Summary: 57 instances, 53 unique

Text	Count
TODO: Figure out how to handle single-batch datasets.	1
# TODO: Consider tensorflow_estimator/python/estimator/util.py.	1
# TODO: Allow user to disable ensemble workers. For example, when there	1
# TODO: Consider extracting a common function to use here and in	1
# TODO: Replace with adanet.Estimator. Currently this just verifies	1
# TODO: Pruning the previous subnetwork may require more metadata	1
# TODO: Make the "config" argument optional using introspection.	1
# TODO: Add checks to make sure phases are valid.	1
TODO: Best ensemble index should always be static during EVAL	1
# TODO: Switch back to TF 2.0 once the distribution bug is fixed.	1
# TODO: Support steps parameter.	1
TODO: Make this code compatible with TPU estimators.	1
# TODO: Run writer.flush() at Session end.	1
# TODO: Remove setters and getters.	1
# TODO: Magic number. Investigate whether there is a	1
# TODO: Add some way to check that work_units has to be called	1
# TODO: Move adanet_loss from subnetwork report to a new	1
# TODO: Consider moving builder mode logic to ensemble_builder.py.	1
# TODO: Currently, TPUEstimator has no global_step set when	1
# TODO: How do we enforce that save_model is called only once per	1
# TODO: Replace candidates with ensemble_specs.	1
# TODO: _ScopedSummary and _ScopedSummaryV2 share a lot of the same	1
# TODO: Add Unit tests for the ndims == 3 path.	1
# TODO: Encapsulate conversion and serialization of a	1
# TODO: Add ability to choose the problem type: regression,	1
# TODO: Handle TPU embedding variables.	1
# TODO: Prevent checkpoints that are currently being	1
TODO: Provide the missing functionality detailed below.	1
# TODO: Find a better way to get all models than to pass in a	1
# TODO: Consume arbitary fit inputs.	1
# TODO: Refactor architecture building logic to its own module.	1
# TODO: Consider using `RunConfig.replace` with the new device_fn,	1
# TODO: Add support for `bad_color` arg.	1
# TODO: Confirm that it makes sense to have global step of 0.	1
# TODO: Remove uses of Evaluator once AdaNet Ranker is implemented.	1
# TODO: Figure out a cleaner way to handle this.	1
# TODO: Eliminate need for candidates.	2
# TODO: Merge CandidateBuilder into SubnetworkManager.	1
# TODO: Add more detailed documentation.	2
# TODO: Should architecture.subnetworks be sorted by iteration	1
# TODO: Add more details documentation.	1
# TODO: Switch optimizers to tf.keras.optimizers.Adam once the	2
# TODO: Optional code organization suggestion:	1
# TODO: Replace with variance complexity measure.	1
# TODO: Do summaries need to be reduced before writing?	1
# TODO: Re-enable once we know that evaluation won't	1
# TODO: Extract output shapes from submodels instead of passing in	1
# TODO: Validate name matches ^[A-Za-z0-9_.\\-/]*$	1
# TODO: Consider making these messages be some kind of Enum.	1
# TODO: Allow better customization of TensorBoard log_dir.	2
# TODO: Find a better way to ensure work_units only gets called	1
# TODO: Handle hook created variables.	1
TODO: Remove iteration_number from the argument of this method.	1