Method test_weights

test/network/test_connections.py:73–164

(self, conn_type, shape_a, shape_b, shape_w, *args, **kwargs)

                assert d == torch.device("cuda:0")

    def test_weights(self, conn_type, shape_a, shape_b, shape_w, *args, **kwargs):
        print("Testing:", conn_type)
        time = 100
        weights = [None, torch.Tensor(*shape_w)]
        wmins = [
            -np.inf,
            0,
            torch.zeros(*shape_w),
            torch.zeros(*shape_w).masked_fill(
                torch.bernoulli(torch.rand(*shape_w)) == 1, -np.inf
            ),
        ]
        wmaxes = [
            np.inf,
            0,
            torch.ones(*shape_w),
            torch.randn(*shape_w).masked_fill(
                torch.bernoulli(torch.rand(*shape_w)) == 1, np.inf
            ),
        ]
        update_rule = kwargs.get("update_rule", None)
        for w in weights:
            for wmin in wmins:
                for wmax in wmaxes:
                    ### Conditional checks ###
                    # WeightDependentPostPre does not handle infinite ranges
                    if (
                        (torch.tensor(wmin, dtype=torch.float32) == -np.inf).any()
                        or (torch.tensor(wmax, dtype=torch.float32) == np.inf).any()
                    ) and update_rule == WeightDependentPostPre:
                        continue

                    # Rmax only supported for Connection & LocalConnection
                    elif (
                        not (conn_type == Connection or conn_type == LocalConnection)
                        and update_rule == Rmax
                    ):
                        return

                    # SparseConnection isn't supported for wmin\\wmax
                    elif (conn_type == SparseConnection) and not (
                        (torch.tensor(wmin, dtype=torch.float32) == -np.inf).all()
                        and (torch.tensor(wmax, dtype=torch.float32) == np.inf).all()
                    ):
                        continue

                    print(
                        f"- w: {type(w).__name__}, "
                        f"wmin: {type(wmax).__name__}, wmax: {type(wmax).__name__}"
                    )
                    if kwargs.get("update_rule") == Rmax:
                        l_a = SRM0Nodes(
                            shape=shape_a, traces=True, traces_additive=True
                        )
                        l_b = SRM0Nodes(
                            shape=shape_b, traces=True, traces_additive=True
                        )
                    else: