Rationale for .predict being a method of Image class rather than Learner?

HamsterHuey · October 28, 2018, 5:30pm

The _predict method is defined in learner.py in the vision module and then bound to the Image class.

fastai/fastai/blob/7d681372cf8aa1f37a05fd161b655c8f58d1a07f/fastai/vision/learner.py#L127


    plt.xlabel('Predicted')


def most_confused(self, min_val:int=1)->Collection[Tuple[str,str,int]]:
    "Sorted descending list of largest non-diagonal entries of confusion matrix"
    cm = self.confusion_matrix()
    np.fill_diagonal(cm, 0)
    res = [(self.data.classes[i],self.data.classes[j],cm[i,j])
            for i,j in zip(*np.where(cm>min_val))]
    return sorted(res, key=itemgetter(2), reverse=True)


def _predict(img, learn):
img = apply_tfms(learn.data.valid_ds.tfms, img, **learn.data.valid_ds.kwargs)
ds = TensorDataset(img.data[None], torch.zeros(1))
dl = DeviceDataLoader.create(ds, bs=1, shuffle=False, device=learn.data.device, tfms=learn.data.valid_dl.tfms,
                             num_workers=0)
return get_preds(learn.model, dl, cb_handler=CallbackHandler(learn.callbacks, []))[0][0]


Image.predict = _predict

In general, I am curious as to why the predict method was exposed in the Image class as opposed to the Learner or ConvLearner class? It seems a lot more intuitive to expect a trained model in Learner to have a .predict() method that takes in an image (or array of Images) rather than to expect an Image class to have a .predict() method that takes in a learner object? Perhaps I am missing some subtle aspect of how things are implemented.

Cheers

jeremy · October 28, 2018, 5:34pm

Yes I agree - I was looking at this this morning and I think it’s a mistake. Unless @sgugger convinces me otherwise, expect to see this changed later today.

devforfu · October 29, 2018, 12:27pm

Agree, it was a bit confusing for me as well when I was trying to run the model against custom data. If we take into account other frameworks, like, sklearn or keras, they always provide various prediction methods for their models.