Skip to content

Latest commit

 

History

History
88 lines (79 loc) · 9.17 KB

CHANGELOG.md

File metadata and controls

88 lines (79 loc) · 9.17 KB

v0.2.1 (March 26, 2024)

Code license changed to MIT License.

New features

Classes and functions

  • Rotate90: New transform to rotate image by 90 degrees
  • Trainer: Option for validating on loss value.

Tools

  • infer.py: New optional cmd argument to rotate image before inference.

Minor Fixes

  • Change UAV dataset hosting (now in ULiège Open Data Repository)
  • It is now possible to resume a training session when a validation frequency > 1 is used.
  • Solving the problem of continuously increasing training time over epochs.

Commits

Alexandre-Delplanque (19):

  • 0fb5a68 - Merge branch 'feature' into main
  • 506e1e1 - chore: update CHANGELOG.md
  • c101dca - fix: update code version
  • 031cc1a - chore: minor change
  • 4a2c68a - fix: update code version
  • b424c5b - Merge pull request #3 from Alexandre-Delplanque/new-license
  • ce16958 - chore: update CHANGELOG.md
  • d2b9078 - chore: final licence changes
  • f063a82 - chore: switch to MIT License
  • d6b3ffb - fix: change how images are rotated
  • c67404b - feat: add Rotate90 transform
  • 39e317a - fix: increasing training time over epochs
  • 4baeab1 - fix: resume training when using valid_freq
  • d86ec37 - fix: new link for UAV dataset #2
  • 53dd60a - fix: weights to device when not None (_ssim_loss)
  • b082fc4 - feat: rotation option for infer.py tool
  • c0bfc31 - fix: only CSV's images when -all arg is False
  • addbc9b - Merge branch 'feature' of https://github.com/Alexandre-Delplanque/Herd-Net into feature
  • d480319 - feat: option for evaluation with loss in Trainer

v0.2.0 (March 29, 2023)

New features

Classes and functions

  • CustomLogger : Argument to disable logging to CSV files (use to much memory).
  • Trainer: Arguments to set the validation frequency during training (valid_freq) and to choose whether to save logs to CSV files (csv_logger).
  • HerdNet: New method for reshaping classes (reshape_classes()), useful for loading pre-trained parameters.
  • FolderDataset: New flag (from_folder) in self.data attribute.

Python modules

  • sampler.py: New python module for hosting samplers for data loading.

Tools

  • train.py: New keys: wandb_run, model.freeze (HerdNet only), datasets.class_def, datasets.sampler and training_settings.valid_freq. Now use the class definition (i.e., datasets.class_def) to make sure the labels match the species names.
  • test.py: New keys: wandb_run and dataset.class_def. Now use the class definition (i.e., dataset.class_def) 1) to make sure the labels match the species names, and 2) for plotting precision-recall curves, saving the detections, the metrics and the confusion matrix.

Commits

Alexandre-Delplanque (17):

  • ff94a5e - chore: update README.md
  • 03ccd66 - version: update version number and modified date
  • 4791624 - chore: add CHANGELOG.md
  • eb159c7 - docs: create doc folder and update configs md
  • f18f9f9 - feat: add new keys to configs
  • 7f91167 - feat: add sampler option in train.py tool
  • bea0fc4 - feat: samplers.py - hosts samplers for dataloading
  • d883b49 - feat: save classes, mean, std in PTH files
  • 8189dc7 - feat: add 'from_folder' flag in FolderDataset data
  • 907a221 - feat: +class def., +labeled results, +conf. matrix
  • 7b4bc74 - feat: +class def., +validation freq., -cross-val
  • 884673d - fix: add head_conv attribute to HerdNet module
  • 7b88af8 - feat: valid_freq and csv_logger args (Trainer)
  • d88ff21 - feat: add option to disable csv logs
  • 7071569 - fix: PointsToMask one-hot encoding option
  • 782d877 - fix: FocalLoss, avoid NaN when output is 0 or 1
  • deec190 - add article reference

v0.1.0 (January 23, 2023)

Initial version of the code, used for producing the results of the reference paper "From Crowd to Herd Counting: How to Precisely Detect and Count African Mammals using Aerial Imagery and Deep Learning?".

Commits

Alexandre-Delplanque (11):