[MRG] Random sampler #295
[MRG] Random sampler #295
Conversation
|
I tried my fonction on dataset. It returns sampling as i want. But i'm not sure that is what we want. I choose a random number of subject, I choose a random class (every class has the same probabilitie to be chosen) and I create the sequence with this window. Because all the class have the same probabilitie i thought it was balanced. |
|
I mean does it enable you to get better performance on real data?
does it lead to a running time increase when learning?
basically is it useful according to you and if so is it a free lunch in
running time
… |
Codecov Report
@@ Coverage Diff @@
## master #295 +/- ##
==========================================
+ Coverage 80.38% 80.59% +0.20%
==========================================
Files 49 49
Lines 3085 3123 +38
==========================================
+ Hits 2480 2517 +37
- Misses 605 606 +1 |
braindecode/samplers/base.py
Outdated
| @@ -81,6 +87,10 @@ def __iter__(self): | |||
| def n_recordings(self): | |||
| return self.info.shape[0] | |||
|
|
|||
| @property | |||
There was a problem hiding this comment.
Might be actually safer to supply n_classes explicitly? Otherwise you may run into problems:
- if in some subset not all classes are present (imagine small evaluation subset in debugging experiment)
- targets somewhere else as in Discrete and synchronized targets support #261 may not work.
There was a problem hiding this comment.
The first point is fixed in the latest commit. I'm not sure about the second point though - the sampler requires categorical targets in the BaseConcatDataset's metadata attribute. What use case exactly did you have in mind @robintibor ?
|
I added tests and uniformized the implementation so it follows the other samplers' logic more closely. Next step is to test #282 with this sampler and make sure we get better results than with naive sampling of sequences. |
Co-authored-by: Alexandre Gramfort <alexandre.gramfort@m4x.org>
Co-authored-by: Alexandre Gramfort <alexandre.gramfort@m4x.org>
Co-authored-by: Alexandre Gramfort <alexandre.gramfort@m4x.org>
Co-authored-by: Alexandre Gramfort <alexandre.gramfort@m4x.org>
Co-authored-by: robintibor <robintibor@gmail.com>
- modifying _init_info to check for required keys in metadata when creating info dataframe - reviewing computation of start_ind - improving docstrings - adding tests
hubertjb
force-pushed
the
random-sampler
branch
from
69734db
to
9a0b77c
Compare
|
Thanks @robintibor ! The validation of the USleep architecture might lead to some minor changes, but otherwise I think this is ready to be merged. |
|
Great, merged! |
Add a new sampler inspired by USleep article. We define a fixed number of sequence. For each sequence, a window is chosen for a random recording and a random class. The sequence is created around this window with a random position. Each class has the same probability to be chosen. So this method allow to equitably represent the classes despite their difference in proportion.
@agramfort @hubertjb @l-omar-chehab