Error in model.fit() about deep-listening

[jaron]

Hi Thomas!

I think I see your problem - this particular part of the notebook requires the full UrbanSound8k dataset in the form of numpy objects in the directory data/us8k-np-cnn. I haven't checked that data into the Git repo as these files are huge (2GB).

You can generate that data, using the code in the section "Saving Extracted Features (optional)" in Notebook 3 - although I've just noticed the version I checked in creates only 3 folds in directory "data/us8k-np-cnn-mini", rather than the full 10 folds "data/us8k-np-cnn", so I've just committed a fix for that.

But if you do have all the data, you should see something like this...

*** Train on (1, 2, 3) Validate on 4 Test on 5 ***
fold1 features:  (5446, 60, 41, 2)
fold2 features:  (5388, 60, 41, 2)
fold3 features:  (5852, 60, 41, 2)
fold4 features:  (6048, 60, 41, 2)
fold5 features:  (5689, 60, 41, 2)
Building model...
Training model...
Train on 16686 samples, validate on 6048 samples
Epoch 1/2
16686/16686 [==============================] - 71s - loss: 2.2278 - acc: 0.2077 - val_loss: 1.9349 - val_acc: 0.4157
Epoch 2/2
16686/16686 [==============================] - 67s - loss: 1.9322 - acc: 0.3254 - val_loss: 1.7717 - val_acc: 0.3747
Evaluating model...
ROC: 0.808

Note the feature sizes of each of the folds. You can check the data exists as expected by looking at what Keras reports. I see in your screenshot it says "Train on 0 samples, validate on 0 samples" - whereas in mine it says "Train on 16686 samples, validate on 6048 samples".

If you'd like me to send you the numpy data objects by WeTransfer, drop me an email at [email protected]

Hope that helps!