6d3ec6c28f32560121d93b8fcab2cf7924f59dc5,thinc/layers/pytorchwrapper.py,,forward,#Any#Any#Any#,27

Before Change

    else:
        pytorch_model.eval()
        x_args, x_kwargs = model.prepare_input(x_data, is_update=False)
        with torch.no_grad():
            y_var = pytorch_model(*x_args, **x_kwargs)
        self._model.train()
        return model.prepare_output(y_var)

    y = model.prepare_output(y_var)

    def backward_pytorch(dy_data):
        d_args, d_kwargs = model.prepare_backward_input(dy_data, y_var)
        torch.autograd.backward(*d_args, **d_kwargs, retain_graph=True)

After Change

    pytorch_model = model.shims[0]

    X_torch = xp2torch(X, requires_grad=is_train)
    Y_torch, torch_backprop = pytorch_model((X_torch,), {}, is_train)
    Y = torch2xp(Y_torch)

    def backprop(dY):

Instances