[WIP] Include alignment steps in coffeine pipeline #58
Conversation
apmellot
changed the title
coffeine/transfer_learning.py
Outdated
| Parameters | ||
| ---------- | ||
| metric : str, default='riemann' | ||
| The metric to compute the mean. |
coffeine/transfer_learning.py
Outdated
| def __init__(self, domains, metric='riemann'): | ||
| self.domains = domains | ||
| self.metric = metric | ||
| self.re_center = TLCenter('target_domain', metric=self.metric) |
There was a problem hiding this comment.
This should not be put in the init
Init should just set the init parameters as attributes
coffeine/transfer_learning.py
Outdated
| def fit(self, X, y): | ||
| X = _check_data(X) | ||
| _, y_enc = encode_domains(X, y, self.domains) | ||
| self.means = self.re_center.fit(X, y_enc).recenter_ |
There was a problem hiding this comment.
re_center_ as it's set and modified in the fit
And means_
coffeine/transfer_learning.py
Outdated
| self.metric = metric | ||
| self.re_scale = TLStretch('target_domain', | ||
| center_data=True, | ||
| metric=self.metric) |
coffeine/transfer_learning.py
Outdated
| y: ndarray, shape (n_matrices,) | ||
| Labels for each matrix. | ||
| domains: ndarray, shape (n_matrices,) | ||
| Domains for each matrix. |
coffeine/pipelines.py
Outdated
| else: | ||
| steps = [ | ||
| projection(**projection_params_) | ||
| ] + alignment_steps + [ |
There was a problem hiding this comment.
Can we be sure that this not buggy? We need to make sure that for every column (‘name’) independent instances are instantiated. Otherwise you risk some stateful behavior.
coffeine/pipelines.py
Outdated
| if 're-center' in alignment: | ||
| alignment_steps.append(ReCenter(domains=domains)) | ||
| if 're-scale' in alignment: | ||
| alignment_steps.append(ReScale(domains=domains)) |
There was a problem hiding this comment.
See comment above. We would use the same ReCenter/ReScale instances for every column of the data frame based on this code here, right?
There was a problem hiding this comment.
I don't understand, isn't it the same thing as what is done a few lines above for the projection and vectorization steps?
Also in practice I checked and each frequency band is aligned independently.
There was a problem hiding this comment.
It’s not the same as we instantiate the classes independently for every column. We can check of course if this is practically unneccessary by now because of internal cloning etc. But if you want to be consistent with the tested / save code you can just adopt the same pattern as for the other projection/vectorization steps. You currently instantiate the classes globally and plug in the same input column by column.
There was a problem hiding this comment.
Ok I see, let's do it as projection/vectorization.
|
Some missing milestones:
|
|
Test failing is related to this issue in pyRiemann: pyRiemann/pyRiemann#257 |
@apmellot what’s your philosophy here? Fixing it in PyRiemann and waiting for that? Implementing a patch locally and then use PyRiemann in the future once it is fixed there? |
The PR is open in pyRiemann, we can wait for it to be merged. |
|
We will need a patch to support earlier versions of PyRiemann - otherwise we depend on master branch. We can add a checker that sees the pyriemann version and depending on the version supplies our patched function or uses the upstream one. |
For now, I only added the re-centering and the re-scaling step between projection and vectorization. I tested it with TUAB and LEMON and the results are as expected.
Next, I want to add the rotation correction step still using pyRiemann, create an example notebook and make nice docstrings.