opennsfw2

• Fix image overlay for GradCAM when using YAHOO preprocessing. The YAHOO preprocessing actually excludes from the model input certain pixels from each edge of the image, the GradCAM visualisation should take this into account.

• Fix type annotation for PIL.Image.Image.
• Use a fixed layer name global_average_pooling2d for the GlobalAveragePooling2D layer, to avoid Keras assigning a different name upon multiple model instantiations.

• For Keras 2 compatibility, a generic type var KerasTensor is defined instead of using keras.KerasTensor directly, as the latter is only available in Keras 3.
• Relax version requirement for Keras in setup config

• Allow passing a model name to the make_open_nsfw_model function. This will be used as the name attribute of the returned Keras model object. The default name is opennsfw2.
• Fix logic for updating the internal global model variable.

• Use Keras 3 instead of Keras Core
• Drop support for Python 3.8; add support for Python 3.11

• Improve internal batching logic for predicting from image paths.

• Improve support for using the APIs in a loop, in particular, reduce the memory usage due to re-creating the model object.

• Switch from tf.keras to keras_core
• Include test run for JAX backend
• Update README
• Other minor fixes

• Fixed model inference usage to avoid potential TF mem-leak

• Remove grad_cam_height and grad_cam_width settings, the output grad-cam image will always be the same size as the input image
• Default alpha is 0.8

• Add progress bar for predicting video frames.
• Set all default batch sizes to 8.

Add aggregation support for predicting video frames.

• Drop support for Python 3.7.
• Use numpy.typing from NumPy 1.22.0.
• Use Pillow 9.0.0, note that this leads to some small floating point differences to the results.
• Fix Grad-CAM tools.

Add matplotlib to required packages.

New features:

• predict_image and predict_images can now also produce Grad-CAM plots that highlight the important regions which lead to the predictions.

Clean up code for model architecture to improve readability.

Fix mypy re-export errors

Save output video, if enabled, as MP4 instead of AVI.

New:

• predict_video_frames for getting the NSFW probability of each video frame
• predict_image for getting the NSFW probability of a single image

Breaking changes:

• predict is removed, and replaced by predict_images, which returns a list of NSFW probabilities of the input images

Update dependencies.