transcv 0.0.5

A library for creating vision transformers for computer vision

transcv

transcv creates vision transformers for visual recognition which can be pre-trained using self-supervised learning

Acknowledgement

1. Ross Wightman's repository for vision transformer
2. Hugging Face
3. DETR repository
4. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
5. fastai
6. nbdev

Install

pip install transcv

Along with transcv, we also need fastai and nbdev. So, it is recommended to use :

pip install fastai nbdev transcv -q --upgrade

How to use

An example for using the VisRecTrans class for getting a custom ViT model :

from transcv.visrectrans import VisRecTrans

vis_rec_ob = VisRecTrans('vit_small_patch16_224', 10, False)
model = vis_rec_ob.create_model()
vis_rec_ob.initialize(model)
embed_callback = vis_rec_ob.get_callback()

Now, the model, along with the embed_callback, can be used with the Learner class, of fastai, and can be fine-tuned on any image classification dataset.

An example for using the SwinT class for building a Swin transformer model :

from transcv.swin import SwinT

swint_ob = SwinT('swin_base_patch4_window7_224', pretrained = False, num_classes = 10)
swin_model = swint_ob.get_model()
assert isinstance(swin_model, nn.Sequential)

Now, the swin_model can be used with the Learner class, of fastai, and can be fine-tuned on any dataset for visual recognition task.

For a detailed description of the classes and methods, please refer to the documentation.