lr=5e-4 --warmup-lr=1e-4 --min-lr=1e-5 --num_workers=4 --batch-size=4 --backbone_name=hrnet_w32 --epochs=500 --loss_alpha=0.5 --criterion_name=Task2BaselineLoss --save_checkpoint=1 python3 task2.py --infer --resume=task2/20210807073623_Point_Exist_Branch2_hrnet_w32_AEDWithClsLossV4_bs16/checkpoint_max_acc.pth --backbone_name=hrnet_w32 todo 1. loss design done 2. try hrnet 3. 黄斑 先裁掉 上下1/4 然后 提取绿色通道 然后对绿色通道clahe 然后闭包运算消除血管 ### task3 https://blog.csdn.net/qq_40511157/article/details/102770108 使用 smp() ablu() ttach(https://github.com/qubvel/ttach) https://neptune.ai/blog/image-segmentation-tips-and-tricks-from-kaggle-competitions add TransUnet python3 task3.py --batch-size=8 --model_name=TransUnet --warmup-lr=1e-4 --lr=5e-4 --min-lr=1e-5 --encoder=R50-ViT-B_16 --criterion=DiceLoss --save_checkpoint=1 --epochs=1000 python3 -m torch.distributed.launch --nproc_per_node=2 --use_env task3.py --batch-size=8 --model_name=TransUnet --warmup-lr=1e-4 --lr=5e-4 --min-lr=1e-5 --encoder=R50-ViT-B_16 --criterion=DiceLoss --save_checkpoint=1 --epochs=1000 python3 task3.py --batch-size=8 --model_name=Unet --warmup-lr=1e-4 --lr=5e-4 --min-lr=1e-5 --encoder=resnet34 --criterion=DiceLoss 0 python3 task3.py --batch-size=16 --model_name=Unet --warmup-lr=1e-4 --lr=5e-4 --min-lr=1e-5 --encoder=resnet34 --criterion=TverskyLoss --save_checkpoint=1 --epochs=300 DETECTION th=0.4->0.6 score 上升 县 python3 task3_det task3 task3/Key_7.79763_20210810033944_UnetPlusPlus_resnext50_32x4d_DiceLoss_bs16 thresh 0.3 7.57021 thresh 0.5 7.67314 +tta 6.51454 thresh 0.5 b4 +tta -0.8 180104 横轴 det th | th01 | 06 | 0.7 | 08 | 1 | | | ---- | ------- | --- | ----------------------- | -------------------- | - | | 0.5 | 8.23519 | | 8.13648(有一个bad case) | 8.10937(159bad case) | | | 0.9 | | | | | | det 06 th 05 调整测试时的tranform 结果8.23562 略高 从实验结果来看 提升cup分割精度对最终score提升比较关键 即小物体分割 **activation** ![image-20210826142734120](https://i.loli.net/2021/08/26/sb7W3tUENvdD9Qh.png) 最长 灰线 默认activation **蓝色 none** 黄色 softmax2d **红色（与蓝色重合） tanh** activation none/tanh softmax2d loss会高 但是score 一致 **Attention** ``` # combine seg use thresh 0.7 # sep disc train 04 test 07 # sep cup train 02 test 04 ``` disc 对test det th 调大 从0.7-》0.8 ![x](https://gitee.com/chenx2ovo/picgo/raw/master/img/image-20210903114731859.png) 0.8-》0.9 上面的图标号137 调到0.9之后没有上上面的情况了 CUP train det th=-0.2 相当于在原有bbox的基础上宽度为原先的0.8 训练效果很差 不管是loss 还是score 都没有之前的水平高 尝试 det th0.2 todo 1. 先验知识计算loss 2. combo loss/主动轮廓能量最小化损失函数 3. 增加分类分支 增加任务难度 减缓收敛速度 预期使得模型收敛到更高的水平 https://journals.sagepub.com/doi/full/10.1177/1729881420921676 ``` https://ieeexplore.ieee.org/document/9098698 https://github.com/JunMa11/SegLoss ``` ### model fusion ['adv_inception_v3', 'cait_m36_384', 'cait_m48_448', 'cait_s24_224', 'cait_s24_384', 'cait_s36_384', 'cait_xs24_384', 'cait_xxs24_224', 'cait_xxs24_384', 'cait_xxs36_224', 'cait_xxs36_384', 'coat_lite_mini', 'coat_lite_small', 'coat_lite_tiny', 'coat_mini', 'coat_tiny', 'convit_base', 'convit_small', 'convit_tiny', 'cspdarknet53', 'cspresnet50', 'cspresnext50', 'deit_base_distilled_patch16_224', 'deit_base_distilled_patch16_384', 'deit_base_patch16_224', 'deit_base_patch16_384', 'deit_small_distilled_patch16_224', 'deit_small_patch16_224', 'deit_tiny_distilled_patch16_224', 'deit_tiny_patch16_224', 'densenet121', 'densenet161', 'densenet169', 'densenet201', 'densenetblur121d', 'dla34', 'dla46_c', 'dla46x_c', 'dla60', 'dla60_res2net', 'dla60_res2next', 'dla60x', 'dla60x_c', 'dla102', 'dla102x', 'dla102x2', 'dla169', 'dm_nfnet_f0', 'dm_nfnet_f1', 'dm_nfnet_f2', 'dm_nfnet_f3', 'dm_nfnet_f4', 'dm_nfnet_f5', 'dm_nfnet_f6', 'dpn68', 'dpn68b', 'dpn92', 'dpn98', 'dpn107', 'dpn131', 'eca_nfnet_l0', 'eca_nfnet_l1', 'eca_nfnet_l2', 'ecaresnet26t', 'ecaresnet50d', 'ecaresnet50d_pruned', 'ecaresnet50t', 'ecaresnet101d', 'ecaresnet101d_pruned', 'ecaresnet269d', 'ecaresnetlight', 'efficientnet_b0', 'efficientnet_b1', 'efficientnet_b1_pruned', 'efficientnet_b2', 'efficientnet_b2_pruned', 'efficientnet_b3', 'efficientnet_b3_pruned', 'efficientnet_b4', 'efficientnet_el', 'efficientnet_el_pruned', 'efficientnet_em', 'efficientnet_es', 'efficientnet_es_pruned', 'efficientnet_lite0', 'efficientnetv2_rw_m', 'efficientnetv2_rw_s', 'ens_adv_inception_resnet_v2', 'ese_vovnet19b_dw', 'ese_vovnet39b', 'fbnetc_100', 'gernet_l', 'gernet_m', 'gernet_s', 'ghostnet_100', 'gluon_inception_v3', 'gluon_resnet18_v1b', 'gluon_resnet34_v1b', 'gluon_resnet50_v1b', 'gluon_resnet50_v1c', 'gluon_resnet50_v1d', 'gluon_resnet50_v1s', 'gluon_resnet101_v1b', 'gluon_resnet101_v1c', 'gluon_resnet101_v1d', 'gluon_resnet101_v1s', 'gluon_resnet152_v1b', 'gluon_resnet152_v1c', 'gluon_resnet152_v1d', 'gluon_resnet152_v1s', 'gluon_resnext50_32x4d', 'gluon_resnext101_32x4d', 'gluon_resnext101_64x4d', 'gluon_senet154', 'gluon_seresnext50_32x4d', 'gluon_seresnext101_32x4d', 'gluon_seresnext101_64x4d', 'gluon_xception65', 'gmixer_24_224', 'gmlp_s16_224', 'hardcorenas_a', 'hardcorenas_b', 'hardcorenas_c', 'hardcorenas_d', 'hardcorenas_e', 'hardcorenas_f', 'hrnet_w18', 'hrnet_w18_small', 'hrnet_w18_small_v2', 'hrnet_w30', 'hrnet_w32', 'hrnet_w40', 'hrnet_w44', 'hrnet_w48', 'hrnet_w64', 'ig_resnext101_32x8d', 'ig_resnext101_32x16d', 'ig_resnext101_32x32d', 'ig_resnext101_32x48d', 'inception_resnet_v2', 'inception_v3', 'inception_v4', 'legacy_senet154', 'legacy_seresnet18', 'legacy_seresnet34', 'legacy_seresnet50', 'legacy_seresnet101', 'legacy_seresnet152', 'legacy_seresnext26_32x4d', 'legacy_seresnext50_32x4d', 'legacy_seresnext101_32x4d', 'levit_128', 'levit_128s', 'levit_192', 'levit_256', 'levit_384', 'mixer_b16_224', 'mixer_b16_224_in21k', 'mixer_b16_224_miil', 'mixer_b16_224_miil_in21k', 'mixer_l16_224', 'mixer_l16_224_in21k', 'mixnet_l', 'mixnet_m', 'mixnet_s', 'mixnet_xl', 'mnasnet_100', 'mobilenetv2_100', 'mobilenetv2_110d', 'mobilenetv2_120d', 'mobilenetv2_140', 'mobilenetv3_large_100', 'mobilenetv3_large_100_miil', 'mobilenetv3_large_100_miil_in21k', 'mobilenetv3_rw', 'nasnetalarge', 'nf_regnet_b1', 'nf_resnet50', 'nfnet_l0', 'pit_b_224', 'pit_b_distilled_224', 'pit_s_224', 'pit_s_distilled_224', 'pit_ti_224', 'pit_ti_distilled_224', 'pit_xs_224', 'pit_xs_distilled_224', 'pnasnet5large', 'regnetx_002', 'regnetx_004', 'regnetx_006', 'regnetx_008', 'regnetx_016', 'regnetx_032', 'regnetx_040', 'regnetx_064', 'regnetx_080', 'regnetx_120', 'regnetx_160', 'regnetx_320', 'regnety_002', 'regnety_004', 'regnety_006', 'regnety_008', 'regnety_016', 'regnety_032', 'regnety_040', 'regnety_064', 'regnety_080', 'regnety_120', 'regnety_160', 'regnety_320', 'repvgg_a2', 'repvgg_b0', 'repvgg_b1', 'repvgg_b1g4', 'repvgg_b2', 'repvgg_b2g4', 'repvgg_b3', 'repvgg_b3g4', 'res2net50_14w_8s', 'res2net50_26w_4s', 'res2net50_26w_6s', 'res2net50_26w_8s', 'res2net50_48w_2s', 'res2net101_26w_4s', 'res2next50', 'resmlp_12_224', 'resmlp_12_distilled_224', 'resmlp_24_224', 'resmlp_24_distilled_224', 'resmlp_36_224', 'resmlp_36_distilled_224', 'resmlp_big_24_224', 'resmlp_big_24_224_in22ft1k', 'resmlp_big_24_distilled_224', 'resnest14d', 'resnest26d', 'resnest50d', 'resnest50d_1s4x24d', 'resnest50d_4s2x40d', 'resnest101e', 'resnest200e', 'resnest269e', 'resnet18', 'resnet18d', 'resnet26', 'resnet26d', 'resnet34', 'resnet34d', 'resnet50', 'resnet50d', 'resnet51q', 'resnet101d', 'resnet152d', 'resnet200d', 'resnetblur50', 'resnetrs50', 'resnetrs101', 'resnetrs152', 'resnetrs200', 'resnetrs270', 'resnetrs350', 'resnetrs420', 'resnetv2_50x1_bit_distilled', 'resnetv2_50x1_bitm', 'resnetv2_50x1_bitm_in21k', 'resnetv2_50x3_bitm', 'resnetv2_50x3_bitm_in21k', 'resnetv2_101x1_bitm', 'resnetv2_101x1_bitm_in21k', 'resnetv2_101x3_bitm', 'resnetv2_101x3_bitm_in21k', 'resnetv2_152x2_bit_teacher', 'resnetv2_152x2_bit_teacher_384', 'resnetv2_152x2_bitm', 'resnetv2_152x2_bitm_in21k', 'resnetv2_152x4_bitm', 'resnetv2_152x4_bitm_in21k', 'resnext50_32x4d', 'resnext50d_32x4d', 'resnext101_32x8d', 'rexnet_100', 'rexnet_130', 'rexnet_150', 'rexnet_200', 'selecsls42b', 'selecsls60', 'selecsls60b', 'semnasnet_100', 'seresnet50', 'seresnet152d', 'seresnext26d_32x4d', 'seresnext26t_32x4d', 'seresnext50_32x4d', 'skresnet18', 'skresnet34', 'skresnext50_32x4d', 'spnasnet_100', 'ssl_resnet18', 'ssl_resnet50', 'ssl_resnext50_32x4d', 'ssl_resnext101_32x4d', 'ssl_resnext101_32x8d', 'ssl_resnext101_32x16d', 'swin_base_patch4_window7_224', 'swin_base_patch4_window7_224_in22k', 'swin_base_patch4_window12_384', 'swin_base_patch4_window12_384_in22k', 'swin_large_patch4_window7_224', 'swin_large_patch4_window7_224_in22k', 'swin_large_patch4_window12_384', 'swin_large_patch4_window12_384_in22k', 'swin_small_patch4_window7_224', 'swin_tiny_patch4_window7_224', 'swsl_resnet18', 'swsl_resnet50', 'swsl_resnext50_32x4d', 'swsl_resnext101_32x4d', 'swsl_resnext101_32x8d', 'swsl_resnext101_32x16d', 'tf_efficientnet_b0', 'tf_efficientnet_b0_ap', 'tf_efficientnet_b0_ns', 'tf_efficientnet_b1', 'tf_efficientnet_b1_ap', 'tf_efficientnet_b1_ns', 'tf_efficientnet_b2', 'tf_efficientnet_b2_ap', 'tf_efficientnet_b2_ns', 'tf_efficientnet_b3', 'tf_efficientnet_b3_ap', 'tf_efficientnet_b3_ns', 'tf_efficientnet_b4', 'tf_efficientnet_b4_ap', 'tf_efficientnet_b4_ns', 'tf_efficientnet_b5', 'tf_efficientnet_b5_ap', 'tf_efficientnet_b5_ns', 'tf_efficientnet_b6', 'tf_efficientnet_b6_ap', 'tf_efficientnet_b6_ns', 'tf_efficientnet_b7', 'tf_efficientnet_b7_ap', 'tf_efficientnet_b7_ns', 'tf_efficientnet_b8', 'tf_efficientnet_b8_ap', 'tf_efficientnet_cc_b0_4e', 'tf_efficientnet_cc_b0_8e', 'tf_efficientnet_cc_b1_8e', 'tf_efficientnet_el', 'tf_efficientnet_em', 'tf_efficientnet_es', 'tf_efficientnet_l2_ns', 'tf_efficientnet_l2_ns_475', 'tf_efficientnet_lite0', 'tf_efficientnet_lite1', 'tf_efficientnet_lite2', 'tf_efficientnet_lite3', 'tf_efficientnet_lite4', 'tf_efficientnetv2_b0', 'tf_efficientnetv2_b1', 'tf_efficientnetv2_b2', 'tf_efficientnetv2_b3', 'tf_efficientnetv2_l', 'tf_efficientnetv2_l_in21ft1k', 'tf_efficientnetv2_l_in21k', 'tf_efficientnetv2_m', 'tf_efficientnetv2_m_in21ft1k', 'tf_efficientnetv2_m_in21k', 'tf_efficientnetv2_s', 'tf_efficientnetv2_s_in21ft1k', 'tf_efficientnetv2_s_in21k', 'tf_inception_v3', 'tf_mixnet_l', 'tf_mixnet_m', 'tf_mixnet_s', 'tf_mobilenetv3_large_075', 'tf_mobilenetv3_large_100', 'tf_mobilenetv3_large_minimal_100', 'tf_mobilenetv3_small_075', 'tf_mobilenetv3_small_100', 'tf_mobilenetv3_small_minimal_100', 'tnt_s_patch16_224', 'tresnet_l', 'tresnet_l_448', 'tresnet_m', 'tresnet_m_448', 'tresnet_m_miil_in21k', 'tresnet_xl', 'tresnet_xl_448', 'tv_densenet121', 'tv_resnet34', 'tv_resnet50', 'tv_resnet101', 'tv_resnet152', 'tv_resnext50_32x4d', 'twins_pcpvt_base', 'twins_pcpvt_large', 'twins_pcpvt_small', 'twins_svt_base', 'twins_svt_large', 'twins_svt_small', 'vgg11', 'vgg11_bn', 'vgg13', 'vgg13_bn', 'vgg16', 'vgg16_bn', 'vgg19', 'vgg19_bn', 'visformer_small', 'vit_base_patch16_224', 'vit_base_patch16_224_in21k', 'vit_base_patch16_224_miil', 'vit_base_patch16_224_miil_in21k', 'vit_base_patch16_384', 'vit_base_patch32_224', 'vit_base_patch32_224_in21k', 'vit_base_patch32_384', 'vit_base_r50_s16_224_in21k', 'vit_base_r50_s16_384', 'vit_huge_patch14_224_in21k', 'vit_large_patch16_224', 'vit_large_patch16_224_in21k', 'vit_large_patch16_384', 'vit_large_patch32_224_in21k', 'vit_large_patch32_384', 'vit_large_r50_s32_224', 'vit_large_r50_s32_224_in21k', 'vit_large_r50_s32_384', 'vit_small_patch16_224', 'vit_small_patch16_224_in21k', 'vit_small_patch16_384', 'vit_small_patch32_224', 'vit_small_patch32_224_in21k', 'vit_small_patch32_384', 'vit_small_r26_s32_224', 'vit_small_r26_s32_224_in21k', 'vit_small_r26_s32_384', 'vit_tiny_patch16_224', 'vit_tiny_patch16_224_in21k', 'vit_tiny_patch16_384', 'vit_tiny_r_s16_p8_224', 'vit_tiny_r_s16_p8_224_in21k', 'vit_tiny_r_s16_p8_384', 'wide_resnet50_2', 'wide_resnet101_2', 'xception', 'xception41', 'xception65', 'xception71'] ['resnet18', 'resnet34', 'resnet50', 'resnet101', 'resnet152', 'resnext50_32x4d', 'resnext101_32x4d', 'resnext101_32x8d', 'resnext101_32x16d', 'resnext101_32x32d', 'resnext101_32x48d', 'dpn68', 'dpn68b', 'dpn92', 'dpn98', 'dpn107', 'dpn131', 'vgg11', 'vgg11_bn', 'vgg13', 'vgg13_bn', 'vgg16', 'vgg16_bn', 'vgg19', 'vgg19_bn', 'senet154', 'se_resnet50', 'se_resnet101', 'se_resnet152', 'se_resnext50_32x4d', 'se_resnext101_32x4d', 'densenet121', 'densenet169', 'densenet201', 'densenet161', 'inceptionresnetv2', 'inceptionv4', 'efficientnet-b0', 'efficientnet-b1', 'efficientnet-b2', 'efficientnet-b3', 'efficientnet-b4', 'efficientnet-b5', 'efficientnet-b6', 'efficientnet-b7', 'mobilenet_v2', 'xception', 'timm-efficientnet-b0', 'timm-efficientnet-b1', 'timm-efficientnet-b2', 'timm-efficientnet-b3', 'timm-efficientnet-b4', 'timm-efficientnet-b5', 'timm-efficientnet-b6', 'timm-efficientnet-b7', 'timm-efficientnet-b8', 'timm-efficientnet-l2', 'timm-tf_efficientnet_lite0', 'timm-tf_efficientnet_lite1', 'timm-tf_efficientnet_lite2', 'timm-tf_efficientnet_lite3', 'timm-tf_efficientnet_lite4', 'timm-resnest14d', 'timm-resnest26d', 'timm-resnest50d', 'timm-resnest101e', 'timm-resnest200e', 'timm-resnest269e', 'timm-resnest50d_4s2x40d', 'timm-resnest50d_1s4x24d', 'timm-res2net50_26w_4s', 'timm-res2net101_26w_4s', 'timm-res2net50_26w_6s', 'timm-res2net50_26w_8s', 'timm-res2net50_48w_2s', 'timm-res2net50_14w_8s', 'timm-res2next50', 'timm-regnetx_002', 'timm-regnetx_004', 'timm-regnetx_006', 'timm-regnetx_008', 'timm-regnetx_016', 'timm-regnetx_032', 'timm-regnetx_040', 'timm-regnetx_064', 'timm-regnetx_080', 'timm-regnetx_120', 'timm-regnetx_160', 'timm-regnetx_320', 'timm-regnety_002', 'timm-regnety_004', 'timm-regnety_006', 'timm-regnety_008', 'timm-regnety_016', 'timm-regnety_032', 'timm-regnety_040', 'timm-regnety_064', 'timm-regnety_080', 'timm-regnety_120', 'timm-regnety_160', 'timm-regnety_320', 'timm-skresnet18', 'timm-skresnet34', 'timm-skresnext50_32x4d', 'timm-mobilenetv3_large_075', 'timm-mobilenetv3_large_100', 'timm-mobilenetv3_large_minimal_100', 'timm-mobilenetv3_small_075', 'timm-mobilenetv3_small_100', 'timm-mobilenetv3_small_minimal_100', 'timm-gernet_s', 'timm-gernet_m', 'timm-gernet_l']"