downstream_mmdet

About code isolation

This downstream_mmdet is isolated from pre-training codes. One can treat this downstream_mmdet as an independent codebase 🛠️.

Fine-tuned ConvNeXt-B weights, log files, and performance

[weights (pre-trained by SparK)] [weights (fine-tuned on COCO)] [log.json] [log]

Installation MMDetection with commit 6a979e2 before fine-tuning ConvNeXt on COCO

We refer to the codebases of ConvNeXt and Swin-Transformer-Object-Detection. Please refer to README.md for installation and dataset preparation instructions.

Note the COCO dataset folder should be at downstream_mmdet/data/coco. The folder should follow the directory structure requried by MMDetection, which should look like this:

downstream_mmdet/data/coco:
    annotations/:
        captions_train2017.json  captions_val2017.json
        instances_train2017.json  instances_val2017.json
        person_keypoints_train2017.json  person_keypoints_val2017.json
    train2017/:
        a_lot_images.jpg
    val2017/:
        a_lot_images.jpg

Training

To train a detector with pre-trained models, run:

# single-gpu training
python tools/train.py <CONFIG_FILE> --cfg-options model.pretrained=<PRETRAIN_MODEL> [other optional arguments]

# multi-gpu training
tools/dist_train.sh <CONFIG_FILE> <GPU_NUM> --cfg-options model.pretrained=<PRETRAIN_MODEL> [other optional arguments] 

For example, to train a Mask R-CNN model with a SparK pretrained ConvNeXt-B backbone and 4 gpus, run:

tools/dist_train.sh configs/convnext_spark/mask_rcnn_convnext_base_patch4_window7_mstrain_480-800_adamw_3x_coco_in1k.py 4 \
  --cfg-options model.pretrained=/some/path/to/official_convnext_base_1kpretrained.pth

The Mask R-CNN 3x fine-tuning config file can be found at configs/convnext_spark. This config is basically a copy of https://github.com/facebookresearch/ConvNeXt/blob/main/object_detection/configs/convnext/mask_rcnn_convnext_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco_in1k.py.

Inference

# single-gpu testing
python tools/test.py <CONFIG_FILE> <DET_CHECKPOINT_FILE> --eval bbox segm

# multi-gpu testing
tools/dist_test.sh <CONFIG_FILE> <DET_CHECKPOINT_FILE> <GPU_NUM> --eval bbox segm

Acknowledgment

We appreciate these useful codebases:

• MMDetection
• ConvNeXt
• Swin-Transformer-Object-Detection