Scribd
Upload
7 views10 pages

Flower Classification Homework Guide

The document outlines a homework assignment for flower classification using a dataset of over 400 images across 9 classes, emphasizing the importance of not using original dataset labels. It suggests building a convolutional neural network and improving performance through various methods, including semi-supervised learning and pseudo-labeling. Submission requirements include a CSV file for predictions and an ipynb file for code, with a deadline of October 17 at 11:59 PM.

Uploaded by

hansviary011202
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
7 views10 pages

Flower Classification Homework Guide

The document outlines a homework assignment for flower classification using a dataset of over 400 images across 9 classes, emphasizing the importance of not using original dataset labels. It suggests building a convolutional neural network and improving performance through various methods, including semi-supervised learning and pseudo-labeling. Submission requirements include a CSV file for predictions and an ipynb file for code, with a deadline of October 17 at 11:59 PM.

Uploaded by

hansviary011202
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd

SDSC 4016

HOMEWORK 2
DESCRIPTION
TASK – FLOWER
CLASSIFICATION
TASK – FLOWER
CLASSIFICATION
• The images are collected from the national flower dataset classified into 9
classes.
• The dataset includes 400+ images for each category in training set. We
split the dataset into labeled training set, unlabeled training set, and
validation set.
• DO NOT utilize the original dataset or labels.
– This is cheating.
TASK – FLOWER
CLASSIFICATION logits probability

-0.25 0.004
+3.02 0.117
+0.56 softmax 0.010
-3.90 0.000
CNN -0.01 0.006
Model +0.25 0.007
-0.47 0.004
+5.00 0.850
-1.32 0.002
HINTS

• Build a convolutional neural network using labeled images with provided


codes.
• Improve the performance using labeled images with different model
architectures or data augmentations.
• Improve the performance with additional unlabeled images in the training
set, such as 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.
PSEUDO-LABELS FILTERING
logits probability

-0.25 0.004
+3.02 0.117
+0.56 softmax 0.010
-3.90 0.000
CNN -0.01 0.006
Model +0.25 0.007
-0.47 0.004
+5.00 0.850
-1.32 0.002

Pseudo-label=7

Confident enough?
KAGGLE SUBIMISSION
FORMAT
• The submitted predictions should be in CSV format.
• The first row is “ID, Category”
• The rest of rows are “{id}, {predictions}” (e.g., 0005, 8)

ID Category
0001 0
0002 7
0003 4
0004 5
SUBMISSIONS

• Submit: one CSV file (your prediction) to Kaggle, one ipynb file (your code, one submission per group) to
canvas.
• Kaggle: [Link]
• Please join through this link.
• You should edit your team name in Kaggle before submit: E.g., Team1
• You should rename your ipynb file with your group name when submitting on canvas: E.g., [Link]
• The ranking of the public and private learderboard (using 75% of the test data) on kaggle, and your codes
are considered for final scoring. The private leaderboard will be published after the HW Deadline.
• Please join the competition with your cityu default email account, for example, xxx-c@[Link]
• Please refer to the announcement on Canvas
• Deadline: Oct 17 11:59 PM
USEFUL RESOURCES

• Semi-supervised learning
– MixMatch: [Link]
– Noisy student: [Link]
• Pytorch
– [Link]
• Torchvision
– [Link]

You might also like