jcbrtl / deep_learning

Collection of my work related to deep learning in a single place.

Deep Learning

Collection of my work related to deep learning in a single place.

https://kushajveersingh.github.io/blog/

Repository Structure

1. graph_machine_learning

   1. what_can_neural_networks_reason_about - Implementation of paper What can Neural Networks Reason About. Check my blog post for the paper summary. Results from one of the experiments presented in the paper are reproduced. twitter-card

2. projects

   1. Semi-supervised parking lot detection -> My submission for Techgig Code Gladiators 2019 AI theme competition that won 1st place at the final.
   2. Waste Seggregation using trashnet -> Contains the code to train models for trashnet and then export them using ONNX. It was part of a bigger project where we ran these models on Rasberry Pi, which controlled wooden planks to classify the waste into different categories (code for rasberry pi not included here).

3. paper_implementations

   1. Photorealistic Style Transfer - Implementation of High Resolution Network for Photorealistic Style Transfer
   2. SPADE-PyTorch - Implementation of Semantic Image Synthesis with Spatially-Adaptive Normalization (SPADE)
   3. Weight Standardization - Implementation of Weight Standardization. Tested using cyclic learning.
   4. Training AlexNet with tips and checks on how to train CNNs - A PyTorch tutorial on how to create an image classifier.
   5. Study of Mish activation function in transfer learning with code and discussion - Implementation of Mish: A Self Regularized Non-Monotonic Neural Activation Function.
   6. Number of bins of a Histogram - Notebook discussing three techniques of choosing the bin size in histograms.
   7. Multi Sample Dropout - Implementation of Multi-Sample Dropout for Accelerated Training and Better Generalization.
   8. Data Augmentation in Computer Vision - Notebook implementing single image data augmentation techniques using just Python.
   9. How to deal with outlier - Notebook discussing ways to deal with outliers.
   10. Leslie N. Smith papers notebook - Jupyter notebook discussing cyclic learning and ways by which we can choose hyperparameter values by looking at valid loss graph.

4. notes - My notes when learning some things.

   • Cython
   • dask
   • matplotlib
   • PIL
   • seaborn
   • vim
   • xarray
   • Python - Data model + James Powell talks

5. random - Random scripts and os setup instructions.

   1. os_setup_instructions - My setup guide for Windows 10 and Ubuntu 20.04 (needs to be updated a bit).
   2. cifar10_data_script.py - To convert CIFAR10 data to numpy array
   3. save_torchvision_models_to_disk -> a quick script to download all pytorch models
   4. unscramble_android_game -> Python script to solve the unscramble android game. Complexity is exponential to generate all the substrings, and a dictionary is used to check for valid words.