Skip to content

Latest commit

 

History

History
110 lines (77 loc) · 3.04 KB

README.md

File metadata and controls

110 lines (77 loc) · 3.04 KB

Quick Start

Linear Evaluation

Method 1: Extract features（recommend）

This method is highly efficient.

It first extracts data features with pretrained models, then performs linear eval on the features.

This means there is no data augments and all feature layers are frozen.

Step1: Extract data features with pretrained model.

sh benchmarks/tools/dist_extract.sh \
        ${EXTRACT_DATASET_CONFIG} \
        ${NUM_GPUS} \
        ${FEATURE_DIR} \
        --checkpoint ${CHECKPOINT}
Arguments
  • EXTRACT_DATASET_CONFIG:the config path of extract data features, refer to extract_dataset_configs.
  • NUM_GPUS: number of gpus
  • FEATURE_DIR:your path to save output features
  • CHECKPOINT: the export checkpoint file of a selfsup model named as epoch_*_export.pt.

Examples:

Export model please reference to ssl.md, or you can test with our default export model [swav_restnet50](http://pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com/EasyCV/modelzoo/selfsup/swav_r50/epoch_200_export.pt -O ./pretrained_models/swav_rn50/epoch_200_export.pt).

sh benchmarks/tools/dist_extract.sh \
        benchmarks/extract_dataset_configs/imagenet.py \
        8 \
        ./linear_eval/imagenet_features \
        --checkpoint ./pretrained_models/swav_rn50/epoch_200_export.pt

Step2: Do linear evaluation with features.

sh tools/dist_train.sh \
        ${FEATURE_LINEAR_EVAL_CONFIG_PATH} \
        ${NUM_GPUS} \
        --work_dir ${WORK_DIR}
Arguments

  • FEATURE_LINEAR_EVAL_CONFIG_PATH:the config path of linear eval with features.

    Reference to benchmarks/selfsup/classification/imagenet,edit feature path to your local or oss path.

  • NUM_GPUS: number of gpus

  • WORK_DIR:your path to save models and logs

Examples:

Edit data_root in ${FEATURE_LINEAR_EVAL_CONFIG_PATH} to your own ${FEATURE_DIR} path.

sh tools/dist_train.sh \
        benchmarks/selfsup/imagenet/swav_r50_feature.py \
        8 \
        --work_dir ./linear_eval

Method 2: Finetune with fc layer

sh tools/dist_train.sh \
        ${LINEAR_EVAL_CONFIG_PATH} \
        ${NUM_GPUS} \
        --work_dir ${WORK_DIR}
        --load_from ${LOAD_FROM}
Arguments
  • LINEAR_EVAL_CONFIG_PATH: the config path of linear eval
  • NUM_GPUS: number of gpus
  • WORK_DIR: your path to save models and logs
  • LOAD_FROM: the pretrained checkpoint file of a selfsup model named as epoch_*.pth.

Examples:

Edit the data_root in the ${LINEAR_EVAL_CONFIG_PATH} to your own data path.

${LOAD_FROM} can use your own pretrained model, or you can test with our default pretrained model above.

sh tools/dist_train.sh \
        benchmarks/selfsup/classification/imagenet/resnet50_8xb32_100e_finetune.py \
        8 \
        --work_dir ./linear_eval
        --load_from ./pretrained_models/swav_rn50/epoch_200.pth