Minor fixes

• Update upstream nltools dependency
• fixes issues related to sklearn API updates (e.g #185)

0.6.1

Notes

This version drops support for Python 3.7 and fixes several dependency related issues:

• #162
• #176
• We can now handle images with an alpha-channel by just grabbing the RGB channels (typically in png files)
• Update minimum scikit-learn version requirement to ensure our viz models are loaded correctly
• Soft-pin numexpr version until this upstream pandas issue is fixed

0.6.0

Notes

This is a large model-update release. Several users noted issues with our AU models due to problematic HOG feature extraction. We have now retrained all of our models that were affected by this issue. This version will automatically download the new model weights and use them without any additional user input.

Detector Changes

We have made the decision to make video processing much more memory efficient at the trade-off of increased processing time, e.g. #139. Previously py-feat would load all frames into RAM and then process them. This was problematic for large videos and would cause kernel panics or system freezes. Now, py-feat will lazy-load video-frames one at a time, which scales to videos of any length or size assuming that your system has enough RAM to hold a few frames in memory (determined by batch_size). However, this also makes processing videos a bit slower and GPU benefits less dramatic. We have made this trade-off in favor of an easier end-user experience, but will be watching torch's VideoReader implementation closely and likely use that in future versions.

Fixes

• #139
• #163
• #165
• #171

Notes

This is a maintenance release that addresses multiple under-the-hood issues with py-feat failing when images or videos contain 0 faces. It addresses the following specific issues amongst others and is recommended for all users:

• #153
• #155
• #158
• #160

Notes

This is a large overhaul and refactor of some of the core testing and API functionality to make future development, maintenance, and testing easier. Notable highlights include:

• tighter integration with torch data loaders
• dropping opencv as a dependency
• experimental support for macOS m1 GPUs
• passing keyword arguments to underlying torch models for more control

Detector Changes

New

• you can now pass keyword arguments directly to the underlying pytorch/sklearn models on Detector initialization using dictionaries. For example you can do: detector = Detector(facepose_model_kwargs={'keep_top_k': 500}) to initialize img2pose to only use 500 instead of 750 features
• all .detect_* methods can also pass keyword arguments to the underlying pytorch/sklearn models, albeit these will be passed to their underlying __call__ methods
• SVM AU model has been retrained with new HOG feature PCA pipeline
• new XGBoost AU model with new HOG feature PCA pipeline
• .detect_image and .detect_video now display a tqdm progressbar
• new skip_failed_detections keyword argument to still generate a Fex object when processing multiple images and one or more detections fail

Breaking

• the new default model for landmark detection was changed from mobilenet to mobilefacenet.
• the new default model for AU detection was changed to our new xgb model which gives continuous valued predictions between 0-1
• remove support for fer emotion model
• remove support for jaanet AU model
• remove support for pnp facepose detector
• drop support for reading and manipulating Affectiva and FACET data
• .detect_image will no longer resize images on load as the new default for output_size=None. If you want to process images with batch_size > 1 and images differ in size, then you will be required to manually set output_size otherwise py-feat will raise a helpful error message

Fex Changes

New

• new .update_sessions() method that returns a copy of a Fex frame with the .sessions attribute updated, making it easy to chain operations
• .predict() and .regress() now support passing attributes to X and or Y using string names that match the attribute names:
  • 'emotions' use all emotion columns (i.e. fex.emotions)
  • 'aus' use all AU columns (i.e. fex.aus)
  • 'poses' use all pose columns (i.e. fex.poses)
  • 'landmarks' use all landmark columns (i.e. fex.landmarks)
  • 'faceboxes' use all facebox columns (i.e. fex.faceboxes)
  • You can also combine feature groups using a comma-separated string e.g. fex.regress(X='emotions,poses', y='landmarks')
• .extract_* methods now include std and sem. These are also included in .extract_summary()

Breaking

• All Fex attributes have been pluralized as indicated below. For the time-being old attribute access will continue to work but will show a warning. We plan to formally drop support in a few versions
  • .landmark -> .landmarks
  • .facepose -> .poses
  • .input -> .inputs
  • .landmark_x -> .landmarks_x
  • .landmark_y -> .landmarks_y
  • .facebox -> .faceboxes

Development changes

• test_pretrained_models.py is now more organized using pytest classes
• added tests for img2pose models
• added more robust testing for the interaction between batch_size and output_size

General Fixes

• data loading with multiple images of potentially different sizes should be faster and more reliable
• fix bug in resmasknet that would give poor predictions when multiple faces were present and particularly small
• #150
• #149
• #148
• #147
• #145
• #137
• #134
• #132
• #131
• #130
• #129
• #127
• #121
• #104

This is version-breaking release! See the full changelog here: https://py-feat.org/pages/changelog.html

Fix import error due to missing init.
Deployed 0.3.7 to Pypi

Trigger Zenodo release