Skip to content

04_assignment.md

Latest commit

 

History

History
147 lines (93 loc) · 4.52 KB

04_assignment.md

File metadata and controls

147 lines (93 loc) · 4.52 KB

Assignment 4: CIFAR 10 and ML5.js

For this assignment, you'll need to accomplish these three items:

  • Run the CIFAR 10 project and save the results
  • Watch MIT's CNN: Computer Vision
  • Implement one of the projects using ML5.js
    • You can choose any of the ML5.js projects
      These are in increasing order of difficulty:
      • Applying Style Transfer
      • Mustachio or Mexican Sombrero?
      • Image classifier
      • Training Style Transfer

1. CIFAR 10

The CIFAR 10 data set is located at:
https://www.cs.toronto.edu/~kriz/cifar.html

In class, we have seen cifar 10 in action with convnet.js
https://cs.stanford.edu/people/karpathy/convnetjs/demo/cifar10.html

Follow the steps for training the CIFAR 10
https://pytorch.org/tutorials/beginner/blitz/cifar10_tutorial.html

Deliverables: "04_cifar.txt" It should be placed in your github, folder /hw4

Note: If you have Windows, you can still run the Jupyter notebook using Google Colab
https://colab.research.google.com/github/pytorch/tutorials/blob/gh-pages/_downloads/cifar10_tutorial.ipynb

A) Copy/paste all the output from running this tutorial

It should contains lines like the following (but many more lines)

[1,  2000] loss: 2.201
[1,  4000] loss: 1.875
... (more lines)

Accuracy of the network on the 10000 test images:

Accuracy of plane : 57 %
Accuracy of   car : 61 %
Accuracy of  bird : 26 %
... (more lines)

B) Re-train the model for a longer time. This can be achieved by increasing the number of epochs, say from range(2) to rnge(10) and find out how much improvement we get for the model and the class predictions.

for epoch in range(10):  # loop over the dataset multiple times

Copy/paste all the iteratons with the predicted loss

It will contains lines similar to: [3, 2000] loss: 1.039 [3, 4000] loss: 1.066 [3, 6000] loss: 1.049 ... (more lines)

Accuracy of the network on the 10000 test images: 60 %

Accuracy of plane : 68 %
Accuracy of   car : 65 %
Accuracy of  bird : 44 %
... (more lines)

C) After you execute all the python code using ipython, save your ipython session

%save my-cifar-pytorch.py 1-20 # this will save the first 20 lines, maybe you need more lines?

2. Watch MIT's CNN: Computer Vision

https://www.youtube.com/watch?v=H-HVZJ7kGI0&index=1&list=PLtBw6njQRU-rwp5__7C0oIVt26ZgjG9NI

Questions

  • What's a Convolution?
  • What's ReLU?
  • What's Upsampling?

3. ML5.js Projects

A. Implementing Style Transfer

What's Style Transfer?
See: https://github.com/NVIDIA/FastPhotoStyle

This is an example for Fast Style Transfer with source code
https://ml5js.org/docs/style-transfer-image-example

Deliverables

Feature Enhancements (optional)

  • Support for image file upload (need a web server app)
  • Support for drag n drop of images
  • Support custom styles by using a custom model derived from the project D) Training Style Transfer

B. Mustachio or Mexican Sombrero?

Modify the Posenet example, so instead of 7 red dots, place a sombrero on top of someone's head. Or place a mustache or beard on the face (below the eyes and nose) Or be creative and try a baseball cap or some Dracula teeth https://ml5js.org/docs/posenet-webcam

C. Image Classifier

Follow the steps to create your image classifier using ml5.js
https://ml5js.org/docs/training-introduction

Youtube: The Coding Train

D. Training Style Transfer

https://ml5js.org/docs/training-styletransfer

It requires running on a GPU system (like paperspace.com) And process 15 GB of image data from the COCO dataset. http://cocodataset.org/#home

The specific COCO data set gets download when running setup.sh in a bash shell.
https://github.com/ml5js/training-styletransfer/blob/master/setup.sh

For Windows, you can install a Bash shell.
https://www.windowscentral.com/how-install-bash-shell-command-line-windows-10

Using a Titan X, it took 4 to 6 hours for training. On a Mac, it can take several months running with a CPU.

If interested, consider using paperspace.com, pricing is per hour.
https://blog.paperspace.com/creating-your-own-style-transfer-mirror/