[SOLVED] MachineLearning Homework 3-Image Classification

Objective

1. Solve image classification with convolutional neural networks.
2. Improve the performance with data augmentations.
3. Understand how to utilize unlabeled data and how it benefits.

Task – Food Classification

Task – Food Classification

• ●  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.

Task – Food Classification

logits
(not normalized)

-0.25 +3.02

+0.56 softmax -3.90

-0.01 +0.25 -0.47 +5.00 -1.32 +1.14 -0.28

probability (normalized)

CNN model

0.004 0.114 0.009 0.001 0.005 0.007 0.003 0.831 0.001 0.017 0.004

Requirements

● This homework is in three levels: ○ Easy

○ Medium

Kaggle link: here

○ 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.
• ●  E.g., generate pseudo-labels for unlabeled data and train with them.

training

labeling

filtering

labeled data

CNN model

unlabeled data

subset of data

combining

Pseudo-labels

labeled data

training labeling

CNN unlabeled model data

subset of data

combining

logits probability (not normalized) (normalized)

pseudo-label = 7

CNN model

-0.25 +3.02

+0.56 softmax -3.90

-0.01 +0.25 -0.47 +5.00 -1.32 +1.14 -0.28

0.004 0.114 0.009 0.001 0.005 0.007 0.003 0.831 0.001 0.017 0.004

confident enough?

filtering

yes

Kaggle Submission Format

• ●  The submitted predictions should be in CSV format.
• ●  The first row is “Id, Category”
• ●  The rest of rows are “{id}, {prediction}” (e.g., 0005, 8)
• ●  There should be (3347 + 1) rows in total.

Grading Policy

• ●  Public simple baseline: +1pt
• ●  Public medium baseline: +1pt
• ●  Public strong baseline: +1pt
• ●  Private simple baseline: +1pt
• ●  Private medium baseline: +1pt
• ●  Private strong baseline: +1pt
• ●  Submit your code: +4pt

Code Submission

● Submit your code via NTU COOL. <student_id>_hw3.zip

● DO

• ○  Specify the source of your code. (You may refer to Academic Ethics Guidelines)
• ○  Organize your code and make it easy to read (not necessary).

• ●  DO NOT
  • ○  Submit empty or garbage files.
  • ○  Submit the dataset or model.
  • ○  Compress your codes into other formats like .rar or .7z and simply rename it to .zip.
• ●  If we find you cheating or your code problematic, you will be punished.

○ Course final score * 0.9 for the first time, or fail the course otherwise.

Bonus

• ●  If you successfully get 10 pts:
  • ○  Your code will be made public to students.
  • ○  You can submit a report in PDF format briefly describing what you

    have done (in English, less than 100 words) for extra 0.5 pts.

  • ○  Reports will also be made public to students.
• ●  Report template