UCL DRL Coursework 1


jupyter nbconvert --to latex ./LFPY4-CW-pythonNB.ipynb
xelatex ./LFPY4-CW-pythonNB.tex

LFPY4-CW-pythonNB.ipynb

2282.5KB

LFPY4-CW-pythonNB.pdf

2703.1KB

I've had a query regarding the meaning of the "root_dir" parameter for the Dataset class. The root_dir should be a string which, points to the directory containing the folders "train", "val" and "test" i.e., it should be the parent directory of the "train", "val" and "test" directories.

Additionally, I've had a query regarding altering code such which has not been explicitely labelled as "YOUR CODE HERE" e.g., altering the inputs to the Dataclass __init__ method. Provided you clearly commented why you have altered something, it is okay to alter code not labelled as "YOUR CODE HERE" however, if your code becomes illegible and the markers cannot understand what or why you have done something you may loose marks so please be very explicit about what you are doing.

Minimum requirements for the baseline model/over tuning the baseline

In the question, there is a space to explain the design choices of your baseline model. I encourage you to explicitely link the statements you make to the guidance provided in the question about what makes a good baseline model;

Pdf for coursework submission

The pdf required for your coursework submission should be created by converting the notebook that you also submit to a pdf. This can be done by, for example, file->print->as pdf in Google Collab. Please save the pdf having run your entire notebook end to end so that the pdf contains the cell outputs. This pdf is used as a back up when marking your work incase the notebook file becomes corrupted.

WandB outputs as evidence

For questions asking for empiricial evidence from results, it is fine to screen shot wandb outputs and load them into the notebook!

Baselinee

I have had a number of questions regarding what defines a "baseline" and how to know if your model is "good enough" with respect to the coursework. I thought I would write a clarification for everyone. Both the concept of a "baseline" and a model being "good enough" are fuzzy and this coursework was explicitely designed to expose this side of model development. I encourage everyone to look at lecture 2 where we discussed both the concept of a baseline and model evaluations.

Criteria


1. Data Analysis: 5 marks
2. Data Preparation: 5 marks
3. Training a Baseline: 30 marks
4. Improving the Baseline: 50 marks
5. Evaluating on the Test Set: 10 marks

Pretrained models are __NOT__ allowed

We recommend using weights and biases to log your training runs

Your code and results and discussions should be concise, well-presented, and easy to read. Each question has a certain portion of marks going towards this.

Ensure you correctly use the train, validation, and test set throughout. You should only ever use the test set once - for the final evaluation.

Consider saving your models so you can reload previous models for the final evaluation

Ensure it is clear to the reader what any plots / figures are presenting. I.e., label axes, include titles, ensure it is clear what experiment it is from (what model / design choices, etc.)