Deep Learning IndexError: too many indices for array - python-2.7

Question

I am trying to train the system on some data, Sound_Fc is a 16X1 float array.
for i in range(0,26983):
Block_coo = X[0,i]
Fc = Block_coo[4]
Sound_Fc = Fc[:,0]
Vib_Fc = Fc[:,1]
y = np.matrix([[1.0],[2],[3],[4],[5],[6],[7],[8],[9],[10],[11],[12],[13],[14],[15],[16]])
testX, trainY, testY) = train_test_split(
Sound_Fc, y, test_size = 0.33, random_state=42)
dbn = NeuralNet(
layers=[
('input', layers.InputLayer),
('hidden', layers.DenseLayer),
('output', layers.DenseLayer),
],
input_shape = (None, trainX.shape[0]),
hidden_num_units=8,
output_num_units=4,
output_nonlinearity=softmax,
update=nesterov_momentum,
update_learning_rate=0.3,
update_momentum=0.9,
regression=False,
max_epochs=5,
verbose=1,
)
dbn.fit(trainX,trainY)
But I'm getting this error
Warning (from warnings module):
File "C:\Users\Essam Seddik\AppData\Roaming\Python\Python27\site-packages\sklearn\cross_validation.py", line 399
% (min_labels, self.n_folds)), Warning)
Warning: The least populated class in y has only 1 members, which is too few. The minimum number of labels for any class cannot be less than n_folds=5.
Traceback (most recent call last):
File "C:\Essam Seddik\Deep Learning Python Tutorial\DNV_DeepLearn.py", line 77, in <module>
dbn.fit(trainX,trainY)
File "C:\Python27\lib\site-packages\nolearn-0.6adev-py2.7.egg\nolearn\lasagne\base.py", line 293, in fit
self.train_loop(X, y)
File "C:\Python27\lib\site-packages\nolearn-0.6adev-py2.7.egg\nolearn\lasagne\base.py", line 300, in train_loop
X, y, self.eval_size)
File "C:\Python27\lib\site-packages\nolearn-0.6adev-py2.7.egg\nolearn\lasagne\base.py", line 401, in train_test_split
kf = StratifiedKFold(y, round(1. / eval_size))
File "C:\Users\Essam Seddik\AppData\Roaming\Python\Python27\site-packages\sklearn\cross_validation.py", line 416, in __init__
label_test_folds = test_folds[y == label]
IndexError: too many indices for array
I tried xrange instead of range, and y=list() instead of the defined y. I tried also small numbers in the for loop range like 5, 10 and 100 instead of 26983.
I tried np.array and np.ndarray and np.atleast_2d. Nothing works !

Answer 1

At every iteration of that loop you are overwriting Sound_Fc. So, at the end of the loop, the value of Sound_Fc is X[0,26982][4][:,0]. You are also overwriting y with the same value over and over again at each iteration of the loop, it is basically a vector with values from 1 to 16. Basically, your total data is 16 points, and the y value of each is a unique value (something between 1 and 16). Then you split this into training and test data, so you make 5 of these 16 points your test set, and 11 of them your training set. With only a single example for each observed y output, your network model is complaining that it cannot extract enough information to predict these y values in the future.
If I understand correctly, instead of overwriting Sound_Fc and y at each iteration, you want to append them to growing x and y vectors. You can do this by using vstack which vertically stacks numpy arrays. Replace the following with that loop:
Sound_Fc = np.vstack( [X[0,i][4][:,0] for i in range(26983)] )
y = np.vstack([np.matrix(range(1,17)).T for i in range(26983)])
Before, the Sound_Fc had the shape (16,1) when you used it as your feature vector. Now it will have the shape (431728, 1). That number is 26983*16, since you're stacking 26983 vectors that have 16 elements each. Your y will have the shape (431728, 1).
[X[0,i][4][:,0] for i in range(26983)] creates a list of 26983 elements, each element is a (16,1) shape numpy array. np.vstack stacks them vertically to get a single, tall, (431728, 1) array. This is your feature vector.
np.matrix(range(1,17)) creates a matrix with the elements 1 to 16. This is of shape (1, 16). By taking its transpose with .T, we make it vertical and now its shape is (16,1). Again, we make a list from 26983 of these, and vstack them to get a (431728, 1) shape vector that basically goes from 1 to 16 and then goes to 1 again and to 16 again, basically repeats that 1-16 pattern over and over again. This is your output vector. Now, for each output (let's say 8, for instance), you have 26983 data points to learn from (well it will be 17809 once you split .66 of all this to be your training set) . Now your model will not complain about not having enough examples for a specific y output.
There might of course be other errors related to other stuff (I can't see your data -- I don't know what's in that big X).