Error with numpy variable Xarray

[TheodoreG]

Hello Thomas, I am running your BAD code. I have use the Matlab to change the ff1010bird dataset into .mat files and I set the feature_type='fbank_d_dd' in your create_hdf5_ff1010bird_public.py code.

However, after the script loaded .mat files the Xarray variable in your code has shape (7690, 200, 56), which is 3 dimensional.
But then in line 154 Xarray = np.ndarray.transpose(Xarray, (0, 3, 1, 2)) you treat this variable as a 4 dimensional array and it get error ValueError: axes don't match array.

Could you please help me with this problem? Thanks very much

[topel]

Hi,
indeed the data should be 4-d arrays as expected by Lasagne (Theano) to run 2-d conv layers: (nb_samples, nb_channels, width, height)

• If you use fbank+delta+delta-delta (fbank_d_dd), you need to reshape your arrays so that nb_channels=3
• If you use static fbank (fbank), then nb_channels should be equal to 1

If the Xarray variable has shape (7690, 200, 56), it means that you're using 56 static fbank coefficients, and you should add an extra dimension at index 1:

Xarray = Xarray[:, np.newaxis, :, :] in order to get a (7690, 1, 200, 56) shape

[TheodoreG]

Thank you so much! I am now running the whole training process and I got following result:
**** Epoch 12 of 30 took 185.667s
lr: 0.002415750 train loss: 0.276738 validation loss: 0.323787 acc: 88.541666 auc: 90.664391 ratio: 85.468974
valid set CM:
[[262 10]
[ 34 78]]
[[96.3 3.7]
[30.4 69.6]]
best auc score update: 90.66 %
current lr: 0.001207875000

Which is close to the leaderboard results.