PASSL implements MoCoBYOL, a way of connecting MoCo and BYOL for self-supervised learning with a multiple loss.
- See INSTALL.md
epochs | official results | passl results | Backbone | Model | |
---|---|---|---|---|---|
MoCo-BYOL | 300 | 71.56 | 72.10 | ResNet-50 | download |
python tools/train.py -c configs/moco_byol/moco_byol_r50_IM.yaml
python -m paddle.distributed.launch --gpus="0,1,2,3,4,5,6,7" tools/train.py -c configs/moco_byol/moco_byol_r50_IM.yaml
Pretraining models with 300 epochs can be found at mocobyol_r50_ep300_ckpt.pdparams
Note: The default learning rate in config files is for 8 GPUs. If using differnt number GPUs, the total batch size will change in proportion, you have to scale the learning rate following new_lr = old_lr * new_ngpus / old_ngpus
.
python tools/extract_weight.py mocobyol_r50_ep300_ckpt.pdparams --output ${WEIGHT_FILE}
- Support PaddleClas
Convert the format of the extracted weights to the corresponding format of paddleclas to facilitate training on paddleclas
python tools/passl2ppclas/convert.py --type res50 --checkpoint ${CHECKPOINT} --output ${WEIGHT_FILE}
Note: It must be ensured that the weights are extracted
python -m paddle.distributed.launch --gpus="0,1,2,3,4,5,6,7" tools/train.py -c configs/moco_byol/moco_byol_50_IM_clas.yaml --pretrained ${WEIGHT_FILE}
python -m paddle.distributed.launch --gpus="0,1,2,3,4,5,6,7" tools/train.py -c configs/moco_byol/moco_byol_r50_IM_clas.yaml --load ${CLS_WEGHT_FILE} --evaluate-only
The trained linear weights in conjuction with the backbone weights can be found at MoCoBYOL