代写ECE4179 - Lab 3 Multi-Layer Perceptrons代写数据结构语言程序

日期：2024-08-28 06:36

ECE4179 - Lab 3

Multi-Layer Perceptrons (MLP)

This lab is about understanding and applying PyTorch-Lightning to simple multi-layer percep- tron tasks. By the end of the lab, you will have developed simple MLPs for multi class classification tasks. The following are the four tasks that you should complete.

•  Task 1: Approximate the sine function using a MLP by following through the framework for implementing a deep learning model - from creating custom dataset to designing the shallow MLP model, and lastly to train and evaluate model performance.

•  Task 2: Applying MLP for classification task. The model used here will remain shallow, and you will be following through the same framework of applying the MLP model to solve the multi-class classification problem.

•  Task 3: Improving model performance from task 2’s MLP model by adding additional hidden layers.

The learning outcomes for this lab are:

•  Familiarising yourself with PyTorch and PyTorch-Lightning

•  Understanding implementation of MLPs and network training.

•  Analysing and describing accuracy and loss plots.

•  Understanding the implications of different MLP results.

Note: Most of the lab does not require you to use the Numpy library.  You will be using Pytorch/Pytorch-Lightning in-built methods to create your tensors instead of Numpy. Follow the Pytorch/Pytorch-Lightning videos for more information.

Introduction.

The fundamentals of deep learning models start off with understanding multilayer perceptrons and the training processes required. There are additional hyper-parameters that we can choose and tune, such as the type of optimizers, learning rate, activation functions, and model architecture to name a few. The ability of activation functions to model non-linearities is what allows MLPs to be able to generalise better than traditional machine learning models.

The submission deadline is 30th of August (Friday) 4:30 PM AEST.

It is recommended that you go through the following videos/documents prior to attending your lab 3:

1.  Begin by reading through this document. This document contains all the relevant information for lab 3

2.  Follow the lessons on PyTorch/PyTorch-Lightning

3.  Watch the introductory video for this lab

4.  You may choose to use GoogleColab for this lab so that you can utilise the free GPU provided.

In the lab and in your own time, you will be completing the lab 3 notebook by going through this document and the provided notebooks.

Task 1: Approximate the Sine Function [40%]

This section include show to develop and evaluate a small neural network for function approxima- tion. You will approximate the sine function with a MLP. The following tasks need to be completed:

•  1.1 Data: Create custom dataset and dataloaders

•  1.2 Model: Design a Shallow MLP model

•  1.3 Train & Evaluate: Train and evaluate model’s performance

•  1.4 Visualise and Analyse the Experimental Results

The first task in this section aims to setup a dataloader class for the noisy sine data.  Task 1.2 requires you to design a shallow linear MLP model with the following structure/hyper-parameters:

Table 1: Hyper parameters for section 1

After constructing the MLP model, you can train the model using the Trainer constructor as it has useful built-in methods such as logging results, model checkpointing, training and validation loop, early-stopping.

Ensure you answer the discussion questions at the end of the task as well.

Task 2: MLP for classification of the Covertype dataset [30%]

In this task, you will use your knowledge learned from Task 1 to build another shallow MLP to perform a classification task on a forest cover type dataset.  This dataset classifies the cover type into 7 classes by considering 54 different features. More descriptions can be found in the notebook. The tasks need to be completed are:

•  2.1 Data: Load the CoverType dataset from sci-kit learn

•  2.2 Model: Design the Shallow MLP model

•  2.3 Train & Evaluate: Train and evaluate model’s performance in different secarios

•  2.4 Visualise the Results

•  2.5 Check the performance in the test dataset

For this Task 2.1, the dataset should be loaded from the existing database that is managed by sci-kit learn. Most instructions are given in the notebook, and please make sure you understand the code.  Then in Task 2.2, you will design a shallow MLP which is similar to but slightly different from the one in Task 1. The general structure of the network is shown below:

After the training is done, you will again do some analysis to the results.   And answer the discussion questions in the end.

Task 3: Deep MLP for classification of the Covertype dataset [30%]

Now in this task, you will implement a Deep MLP which has more hidden layers to again do classification on the CoverType dataset. The tasks should be completed are:

•  3.1 Model: Design a Deep MLP model

•  3.2 Train & Evaluate: Train and evaluate model’s performance

•  3.3 Visualise the Results

Since the data is already loaded, you can start from design the model, the structure of the model should be similar to the one below:

The training process would be similar to previous tasks, and you will be asked to complete some analysis task.

And again, do not forget to complete the discussion questions at the end of task.

Hopefully this lab has given you insight in training a simple model via the PyTorch-Lightning framework, and provided realistic scenarios in which optimizing hyper-parameters and considering data augmentation is required for any deep learning problem!