GCViT: Global Context Vision Transformer

Tensorflow 2.0 Implementation of GCViT

This library implements GCViT using Tensorflow 2.0 specifally in tf.keras.Model manner to get PyTorch flavor.

Model

• Architecture:

• Local Vs Global Attention:

Result

Official codebase had some issue which has been fixed recently (27 July 2022). Here's the result of ported weights on ImageNetV2-Test data,

Installation

pip install -U gcvit
# or
# pip install -U git+https://github.com/awsaf49/gcvit-tf

Usage

Load model using following codes,

from gcvit import GCViTTiny
model = GCViTTiny(pretrain=True)

Simple code to check model's prediction,

from skimage.data import chelsea
img = tf.keras.applications.imagenet_utils.preprocess_input(chelsea(), mode='torch') # Chelsea the cat
img = tf.image.resize(img, (224, 224))[None,] # resize & create batch
pred = model(img).numpy()
print(tf.keras.applications.imagenet_utils.decode_predictions(pred)[0])

Prediction:

[('n02124075', 'Egyptian_cat', 0.9194835),
('n02123045', 'tabby', 0.009686623), 
('n02123159', 'tiger_cat', 0.0061576385),
('n02127052', 'lynx', 0.0011503297), 
('n02883205', 'bow_tie', 0.00042479983)]

For feature extraction:

model = GCViTTiny(pretrain=True)  # when pretrain=True, num_classes must be 1000
model.reset_classifier(num_classes=0, head_act=None)
feature = model(img)
print(feature.shape)

Feature:

(None, 512)

For feature map:

model = GCViTTiny(pretrain=True)  # when pretrain=True, num_classes must be 1000
feature = model.forward_features(img)
print(feature.shape)

Feature map:

(None, 7, 7, 512)

Live-Demo

• For live demo on Image Classification & Grad-CAM, with ImageNet weights, click powered by 🤗 Space and Gradio. here's an example,

Example

For working training example checkout these notebooks on Google Colab & Kaggle .

Here is grad-cam result after training on Flower Classification Dataset,

To Do

• New updated weights have been added.
• Working training example in Colab & Kaggle.
• GradCAM showcase.
• Gradio Demo.
• Build model with tf.keras.Model.
• Port weights from official repo.
• Support for TPU.

Acknowledgement

• GCVit (Official)
• Swin-Transformer-TF
• tfgcvit
• keras_cv_attention_models

Citation

@article{hatamizadeh2022global,
  title={Global Context Vision Transformers},
  author={Hatamizadeh, Ali and Yin, Hongxu and Kautz, Jan and Molchanov, Pavlo},
  journal={arXiv preprint arXiv:2206.09959},
  year={2022}
}