When is it ok to not call the `super` constructor on extending (nn.)Module

bgraysea · January 14, 2021, 10:25am

When is it ok to not call the super constructor on extending (nn.)Module

From https://pytorch.org/docs/stable/generated/torch.nn.Module.html the example is

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()

However; I don’t see any call to super in

github.com

fastai/fastai/blob/3666130b6c5afb12c04d5e8e08252d7393c2b2c0/fastai/test_utils.py#L31


        return TensorDataset(x, a*x + b + 0.1*torch.randn(bs*n, 1))
    train_ds = get_data(n_train)
    valid_ds = get_data(n_valid)
    device = default_device() if cuda else None
    train_dl = TfmdDL(train_ds, bs=bs, shuffle=True, num_workers=0)
    valid_dl = TfmdDL(valid_ds, bs=bs, num_workers=0)
    return DataLoaders(train_dl, valid_dl, device=device)
# Cell
class RegModel(Module):
    def __init__(self): self.a,self.b = nn.Parameter(torch.randn(1)),nn.Parameter(torch.randn(1))
    def forward(self, x): return x*self.a + self.b
# Cell
@delegates(Learner.__init__)
def synth_learner(n_trn=10, n_val=2, cuda=False, lr=1e-3, data=None, model=None, **kwargs):
    if data is None: data=synth_dbunch(n_train=n_trn,n_valid=n_val, cuda=cuda)
    if model is None: model=RegModel()
    return Learner(data, model, lr=lr, loss_func=MSELossFlat(),
                   opt_func=partial(SGD, mom=0.9), **kwargs)

class RegModel(Module):
    def __init__(self): self.a,self.b = nn.Parameter(torch.randn(1)),nn.Parameter(torch.randn(1))

Is the Module in RegModel(Module) the same as nn.Module?
help(Module) gives alot of output; and looks like its directly from pytorch

bgraysea · February 12, 2021, 4:02pm

Haha; the only difference between nn.Module (pytorch) and Module (fastai) is explained - https://github.com/fastai/fastai/blob/aaccca299b0dbf36a8fcf12d2f47430821f10a7d/fastai/torch_core.py#L569

class Module(nn.Module, metaclass=PrePostInitMeta):
    "Same as `nn.Module`, but no need for subclasses to call `super().__init__`"