EVA 4.0 (Extensive Vision AI)
This contains the solutions to the assignments given in The School of AI - Extensive Vision AI 4.0
Website: https://satyajitghana.github.io/TSAI-DeepVision-EVA4.0/
[i'd prefer to use the website link]
-
This describes the very basics of Python, i would recommend binge watching Raymond Hettinger’s YouTube Videos, as much as you can, also learn list comprehension, slicing, partial, functools, functional programming, classes, MRO, decorators, lambdas, python 3 typing
Also here i learnt the basics of how a Neural Network learns, how the different channels are formed, what are kernels, how does the DNN make sense of the input it gets, and more !
Have you ever wondered, like why do we only use 3x3 kernels ? i had always wondered, why not 5x5 ? 7x7 ? in this session i learnt why, and how does even the kernel size matter ? why is it only odd numbers ? and why only squares ? and why is 3x3 kernel used twice same as a 5x5 kernel, what is receptive field ?
-
This describes the basic neural network architecure, see its page for details
-
Basic Pytorch architecture for working with neural networks, introduces you to
nn.Module, optimizers, forward and backward pass, datasets, how to apply simple augmentation. -
Here we had to train MNIST to get 99.4% accuracy with some given contraints, was quite fun to do this, got me out of the noob shell of what a neural network actually is and what it does, because i had to write code manually, not copy paste any more.
-
This was very important to realise the basic steps required to make a neural network and then go on to optimize it, to get the perfect model size and accuracy, very important, please see its documentation, and all the notebooks
-
This assigment focuses on the importance of normalization and regularization in neural networks, aim was to also get > 99.4% accuracy in less than 40 epochs for MNIST with limited model parameters
-
Here i was introduced to various convolution types you can add in the network, we were given a custom network which we had to implement, also we had to make a custom library of python to support and ease the process of NN building and training, this was quite a eye openener for me, i made my first python package !, also i learnt the value of time contraints, since we have to finish the assignments within a week, along with my college work it was burdening, or maybe i was procrastinating?
-
Here i was introduced to ResNet18 and CIFAR dataset with aim to get 85% test accuracy
-
This was amazing, i was introduced to GradCAM, and i myself learnt about saliency map, this was something i also used in my college group project, it’s used to diagnosing if the model is actually learning anything from the dataset. Quite a nifty feature, eh ?
-
Here i had to use LR Finder to find the best learning rate for the model, i also used OneCyclePolicy for faster convergence. i reached 92.03% accuracy in CIFAR using ResNet18
-
A custom architecture was implemented here, along with that i used OneCycleLR to improve the convergence of the NN
-
Here we had to do image annotation and collect doggo dataset.
Also i had to make a custom dataloader and dataset class in PyTorch to support TinyImageNet, and then train a model to reach 60% accuracy, which i failed to, but still i learnt from my mistakes ! the one who doesnt make mistakes does really make anything does he ?
-
I was introduced to YOLO, something i always wondered how it exactly worked, and why is it called YOLO ? now i understand why. and people complain about the FPS of YOLO, now i know why, they dont pay attention to the anchor boxes, which is very important and its different for different datasets.
I made a custom Bugs Bunny detector using the YOLO architecture :) i like my detector :) i like bugs bunny :) i custom collected bugs bunny images, self annotated about 600 images of my bunny and trained the network for like 5 hours on a Tesla P100
-
I was introduced to RCNN family and SSD, but that wasnt the main thing here.
The main thing was that i created a dataset containing 1.2M images ! do you understand how crazy that is ? this was a pre-req for the upcoming CapStone project which will be to do object segmentation and also monocular depth estimation both at the same time.
-
This was the final project, please see its README, i spent a lot of time writing code and documenting it.
I’m sure you’ll be impressed by this one ;)