Objective
Solve image classification with convolutional neural networks.
Improve the performance with data augmentations.
Understand how to utilize unlabeled data and how it benefits.
The images are collected from the food-11 dataset classified into 11 classes.
The dataset here is slightly modified:
Training set: 280 * 11 labeled images + 6786 unlabeled images
Validation set: 60 * 11 labeled images
Testing set: 3347 images
DO NOT utilize the original dataset or labels.
○ This is cheating.
Kaggle link: here
Requirements
This homework is in three levels:Easy
○ Medium
○ Hard
You can easily finish the easy level by running the example code.
For the rest, we recommend you start with the same code.We already prepared some TODO blocks for you.
DO NOT pre-train your model on other datasets.
If you use some well-known model architecture (e.g., ResNet), make sure NOT to load pre-trained weights as initialization.
Requirements - Easy
Build a convolutional neural network using labeled images with provided codes.
Public simple baseline: 44.862 (accuracy, %)
Requirements - Medium
Improve the performance using labeled images with different model architectures or data augmentations.
Public medium baseline: 52.807 (accuracy, %)
You can achieve the baseline by adding a few lines to the example code.
Requirements - Hard
Improve the performance with additional unlabeled images.
Public strong baseline: 82.138 (accuracy, %)
Do it on your own (by finishing TODO blocks in the example code).
Using unlabeled testing data here is allowed.
Hint: semi-supervised learning, self-supervised learning
Semi-supervised Learning
There are many ways to do semi-supervised learning.
g., generate pseudo-labels for unlabeled data and train with them.
Pseudo-labels
Useful Resources
Semi-supervised learninghttps://speech.ee.ntu.edu.tw/~tlkagk/courses/ML_2016/Lecture/semi%20(v3).pdf
○ https://www.youtube.com/watch?v=fX_guE7JNnY&ab_channel=Hung-yiLee
○ MixMatch: https://arxiv.org/abs/1905.02249 ○ Noisy student: https://arxiv.org/abs/1911.04252
PyTorchhttps://pytorch.org/
Torchvisionhttp://pytorch.org/vision/stable/index.html
