README.md

Autoencoder: Efficiency vs Tradeoffs

Update Logs:

The reproduction is still ongoing. Please refer the following logs for current and future updates:

1. Individual implimentation using official and partial references as mentioned in the paper is updated.
2. Single-Click result implimentation is ongoing.
3. Other autoencoder models will be added in the future.

Overview

This repository introduces the paper "A Comprehensive Study of Auto-Encoders for Anomaly Detection: Efficiency and Trade-Offs" by Asif Ahmed Neloy and Dr. Max Turgeon.

The following autonencoders are reproduced in this repository:

• Denoising Auto-Encoder (DAE)
• Sparse Auto-Encoder (SAE)
• Contractive Auto-Encoder (CAE)
• Variational Auto-Encoder (VAE)
• Conditional Variational Auto-Encoder (CVAE)
• beta-VAE
• Adversarial Variational Auto-Encoder (adVAE)
• Importance Weighted Auto-Encoder (IWAE)
• Probabilistic Auto-Encoder (PAE)
• Robust Deep Auto-Encoders (RDA)
• Vector Quantised-Variational Auto-Encoder (VQ-VAE)

Reproducibility

The reproducibility process is divided into two process-

1. Reproduce individual results using the Official repos or Particial reference repos
2. Reproduce all results using the modified models.

Install Requirements

The requirements.txt file contains python packages for reproducing all models. There is no individual requirements.txt file.

Prerequisite:

• Anaconda
• Python 3.7.11
• Windows 10/11 (also can be reproduced using any linux machine. However, the package requirements needs to be adjusted)
• GPUs (PyTorch and TensorFlow are created using gpu models)

Installation steps:

• create new env: conda create -name autoencoder python=3.7 -y
• activate env: conda activate autoencoder
• open the directory and install the requirements.txt: pip install -r requirements.txt

Individual Results

1. DAE

1. Navigate to the Directory: \Partial References\DAE\
2. Run: python main.py

2. SAE

1. Navigate to the Directory: \Partial References\SAE\
2. Run: python main.py
3. Optional Result: python sparseae_generic.py

3. CAE

1. Navigate to the Directory: \Partial References\CAE\
2. Run: python main.py
3. Optional: Change parameters in Fashion-MNIST implementation python main_FashionMNIST.py

4. VAE

1. Navigate to the Directory: Official Repos\VAE

2. Run: python run_main.py --dim_z 20

3. Required Arguments --dim_z: Dimension of latent vector. Default: 20

4. Optional:

• --results_path: File path of output images. Default: results
• --add_noise: Boolean for adding salt & pepper noise to input image. Default: False
• --n_hidden: Number of hidden units in MLP. Default: 500
• --learn_rate: Learning rate for Adam optimizer. Default: 1e-3
• --num_epochs: The number of epochs to run. Default: 20
• --batch_size: Batch size. Default: 128
• --PRR: Boolean for plot-reproduce-result. Default: True
• --PRR_n_img_x: Number of images along x-axis. Default: 10
• --PRR_n_img_y: Number of images along y-axis. Default: 10
• --PRR_resize_factor: Resize factor for each displayed image. Default: 1.0
• --PMLR: Boolean for plot-manifold-learning-result. Default: False
• --PMLR_n_img_x: Number of images along x-axis. Default: 20
• --PMLR_n_img_y: Number of images along y-axis. Default: 20
• --PMLR_resize_factor: Resize factor for each displayed image. Default: 1.0
• --PMLR_n_samples: Number of samples in order to get distribution of labeled data. Default: 5000

5. CVAE

1. Navigate to the Directory: \Partial References\CVAE\
2. Train and generate CVAE result: python main.py
3. Train and generate VAE result: python VAE.py
4. Testing Results: python test_generation.py

6. beta-VAE

1. Navigate to the Directory: Official Repos\Beta_VAE
2. Train: python train.py --z_dim 10 --beta 0.1 --dataset mnist --model conv_mnist --experiment_name z10_beta0.1_mnist_conv
3. Alternative: train.py --z_dim 10 --beta 0.1 --dataset mnist --model mlp_mnist --experiment_name z10_beta0.1_mnist_mlp

7. adVAE

1. Go to the directory: cd Official Repos\adVAE

2. Train model: python run.py

3. optional parameters:

   • --datnorm, type=bool, default=True, help='Data normalization'
   • --z_dim, type=int, default=128, help='Dimension of latent vector'
   • --mx, type=int, default=1, help='Positive margin of MSE target.'
   • --mz, type=int, default=1, help='Positive margin of KLD target.'
   • --lr, type=int, default=1e-4, help='Learning rate for training'
   • --epoch, type=int, default=100, help='Training epoch'
   • --batch, type=int, default=32, help='Mini batch size'

8. IWAE

1. Navigate to the Directory: \Partial References\IWAE

2. For implementing the result mentioned in the reference paper: python experiments.py --model [model] --dataset [dataset] --k [k] --layers [l] (where model is vae or iwae; dataset is one of BinFixMNIST, MNIST, OMNI; k is 1, 5, or 50; l is 1 or 2.)

3. Optional implementation: python experiments.py --exp iwae_to_vae

4. Optional: python download_mnist.py: Download MNIST Dataset. Default: mnist

9. PAE

1. Navigate to the Directory: cd \Official Repos\PAE
2. Run the python file: python main.py --helpfull
3. Optional Training: python main.py

10. RDA

1. Open the Directory: cd \Official Repos\RDA
2. Train and regenerate: python main.py
3. Compare to VAE: python VAE_SRDA.py

11. VQ-VAE

1. Navigate to the Directory: Official Repos\VQ_VAE

2. Run: python main.py --MNIST

3. Required Arguments --dataset: Dataset. Default: CIFAR-10

4. Optional:

• --batch_size, type=int, default=32*)
• --n_updates, type=int, default=5000)
• (--n_hiddens, type=int, default=128)
• (--n_residual_hiddens, type=int, default=32)
• (--n_residual_layers, type=int, default=2)
• (--embedding_dim, type=int, default=64)
• (--n_embeddings, type=int, default=512)
• (--beta, type=float, default=.25)
• (--learning_rate, type=float, default=3e-4)
• (--log_interval, type=int, default=50)

All Results

ALL RESULTS is implemented to visualize the ROC-AUC, Latent Space Representation and Manifolds for MNIST, FASHION_MNIST datasets for all models regardless of the training parameters and hyper-parameters. (On-going)

License