fastai-p1

Practical Deep Learning For Coders, Part 1 fast-ai site

HW Description

1. HW01: Image classification with CNN (Glass vs Plastic)
   • Data was downloaded using google image search + chrome downloader extension. To help chrome download plugins to download image easier, don't forget to do "Advance Search" and set the "usage rights" to "free to use or share".
   • Results: 96.7% (1 mistake)
2. HW02: Image multi-label classification kaggle plant-seedlings-classification
   • Data was downloaded using Kaggle CLI and added to './data/<comp_name>'
   • Best Results: 0.98236 (kaggle)
3. HW03: Melbourne Housing Price Prediction data source
   • Data was downloaded using Kaggle CLI and added to './data/melbourne-housing'
   • Best R2 result: ~80%
4. HW04a: Pytorch Tutorial link
5. HW04b: Basic Sentiment Analysis using Pytorch Text
   • Data was downloaded from link and added to './data/sentiment'
   • Best Result for IMDB: 0.9889
6. HW05: (Stochastic) Gradient Descent and ADAM Optimizer implementation
   • SGD was based from fastai notebook
   • ADAM was based from arxiv paper
7. HW06a: Char level RNN based from link
   • Tried producing text learned from the Psalms (Bible, KJV).
   • The result was not that well. The output keeps on cycling through the same words.
8. HW06b: Char level LSTM based from the fast-ai RNN notebook (lesson 6)
   • Tried producing text learned from the Psalms (Bible, KJV).
   • The results were a lot better, but there is still areas for improvement.
9. HW06c: Char level LSTM based from the fast-ai RNN notebook (lesson 6)
   • Tried producing text learned from the Psalms (Bible, KJV).
   • The difference with HW06b is that I used a shorter bptt (bptt = 10 instead of 100)
   • The results is surprisingly good compared to bptt = 100.

Setting Up Work Environment

1. Google instance with GPU link
2. Download kaggle data using Kaggle CLI link
3. Extract .7z with "7z e ", .tar with "tar -xvf ", .zip with "unzip