Andrew Ferlitsch
Andrew Ferlitsch
**REQUIRED:** Add a summary of your PR here, typically including why the change is needed and what was changed. Include any design alternatives for discussion purposes. This fixes issue #850...
This implementation has the same error (missing ZeroPadding2D after stem) that the version in keras-application had, which has been fixed. See issue: https://github.com/keras-team/keras-applications/issues/30
The image augmentation should be extended to do a change in perspective. This is not simple skew, but a 3D transformation of viewing the image from a different angle. The...
Currently, when processing a collection, an Image() object is instantiated per image. Each time the Image() object has to parse parameters and examine the data to determine settings. Most of...
A new mode for feeding should pull the machine learning ready data from storage as it is being feed, and not keep the dataset in memory. A proposed name for...
For the memoryless enhancement, further feeding performance can be achieved through pipelining using a worker thread. Instead of reading/transform/feed sequentially, as the foreground one is being prep/feed a background thread...
There should be a mode using worker threads (for multi-core CPUs) to parallel process groups of images into machine learning ready data (vs. serially).
The image augmentation feature should be extended to include grayscale. Actually, I would prefer that instead of simple grayscale that it did night-time darkening.
The split property to Images class should be extended to support stratification, similar to sci-learn: http://scikit-learn.org/stable/modules/generated/sklearn.model_selection.train_test_split.html
Acronyms like U.S. should be preserved (i.e., US) and if recognized converted to their proper name (e.g., United States), when being preprocessed by Words class.