Deep Vision

Computer-Vision related research papers implemented using PyTorch, from scratch.

• /arch/* contains various CNN architectures
• /semantic_segmentation/* contains CNN models related to the task of semantic segmentation
• /gan/* contains models for generative adversarial networks

The following list of papers have been currently implemented:

Architectures

Model	Year	Paper
LeNet-5	1998	Gradient-Based Learning Applied to Document Recognition
AlexNet	2012	ImageNet Classification with Deep Convolutional Neural Networks
ZFNet	2013	Visualizing and Understanding Convolutional Networks
VGG-16	2014	Very Deep Convolutional Networks for Large-Scale Image Recognition
GoogLeNet	2014	Going Deeper with Convolutions
ResNet	2015	Deep Residual Learning for Image Recognition
Inception-v2	2015	Rethinking the Inception Architecture for Computer Vision
Xception	2016	Xception: Deep Learning with Depthwise Separable Convolutions
MobileNet	2017	MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications

Semantic Segmentation

Model	Year	Paper
FCN	2014	Fully Convolutional Networks for Semantic Segmentation
SegNet	2015	SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation
U-Net	2015	U-Net: Convolutional Networks for Biomedical Image Segmentation
PSPNet	2015	Pyramid Scene Parsing Network
ENet	2016	ENet: A Deep Neural Network Architecture for Real-Time Semantic Segmentation
ICNet	2017	ICNet for Real-Time Semantic Segmentation on High-Resolution Images

Generative Adversarial Networks (GANs)

Model	Year	Paper
GAN	2014	Generative Adversarial Networks
DCGAN	2015	Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks
Pix2Pix	2016	Image-to-Image Translation with Conditional Adversarial Networks
WGAN	2017	Wasserstein GAN

Training / Configuring the Models

The only dependencies required are

• torch (v1.9.0 or similar)
• numpy (v1.20.1 or similar)
• tqdm

NOTE #1: Many of the models use external datasets, some of which have not been pushed to this repository due to the sheer amount of size. Cross-check /[model_class]/[model]/dataset/ directory and see the *.py file for information about the dataset that needs to be used. Alternatively, you can also use any relevant dataset of your choice, but it's upto you to write the dataset class and change the required parameters (more on it below).

Follow the steps to train the model:

• Change to the relevant model directory: $ cd /[model_class]/[model]/
• Change/Download the dataset, if necessary and place the contents in dataset/ directory.
• Update main.py and/or train.py (necessary if using a different dataset)
• Execute main.py: $ python3 main.py

Disclaimer

Some of the implementations differ from the corresponding paper's mentioned configurations (especially the older papers) by a small amount. Wherever such differences are present, it is mentioned in the comments.