A walk with fastai2 - Tabular - Study Group and Online Lectures Megathread

muellerzr · March 22, 2020, 5:14pm

The recording is available in this thread here (if you may not have access to this now that is okay). A basic summary is that when we call learn.get_preds (which predict does), we can pass in decoded=True. What this entails is that our loss function will decode the values. Why and how this could be different is say with CrossEntropy, we’ll actually run a softmax regularly, but our decoded values will also perform an argmax. Converting it to an actual class name is done through the dataloader’s decode_batch() function

github.com

fastai/fastai2/blob/1dfa40098a0892a67a1f4de548bd76e0d32ec1ec/fastai2/layers.py#L301


        if targ.dtype in [torch.int8, torch.int16, torch.int32]: targ = targ.long()
        if self.flatten: inp = inp.view(-1,inp.shape[-1]) if self.is_2d else inp.view(-1)
        return self.func.__call__(inp, targ.view(-1) if self.flatten else targ, **kwargs)


# Cell
@delegates(keep=True)
class CrossEntropyLossFlat(BaseLoss):
    "Same as `nn.CrossEntropyLoss`, but flattens input and target."
    y_int = True
    def __init__(self, *args, axis=-1, **kwargs): super().__init__(nn.CrossEntropyLoss, *args, axis=axis, **kwargs)
    def decodes(self, x):    return x.argmax(dim=self.axis)
    def activation(self, x): return F.softmax(x, dim=self.axis)


# Cell
@delegates(keep=True)
class BCEWithLogitsLossFlat(BaseLoss):
    "Same as `nn.CrossEntropyLoss`, but flattens input and target."
    def __init__(self, *args, axis=-1, floatify=True, thresh=0.5, **kwargs):
        super().__init__(nn.BCEWithLogitsLoss, *args, axis=axis, floatify=floatify, is_2d=False, **kwargs)
        self.thresh = thresh