An Implementation of Transformer in Transformer in TensorFlow for image classification, attention inside local patches

Last update: Jul 25, 2022

Overview

Transformer-in-Transformer

An Implementation of the Transformer in Transformer paper by Han et al. for image classification, attention inside local patches. Transformer in Transformer uses pixel level attention paired with patch level attention for image classification, in TensorFlow.

PyTorch Implementation

Installation

Run the following to install:

pip install tnt-tensorflow

Developing tnt-tensorflow

To install tnt-tensorflow, along with tools you need to develop and test, run the following in your virtualenv:

git clone https://github.com/Rishit-dagli/Transformer-in-Transformer.git
# or clone your own fork

cd tnt
pip install -e .[dev]

Usage

import tensorflow as tf
from tnt import TNT

tnt = TNT(
    image_size=256,  # size of image
    patch_dim=512,  # dimension of patch token
    pixel_dim=24,  # dimension of pixel token
    patch_size=16,  # patch size
    pixel_size=4,  # pixel size
    depth=5,  # depth
    num_classes=1000,  # output number of classes
    attn_dropout=0.1,  # attention dropout
    ff_dropout=0.1,  # feedforward dropout
)

img = tf.random.uniform(shape=[5, 3, 256, 256])
logits = tnt(img) # (5, 1000)

Want to Contribute 🙋‍♂️ ?

Awesome! If you want to contribute to this project, you're always welcome! See Contributing Guidelines. You can also take a look at open issues for getting more information about current or upcoming tasks.

Want to discuss? 💬

Have any questions, doubts or want to present your opinions, views? You're always welcome. You can start discussions.

Citation

@misc{han2021transformer,
      title={Transformer in Transformer}, 
      author={Kai Han and An Xiao and Enhua Wu and Jianyuan Guo and Chunjing Xu and Yunhe Wang},
      year={2021},
      eprint={2103.00112},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

License

Copyright 2020 Rishit Dagli

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.

Comments

Add Unit Tests
The tests should check for the rank and shape of the output tensors, the test should override tf.test.TestCase base class.

[x] #15

[x] #16

[x] #18

[x] #17

Feel free to take inspiration from:

https://github.com/Rishit-dagli/Fast-Transformer/blob/main/fast_transformer/test_fast_transformer.py

For parametrization feel free to follow https://stackoverflow.com/a/34094/11878567, can be used in the exact same way with subTest in TensorFlow

enhancement good first issue
opened by Rishit-dagli 3
Update Workflows to run tests
This issue follows #11

Update GitHub Workflows to:

[ ] Run Tests before uploading to PyPI

[ ] Create a workflow to run tests on commits

Feel free to take inspiration from https://github.com/Rishit-dagli/Fast-Transformer/tree/main/.github/workflows
enhancement good first issue
opened by Rishit-dagli 0

Creates an Attention layer

Verify output shapes just from the attention layer:

import tensorflow as tf
Attention(dim=256)(tf.random.normal([3,256,256]))

# <tf.Tensor: shape=(3, 256, 256), dtype=float32,

Closes #3

opened by Rishit-dagli 0

Put together a TNT class

Verify shapes:

tnt = TNT(
    image_size=256,  # size of image
    patch_dim=512,  # dimension of patch token
    pixel_dim=24,  # dimension of pixel token
    patch_size=16,  # patch size
    pixel_size=4,  # pixel size
    depth=5,  # depth
    num_classes=1000,  # output number of classes
    attn_dropout=0.1,  # attention dropout
    ff_dropout=0.1,  # feedforward dropout
)

img = tf.random.uniform(shape=[1, 3, 256, 256])
print(tnt(img).shape)

# (1, 1000)
```

opened by Rishit-dagli 0

Create an Attention layerr

Verify output shapes just from the attention layer:

import tensorflow as tf
Attention(dim=256)(tf.random.normal([3,256,256]))

# <tf.Tensor: shape=(3, 256, 256), dtype=float32,

opened by Rishit-dagli 0

Create a PreNorm layer

Verify output shapes from this layer:

import tensorflow as tf
PreNorm(dim=1, fn=tf.keras.layers.Dense(5))(tf.random.normal([10, 1]))

# <tf.Tensor: shape=(10, 1), dtype=float32,

opened by Rishit-dagli 0

Releases(v0.2.0)

v0.2.0(Feb 2, 2022)
This is an interesting release for the project, including a pre-trained model on ImageNet, reproducibility of paper results, tests, and end-to-end training.

✅ Bug Fixes / Improvements

Create an end-to-end training example demonstrating how to train a TNT model for image classification through a custom training loop on the TF Flowers dataset (#14)

Pre-trained model to reproduce the paper results have been made available (in this release as well as on TensorFlow Hub)

Create an off-the-shelf inference example, that highlights how you can directly use the pre-trained model made available

Unit Tests for the Attention class (#19)

Unit Tests for the main TNT class (#20)

Full Changelog: https://github.com/Rishit-dagli/Transformer-in-Transformer/compare/v0.1.0...v0.2.0
Source code(tar.gz)
Source code(zip)
tnt_s_patch16_224.tar.gz(84.42 MB)
v0.1.0(Dec 3, 2021)
This is the initial release of TNT TensorFlow and implements Transformers in Transformers as a subclassed TensorFlow model.

Classes

Attention: Implements attention as a TensorFlow Keras Layer making some modifications.

PreNorm: Normalize the activations of the previous layer for each given example in a batch independently and apply some function to it, implemented as a TensorFlow Keras Layer.

FeedForward: Create a FeedForward neural net with two Dense layers and GELU activation, implemented as a TensorFlow Keras Layer.

TNT: Implements the Transformers in Transformers model using all the other classes, and converts to logits. Implemented as a TensorFlow Keras Model.

Source code(tar.gz)
Source code(zip)
tnt_s_patch16_224.tar.gz(84.42 MB)

Owner

Rishit Dagli

High School,TEDx,2xTED-Ed speaker | International Speaker | Microsoft Student Ambassador | Mentor, @TFUGMumbai | Organize @KotlinMumbai

GitHub Repository https://pypi.org/project/tnt-tensorflow

On the Complementarity between Pre-Training and Back-Translation for Neural Machine Translation (Findings of EMNLP 2021))

PTvsBT On the Complementarity between Pre-Training and Back-Translation for Neural Machine Translation (Findings of EMNLP 2021) Citation Please cite a

10 Nov 25, 2022

An Implementation of Transformer in Transformer in TensorFlow for image classification, attention inside local patches

Related tags

Overview

Transformer-in-Transformer

Installation

Developing tnt-tensorflow

Usage

Want to Contribute 🙋‍♂️ ?

Want to discuss? 💬

Citation

License

Comments

Add Unit Tests

Update Workflows to run tests

Creates an Attention layer

Put together a TNT class

Create an Attention layerr

Create a PreNorm layer

Releases(v0.2.0)

v0.2.0(Feb 2, 2022)

✅ Bug Fixes / Improvements

v0.1.0(Dec 3, 2021)

Classes

Owner

Rishit Dagli

On the Complementarity between Pre-Training and Back-Translation for Neural Machine Translation (Findings of EMNLP 2021))

[SIGMETRICS 2022] One Proxy Device Is Enough for Hardware-Aware Neural Architecture Search

This repo contains the pytorch implementation for Dynamic Concept Learner (accepted by ICLR 2021).

Optimizers-visualized - Visualization of different optimizers on local minimas and saddle points.

Distributionally robust neural networks for group shifts

This is the official PyTorch implementation for "Mesa: A Memory-saving Training Framework for Transformers".

Deep Multi-Magnification Network for multi-class tissue segmentation of whole slide images

Pytorch ImageNet1k Loader with Bounding Boxes.

Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising

TensorFlow implementation of "Variational Inference with Normalizing Flows"

Implementation of paper "DCS-Net: Deep Complex Subtractive Neural Network for Monaural Speech Enhancement"

Welcome to The Eigensolver Quantum School, a quantum computing crash course designed by students for students.

🌎 The Modern Declarative Data Flow Framework for the AI Empowered Generation.

A Python wrapper for Google Tesseract

Using multidimensional LSTM neural networks to create a forecast for Bitcoin price

Text Extraction Formulation + Feedback Loop for state-of-the-art WSD (EMNLP 2021)

TensorFlow ROCm port

Code for our paper "Interactive Analysis of CNN Robustness"

GLODISMO: Gradient-Based Learning of Discrete Structured Measurement Operators for Signal Recovery

Udacity's CS101: Intro to Computer Science - Building a Search Engine