Interpreting Vision Models: A Comparative Study of MLPs and CNNs

Overview

This repository contains code and resources related to our course project for Deep Learning, offered in the Fall Semester 2023 at ETH Zurich. The pre-trained Convolutional Neural Network (CNN) models utilized in this project were obtained from the PyTorch_CIFAR10 GitHub repository. Additionally, the pre-trained Multilayer Perceptron (MLP) model can be found in the scaling_mlps GitHub repository.

The weights for the CNN models (VGG-13bn, ResNet50, and DenseNet169) are available for download from this Google Drive Link. Please save these weights under src/CNN_models/state_dicts/.

Environment Setup

To replicate our experiments, follow these steps to create and activate a virtual environment:

For Windows:

python -m venv ./interpret_MLP
interpret_MLP\Scripts\activate

For MacOS and Linux:

python -m venv ./interpret_MLP
source interpret_MLP/bin/activate

Install the required packages using the following command:

pip install -r requirements.txt

Note: We run our experiments on Python version 3.9.18 (as included in requirements.txt). However, we recommend users to use the correct python executable to initiate the virtual environment

Activation Maximization

In the src/Act_Max folder, you will find all the files necessary to reproduce the "Activation Maximization" results corresponding to Section 3.1 of our project.

Activation Maximization:

Adjust the working directory at the top of each file to point to the /src folder. Explore Act_Max_VGG_13.ipynb and Act_Max_B_12_W_1024.ipynb notebooks to experiment with hyperparameters and regularization techniques.

For running Act_Max_B_12_W_1024.ipynb on a CPU-only machine, modify the load() function in your_env/lib/python3.*/site-package/torch/serialization.py to have map_location: MAP_LOCATION = 'cpu'. Restart the kernel and import the necessary packages. This step is not necessary for GPU machines.

MLP

Conduct a hyperparameter search using hyper_search.py. Set the working directory manually. Visualize results with the visualization.ipynb notebook. Images created by hyper_search.py are available here.

Produce the CIFAR-10 dataset with images transformed into randomly ordered tiles with mixed_up_cifar_10.py. Set the working directory manually. Store generated datasets under src/cifar10_datasets/.

Alternatively, download datasets from this Google Drive Link and save them under src/cifar10_datasets/.

You can evaluate the accuracy of the MLP and VGG models together with FGSM-perturbed dataset in the following section.

Fast Gradient Sign Attack (FGSM):

In the src/FGSM folder, you will find all the files necessary to reproduce the results of "Adversarial Attacks" corresponding to Section 3.2 of our project.

Adjust the working directory at the top of each file. Use FGSM_CNN_CIFAR10.py and FGSM_MLP_CIFAR10.py to create adversarial datasets for each model. Manually set the model and file name. Store generated datasets under src/cifar10_datasets/.

Alternatively, download datasets from this Google Drive Link and save them under src/cifar10_datasets/.

Evaluate CNNs on adversarial datasets and the original CIFAR10 dataset using CIFAR10_CNN.ipynb. For MLP, use CIFAR10_MLP.ipynb.

Dictionary Learning

The src/Dictionary_learning folder contains scripts to reproduce our project's "Dictionary Learning with Sparse Autoencoders" section (3.3).

1. Collect last-layer activations for MLP and VGG using harvest_activations_mlp.ipynb and harvest_activations_vgg.ipynb. Adjust the working directory at the top of each file.

   • Optionally, download activations from the Google Drive at the end of this ReadMe (acts_B_12-Wi_1024_cifar10_test_postskip.h5 and acts_VGG13_bn_cifar10_test.h5, respectively). Perhaps confusingly, the sparse autoencoders (SAEs) are trained on activations elicited by images from the CIFAR10 test set.

2. Run train_autoencoder_for_mlp.py and train_autoencoder_for_vgg.py to train the sparse autoencoders on the collected activations. You will have to change the location of the activations file at line 61.

   • If you prefer not to train the models yourself, you can download the model parameters from the Google Drive at the bottom of this ReadMe (SAE_100_epochs_bs_32_CIFAR10_test_B_12-Wi_1024_postskip.pt and SAE_100_epochs_bs_32_CIFAR10_test_vgg_bn13.pt, respectively).

3. Download the images from the CIFAR10 training set, together with their targets from the Google Drive at the bottom of this ReadMe (targs_cifar10_train.h5 and ims_cifar10_train.h5)

   • This is important to reproduce our feature analysis results because the images are in the same order as when we first did the analysis.

After setting the working directory in the files, run feature_analysis_mlp.ipynb and feature_analysis_vgg.ipynb to conduct feature analysis

All the data loaded by the scripts can be found here: Google Drive Link. Place all files you download in the dictionary_learning folder.