Downloads · 30 days
33
7% of all-time downloads
learning_to_tokenize_in_ViT
keras-io/learning_to_tokenize_in_ViT
learning_to_tokenize_in_ViT is a image classification model from keras-io. Use it when you need a label for an image. It is set up for tf-keras.
The full credit goes to: Aritra Roy Gosthipaty, Sayak Paul
All-time downloads
480
Public
Repo size
10.1 MB
Likes
1
Public
Repo makeup
Click a slice to open those files.
.data-00000-of-000015.4 MB · 54%
At a glance
- Task
- Image Classification
- Library
- tf-keras
- Access
- Public
- Created
- Jul 10, 2022
- Updated
- Jul 5, 2024
- SHA
- 83bdaf31
Spaces
Model card
From the Hugging Face model README
Tensorflow Keras implementation of : Learning to tokenize in Vision Transformers
The full credit goes to: Aritra Roy Gosthipaty, Sayak Paul
Short description:
ViT and other Transformer based architectures break down images into patches. As we increase the resolution of the images, the number of patches increases as well. To tackle this, Ryoo et al. introduced a new module called TokenLearner which can help reduce the number of patches used. The full paper can be found here
Model and Dataset used
The Dataset used here is CIFAR-10. The model used here is a mini ViT model with the TokenLearner module.
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
| Hyperparameters | Value |
|---|---|
| name | AdamW |
| learning_rate | 0.0010000000474974513 |
| decay | 0.0 |
| beta_1 | 0.8999999761581421 |
| beta_2 | 0.9990000128746033 |
| epsilon | 1e-07 |
| amsgrad | False |
| weight_decay | 9.999999747378752e-05 |
| exclude_from_weight_decay | None |
| training_precision | float32 |
Training Metrics
After 20 Epocs, the test accuracy of the model is 55.9% and the Top 5 test accuracy is 95.06%
Model Plot
<details> <summary>View Model Plot</summary>
- Task
- Image Classification
- Library
- tf-keras
- Created
- Jul 10, 2022
- Updated
- Jul 5, 2024