Summary: 312 instances, 247 unique

Text	Count
// TODO: gl_sframe_range::end() should be a const method.	2
// TODO: List of todo's for this file	18
// TODO: Consider converting float_array values to flex_nd_vec, but we must	1
// TODO: Should the metric implementations themselves (or	1
// TODO: What feedback can we give if the user requests a batch size that	1
// TODO: Returns 4GB as that makes sure default batch size is used.	1
// TODO distinguish dict from SGraph	1
// TODO: Define Optional<T>. (Or require C++17 for std::optional?)	1
// TODO: Accessors describing name inputs and expected shapes.	1
// TODO: Check that k exists in eval_map.	1
// TODO: Also verify that activation shape here is [1024, 13, 13]?	1
# TODO: decorrelate reducer_override (which is tied to CPython's	1
// TODO implement other alignments	1
// TODO: fill this out.  This should get much more sophisticated.	1
// TODO - allow other font properties	1
// TODO -- pick the limited set of categories intelligently (n most popular	1
* \TODO: We *could* templatize this around the column type, allowing this to	1
// TODO: This implementation would ideally just post-process the return value of	1
// TODO apply font weights in non-macOS	1
// TODO: write function to check valid data spec	1
// TODO: better format	1
// TODO: refactor for multiple indicies	1
// TODO: Clean up the relationship between the save/load interface defined here	1
// TODO: Style transfer backends should support both training and inference.	1
# @TODO: Should be able to automatically choose number of iterations	1
// TODO: Move this model-spec generation code into a separate file, ideally	1
/* TODO: const float_array_map& config if needed */	3
# TODO: https://github.com/apple/turicreate/issues/2672	1
// @TODO: think about rating data later	1
// TODO: implement `get_similar_items` used with Image Saliency for OD	1
// -- TODO: Add object detection network	1
// TODO: This is broken on Windows, but is only used by RPC, so putting it off	1
//TODO: Figure out how to get error string on Windows	1
# FIXME: add tests	1
// TODO: Have MPS path use these parameters, instead	1
// TODO: Automatically infer the image column name, or throw error if you	1
// @TODO: think about side data later	1
// TODO: Expose seed parameter publicly.	1
# TODO: automatically tune this.	1
// TODO: Call object_detector::train from here.	1
* TODO: Fill in details	1
/* TODO: Render Labels on Images */	1
// TODO: The semantics below are adopted from Core Image.	1
// TODO: Estimate the size of the sframe so that we could decide number of	1
// TODO: set node_fwd.flag to control training vs. inference mode once we have graphs for both	1
// TODO: refactor to share code with the non-browser case,	2
* TODO:	1
* TODO This can be more intelligent as required. For now, it is kinda dumb.	1
// TODO: Adopt NCHW.	1
// TODO exception handling	1
// TODO: Operations such as reshape, slice, etc.?	1
// TODO: This hack should not be required.	1
// TODO - for now, skip values with nan/inf	1
size_t bin = hash64(seeds[j] ^ i) % num_bins;  // TODO: bit mask	4
/* @TODO: Split all backgrounds into as many chunks as there are cores	1
// TODO: Move these helper functions to st_model_trainer.cpp	1
// TODO: previously, we had a check for if (endltype(f) == endltype(std::endl))	1
// TODO: fix all warnings and make these errors.	1
//TODO: for now	1
// TODO: Investigate the number of copies a tensor endures end-to-end.	1
// TODO: Add a mutable float_array interface so we can validate size.	1
TODO:	1
# Now, set up the environment.  TODO: move these in to actual	1
// TODO: Investigate higher-quality downsampling? The original Python toolkit	1
// TODO: Someday, this will all be an implementation detail of each	2
// TODO: can it be set by the available RAM??	1
// TODO: Make less ugly!	14
# TODO: replace all these with in-build macros based on standard macros in Availability.h	1
//TODO: Make specific errors available	1
std::string output_type); // TODO: This should be const	1
* TODO: Add proper arguments to create_drawing_classifier	1
// TODO: Should these also be smoothed?	1
// TODO: below are not available from NSCursor, images are needed	1
* \TODO Factor out the shared structure for data iterators	1
* TODO: Add some comments hare about what the base class is supposed to be.	1
// TODO: Recycle these allocations.	1
// TODO: A struct instead of a map would be nice, too.	1
// TODO: Eventually this should also support computing loss from labels, when	1
// -- TODO: Add activity classifier network	1
// TODO: Change for testing	1
// TODO: Once we adopt an asynchronous API, we can let this "double buffering"	1
// TODO: Migrate to neural_net::float_array_map	1
// Helper class for extracting rules // TODO REFACTOR	1
// TODO: Replace model_spec with a Checkpoint class that encapsulates	1
// TODO: yes 20 is a magic number.	1
// TODO: Make some of these parameters to the model	1
# TODO: Refactor Instance Norm	1
// TODO: Remove resize_only by passing all the augmentation options	1
* \TODO: This should be a flex_list to accomodate integer labels!	1
/** TODO: Figure out a better solution to having `num_samples` be a	1
// TODO: write function to check valid image spec	1
// TODO: memory pool implementations, etc.	1
// TODO: When MLFoundation supports training across multiple GPUs, ensure	1
* TODO: Refactor this and perplexity to share code.	1
gl_sframe training_data_;  // TODO: Avoid storing gl_sframe AND data_iterator.	3
// TODO: The original Python code path allowed users to specify no validation	1
// TODO: Remove this vestigial macro invocation once the dependency on cppipc	1
// TODO: Dispatch augmentation to a separate thread/queue.	1
// TODO: Change if vega tooltips have log level options	1
// TODO: Someday we will remove this brittle dependency on names by	2
// TODO: Redo this function!!! There are tons of in-memory sections.	1
// - there should only be a single column if this is the case.  TODO: Why?	1
// TODO: Parse from an input stream, without loading entire file into memory	1
// TODO: Should this ultimately use std::random_device instead?	1
// TODO: The names should somehow be a parameter of this class.	5
// TODO: Temporary row_number column to do a stable sort, better solution!	1
// TODO - what if it's string? how to tell? can it be anything else?	1
// TODO: convert interface above to use the extensions methods here	2
# TODO: Recycle the ndarray instances we're allocating below with	1
// TODO: Parameters such as graph mode should be more explicit than just a	1
// TODO: Iterator needs to support resuming from an offset.	1
// TODO: Move this into DarknetYOLOModelTrainer, since these heuristics are	1
// TODO: Streamline this to prevent all the copies from the various types.	1
// TODO: Remainder of interface: predict, etc.	1
// TODO: resolve these issues at the source level	1
// TODO: add proxy support	1
// TODO this is bad -- we need a non-const Aggregation in order to call	1
// TODO: Figure out the right failure behavior for when denominator is 0.	1
// TODO: These should be exposed in a way that facilitates experimentation.	1
// TODO: convert interface above to use the extensions methods here	6
// TODO: This copy is inefficient. This should be a wrapper around NSData:	1
// TODO: Merge all logical_filter transforms that have identical masks.	1
// TODO: make train asynchronous	1
// TODO: This function should be const.	1
/* TODO: const float_array_map& config if needed */	1
// TODO: This heavyweight shuffle operation introduces spikes into the	2
// TODO: In C++17, we can use std::map::merge to move the table entries	2
// TODO: No, really, replace this legacy interface with one that just accepts	1
# TODO: take care of batch size	1
// TODO: Add estimated disk and memory size to SFrames.	1
// TODO do we need to reset the CTM on each call here? (And reapply our CG->Canvas transformations?)	1
//TODO: Why don't I have to take the address of cancel_handler?	1
// TODO: Get this to take an arbitrary function	1
{"content_feature", content_feature},  // TODO: refactor to take content name and style name	1
// TODO - error handling?	2
// TODO send kill signal to threads	1
// TODO: think about templating to make this simpler.	1
// TODO: optimize this, as some of the range is already sorted	1
// TODO: Potentially plumb `::google::protobuf::MessageLite` to variant	1
* TODO: write what needs to be done for batching	1
// TODO: The original Python implementation also exposed "anchors",	1
TODO: Switch to gl_sarray and the new SDK implementation.	1
// TODO - what if it has values > 32 bit? MLMultiArray only supports 32-bit ints.	1
// TODO: Codify priority levels	1
// TODO: Support additional layers (and further parameterize the above) as	1
// TODO: Should accept model_backend as an optional argument to avoid	1
# TODO: Evaluate method to update lr in set_learning_rate()	1
* TODO: this should	1
// TODO: Expose via forthcoming C-API checkpointing mechanism?	1
// TODO -- what should we assert here instead, to make sure we have enough	1
*  TODO: make it a general function of setting up a bipartite graph	1
// TODO: eventually we should probably have an "undefined" bin	1
# TODO: allow autodetection of light/dark mode.	2
// TODO: Ideally this would not require copying. But we should move away	1
// TODO: Replace these with an instance of TCModelTrainerBackendGraphs?	1
* TODO: Implement batching for multiple images	1
// TODO not implemented yet.	1
// TODO: Expose these from DarknetYOLOCheckpoint.	1
//TODO: Perhaps combine the sframe and the join positions into a struct?	1
// TODO: This next section could benefit from an iterator over floats,	1
// TODO: MANY MANY HEURISTICS	1
// TODO: raise exception	2
// TODO: force has key	2
// TODO: Clean the choice of data_type up	1
// TODO: Only sort the largest num_words	1
// TODO: Encapsulate these network architecture parameters better, to support	1
// TODO: Remove the inheritance from ipc_object_base once the cppipc code has	1
# TODO: early stopping	1
// TODO: Clean up logic surrounding batch size and memory budget.	1
#pragma clang diagnostic ignored "-Woverloaded-virtual"  // TODO: fix these issues below	1
/* TODO: Design Better Error Message Compornent */	1
// TODO: filter annotation  array use `x` as index into the array	1
// TODO Test	2
// TODO: This can be optimized by precomputing this for all zones outsize	1
// TODO: This copy is inefficient. We can construct an NSData that wraps a	1
// TODO: When we support alternative base models, we will have to generalize.	1
// TODO: Move this into object_detection::ModelTrainer once the	1
// TODO -- actually show this with the axes the user asked for	1
// TODO: augment -> train	1
# TODO: Include self.labels: feed_dict['label'] to handle labels from validation set	1
// TODO: Remove this model-specific code once the inference path no longer	1
// TODO BUG? Custom model doesn't seem to allow optional outputs	1
// TODO: write function to check valid vega spec	1
// TODO: Remove this type alias once this class stops inheriting from	1
// TODO: Check that statement	1
# TODO: Converge to NCHW everywhere.	1
x[i] = DenseVector(1); // TODO: Why doesn't 0 work?	1
// TODO: If we standardize on NCHW in the toolkit code, then we can avoid the	1
* TODO Refactor this into subclasses + subfunctions	1
// TODO: find a better way of initializing vertex_data	1
// TODO: Codify priority levels?	2
// TODO: Use an Optional type instead.	1
// TODO: Recycle non-temporary MPSImage instances using image allocators	1
// TODO: Performance, instead of performing a string comparison everywhere,	1
// @TODO: When we make a dylib, we'll need some sophisticated	1
// TODO: add free all matrices and array. Mostly copy_weight_matrices_	1
// TODO: Validate the dictionary keys in compute_properties. Let downstream	1
# TODO we can link the requirements from here when target_link_libraries	1
* TODO: will slowly move away all users of this function to get_lazy_sarray	1
// TODO: This class should be responsible for producing the augmenter itself.	1
// TODO should we be updating _font to reflect the system font we've fallen back to?	1
// TODO - not sure what to do here.	1
// TODO: Remove this method. Just let subclasses define the entire training	1
// TODO: add options.	1
// TODO: Raise exception	1
// TODO: Unify with the logic in supervised_learning_model_base::api_evaluate,	1
//TODO: What happens if a join key (or part of one) is NULL?	1
// TODO: The original Python implementation only sampled the training data for	1
// TODO distinguish list and vector	1
// TODO: revisit the code when we actually have vertex groups	1
// TODO: Images intended for clients to (optionally) read should be	1
// TODO: Remove this eager construction once we stop putting weak pointers in	1
// TODO make add_element take a sframe_row::row_reference instead	1
// TODO: optimize this.	1
//TODO: Implement this functionality in Windows.	1
std::string output_type);  // TODO: This should be const	1
// TODO: Remove this proxy subclass once the dependency on cppipc has been	1
// TODO: This function should be const	3
//TODO: This is a HUGE hack. From the Python side, a list of ints is turned	1
// TODO - what should this do on other platforms?	1
// TODO: If gl_sframe_range::end() were a const method, we wouldn't need to	2
/* TODO: Indicator to Show how many Images were Annotated */	1
* TODO: This code is *very* similar to the code in supervised_learning.cpp.	1
// TODO: add multiple batch sizes	2
// TODO: Where is the right place for this? Probably not here...	1
std::vector<float> hwc_data(chw_array.size());  // TODO: Avoid initializing	1
// TODO: Adopt EncodedBatch instead.	1
# @TODO: Think about how to bypass the md5 checksum if the user wants to	1
const static size_t NUM_THREADS = 6; // handle requests on 6 threads - TODO optimize this number	1
// TODO: Factor out more shared code from FillLossLabelsBatch and the below?	1
//TODO: Break this into its own function	1
// TODO: Just save this value before writing it to disk.	1
// TODO: modifications to the cursor attribute should actually be applied	1
<< std::endl;  // TODO: finish	1
// TODO Implement + Test	1
// TODO: implement `background_work` used with Image Saliency for OD	1
// TODO: refactor code to be more readable with loops	2
// TODO: The original Python code path only runs the model selector on at most	1
#pragma clang diagnostic ignored "-Woverloaded-virtual"  // TODO: fix this issue	1
//TODO: This is what will run out of memory first when scaling up	1
// TODO: implement `cast_annotations` used with Image Saliency for OD	1
// TODO: remove 'is_dense' once the ball tree is updated to use distance components	1
// TODO - do we need to do anything here? perhaps not.	2
// TODO: Investigate parallelizing this file I/O.	1
// TODO: Implement "report_by_class" using the standard evaluation framework.	1
//TODO: This has a race condition that I'm ignoring as I don't think it will	1
// TODO: Remove this method. It is only called by the base class implementation	1