[ENH] OWTSNE: Offload computation to separate thread

[pavlin-policar]

Issue

The t-SNE widget previously performed all computation inside the main thread, while simulating a thread i.e. it would pause every once in a while and handle user action. For larger data sets, this meant the widget would freeze for a while until the widget received control again.

Description of changes

The widget now offloads all of the computation into a separate thread, and the widget is only responsible for preparing a Task specification, which is then handled by the other thread.

Other changes include:

• The widget now remembers partial results better, i.e. changing the exaggeration parameter doesn't recompute all the nearest neighbours again.
• More descriptive status message and the progress bar now takes into account all the steps of the procedure.
• Since t-SNE is deterministic, once the optimization is complete, pressing "Start" does nothing now, because the results will invariably the same.
• The PCA slider now changes its maximum value based on the number of features in the data.
• Add Info thing to indicate that the settings have been changed and the user should rerun the optimization.
• I've made the t-SNE API a bit more granular so that it's possible to gain access to the affinities (nearest neighbors). In principle, this is a bad idea, because this means the widget is tied to my openTSNE implementation, but I think it's justified because 1. in case we ever opt for a different implementation, removing this from the widget requires little effort and 2. finding the nearest neighbors can take a very long time (up to half of all the computation time for larger data sets) therefore the widget should be smart enough to cache results wherever possible.
• The widget now replots fairly frequently again. I did this because it seems that doing the work in a separate thread and drawing in the main one play well together, and the speed difference between updating often and updating rarely is small.

Includes

• Code changes
• Tests
• Documentation

[codecov]

Codecov Report

Merging #3604 into master will increase coverage by 0.05%.
The diff coverage is 96.64%.

@@            Coverage Diff             @@
##           master    #3604      +/-   ##
==========================================
+ Coverage    84.9%   84.95%   +0.05%     
==========================================
  Files         374      374              
  Lines       68852    69084     +232     
==========================================
+ Hits        58457    58692     +235     
+ Misses      10395    10392       -3

[VesnaT]

I think this line is not needed, since the text is already set in start().

[VesnaT]

Maybe you could write some tests for the TSNERunner. Each its method is easily testable. And it would also be good to check that the embedding is not modified inplace.

[VesnaT]

Some observations:

• OWFile (Iris) -> OWtSNE (Check Preserve global structure -> Info message appears) -> OWFile (Reload file) -> OWtSNE (Info message has disappeared)
• OWFile (Iris) -> OWtSNE
  -> OWDataTable (Select 10 rows) -> OWtSNE (the same widget as above (input subset)) -> OWtSNE (Check Preserve global structure) -> OWDataTable (the same widget as above - reselect 10 rows) -> Crash

Traceback (most recent call last):
  File "/Users/vesna/orange3/Orange/canvas/scheme/widgetsscheme.py", line 1082, in process_signals_for_widget
    widget.handleNewSignals()
  File "/Users/vesna/orange3/Orange/widgets/visualize/utils/widget.py", line 461, in handleNewSignals
    self.graph.update_point_props()
  File "/Users/vesna/orange3/Orange/widgets/visualize/owscatterplotgraph.py", line 554, in update_point_props
    self.update_colors()
  File "/Users/vesna/orange3/Orange/widgets/visualize/owscatterplotgraph.py", line 932, in update_colors
    pen_data, brush_data = self.get_colors()
  File "/Users/vesna/orange3/Orange/widgets/visualize/owscatterplotgraph.py", line 826, in get_colors
    subset = self.master.get_subset_mask()
  File "/Users/vesna/orange3/Orange/widgets/visualize/utils/widget.py", line 482, in get_subset_mask
    dtype=np.bool, count=len(valid_data))
  File "/Users/vesna/orange3/Orange/widgets/visualize/utils/widget.py", line 481, in <genexpr>
    return np.fromiter((ex.id in self.subset_indices for ex in valid_data),
TypeError: unhashable type: 'numpy.ndarray'

• When running tsne with several interruptions I get slightly different results than without interrupting calculation:
  OWFile (Iris) -> OWtSNE (Click Stop/Resume several times)
  -> OWtSNE
  -> Output Data vary

[VesnaT]

The following still happens:

• OWFile (Iris) -> OWtSNE (Check Preserve global structure -> Info message appears) -> OWFile (Reload file) -> OWtSNE (Info message has disappeared)
• OWFile (Iris) -> OWtSNE
  -> OWDataTable (Select 10 rows) -> OWtSNE (the same widget as above (input subset)) -> OWtSNE (Check Preserve global structure) -> OWDataTable (the same widget as above - reselect 10 rows) -> Crash

[pavlin-policar]

@VesnaT I fixed the two issues you found above. I've fixed them in a very simple and hacky way. I tried to find an elegant solution, but any of these would require touching the base widget and then fixing the other widgets.