tekayanidham / basic-image-reconstruction-with-autoencoder Goto Github PK

The training approach is always training the 2 sized bottleneck along with the input custom size. To ensure the non-familiarity with test data for both models, I chose to save the test data fin the training file to be used later in the latent space analysis. The test data file is replaced with every training.

• data/ : contains the data files( downloaded in the training script for torchvision datasets)
• model/autoencoder.py : the structure of the autoencoder
• train.py : the training script for both models, fixed 2 sized bottleneck and custom input sized and saving the test data.
• viz_reconstruction.py : the script for visualizing the reconstruction of the test data comparing the 2 models performance.
• encode_latent_space.py : script to use the encoder part to generate the latent space of from a number of input images (e.g. 1000), the results are saved in .npy numpy file.
• viz_latent_space.py : script to visualize both latent spaces of both models and compare them, in the argument you have the choice of using UMAP or t-SNE.
• viz_kmeans_latent_space.py : script to visualize the clustered latent space using k-means.
• requirements.txt : the requirements file for the project.
• notebook_alternative.ipynb : the notebook for this project instead running scripts internally in terminal.
• model/ : the folder containing the saved models, test data, and latent spaces.

GTSRB (German Traffic Sign Recognition Benchmark)

• Size : 26640
• 43 classes

Training the models:

python train.py --num_epochs 10 --latent_size 256 --batch_size 128 --lr 0.0001

Note : the only required argument is latent_size, the rest are optional(defaults as displayed).

Visualizing the reconstruction:

python viz_reconstruction.py --model_path output/model_latent_size_256.pth --test_data output/test_data.pth

Note : the only required argument is latent_size, the rest are optional(defaults as displayed).

Encoding the latent space:

python encode_latent_space.py --model_path output/model_latent_size_256.pth --test_data output/test_data.pth --num_samples 1000

Note : the only required argument is latent_size, the rest are optional(defaults as displayed).

Visualizing the latent space:

python viz_latent_space.py --latent_space output/latent_space_size_256.npy --vis_type umap

Vizualizing clustered umap reduced latent space with k-means:

python viz_kmeans_latent_space.py --latent_space output/latent_space_size_256.npy --num_clusters 43