Code

Topic recognition#469

Code

Open

4lon wants to merge 63 commits intoshakes76:topic-recognitionshakes76/PatternFlow:topic-recognitionfrom 4lon:topic-recognition4lon/PatternFlow:topic-recognitionCopy head branch name to clipboard

+448-1

Commits

Commits on Oct 11, 2022

• Initial commit and filling out basic project structure.

  4lon

  committed
  d4943af

  Copy full SHA for d4943af

• Added basic dataset class that loads all images in directory and return samples for learning.

  4lon

  committed
  32876fb

  Copy full SHA for 32876fb

• Added template for VAE

  4lon

  committed
  0c9dc73

  Copy full SHA for 0c9dc73

• Corrected error in dataset class location, moved from train to dataset.

  4lon

  committed
  0d501b0

  Copy full SHA for 0d501b0

Commits on Oct 14, 2022

• Added Vector Quantizer

  4lon

  committed
  4176432

  Copy full SHA for 4176432

• Updated encoder and decoders

  4lon

  committed
  067b1fc

  Copy full SHA for 067b1fc

• Fixed wrong layer errors and added vqvae model, likely need to fix padding issue.

  4lon

  committed
  d9e43b3

  Copy full SHA for d9e43b3

Commits on Oct 20, 2022

• Updated to use pytorch embeddings and fleshed out vector quantizer.

  4lon

  committed
  5cb7a40

  Copy full SHA for 5cb7a40

• Cleanup

  4lon

  committed
  662ef51

  Copy full SHA for 662ef51

• Cleanup and redid vq_vae class to add more individual functions. This should be useful later for implementation and prediction.

  4lon

  committed
  d2611f7

  Copy full SHA for d2611f7

• Updated to make dataloaders in dataset.py

  4lon

  committed
  eb289e9

  Copy full SHA for eb289e9

• Basis of training script setting up

  4lon

  committed
  dbea67a

  Copy full SHA for dbea67a

• Built out saving model and modularised functions

  4lon

  committed
  43cc875

  Copy full SHA for 43cc875

• Fixed up incorrect dataset formatting and loading of imgs

  4lon

  committed
  6a60fba

  Copy full SHA for 6a60fba

• Fleshed out training function

  4lon

  committed
  40bb45c

  Copy full SHA for 40bb45c

• Fleshed out testing function

  4lon

  committed
  eecfde1

  Copy full SHA for eecfde1

• Added image generation and saving to tensorboard for visual analysis

  4lon

  committed
  aa65485

  Copy full SHA for aa65485

• Refactored modules to include original functions file and updated dataset location.

  4lon

  committed
  9ef9451

  Copy full SHA for 9ef9451

Commits on Nov 7, 2022

• Increase HPC utilisation

  4lon

  committed
  670cc67

  Copy full SHA for 670cc67

• add checks

  4lon

  committed
  3c7ebce

  Copy full SHA for 3c7ebce

• Increased Epochs

  4lon

  committed
  968ad8a

  Copy full SHA for 968ad8a

• Original pytorch attempt was not working, start of tensorflow attempt, reimplemented the dataset, however this current implementation is loading every image into an array, that's far from ideal so …

  Show description for cd008ee
  4lon

  committed
  cd008ee

  Copy full SHA for cd008ee

• Old train model, not relevant anymore but good for looking back and not ready to remove

  4lon

  committed
  eee3087

  Copy full SHA for eee3087

• Redesigned VQVAE modules for tensorflow. This is a simpler design overall however time is limited. If I have time I intend to work back to this and improve the model

  4lon

  committed
  65a76a4

  Copy full SHA for 65a76a4

• Reimplemented basic training method, but this one does not evaluate against other data partitions or save sample outputs.

  4lon

  committed
  c54bd95

  Copy full SHA for c54bd95

• Fixed some errors encountered when trying to train because of tracker dictionary, and updated directory to dataset.

  4lon

  committed
  00490b4

  Copy full SHA for 00490b4

• Added plotting of results including losses and reconstruction post learning

  4lon

  committed
  09ca20c

  Copy full SHA for 09ca20c

• Updated directory

  4lon

  committed
  53811cf

  Copy full SHA for 53811cf

• Fixed missing param

  4lon

  committed
  bd443ee

  Copy full SHA for bd443ee

• Train on more images

  4lon

  committed
  120f5e1

  Copy full SHA for 120f5e1

• Updated optional param

  4lon

  committed
  d7400d1

  Copy full SHA for d7400d1

• Imporved plotting

  4lon

  committed
  07950f1

  Copy full SHA for 07950f1

• Removed print statement

  4lon

  committed
  e65e287

  Copy full SHA for e65e287

• Fixed image formatting

  4lon

  committed
  36a4200

  Copy full SHA for 36a4200

• New dataset dtype to remove reconstruction issues.

  4lon

  committed
  ba208a4

  Copy full SHA for ba208a4

• Reduced batch size because of memory running out in new datatype (going from 16bits to 32 bits must mean smaller batch size required.)

  4lon

  committed
  178c251

  Copy full SHA for 178c251

• Reduced batch size again.

  4lon

  committed
  f6eea6d

  Copy full SHA for f6eea6d

• Changed data type because space requirement was unachievable

  4lon

  committed
  ed9a611

  Copy full SHA for ed9a611

• Built basic masked convolution layer and limited dataset size (memory issues)

  4lon

  committed
  d32fda4

  Copy full SHA for d32fda4

• Implemented residual block for pixelcnn

  4lon

  committed
  12bbfa6

  Copy full SHA for 12bbfa6

• Implemented basic pixel cnn architecture but not sure it works.

  4lon

  committed
  31bcdef

  Copy full SHA for 31bcdef

• Added pixelcnn training to learn codebook production

  4lon

  committed
  c0ede93

  Copy full SHA for c0ede93

• Updated plotting to also save pixel cnn performance

  4lon

  committed
  cd4838a

  Copy full SHA for cd4838a

• Updated training with ssim metrics

  4lon

  committed
  3660c53

  Copy full SHA for 3660c53

• reduced dataset size

  4lon

  committed
  b1e9fda

  Copy full SHA for b1e9fda

• reduced dataset size

  4lon

  committed
  fafcd18

  Copy full SHA for fafcd18

• reduced dataset size

  4lon

  committed
  b9d4ce0

  Copy full SHA for b9d4ce0

• reduced dataset size

  4lon

  committed
  2de0a80

  Copy full SHA for 2de0a80

• Added sample model outputs

  4lon

  committed
  570a4b6

  Copy full SHA for 570a4b6

• Added better sample model outputs

  4lon

  committed
  df901ca

  Copy full SHA for df901ca

• Refactored training to seperate pixel cnn and vqvae in case of seperate training (this helps quite a bit with the memory problems of the dataset implementation.

  4lon

  committed
  14d2ebb

  Copy full SHA for 14d2ebb

• Fleshing out report

  4lon

  committed
  b193525

  Copy full SHA for b193525

• Fleshing out report

  4lon

  committed
  a15a457

  Copy full SHA for a15a457

• Add brain generating function but doesn't really work.

  4lon

  committed
  19d4895

  Copy full SHA for 19d4895

• Removed old attempt in pytorch and added snapshots of inference.

  4lon

  committed
  9de9e81

  Copy full SHA for 9de9e81

• Fixed up incorrect function call from refactor and add SSIM evaulation and recontruction to predict.py

  4lon

  committed
  3fb1dfc

  Copy full SHA for 3fb1dfc

• Finalising report

  4lon

  committed
  fcbbb82

  Copy full SHA for fcbbb82

• Fixed reconstruction a bit

  4lon

  committed
  0a71411

  Copy full SHA for 0a71411

Commits on Nov 24, 2022

• Merge branch 'topic-recognition' into topic-recognition

  4lon

  authored
  0231d01

  Copy full SHA for 0231d01

• Delete recognition/44801582_OASIS_VAE/samples/pixelcnn_model directory

  4lon

  authored
  8be7c8e

  Copy full SHA for 8be7c8e

• Delete recognition/44801582_OASIS_VAE/samples/vqvae_model directory

  4lon

  authored
  67ca8b2

  Copy full SHA for 67ca8b2

• Delete pixelcnn_model_weights.h5

  4lon

  authored
  949fd52

  Copy full SHA for 949fd52

• Delete vqvae_model_weights.h5

  4lon

  authored
  308597f

  Copy full SHA for 308597f