iKrishneel / detectron2_timm

A simple wrapper library for binding timm models as detectron2 backbones

detectron2_timm

A Simple wrapper for binding the models in timm library into detectron2 backbone for training two-stage detectors using detection2. This simple library, does not require and changes to either timm models or detectron2. Whichever model is found in timm.models.list_models() is automatically binded to the detectron2 backbone (including FPN).

Examples

The following models are already tested.

• xcit → check the xcit.ipynb which shows example of using this library with the pretrained xcit models for evaluation on coco2017 dataset.
• swin → check the swin.ipynb which shows example of using this library with the pretrained swin models for evaluation on coco2017 dataset.

Configs

The configuration of the binded backbone model from timm is specified in the extend config. In the config you specify laters from which to extract the features, rename this feature extractor layers and also which layers to remove from the timm model.

The extended config has the following attributes:

cfg.MODEL.BACKBONE.NAME = "build_xcit_small_12_p8_224_fpn_backbone"       # name of the model
cfg.MODEL.BACKBONE.FREEZE_AT = 0                                          # freeze at which layer

# backbone config
cfg.MODEL.BACKBONE.CONFIG.REMOVE_LAYERS = []                              # layers to remove from the timm model
cfg.MODEL.BACKBONE.CONFIG.OUT_FEATURES = []                               # layers in timm model from which to extract the features
cfg.MODEL.BACKBONE.CONFIG.REMAPS = []                                     # name of the output features, the order must be same as out_features
cfg.MODEL.BACKBONE.CONFIG.STRIDES = []                                    # strides of each output features, the order must be same out_features
cfg.MODEL.BACKBONE.CONFIG.PRETRAINED = False                              # init with pretrained model   

# input
cfg.INPUT.FIXED_INPUT_SIZE = True                                         # model has fix size input, eg. for transformers

Model Name Scheme

The binding backbone model uses similar naming scheme as in detectron2. Basically build_ and _backbone are appendend before and after the timm model name. Example for the model xcit_small_12_p8_224 in timm, the binding backbone without FPN name is

build_xcit_small_12_p8_224_backbone

and the model with FPN has _fpn appended before _backbone

build_xcit_small_12_p8_224_fpn_backbone

Using the Backbone

The binded backbone are added the global BACKBONE_REGISTRY dict of detectron2. To use any of the model, just import the backbone

# example
from detectron2_timm import build_xcit_small_12_p8_224_fpn_backbone

Dependencies

The library is tested with the following dependencies

einops == 0.3.0
torch == 1.8.0
torchvision == 0.9.0
timm == 0.4.13   # for xcit models

Train and Evaluation

The training and evaluation are same as that in detectron2 except that you should use the train_net.py which will have all the binding models.

Train

To train the xcit model

$ python tools/train_net.py --config-file config/xcit/xcit_small_12_p8_224_fpn.yaml --num-gpus 4

Test

To test the xcit model

$ python tools/train_net.py --config-file config/xcit/xcit_small_12_p8_224_fpn.yaml --num-gpus 4 --eval-only MODEL.WEIGHTS ./logs/mrcnn_xcit_small_12_p8.pth

⚠️ Warning

Only few models in timm (see example) are tested so far. Other models might require some changes or fixes.