Nilearn: Error using SpaceNetRegressor

1

Hi all,

Encountering a strange error using the SpaceNetRegressor fit function. My input is a list of images y_train and an array of scalars X_train. Each are the same length. All the images are correct and the same dimensions.

Here’s my (simplified) code:

decoder = SpaceNetRegressor(mask=V_img, penalty=params_snet['penalty'],
                        eps=params_snet['eps'],  # prefer large alphas
                        memory=params_snet['memory'])

istrain = train_set[:,f]
y_train = Xs[glm][istrain][:,0]
X_train = np.array(V_betas)
X_train = list(X_train[istrain])

print(len(X_train))
print(y_train.size)
print(X_train[0].shape)

decoder.fit(X_train, y_train)

The output is:

120
120
(182, 218, 182)
[NiftiMasker.fit] Loading data from None
[NiftiMasker.fit] Resampling mask
/Users/lpzatr/anaconda3/lib/python3.6/site-packages/nilearn/_utils/cache_mixin.py:291: UserWarning: memory_level is currently set to 0 but a Memory object has been provided. Setting memory_level to 1.
  warnings.warn("memory_level is currently set to 0 but "
/Users/lpzatr/anaconda3/lib/python3.6/site-packages/nilearn/decoding/space_net.py:195: RuntimeWarning: divide by zero encountered in log10
  return np.logspace(np.log10(alpha_min), np.log10(alpha_max),
/Users/lpzatr/anaconda3/lib/python3.6/site-packages/numpy/core/function_base.py:117: RuntimeWarning: invalid value encountered in double_scalars
  delta = stop - start
/Users/lpzatr/anaconda3/lib/python3.6/site-packages/nilearn/decoding/proximal_operators.py:20: RuntimeWarning: invalid value encountered in maximum
  shrink[y_nz] = np.maximum(1 - alpha / np.abs(y[y_nz]), 0)
/Users/lpzatr/anaconda3/lib/python3.6/site-packages/nilearn/decoding/space_net.py:274: RuntimeWarning: invalid value encountered in reduce
  if w.ptp() == 0:

...

~/anaconda3/lib/python3.6/site-packages/scipy/stats/mstats_basic.py in spearmanr(x, y, use_ties)
    457     df = n-2
    458     if df < 0:
--> 459         raise ValueError("The input must have at least 3 entries!")
    460 
    461     # Gets the ranks and rank differences

ValueError: The input must have at least 3 entries!

Any idea where to look for issues? No idea why the input is less than 3 entries at this point.

2

I wonder if the strange behavior (which I might agree calling a bug) comes from the fact that you are giving a list of images. Do you reproduce the problem with a 4D images?

3

Hi Gael,

Thanks for your quick response. I ran this using a 4D image (assuming the order is x,y,z,subj?) and it seemed to get a bit further, but ultimately gave the same error:

istrain = train_set[:,f]
y_train = Xs[glm][istrain][:,0]
X_train = V_betas[:,:,:,istrain]

X_train_img = nib.Nifti1Image(X_train, V_img.affine, V_img.header)

print(y_train.size)
print(X_train.shape)

decoder.fit(X_train_img, y_train)

Output:

120
(182, 218, 182, 120)
[NiftiMasker.fit] Loading data from None
[NiftiMasker.fit] Resampling mask
/Users/lpzatr/anaconda3/lib/python3.6/site-packages/nilearn/_utils/cache_mixin.py:291: UserWarning: memory_level is currently set to 0 but a Memory object has been provided. Setting memory_level to 1.
  warnings.warn("memory_level is currently set to 0 but "
[NiftiMasker.transform_single_imgs] Loading data from Nifti1Image(
shape=(182, 218, 182, 120),
affine=array([[  -1.,    0.,    0.,   90.],
       [   0.,    1.,    0., -126.],
       [   0.,    0.,    1.,  -72.],
       [   0.,    0.,    0.,    1.]])
)
[NiftiMasker.transform_single_imgs] Extracting region signals
[NiftiMasker.transform_single_imgs] Cleaning extracted signals

...

~/anaconda3/lib/python3.6/site-packages/scipy/stats/mstats_basic.py in spearmanr(x, y, use_ties)
    457     df = n-2
    458     if df < 0:
--> 459         raise ValueError("The input must have at least 3 entries!")
    460 
    461     # Gets the ranks and rank differences

ValueError: The input must have at least 3 entries!

I’m using Jupyter Notebook so it’s tricky to debug, but I’ll get it set up in an IDE and try and dig further.

4

OK, that’s not the problem.

Looking at the traceback better, the input is three entries in the inner cross-validation loop that is used to set parameters, in particular in the test set there. You are not specifying a cv, are you?

5

Nope, I haven’t set a cv parameter. According to the docs that means it uses a default of 8 folds…

Just ran it with cv=None (so I guess 3 folds) and it gives the same error.

6

@andrew.reid Is it easy for you to give X_train_img, y_train uploading somewhere ?

It helps to diagnose the error.

Thanks

7

Hi!

In the process of doing that, I’ve found the issue. I was using my intercept column (all 1’s) instead of an actual data column as the y input. This cause an np.dot step to return nans and everything failed at that point!

It appears to be running fine now. Thanks for your help anyways… love the nilearn package!

8

OK, good that you diagnosed the problem. We need to add a test to capture such situation and raise a meaningful error.

Good you please add an issue to the nilearn issue tracker to explain the problem.

Thanks!

9

I am in the process of preparing a PR for that.