resnest50 -> 75.58%
数据增强方法:
transforms.Resize((self.BASE_RESIZE_SIZE, self.BASE_RESIZE_SIZE), Image.BILINEAR),
transforms.RandomCrop(self.INPUT_SIZE),
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(degrees=15),
transforms.ColorJitter(brightness=self.BRIGHTNESS, contrast=self.CONTRAST, hue=self.HUE, saturation=self.SATURATION),
transforms.ToTensor(),
self.imagenet_normalization_paramters
学习率:
0.1初始化，不进行warmup 90个epoch，0.3, 0.6, 0.9 * epoch 按ratio=0.1进行线性衰减
| model | accuracy |
|---|---|
| resne50(baseline)256->224 | 76.47% |
| resnet50(baseline)512->448 | 84.55% |
| efnet-b0 256->224 | 77.93% |
| efnet-b1 272->240 | 79.63% |
| efnet-b2 292->260 | 81.22% |
| efnet-b3 352->300 | 83.12% |
| efnet-b4 416->380 | 84.45% |
| efnet-b5 512->456 | 85.64% |
| efnet-b6 600->528 | 87.05% |
| efnet-b7 640->600 | 88.11% |
bash train.sh
其中，horovod可以指定几个node，对应的就是几块GPU
CUDA_VISIBLE_DEVICES=0 python test.py
| model | test acc |
|---|---|
| efnetb2 260 | 41.975 |
| resnet50 224 | 36.11 |
| efnetb2+resnet50_224 ensemble1 | 43.105 |
| efnetb3 300 | 42.395 |
| resnet50 448 | 41.63 |
| efnetb3+resnet50_448 ensemble2 | 47.05 |
| ensemble1+ensemble2 | 47.395 |
- 20201016 update
| model | test acc | trick |
|---|---|---|
| efnetb5 456 | 51.04 | cosine_lr+cutmix+40epoch+no_warmup+32bs+0.1lr |
| r50 448 | 47.64 | cosine_lr+cutmix+40epoch+no_warmup+48bs+0.1lr |
| ensemble r50_efnetb5 | 52.408 | same |
- 20201018 update
| model | test acc | trick |
|---|---|---|
| regnet12gf | 53.578 | cosine_lr+cutmix+40epoch+no_warmup+48bs+0.1lr |
| regnet12gf + efnetb5 | 55.758 | same |
- 20201026 update
| model | test acc | trick |
|---|---|---|
| resnest200 | 56.145 | cosine_lr+cutmix+60epoch+no_warmup+24bs+0.1lr+ls |
| resnest200 + regnet12gf | 58.35% | same |
| resnest200 + regnet12gf + efnetb5 | 58.558% | same |
- 20201111 update
| model | test acc | trick |
|---|---|---|
| resnest101 + resnest200 + resnest 269 | 60.135 | data_clean_v3.log + freeze_fc + focalloss + 40 epoch + cutmix |
| resnest200 data_v1 + resnest200 data_v2 + resnest200 data_v3 | 60.808 | freeze_fc + focalloss + 40 epoch + cutmix |
| ensemble | 61.548 | mean |
- /data/remote/yy_git_code/cub_baseline/dataset/train_accv_clean_v2.log 使用56.145模型进行数据清洗
- /data/remote/yy_git_code/cub_baseline/dataset/train_accv_clean_v3.log 对重复标签的数据进行清洗
- /data/remote/yy_git_code/cub_baseline/dataset/train_accv_clean_v3_remove_small.log remove小于48尺寸的图片
- /data/remote/yy_git_code/cub_baseline/dataset/train_accv_clean_v3_with_repeat_plabel.log 使用resnest系列的3个模型对有二义性的数据生成伪标签，填充到cleanv3里面进行训练。