class KANLayer(nn.Module):
    def __init__(self, in_dim=3, out_dim=2, num=5, k=3, noise_scale=0.1, scale_base=1.0, scale_sp=1.0, base_fun=torch.nn.SiLU(), grid_eps=0.02, grid_range=[-1, 1], sp_trainable=True, sb_trainable=True, device='cpu'):
        super(KANLayer, self).__init__()
        # size 
        self.size = size = out_dim * in_dim
        self.out_dim = out_dim
        self.in_dim = in_dim
        self.num = num
        self.k = k

        # shape: (size, num)
        self.grid = torch.einsum('i,j->ij', torch.ones(size, ), torch.linspace(grid_range[0], grid_range[1], steps=num + 1))
        noises = (torch.rand(size, self.grid.shape[1]) - 1 / 2) * noise_scale / num
        noises = noises.to(device)
        # shape: (size, coef)
        self.coef = torch.nn.Parameter(curve2coef(self.grid, noises, self.grid, k))
        if isinstance(scale_base, float):
            self.scale_base = torch.nn.Parameter(torch.ones(size, ) * scale_base).requires_grad_(sb_trainable)  # make scale trainable
        else:
        self.scale_sp = torch.nn.Parameter(torch.ones(size, ) * scale_sp).requires_grad_(sp_trainable)  # make scale trainable
        self.base_fun = base_fun

        self.mask = torch.nn.Parameter(torch.ones(size, )).requires_grad_(False)
        self.grid_eps = grid_eps
        self.weight_sharing = torch.arange(size)
        self.lock_counter = 0
        self.lock_id = torch.zeros(size)
        self.device = device

    def forward(self, x):
        batch = x.shape[0]
        # x: shape (batch, in_dim) => shape (size, batch) (size = out_dim * in_dim)
        x = torch.einsum('ij,k->ikj', x, torch.ones(self.out_dim, ).to(self.device)).reshape(batch, self.size).permute(1, 0)
        preacts = x.permute(1, 0).clone().reshape(batch, self.out_dim, self.in_dim)
        base = self.base_fun(x).permute(1, 0)  # shape (batch, size)
        y = coef2curve(x_eval=x, grid=self.grid[self.weight_sharing], coef=self.coef[self.weight_sharing], k=self.k, device=self.device)  # shape (size, batch)
        y = y.permute(1, 0)  # shape (batch, size)
        postspline = y.clone().reshape(batch, self.out_dim, self.in_dim)
        y = self.scale_base.unsqueeze(dim=0) * base + self.scale_sp.unsqueeze(dim=0) * y
        y = self.mask[None, :] * y
        postacts = y.clone().reshape(batch, self.out_dim, self.in_dim)
        y = torch.sum(y.reshape(batch, self.out_dim, self.in_dim), dim=2)  # shape (batch, out_dim)
        # y shape: (batch, out_dim); preacts shape: (batch, in_dim, out_dim)
        # postspline shape: (batch, in_dim, out_dim); postacts: (batch, in_dim, out_dim)
        # postspline is for extension; postacts is for visualization
        return y, preacts, postacts, postspline

    def update_grid_from_samples(self, x):
        batch = x.shape[0]
        x = torch.einsum('ij,k->ikj', x, torch.ones(self.out_dim, ).to(self.device)).reshape(batch, self.size).permute(1, 0)
        x_pos = torch.sort(x, dim=1)[0]
        y_eval = coef2curve(x_pos, self.grid, self.coef, self.k, device=self.device)
        num_interval = self.grid.shape[1] - 1
        ids = [int(batch / num_interval * i) for i in range(num_interval)] + [-1]
        grid_adaptive = x_pos[:, ids]
        margin = 0.01
        grid_uniform = torch.cat([grid_adaptive[:, [0]] - margin + (grid_adaptive[:, [-1]] - grid_adaptive[:, [0]] + 2 * margin) * a for a in np.linspace(0, 1, num=self.grid.shape[1])], dim=1)
        self.grid.data = self.grid_eps * grid_uniform + (1 - self.grid_eps) * grid_adaptive
        self.coef.data = curve2coef(x_pos, y_eval, self.grid, self.k, device=self.device)

    def initialize_grid_from_parent(self, parent, x):
        batch = x.shape[0]
        # preacts: shape (batch, in_dim) => shape (size, batch) (size = out_dim * in_dim)
        x_eval = torch.einsum('ij,k->ikj', x, torch.ones(self.out_dim, ).to(self.device)).reshape(batch, self.size).permute(1, 0)
        x_pos = parent.grid
        sp2 = KANLayer(in_dim=1, out_dim=self.size, k=1, num=x_pos.shape[1] - 1, scale_base=0.).to(self.device)
        sp2.coef.data = curve2coef(sp2.grid, x_pos, sp2.grid, k=1)
        y_eval = coef2curve(x_eval, parent.grid, parent.coef, parent.k, device=self.device)
        percentile = torch.linspace(-1, 1, self.num + 1).to(self.device)
        self.grid.data = sp2(percentile.unsqueeze(dim=1))[0].permute(1, 0)
        self.coef.data = curve2coef(x_eval, y_eval, self.grid, self.k, self.device)

    def get_subset(self, in_id, out_id):
        spb = KANLayer(len(in_id), len(out_id), self.num, self.k, base_fun=self.base_fun)
        spb.grid.data = self.grid.reshape(self.out_dim, self.in_dim, spb.num + 1)[out_id][:, in_id].reshape(-1, spb.num + 1)
        spb.coef.data = self.coef.reshape(self.out_dim, self.in_dim, spb.coef.shape[1])[out_id][:, in_id].reshape(-1, spb.coef.shape[1])
        spb.scale_base.data = self.scale_base.reshape(self.out_dim, self.in_dim)[out_id][:, in_id].reshape(-1, )
        spb.scale_sp.data = self.scale_sp.reshape(self.out_dim, self.in_dim)[out_id][:, in_id].reshape(-1, )
        spb.mask.data = self.mask.reshape(self.out_dim, self.in_dim)[out_id][:, in_id].reshape(-1, )

        spb.in_dim = len(in_id)
        spb.out_dim = len(out_id)
        spb.size = spb.in_dim * spb.out_dim
        return spb

    def lock(self, ids):
        self.lock_counter += 1
        # ids: [[i1,j1],[i2,j2],[i3,j3],...]
        for i in range(len(ids)):
            if i != 0:
                self.weight_sharing[ids[i][1] * self.in_dim + ids[i][0]] = ids[0][1] * self.in_dim + ids[0][0]
            self.lock_id[ids[i][1] * self.in_dim + ids[i][0]] = self.lock_counter

    def unlock(self, ids):
        # check ids are locked
        num = len(ids)
        locked = True
        for i in range(num):
            locked *= (self.weight_sharing[ids[i][1] * self.in_dim + ids[i][0]] == self.weight_sharing[ids[0][1] * self.in_dim + ids[0][0]])
        if locked == False:
            print("they are not locked. unlock failed.")
            return 0
        for i in range(len(ids)):
            self.weight_sharing[ids[i][1] * self.in_dim + ids[i][0]] = ids[i][1] * self.in_dim + ids[i][0]
            self.lock_id[ids[i][1] * self.in_dim + ids[i][0]] = 0
        self.lock_counter -= 1








결론 : 죤내 복잡함

- dc official App