Self-supervised learning algorithms provide a way to train Deep Neural Networks in an unsupervised way using contrastive losses

Overview

Self-supervised learning

Paper Conference

CI testing

Self-supervised learning algorithms provide a way to train Deep Neural Networks in an unsupervised way using contrastive losses. The idea is to learn a representation which can discriminate between negative examples and be as close as possible to augmentations and transformations of itself. In this approach, we first train a ResNet on the unlabeled dataset which is then fine-tuned on a relatively small labeled one. This approach drastically reduces the amount of labeled data required, a big problem in applying deep learning in the real world. Surprisingly, this approach actually leads to increase in robustness as well as raw performance, when compared to fully supervised counterparts, even with the same architecture.

In case, the user wants to skip the pre-training part, the pre-trained weights can be downloaded from here to use for fine-tuning tasks and directly skip to the second part of the tutorial which is using the 'ssl_finetune_train.py'.

Steps to run the tutorial

1.) Download the two datasets TCIA-Covid19 & BTCV (More detail about them in the Data section)
2.) Modify the paths for data_root, json_path & logdir in ssl_script_train.py
3.) Run the 'ssl_script_train.py'
4.) Modify the paths for data_root, json_path, pre-trained_weights_path from 2.) and logdir_path in 'ssl_finetuning_train.py'
5.) Run the 'ssl_finetuning_script.py'
6.) And that's all folks, use the model to your needs

1.Data

Pre-training Dataset: The TCIA Covid-19 dataset was used for generating the pre-trained weights. The dataset contains a total of 771 3D CT Volumes. The volumes were split into training and validation sets of 600 and 171 3D volumes correspondingly. The data is available for download at this link. If this dataset is being used in your work, please use [1] as reference. A json file is provided which contains the training and validation splits that were used for the training. The json file can be found in the json_files directory of the self-supervised training tutorial.

Fine-tuning Dataset: The dataset from Beyond the Cranial Vault Challenge (BTCV) 2015 hosted at MICCAI, was used as a fully supervised fine-tuning task on the pre-trained weights. The dataset consists of 30 3D Volumes with annotated labels of up to 13 different organs [2]. There are 3 json files provided in the json_files directory for the dataset. They correspond to having different number of training volumes ranging from 6, 12 and 24. All 3 json files have the same validation split.

References:

1.) Harmon, Stephanie A., et al. "Artificial intelligence for the detection of COVID-19 pneumonia on chest CT using multinational datasets." Nature communications 11.1 (2020): 1-7.

2.) Tang, Yucheng, et al. "High-resolution 3D abdominal segmentation with random patch network fusion." Medical Image Analysis 69 (2021): 101894.

2. Network Architectures

For pre-training a modified version of ViT [1] has been used, it can be referred here from MONAI. The original ViT was modified by attachment of two 3D Convolutional Transpose Layers to achieve a similar reconstruction size as that of the input image. The ViT is the backbone for the UNETR [2] network architecture which was used for the fine-tuning fully supervised tasks.

The pre-trained backbone of ViT weights were loaded to UNETR and the decoder head still relies on random initialization for adaptability of the new downstream task. This flexibility also allows the user to adapt the ViT backbone to their own custom created network architectures as well.

References:

1.) Dosovitskiy, Alexey, et al. "An image is worth 16x16 words: Transformers for image recognition at scale." arXiv preprint arXiv:2010.11929 (2020).

2.) Hatamizadeh, Ali, et al. "Unetr: Transformers for 3d medical image segmentation." arXiv preprint arXiv:2103.10504 (2021).

3. Self-supervised Tasks

The pre-training pipeline has two aspects to it (Refer figure shown below). First, it uses augmentation (top row) to mutate the data and second, it utilizes regularized contrastive loss [3] to learn feature representations of the unlabeled data. The multiple augmentations are applied on a randomly selected 3D foreground patch from a 3D volume. Two augmented views of the same 3D patch are generated for the contrastive loss as it functions by drawing the two augmented views closer to each other if the views are generated from the same patch, if not then it tries to maximize the disagreement. The CL offers this functionality on a mini-batch.

image

The augmentations mutate the 3D patch in various ways, the primary task of the network is to reconstruct the original image. The different augmentations used are classical techniques such as in-painting [1], out-painting [1] and noise augmentation to the image by local pixel shuffling [2]. The secondary task of the network is to simultaneously reconstruct the two augmented views as similar to each other as possible via regularized contrastive loss [3] as its objective is to maximize the agreement. The term regularized has been used here because contrastive loss is adjusted by the reconstruction loss as a dynamic weight itself.

The below example image depicts the usage of the augmentation pipeline where two augmented views are drawn of the same 3D patch:

image

Multiple axial slices of a 96x96x96 patch are shown before the augmentation (Ref Original Patch in the above figure). Augmented View 1 & 2 are different augmentations generated via the transforms on the same cubic patch. The objective of the SSL network is to reconstruct the original top row image from the first view. The contrastive loss is driven by maximizing agreement of the reconstruction based on input of the two augmented views. matshow3d from monai.visualize was used for creating this figure, a tutorial for using can be found here

References:

1.) Pathak, Deepak, et al. "Context encoders: Feature learning by inpainting." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.

2.) Chen, Liang, et al. "Self-supervised learning for medical image analysis using image context restoration." Medical image analysis 58 (2019): 101539.

3.) Chen, Ting, et al. "A simple framework for contrastive learning of visual representations." International conference on machine learning. PMLR, 2020.

4. Experiment Hyper-parameters

Training Hyper-Parameters for SSL:
Epochs: 300
Validation Frequency: 2
Learning Rate: 1e-4
Batch size: 4 3D Volumes (Total of 8 as 2 samples were drawn per 3D Volume)
Loss Function: L1 Contrastive Loss Temperature: 0.005

Training Hyper-parameters for Fine-tuning BTCV task (All settings have been kept consistent with prior UNETR 3D Segmentation tutorial):
Number of Steps: 30000
Validation Frequency: 100 steps
Batch Size: 1 3D Volume (4 samples are drawn per 3D volume)
Learning Rate: 1e-4
Loss Function: DiceCELoss

4. Training & Validation Curves for pre-training SSL

image

L1 error reported for training and validation when performing the SSL training. Please note contrastive loss is not L1.

5. Results of the Fine-tuning vs Random Initialization on BTCV

Training Volumes Validation Volumes Random Init Dice score Pre-trained Dice Score Relative Performance Improvement
6 6 63.07 70.09 ~11.13%
12 6 76.06 79.55 ~4.58%
24 6 78.91 82.30 ~4.29%

Citation

@article{Arijit Das,
  title={Self-supervised learning for medical data},
  author={Arijit Das},
  journal={https://github.com/das-projects/selfsupervised-learning},
  year={2020}
}
Owner
Arijit Das
Data Scientist who is passionate about developing and implementing robust and explainable Machine Learning algorithms.
Arijit Das
PyTorch inference for "Progressive Growing of GANs" with CelebA snapshot

Progressive Growing of GANs inference in PyTorch with CelebA training snapshot Description This is an inference sample written in PyTorch of the origi

320 Nov 21, 2022
An optimization and data collection toolbox for convenient and fast prototyping of computationally expensive models.

An optimization and data collection toolbox for convenient and fast prototyping of computationally expensive models. Hyperactive: is very easy to lear

Simon Blanke 422 Jan 04, 2023
RM Operation can equivalently convert ResNet to VGG, which is better for pruning; and can help RepVGG perform better when the depth is large.

RMNet: Equivalently Removing Residual Connection from Networks This repository is the official implementation of "RMNet: Equivalently Removing Residua

184 Jan 04, 2023
Implementation of association rules mining algorithms (Apriori|FPGrowth) using python.

Association Rules Mining Using Python Implementation of association rules mining algorithms (Apriori|FPGrowth) using python. As a part of hw1 code in

Pre 2 Nov 10, 2021
Lowest memory consumption and second shortest runtime in NTIRE 2022 challenge on Efficient Super-Resolution

FMEN Lowest memory consumption and second shortest runtime in NTIRE 2022 on Efficient Super-Resolution. Our paper: Fast and Memory-Efficient Network T

33 Dec 01, 2022
Unofficial implementation of HiFi-GAN+ from the paper "Bandwidth Extension is All You Need" by Su, et al.

HiFi-GAN+ This project is an unoffical implementation of the HiFi-GAN+ model for audio bandwidth extension, from the paper Bandwidth Extension is All

Brent M. Spell 134 Dec 30, 2022
Multi Camera Calibration

Multi Camera Calibration 'modules/camera_calibration/app/camera_calibration.cpp' is for calculating extrinsic parameter of each individual cameras. 'm

7 Dec 01, 2022
🧑‍🔬 verify your TEAL program by experiment and observation

Graviton - Testing TEAL with Dry Runs Tutorial Local Installation The following instructions assume that you have make available in your local environ

Algorand 18 Jan 03, 2023
Visual Memorability for Robotic Interestingness via Unsupervised Online Learning (ECCV 2020 Oral and TRO)

Visual Interestingness Refer to the project description for more details. This code based on the following paper. Chen Wang, Yuheng Qiu, Wenshan Wang,

Chen Wang 36 Sep 08, 2022
A Neural Net Training Interface on TensorFlow, with focus on speed + flexibility

Tensorpack is a neural network training interface based on TensorFlow. Features: It's Yet Another TF high-level API, with speed, and flexibility built

Tensorpack 6.2k Jan 01, 2023
Code for the Paper "Diffusion Models for Handwriting Generation"

Code for the Paper "Diffusion Models for Handwriting Generation"

62 Dec 21, 2022
PyTorch code for JEREX: Joint Entity-Level Relation Extractor

JEREX: "Joint Entity-Level Relation Extractor" PyTorch code for JEREX: "Joint Entity-Level Relation Extractor". For a description of the model and exp

LAVIS - NLP Working Group 50 Dec 01, 2022
Code for ICCV 2021 paper: ARAPReg: An As-Rigid-As Possible Regularization Loss for Learning Deformable Shape Generators..

ARAPReg Code for ICCV 2021 paper: ARAPReg: An As-Rigid-As Possible Regularization Loss for Learning Deformable Shape Generators.. Installation The cod

Bo Sun 132 Nov 28, 2022
基于YoloX目标检测+DeepSort算法实现多目标追踪Baseline

项目简介: 使用YOLOX+Deepsort实现车辆行人追踪和计数,代码封装成一个Detector类,更容易嵌入到自己的项目中。 代码地址(欢迎star): https://github.com/Sharpiless/yolox-deepsort/ 最终效果: 运行demo: python demo

114 Dec 30, 2022
Official implementation of DreamerPro: Reconstruction-Free Model-Based Reinforcement Learning with Prototypical Representations in TensorFlow 2

DreamerPro Official implementation of DreamerPro: Reconstruction-Free Model-Based Reinforcement Learning with Prototypical Representations in TensorFl

22 Nov 01, 2022
A Python module for the generation and training of an entry-level feedforward neural network.

ff-neural-network A Python module for the generation and training of an entry-level feedforward neural network. This repository serves as a repurposin

Riadh 2 Jan 31, 2022
Discovering Explanatory Sentences in Legal Case Decisions Using Pre-trained Language Models.

Statutory Interpretation Data Set This repository contains the data set created for the following research papers: Savelka, Jaromir, and Kevin D. Ashl

17 Dec 23, 2022
Catbird is an open source paraphrase generation toolkit based on PyTorch.

Catbird is an open source paraphrase generation toolkit based on PyTorch. Quick Start Requirements and Installation The project is based on PyTorch 1.

Afonso Salgado de Sousa 5 Dec 15, 2022
Pytorch implementation of few-shot semantic image synthesis

Few-shot Semantic Image Synthesis Using StyleGAN Prior Our method can synthesize photorealistic images from dense or sparse semantic annotations using

40 Sep 26, 2022
Generative Exploration and Exploitation - This is an improved version of GENE.

GENE This is an improved version of GENE. In the original version, the states are generated from the decoder of VAE. We have to check whether the gere

33 Mar 23, 2022