COMP9417 Assignment Topic 0

Forrest Koch (z3463797)

The source code and results for this project can be obtained from the MAPnet github repository. Data is not available for download, however, a live-demo can be provided upon request.

• development_log.md: This file served as a diary to occasionally log progress throughout the development process. It begins with a brief description of the "Framework" as well as the expected folder structure of datasets. These topics may be of value to the reader for understanding the project.
• README.md: this README file
• report.Rmd: R-markdown file used to write the report
• report.pdf: the pdf version of the report
• references.bib: BibTex reference file
• requirements.txt: the module requirements for this project (can install with pip)
• mapnet/: This folder contains the source code files for the bulk of the project
  • train.py: This is the main run script implementing the training loop and evaluation proceedures. Run with python3 train.py -h to see the options.
  • data.py: This module implements the Dataset class and other helper functions used for data manipulation.
  • model.py: This module implements the torch.nn.Module defining the network architecture to be trained.
  • defaults.py: This file contains the default settings for a few global variables.
• models/: This folder contains information about the various models trained during the experiments. After the filter tests (section 3.5), it was no longer possible to upload the *.dat files to the github repo. Any missing data models can be supplied upon request, however, it measures around 88G in size and could not practically be uploade.
  • Data for each model is contained within it's own timestamped folder (YYYY-MM-DD_HH-MM-SS). Inside each folder will be:
    • arguments.txt: This file details the arguments supplied to the program call
    • loss.csv: This file contains information about the training performance. It is in the form [epoch],[lr],[train err],[test_err]
    • *.dat: These files are saved MAPnet modules. They can be loaded with the torch.load function.
  • lr_trials/: contains the models trained for section 3.2
  • layer_trials/: contains the models trained for section 3.3
  • decay_test/: contains the models train for section 3.4
  • filters_test/: contains the models trained for section 3.5
  • optims/: contains the models trained for section 3.6
  • final_model/: contains the final model used for section 3.7
• results/: contains the spreadsheets used for analysis in the report
  • csvs/lr_comparison: used for section 3.2
  • csvs/layer_comparison: used for section 3.3
  • decay_results.csv: used for section 3.4
  • filters_test.csv: used for section 3.5
  • weight_decay_results.csv: used for section 3.6
  • validation_ages.csv: used for section 3.7