The report is in the root directory, named 10719_Project_Report.pdf
Please change the name in the first line of requirements.yml (otherwise I believe this will change your default env) and then run:
conda env create -f requirements.yml
to create an Anaconda environment for you to run my code. I have to apologize for not being familiar with how to describe my environment more briefly. In fact, I think installing:
- torch==2.0.1 with CUDA (w/o is also OK, but will be slow)
- numpy ==1.26.1
- pandas==1.5.3
- tqdm==4.65.0
- torchsummary==1.5.1 (this is not necessary but can help you visualize the model)
would be enough.
There are four experiments in the report (1-4), corresponding to Figures 2-5, respectively. To run experiment i, please:
cd bats
experiment[i].bat # run the bat
cd ..
python plotter[i].py
This wil give you plots in the plots/experiment[i]/ folder.
Note I had just re-organized the directory and some coding files, if there are any bugs, please let me know and allow me a bit time to fix it. Sorry for that.
The pretrained models are in Google Drive. If you want to pretrain them by yourself, you could use the following code:
# make sure you are in root directory
python pretrain.py --data_name fmnist --goal [accuracy you want] --max_epochs [maximum epochs] --model MLP --mlp_hidden 512 128
This will generate model in models/ folder.
You could also tune the hyper-parameters in each file as you wish. Thank you very much!
Last but not least, many thanks to Jong-Ik Park. His code helped me a lot in working on this project.