Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
* Better sampler (#160) * working verison ish * faster sampler * Remove quickdraw notebooks. Co-authored-by: Owen Vallis <ovallis@google.com> * * Add get_slice to SingleShotMemorySampler * Make get_examples() -> _get_examples() * Update doc strings to refer to classes_per_batch and examples_per_class_per_batch * Update notebook to also use examples_per_class * Update Return types of _get_examples() and generate_batch() to use FloatTensor and IntTensor * Set version to 0.14 * Run workflow on push for master only and on pull_request for all branches. * Add create_index method to enable setting an index outside of the compile call. (#164) fix #163 * * Add sampler_io_cookbook notebook * Fix get_slice in tfds sampler * other small sampler fixes * Remove randomflip from mnist augmentation in sampler cookbook. * Add link to sampler notebook in examples readme. * * Remove tf.gather from memory_samplers._get_examples(). Replace with faster for loop over lists and convert to np.array * Improve docstring coverage * Basic formatting using yamf * Reorder top level imports in __init__ * Update batch info when initializing sampler to better reflect the initial batch size and number of augmenters. * kaggle first * Raise errors if users try to evaluate over an empty index or pass an … (#169) * Raise errors if users try to evaluate over an empty index or pass an empty set of lookupd to the evaluators. * Add hints to index Exceptions. * Make batch size msg clearer. * MultiSampler now samples without replacement unless the class size is < num_examples_per_class. In that case, we warn the user and then sample with replacement. The warning will only be printed once per class. Also fixed a mypy error in the img_augments.py module. * Update SimilarityModel.compile() arg order to be consistent with the Keras Model.compile(). * Enforce that class_mapping is map by calling .get() and move the class_label code into it's own function. * Update hello world to remove class_mapping from viz_neighbors_imgs as we can just use the class ids. * Ensure class mapping in nn_viz requires a dict (#170) * Enforce that class_mapping is map by calling .get() and move the class_label code into it's own function. * Update hello world to remove class_mapping from viz_neighbors_imgs as we can just use the class ids. * Add support for passing distance thresholds and matcher in callbacks. * Update new for 0.14 in README Co-authored-by: Elie Bursztein <github@elie.net>
- Loading branch information