t-SNE Model Understanding #1731
Conversation
Codecov Report
@@ Coverage Diff @@
## main #1731 +/- ##
=========================================
+ Coverage 100.0% 100.0% +0.1%
=========================================
Files 243 243
Lines 19353 19430 +77
=========================================
+ Hits 19344 19421 +77
Misses 9 9
Continue to review full report at Codecov.
|
ParthivNaresh
requested review from
freddyaboulton,
dsherry,
angela97lin,
bchen1116,
chukarsten and
jeremyliweishih
and removed request for
freddyaboulton and
dsherry
There was a problem hiding this comment.
I think all this looks good to me. I'm not 100% clear on tsne, but it looks like dimensionality reduction and when I .show()'ed the figs it seemed consistent.
One thing that other graph tests have done is test that the underlying algorithm's data is present in the fig_dict. ala fig_dict_data['data'][0]['values'] == tsne(params) I think just to be consistent with the others you should probably throw that in here. Everything else looks excellent.
evalml/model_understanding/graphs.py
Outdated
| return X_new | ||
|
|
||
|
|
||
| def graph_t_sne(X, n_components=2, perplexity=30.0, learning_rate=200.0, metric='euclidean', marker_line_width=2, marker_size=7): |
There was a problem hiding this comment.
Not sure if I'm missing something, but I don't see anything graphed in both my Jupyter notebook and jupyter lab, even after I run make installdeps.
Are you able to get it to show in both notebook and lab? Can you post a screenshot of the resulting graph?
Also, it takes over a minute to finish this method, whereas t_sne was fairly quick. Is this the same for you?
There was a problem hiding this comment.
@bchen1116 Should look something like this, and I'm able to get the results instantly but I'm also using a small dataset. I'm sure the time increases exponentially with the dataset size
There was a problem hiding this comment.
@ParthivNaresh Looks great! The one thing I want to resolve before merge is the marker_color=color line in graph_t_sne. I think it's slowing us down. The rest are non-blocking nitpicks on docstrings and a minor test change.
…eryx/evalml into 1709-t-SNE-Model-Understanding
| if not marker_size >= 0: | ||
| raise ValueError("The parameter marker_size must be non-negative") | ||
|
|
||
| X_embedded = t_sne(X, n_components=n_components, perplexity=perplexity, learning_rate=learning_rate, metric=metric, **kwargs) |
There was a problem hiding this comment.
Maybe we should hard-code n_components=2 and remove it as a function parameter? I think it would avoid errors in n_components=1 and I don't know if we should let users pass >=3 components if we're only going to plot the first two anyways?
Fixes #1709