error on running LFDA algorithm

[piaozhx]

hi, I want to use LFDA algorithm for my data but have some confused problem.
firstly, I try the example code in the document, every thing is ok. but when I run this for myself data, It got an error:

  File "unit_test.py", line 245, in person_reid_LFDA
    lfda.fit(train_X, train_Y)
  File "/usr/local/lib/python2.7/dist-packages/metric_learn/lfda.py", line 105, in fit
    which='LA')
  File "/usr/local/lib/python2.7/dist-packages/scipy/sparse/linalg/eigen/arpack/arpack.py", line 1605, in eigsh
    params.iterate()
  File "/usr/local/lib/python2.7/dist-packages/scipy/sparse/linalg/eigen/arpack/arpack.py", line 537, in iterate
    self.ipntr, self.workd, self.workl, self.info)
ValueError: On entry to DLASCL parameter number 4 had an illegal value

in this case, train_X shape is (632, 4096), train_Y shape is (632,), and my code is:

    lfda = LFDA(k=5, dim=128)
    lfda.fit(train_X, train_Y)

Is there any idea for this error?

[perimosocordiae]

Looks like you might have an issue with your scipy installation. What dtypes are train_X and train_Y?

[piaozhx]

@perimosocordiae Thanks for your reply! but I'm sure that train_X and train_Y dtype are both "float32", which should not cause the error.

I think I find the problem, the latest version in pip installation is 0.3.0, but code in here is 0.4.0. I install the 0.4.0 code and It can work well. I think line 142 figure out the ValueError in my case.

But when I want to use the lfda.metric() to get the metric matrix, I get a new error, which seems to think that lfda.metric is a str type.I review the code in line 51, I think this is a bug.

when I replace the "metric" to "_metric", I find that It will raise another confused error:

File "/usr/lib/python2.7/dist-packages/sklearn/base.py", line 141, in _pprint
    this_repr = '%s=%s' % (k, printer(v))
  File "/usr/lib/python2.7/dist-packages/sklearn/base.py", line 277, in __repr__
    offset=len(class_name),),)

I try several different sense, when I don't do the replacement, in other words, let

self.metric = metric # metric is a str

every thing is OK except can't get the metric matrix.

when I do the replace, commit line 131, then it works well, I can get the metric matrix by calling the lfda.metric() but I don't know why it works.

[perimosocordiae]

You're right, this is a bug! It was triggered by the recent efforts to make things more sklearn-compatible, and somehow wasn't caught by the unit tests.

Until a fix is submitted, one workaround is to use the transformer() method and create the metric matrix yourself, in the same way the base class does it: https://github.com/all-umass/metric-learn/blob/238be72c00fde4f4351601ef1ce8c2fcd710c9e3/metric_learn/base_metric.py#L19