ECE371 Neural Networks and Deep Learning
Assignment 1: Image classification by using deep models
Due Date: 23:59, 14
th May, 2025
This assignment aims to train models for flower classification. You can choose either Colab online
environment or local environment. This assignment will worth 15% ofthe final grade. Exercise 1: Fine-tune classification model using MMClassification (50%)
Please complete the fine-tune training based on the pre-training model provided by MMClassification
(https://github.com/open-mmlab/mmpretrain/tree/1.x). You should:
1. Prepare the flower datasets. The flower pictures are provided in flower_dataset.zip. The flower dataset contains flowers from 5 categories: daisy 588, dandelion 556, rose 583, sunflower 536 and tulip 585. Please split the dataset into training set and validation set in a ratio
of 8:2, and organize it into ImageNet format. Detailed steps:
1) Put the training set and validation set under folders named ‘train’ and ‘val’. 2) Create and edit the category name file. Please write all names flower categories into file
‘classes.txt’with each line representing one class. 3) Generate training (optional) and validation sets annotation lists: ‘train.txt’and ‘val.txt’. Each line should contain a filename and its corresponding annotation. Example:
daisy/NAME**.jpg 0
daisy/NAME**.jpg 0
... dandelion/NAME**.jpg 1
dandelion/NAME**.jpg 1
... rose/NAME**.jpg 2
rose/NAME**.jpg 2
... sunflower/NAME**.jpg 3
sunflower/NAME**.jpg 3
... tulip/NAME**.jpg 4
tulip/NAME**.jpg 4
The final file structure should be:
flower_dataset
|--- classes.txt
|--- train.txt
|--- val.txt
| |--- train
| | |--- daisy
|
|
|--- …
--- dandelion
|--- NAME1.jpg
|--- NAME2.jpg
|--- …
--- rose
|--- NAME1.jpg
|--- NAME2.jpg
|--- …
--- sunflower
|--- NAME1.jpg
|--- NAME2.jpg
|--- …
--- tulip
|--- NAME1.jpg
|--- NAME2.jpg
|--- …
val --- daisy
|--- NAME1.jpg
|--- NAME2.jpg
|--- …
--- dandelion
|--- NAME1.jpg
|--- NAME2.jpg
|--- …
--- rose
|--- NAME1.jpg
|--- NAME2.jpg
|--- …
--- sunflower
|--- NAME1.jpg
|--- NAME2.jpg
|--- …
--- tulip
|--- NAME1.jpg
|--- NAME2.jpg
|--- …
This process can be done using Python or other scripting programs. And it can be completed
locally/offline to save the Colab’s time online. Once the dataset has been prepared, please migrate the processed dataset to the project folder, (e.g., ./data). To reduce duplicate uploads, you can Sync the data to google drive
|--- NAME1.jpg
|--- NAME2.jpg
and import it in Colab. 2. Modify the configuration file
Use the _base_ inheritance mechanism to build profiles for fine-tuning, which can be inherited
and modified from any ImageNet-based profile provided by MMClassification. 1) Modify the model configuration. Change the category header to adapt the model to the
number of data categories in our flower dataset. 2) Modify the dataset configuration. Change the data paths for the training set, validation set, the list of dataset annotations, and the category name file. And modify the evaluation method
to use only the top-1 classification error rate. 3) Modify learning rate strategy. Fine-tuning generally uses a smaller learning rate and fewer
training period. Therefore please change them in configuration file. 4) Configuring pre-trained models. Please find the model file corresponding to the original
configuration file from Model Zoo. Then download it to Colab or your local environment
(usually in the checkpointsfolder). Finally you need to configure the path to the pre-trained
model in the configuration file. 3. Complete the finetune training using tools. Please use tools/train.py to fine-tune the model and specify the work path via the work_dir
parameter, where the trained model will be stored. Tune the parameters, or use a different pre-trained model to try to get a higher classification
accuracy. For reference, it is not difficult to achieve classification accuracies above 90% on this
dataset. Exercise 2: Complete the classification model training script (50%)
The provided script main.py is a simple PyTorch implement to classify the flower dataset you’ve
prepared above, but this script is not complete. 1. You’ll be expected to write some code in some code blocks. These are marked at the top of the
block by a #GRADED FUNCTIONcomment, and you’ll write your code in between the ###
START SOLUTION HERE ### and ###END SOLUTION HERE### comments. 2. After coding your function, put your flower datasets flower_dataset to the EX2 folder (EX2/
flower_dataset) and then run this main.py script. 3. If your code is correct, you can obtain the right printed information with loss, learning rate and
accuracy on validation set, and the best model with the highest validation accuracy will be stored
in the Ex2/work_dirfolder. 4. You can modify the configuration or the model in main.pyto beat the original result. (optional)
5. Please write a report with Latex and submit a .pdf file (the main text should not exceed 4
pages, excluding references). Please use this overleaf template https://www.overleaf.com/read/vjsjkdcwttqp#ffc59a . There are detailed report requirements.
Submission requirements:
1. You need to submission all materials to GitHubClassroom. Please create a GitHub account in
advance. . Later we will provide a link of this assignment, click it and you
will get an initial repository containing two folders named: Ex1 with flower_dataset.zipin it, and
Ex2 with main.pyin it. You need to upload all the materials below to your repository:
1) For exercise 1, please put your configuration file and the saved trained model in Ex1;
2) For exercise 2, please put your report, completed script file and the saved trained model
(auto saved in work_dir) in Ex2. 2. Please note that, the teaching assistants may ask you to explain the meaning of the program, to
ensure that the codes are indeed written by yourself. Plagiarism will not be tolerated. We may
check your code. 3. The deadline is 23:59 PM, 14
th May. For each day of late submission, you will lose 10% of your
mark in corresponding assignment. If you submit more than three days later than the deadline, you
will receive zero in this assignment. No late submission emails or message will be replied.
請(qǐng)加QQ：99515681 郵箱：99515681@qq.com WX：codinghelp
