error from metrics=[accuracy] in lesson-4.imdb about fastai

[jph00]

Ah ok that's a different part of the NB to where I was looking. I'll check it out. BTW I check forums much more often than GH so that's the best place to report issues.

[jph00]

OK I've pushed an accuracy_np function to fix this.

[MattKleinsmith]

Regression occurs in the lesson1 notebook accuracy calls, too. (On a deadline atm. Will fix if it exists when I'm free.)

https://github.com/fastai/fastai/blob/master/courses/dl1/lesson1.ipynb

[apiltamang]

@kevindewalt
Might be a cause of a recent change (minor) I made to the 'fit' method. I will look onto it over the weekend. :)

[interogativ]

This is because you are comparing an autograd variable to a tensor. Here's how I got around this:

def accuracy1(preds, targs):
    preds = torch.max(preds, dim=1)[1]
    #print ('preds: {}, {}'.format(preds.shape,type(preds)))
    #print ('targs: {}, {}'.format(targs.shape,type(targs)))
    return (preds==targs.data).float().mean()

[jph00]

I recently changed metrics to use pytorch tensors rather than numpy arrays. So it looks like this is an example of a regression. I'll take a look shortly.

[jph00]

@MattKleinsmith it works for me in in the lesson 1 notebook. Are you sure you have git pull up to date, and are using the latest version of the notebooks?

[MattKleinsmith]

@jph00 My mistake. Pulling and testing now.

[MattKleinsmith]

@jph00 I'm getting:

torch.max received an invalid combination of arguments - 
got (numpy.ndarray, dim=int), but expected one of:
 * (torch.FloatTensor source)
[...]

If you go to lesson1.ipynb and CTRL+F "accuracy(" you'll see that probs, an argument to accuracy, comes from np.mean, making it a numpy array.