Code

s46144603 VQVAE OASIS#478

Code

Open

joshuamcallan wants to merge 21 commits intoshakes76:topic-recognitionshakes76/PatternFlow:topic-recognitionfrom joshuamcallan:topic-recognitionjoshuamcallan/PatternFlow:topic-recognitionCopy head branch name to clipboard

+251

Commits

Commits on Oct 24, 2022

• Loaded OASIS dataset from Cdrive and seperated into train, test and validate - each preprocessed by normalising by factoring by 255.

  joshuamcallan

  committed
  a0d5fda

• Created a simple encoder for the VQVAE to reduce the amount of paramters in the model.

  joshuamcallan

  committed
  5cd4b4b

• Added a simple decoder network for the VQVAE.

  joshuamcallan

  committed
  aa8df4f

Commits on Oct 25, 2022

• Added VectorQuantization class in order to create a discrete latent space.

  joshuamcallan

  committed
  5f130b1

• Added a vqvae model by feeding inputs into the encoder and the encoder through vq_layer for the quantised vae and finally fed the quantised vae through the decoder for the model.

  joshuamcallan

  committed
  7a5e53c

• Added training loop for VQVAE in order to backpropagate and record losses.

  joshuamcallan

  committed
  9ebb927

• Added predicitions using the test set based on the VQVAE model.

  joshuamcallan

  committed
  71c4e5d

• Put the training model into the wrong file (predict.py) so swapped it to the train.py file and also added plots of the losses and metrics during training.

  joshuamcallan

  committed
  ba90602

• Removed the training model from this file and added functionality to compare original vs reconstructed images as well as calculate the structured similarity (SSIM) of these two images).

  joshuamcallan

  committed
  2866fd0

• Added a description of the VQVAE model and how it works. Also added a main title and description of the problem. Some headings were also made.

  joshuamcallan

  committed
  a12454a

• Added more description for implementation of dataset as well as module dependencies and references.

  joshuamcallan

  committed
  8368b7d

• When choosing the random images, it was selecting from the train dataset so fixed this to sample from the test set.

  joshuamcallan

  committed
  27f67a6

• Changed epoch number to 200 to reduce computation time but increases the loss as a result.

  joshuamcallan

  committed
  345b921

• Added an error rate plot as well as side-byside comparisons of original test images vs reconstructed images. Not certain yet that they will show on github however.

  joshuamcallan

  committed
  edd99f0

• Side-by-side comparisons of test images vs reconstructed images as well as loss rate graph.

  joshuamcallan

  committed
  17089a3

• Description of usage of algorithm with this dataset and future datasets.

  joshuamcallan

  committed
  c42dbaa

• Talked about the train, test, validation split and also the SSIM of each reconstructed image shown. All are above 0.6!

  joshuamcallan

  committed
  a109c13

• All OASIS data added.

  joshuamcallan

  committed
  9ad8c16

• Fixed path variables for users convenience to alter path for themselves to obtain OASIS data.

  joshuamcallan

  committed
  df5d9a6

• Deleted a line so the file is class and function based only.

  joshuamcallan

  committed
  793b32d

• Added information about path locations for user to alter themselves as well as the version of python used.

  joshuamcallan

  committed
  5c1008a