Description
Instructions:
In this homework you will practice building a multi-class image classification pipeline to classify daily life
images (CIFAR10), based on the modern neural networks. Please keep the batch size untouched. The goals of
this homework are as follows:
• understand how to use Pytorch to build multi-class classifiers.
• understand the mechanism of convolution in image classification.
• learn the power of non-linearity in modern neural networks.
• implement and apply a fully-connected multi-class image classifier.
• implement and apply a Convolutional Neural Networks (CNN) classifier.
Problem Description.
In this homework, you are asked to implement fully-connected (MLP) and Convolutional Neural Networks
(CNN) image classifier on CIFAR10 dataset. In this task, you only need to perform multi-class classification.
Details of these models could be found in lecture 11 slides. We provide a skeleton code for data loading and
iterations of training data. You are asked to implement the rest of training in Pytorch code. Detailed submission
requirements are written in the final section.
Resources
You can follow the setup instructions at https://pytorch.org/get-started/locally/.
A useful tutorial on learning building CNN at https://pytorch.org/tutorials/beginner/blitz/cifar10_
tutorial.html.
Convolutional functions could be found here: https://pytorch.org/docs/stable/nn.html#convolution-functions.
Data
We use CIFAR10 classification dataset. Pytorch/torchvision has provide a useful dataloader to automatically
download and load the data into batches. We have written the data loader for you as follow. You can find it in
the attached file.
Submitting your assignment
Upload a PDF report using LATEX containing your experiment results to CCLE (remember to include your name
and ID number in the report). Some requirements for your report:
• Limit the number of pages to 3.
• You are asked to implement a 7 layers fully-connected neural networks with ReLu activation function.
The model accuracy should be around 50% percent.
• You are asked to implement a 7 layers convolutional neural networks, 4 convolutional layers and 3 fullyconnected layers, with ReLu activation function. The input dimension of 1st fully-connected layer must
be 4096. The model accuracy should be around 85% percent.
• Describe your 2 model structures including in channels, out channels, stride, kernel size, padding for CNN
layer; input dim, out dim for fully connected layer.
1
• For each of the model, report the (PB
b=1
PDb
d=1
loss(labelb,d,fb(datab,d)
Db
)/B for each training epoch, where B
is the total number of batches, fb is the model after updated by b-th batch and Db is the number of data
points in b-th batch. An epoch is defined as one iteration of all dataset. Essentially, during a training
epoch, you record down the average training loss of that batch after you update the model, and then
report the average of all such batch-averaged losses after one iteration of whole dataset. You could plot
the results as a figure our simply list down. Please at least report 10 epochs.
• Report the final testing accuracy of trained model.
• Please compare results for 2 models (MLP and CNN)).
• Try neural network without non-linear activation functions and discuss your findings.
Also, upload your code in a zip file and show how to run your code in README.
2
