learn_PyTorch

🤔Introduction

This is a project from https://github.com/mrdbourke.

I wanted to learn all about PyTorch, thus I completed this course in 3 weeks which allowed me to learn all sorts of things

💡What did I learn

In tutorial 3:

• We imported data from MNIST and used matplotlib to print out the MNIST data.
  
• Implemented TinyVGG used to test the MNIST data
  
• Separated data into training and testing
  
• Built training and testing loops to run a couple epochs using a CPU and GPU and tried to see which one is faster
  
• Learnt about suitable loss and optimizer functions
  
• Implemented a Confusion Matrix
  
• Extracted data from a Fahion MNIST dataset and tested it on a TinyVGG and used it to see how accurate the model is to determine the type of clothes there is
  

In tutorial 4:

• Learnt how to convert images into numPy arrays and also learnt how to resize images and convert it into tensors
• Ran train and test loops with different number of epochs and saw what were the differences in terms of accuracy and time taken
• Learnt about what num_workers mean
• Learnt about what HWC and CHW means

In Tutorial 5:

• Learnt how to write imprtant functions into separate files for future use
  
• get_data.py was used to unzip the data files from the source
• data_setup.py was used to separate data into test and train
• engine.py was used for training and testing data
• model_builder.py is used to show the TinyVGG implementation
• utils.py is used to save the model

In Tutorial 6:

• Learnt how to resize the training images, normalise it and convert it into tensors to be used by a pre-trained model
• Learnt why we would need to freeze some sections of the model
• Learnt how to find the most wrong datasets and showed it onto a table from Pandas
• Plot graphs to see how close the train and test loss are close to each other and how close the train and test accuracy are

In Tutorial 7:

• Learnt how to use Summary Writer, to save runs of a model
• Tested ml accuracy architecture of efficientnet_b2 and efficientnet_v2_s
• learnt how to use data augmentaton to achieve a larger variety of datasets to prevent overfiiting (which will be talked about later lol)

In Tutorial 8:

• Learnt how to convert images into multiple patches
• Learn how to make a ViT architecture utiizing patch embedding, class tokens, positional embedding, Encoder layers and Mlp head
• Utilized a pretrained model called vit_b_16

In Tutorial 9:

Revision of everything learnt above

🚧Challenges faced

Going trough this course and doing all the exercise in the notebook made me realised how hard learning ML is really like. Technically speaking, I went trough this whole course twice in order to get proper understanding on what's going on, and doing the exercises wasn't easy as well, it took me a really long time in order to understand the important algos that I saved into helper functions. But the more you practice, the better you going to be :)