EdgeCrafter

This page describes how to use EdgeCrafter ECViT models with LightlyTrain for pretraining and distillation.

EdgeCrafter ECViT models are compact Vision Transformers designed for efficient dense prediction. LightlyTrain supports ECViT backbones through the edgecrafter model package.

Note

EdgeCrafter is released under the Apache 2.0 license.

Pretrain an EdgeCrafter Model

EdgeCrafter ECViT backbones can be pretrained with LightlyTrain. For example, you can distill a compact ECViT backbone like this:

import lightly_train

if __name__ == "__main__":
    lightly_train.pretrain(
        out="out/my_experiment",
        data="my_data_dir",
        model="edgecrafter/ecvitt",
        method="distillation",
    )

See Distillation method for more details on pretraining and its configuration options.

Note

EdgeCrafter ECViT backbones currently support RGB images only. Multi-channel input is not supported.

Supported Models

The following ECViT backbone models are supported:

• edgecrafter/ecvitt

• edgecrafter/ecvittplus

• edgecrafter/ecvits

• edgecrafter/ecvitsplus

For examples on how to fine-tune ECViT backbones inside LT-DETR for object detection, see Object Detection.