worst-group-spurious-dfr

    for ax, sz in enumerate(shape):
        if sz == 1 and grad.shape[ax] != 1:
            grad = grad.sum(axis=ax, keepdims=True)
    return grad.reshape(shape)

    for ax, sz in enumerate(shape):
        if sz == 1 and grad.shape[ax] != 1:
            grad = grad.sum(axis=ax, keepdims=True)
    return grad.reshape(shape)


def _keepshape(shape, axis):
    """Shape with the reduced axes set to 1 (so a reduced grad can be reshaped + broadcast)."""
    if axis is None:
        return tuple(1 for _ in shape)
    axes = (axis,) if isinstance(axis, int) else tuple(axis)
    axes = tuple(ax % len(shape) for ax in axes)
    return tuple(1 if i in axes else s for i, s in enumerate(shape))


def _col2im(cols, x_shape, kh, kw, pad, stride, OH, OW):
    """Inverse of _im2col: scatter-add columns back to the (strided, padded) input positions."""
    N, C, H, W = x_shape
    st = int(stride)
    cols_r = cols.reshape(N, C, kh, kw, OH, OW)
    Hp = H + 2 * pad
    Wp = W + 2 * pad
    xp = np.zeros((N, C, Hp, Wp), dtype=np.float64)
    for i in range(kh):
        for j in range(kw):
            xp[:, :, i:i + st * OH:st, j:j + st * OW:st] += cols_r[:, :, i, j, :, :]
    if pad > 0:
        return xp[:, :, pad:pad + H, pad:pad + W]
    return xp

        TODO: implement the topological walk + seeding. (The per-op ``_backward`` closures do
        the local accumulation; here you only orchestrate the order + seed ``self.grad``.)
        """
        raise NotImplementedError("Tensor.backward")

        TODO: implement the topological walk + seeding. (The per-op ``_backward`` closures do
        the local accumulation; here you only orchestrate the order + seed ``self.grad``.)
        """
        if grad is None:
            grad = np.ones_like(self.data)
        self._accum(_as_f64(grad))
        topo: List[Tensor] = []
        visited = set()

        def build(v: "Tensor"):
            if id(v) in visited:
                return
            visited.add(id(v))
            for p in v._parents:
                build(p)
            topo.append(v)

        build(self)
        for v in reversed(topo):
            v._backward()

    def _bw():
        raise NotImplementedError("add backward")  # TODO

    def _bw():
        g = out.grad
        a._accum(_unbroadcast(g, a.data.shape))
        b._accum(_unbroadcast(g, b.data.shape))

    def _bw():
        raise NotImplementedError("sub backward")  # TODO

    def _bw():
        g = out.grad
        a._accum(_unbroadcast(g, a.data.shape))
        b._accum(_unbroadcast(-g, b.data.shape))

    def _bw():
        raise NotImplementedError("mul backward")  # TODO

    def _bw():
        g = out.grad
        a._accum(_unbroadcast(g * b.data, a.data.shape))
        b._accum(_unbroadcast(g * a.data, b.data.shape))

    def _bw():
        raise NotImplementedError("div backward")  # TODO

    def _bw():
        g = out.grad
        a._accum(_unbroadcast(g / b.data, a.data.shape))
        b._accum(_unbroadcast(-g * a.data / (b.data * b.data), b.data.shape))

    def _bw():
        raise NotImplementedError("power backward")  # TODO

    def _bw():
        a._accum(out.grad * pf * a.data ** (pf - 1.0))

    def _bw():
        raise NotImplementedError("maximum backward")  # TODO

    def _bw():
        g = out.grad
        eq = (a.data == b.data).astype(np.float64)
        amask = np.where(a.data > b.data, 1.0, 0.0) + 0.5 * eq
        bmask = np.where(b.data > a.data, 1.0, 0.0) + 0.5 * eq
        a._accum(_unbroadcast(g * amask, a.data.shape))
        b._accum(_unbroadcast(g * bmask, b.data.shape))

    def _bw():
        raise NotImplementedError("minimum backward")  # TODO

    def _bw():
        g = out.grad
        eq = (a.data == b.data).astype(np.float64)
        amask = np.where(a.data < b.data, 1.0, 0.0) + 0.5 * eq
        bmask = np.where(b.data < a.data, 1.0, 0.0) + 0.5 * eq
        a._accum(_unbroadcast(g * amask, a.data.shape))
        b._accum(_unbroadcast(g * bmask, b.data.shape))

    def _bw():
        raise NotImplementedError("where backward")  # TODO

    def _bw():
        g = out.grad
        a._accum(_unbroadcast(g * np.where(c, 1.0, 0.0), a.data.shape))
        b._accum(_unbroadcast(g * np.where(c, 0.0, 1.0), b.data.shape))

    def _bw():
        raise NotImplementedError("exp backward")  # TODO

    def _bw():
        a._accum(out.grad * e)

    def _bw():
        raise NotImplementedError("log backward")  # TODO

    def _bw():
        a._accum(out.grad / a.data)

    def _bw():
        raise NotImplementedError("sqrt backward")  # TODO

    def _bw():
        a._accum(out.grad * 0.5 / s)

    def _bw():
        raise NotImplementedError("abs backward")  # TODO

    def _bw():
        a._accum(out.grad * np.sign(a.data))

    def _bw():
        raise NotImplementedError("clip backward")  # TODO

    def _bw():
        mask = (a.data > lo) & (a.data < hi)
        a._accum(out.grad * mask)

    def _bw():
        raise NotImplementedError("sum backward")  # TODO

    def _bw():
        g2 = np.reshape(out.grad, _keepshape(a.data.shape, axis))
        a._accum(np.broadcast_to(g2, a.data.shape).astype(np.float64).copy())

    def _bw():
        raise NotImplementedError("mean backward")  # TODO

    def _bw():
        g2 = np.reshape(out.grad, _keepshape(a.data.shape, axis)) / n
        a._accum(np.broadcast_to(g2, a.data.shape).astype(np.float64).copy())

    def _bw():
        raise NotImplementedError("max backward")  # TODO

    def _bw():
        mask = (a.data == m).astype(np.float64)
        counts = mask.sum(axis=axis, keepdims=True)
        g2 = np.reshape(out.grad, _keepshape(a.data.shape, axis))
        a._accum(mask * g2 / counts)

    def _bw():
        raise NotImplementedError("min backward")  # TODO

    def _bw():
        mask = (a.data == m).astype(np.float64)
        counts = mask.sum(axis=axis, keepdims=True)
        g2 = np.reshape(out.grad, _keepshape(a.data.shape, axis))
        a._accum(mask * g2 / counts)

    def _bw():
        raise NotImplementedError("transpose backward")  # TODO

    def _bw():
        if axes is None:
            a._accum(np.transpose(out.grad))
        else:
            inv = np.argsort(axes)
            a._accum(np.transpose(out.grad, tuple(inv)))

    def _bw():
        raise NotImplementedError("reshape backward")  # TODO

    def _bw():
        a._accum(out.grad.reshape(a.data.shape))

    def _bw():
        raise NotImplementedError("getitem backward")  # TODO

    def _bw():
        gz = np.zeros_like(a.data)
        np.add.at(gz, idx, out.grad)
        a._accum(gz)

    def _bw():
        raise NotImplementedError("concat backward")  # TODO

    def _bw():
        idxs = np.cumsum(sizes)[:-1]
        parts = np.split(out.grad, idxs, axis=axis)
        for t, part in zip(ts, parts):
            t._accum(part)

    def _bw():
        raise NotImplementedError("stack backward")  # TODO

    def _bw():
        for i, t in enumerate(ts):
            t._accum(np.take(out.grad, i, axis=axis))

    def _bw():
        raise NotImplementedError("matmul backward")  # TODO

    def _bw():
        g = out.grad
        ga = g @ np.swapaxes(b.data, -1, -2)
        gb = np.swapaxes(a.data, -1, -2) @ g
        a._accum(_unbroadcast(ga, a.data.shape))
        b._accum(_unbroadcast(gb, b.data.shape))

    def _bw():
        raise NotImplementedError("relu backward")  # TODO

    def _bw():
        a._accum(out.grad * (a.data > 0.0))

    def _bw():
        raise NotImplementedError("leaky_relu backward")  # TODO

    def _bw():
        a._accum(out.grad * np.where(a.data > 0.0, 1.0, sl))

    def _bw():
        raise NotImplementedError("sigmoid backward")  # TODO

    def _bw():
        a._accum(out.grad * s * (1.0 - s))

    def _bw():
        raise NotImplementedError("tanh backward")  # TODO

    def _bw():
        a._accum(out.grad * (1.0 - t * t))

    def _bw():
        raise NotImplementedError("gelu backward")  # TODO

    def _bw():
        pdf = np.exp(-0.5 * x * x) / np.sqrt(2.0 * np.pi)
        a._accum(out.grad * (cdf + x * pdf))

    def _bw():
        raise NotImplementedError("softmax backward")  # TODO

    def _bw():
        g = out.grad
        dot = (g * s).sum(axis=axis, keepdims=True)
        a._accum(s * (g - dot))

    def _bw():
        raise NotImplementedError("log_softmax backward")  # TODO

    def _bw():
        g = out.grad
        dot = g.sum(axis=axis, keepdims=True)
        a._accum(g - s * dot)

    def _bw():
        raise NotImplementedError("softplus backward")  # TODO

    def _bw():
        sig = 1.0 / (1.0 + np.exp(-bx))
        a._accum(out.grad * sig)

    def _bw():
        raise NotImplementedError("silu backward")  # TODO

    def _bw():
        a._accum(out.grad * (sig + a.data * sig * (1.0 - sig)))

    def _bw():
        raise NotImplementedError("elu backward")  # TODO

    def _bw():
        deriv = np.where(x > 0.0, 1.0, al * ex)
        a._accum(out.grad * deriv)

    def _bw():
        raise NotImplementedError("mish backward")  # TODO

    def _bw():
        sig = 1.0 / (1.0 + np.exp(-x))
        a._accum(out.grad * (t + x * (1.0 - t * t) * sig))

    def _bw():
        raise NotImplementedError("hardtanh backward")  # TODO

    def _bw():
        mask = (a.data > lo) & (a.data < hi)
        a._accum(out.grad * mask)

    def _bw():
        raise NotImplementedError("var backward")  # TODO

    def _bw():
        g2 = np.reshape(out.grad, _keepshape(a.data.shape, axis))
        a._accum(g2 * (2.0 / (n - ddof)) * (a.data - mu))

    def _bw():
        raise NotImplementedError("std backward")  # TODO

    def _bw():
        g2 = np.reshape(out.grad, _keepshape(a.data.shape, axis))
        a._accum(g2 * (a.data - mu) / (n * skeep))

    def _bw():
        raise NotImplementedError("cumsum backward")  # TODO

    def _bw():
        g = out.grad
        a._accum(np.flip(np.cumsum(np.flip(g, axis=axis), axis=axis), axis=axis))

    def _bw():
        raise NotImplementedError("mse_loss backward")  # TODO

    def _bw():
        pred._accum(out.grad * 2.0 * diff / n)

    def _bw():
        raise NotImplementedError("cross_entropy backward")  # TODO

    def _bw():
        y = np.zeros_like(sm)
        y[np.arange(n), t] = 1.0
        logits._accum(out.grad * (sm - y) / n)

    def _bw():
        raise NotImplementedError("group_weighted_ce backward")  # TODO

    def _bw():
        y = np.zeros_like(sm)
        y[np.arange(n), t] = 1.0
        logits._accum(out.grad * (sm - y) * scale[:, None])

    def _bw():
        raise NotImplementedError("reweighted_ce backward")  # TODO

    def _bw():
        y = np.zeros_like(sm)
        y[np.arange(n), t] = 1.0
        logits._accum(out.grad * (sm - y) * (w / wsum)[:, None])

    def _bw():
        raise NotImplementedError("group_dro_loss backward")  # TODO

    def _bw():
        y = np.zeros_like(sm)
        y[np.arange(n), t] = 1.0
        logits._accum(out.grad * (sm - y) * scale[:, None])

    def _bw():
        raise NotImplementedError("logit_adjusted_ce backward")  # TODO

    def _bw():
        y = np.zeros_like(sm)
        y[np.arange(n), t] = 1.0
        logits._accum(out.grad * (sm - y) / n)

    def _bw():
        raise NotImplementedError("focal_loss backward")  # TODO

    def _bw():
        y = np.zeros_like(sm)
        y[np.arange(n), t] = 1.0
        omp = 1.0 - p
        dfdp = g * (omp ** (g - 1.0)) * np.log(pc) - (omp ** g) / pc
        coef = dfdp * p
        grad = coef[:, None] * (y - sm)
        logits._accum(out.grad * grad / n)

    def _bw():
        raise NotImplementedError("irm_penalty backward")  # TODO

    def _bw():
        sx = (sm * x).sum(axis=-1, keepdims=True)
        term = (sm - y) + sm * (x - sx)
        grad = 2.0 * grad_w * (1.0 / n) * term
        logits._accum(out.grad * grad)

    def _bw():
        raise NotImplementedError("gce_loss backward")  # TODO

    def _bw():
        y = np.zeros_like(sm)
        y[np.arange(n), t] = 1.0
        pq = pc ** qf
        grad = -(pq[:, None]) * (y - sm)
        logits._accum(out.grad * grad / n)

    def _bw():
        raise NotImplementedError("vrex_penalty backward")  # TODO

    def _bw():
        grad = (2.0 / K) * (rd - mu)
        r._accum(out.grad * grad.reshape(r.data.shape))

    def _bw():
        raise NotImplementedError("ldam_loss backward")  # TODO

    def _bw():
        y = np.zeros_like(sm)
        y[np.arange(n), t] = 1.0
        logits._accum(out.grad * sc * (sm - y) / n)

    def _bw():
        raise NotImplementedError("spectral_decoupling backward")  # TODO

    def _bw():
        logits._accum(out.grad * lm * x / n)

    def _bw():
        raise NotImplementedError("layernorm backward")  # TODO

    def _bw():
        g = out.grad
        gx = g * gamma.data
        s1 = gx.sum(axis=-1, keepdims=True)
        s2 = (gx * xhat).sum(axis=-1, keepdims=True)
        dx = inv / D * (D * gx - s1 - xhat * s2)
        a._accum(dx)
        gamma._accum(_unbroadcast(g * xhat, gamma.data.shape))
        beta._accum(_unbroadcast(g, beta.data.shape))

    def _bw():
        raise NotImplementedError("batchnorm backward")  # TODO

    def _bw():
        g = out.grad
        gx = g * gamma.data
        s1 = gx.sum(axis=0, keepdims=True)
        s2 = (gx * xhat).sum(axis=0, keepdims=True)
        dx = inv / N * (N * gx - s1 - xhat * s2)
        a._accum(dx)
        gamma._accum(_unbroadcast(g * xhat, gamma.data.shape))
        beta._accum(_unbroadcast(g, beta.data.shape))

    def _bw():
        raise NotImplementedError("batchnorm2d backward")  # TODO

    def _bw():
        g = out.grad
        gC = gamma.data.reshape(1, C, 1, 1)
        gxhat = g * gC
        if training:
            s1 = gxhat.sum(axis=(0, 2, 3), keepdims=True)
            s2 = (gxhat * xhat).sum(axis=(0, 2, 3), keepdims=True)
            dx = inv / m * (m * gxhat - s1 - xhat * s2)
        else:
            dx = gxhat * inv
        x._accum(dx)
        gamma._accum((g * xhat).sum(axis=(0, 2, 3)))
        beta._accum(g.sum(axis=(0, 2, 3)))

    def _bw():
        raise NotImplementedError("rms_norm backward")  # TODO

    def _bw():
        g = out.grad
        gy = g * gamma.data
        s = (gy * x).sum(axis=-1, keepdims=True)
        dx = r * gy - (r ** 3 / D) * x * s
        a._accum(dx)
        gamma._accum(_unbroadcast(g * xhat, gamma.data.shape))

    def _bw():
        raise NotImplementedError("groupnorm2d backward")  # TODO

    def _bw():
        g = out.grad
        gfull = g * g_
        gg = gfull.reshape(N, G, m)
        xhg = xhat.reshape(N, G, m)
        s1 = gg.sum(axis=2, keepdims=True)
        s2 = (gg * xhg).sum(axis=2, keepdims=True)
        dxg = inv / m * (m * gg - s1 - xhg * s2)
        x._accum(dxg.reshape(N, C, H, W))
        gamma._accum((g * xhat).sum(axis=(0, 2, 3)))
        beta._accum(g.sum(axis=(0, 2, 3)))

    def _bw():
        raise NotImplementedError("conv2d backward")  # TODO

    def _bw():
        g = out.grad
        gop = g.reshape(N, Cout, OH * OW)
        bias._accum(g.sum(axis=(0, 2, 3)))
        dWm = np.einsum("nop,nkp->ok", gop, cols)
        weight._accum(dWm.reshape(Cout, Cin, kh, kw))
        dcols = np.einsum("ok,nop->nkp", Wm, gop)
        dx = _col2im(dcols, x.data.shape, kh, kw, pad, st, OH, OW)
        x._accum(dx)

    def _bw():
        raise NotImplementedError("maxpool2d backward")  # TODO

    def _bw():
        g = out.grad
        xr = xd.reshape(N, C, H // k, k, W // k, k)
        mx = xr.max(axis=(3, 5), keepdims=True)
        mask = (xr == mx).astype(np.float64)
        counts = mask.sum(axis=(3, 5), keepdims=True)
        gg = g[:, :, :, None, :, None]
        dx = mask * gg / counts
        x._accum(dx.reshape(N, C, H, W))

    def _bw():
        raise NotImplementedError("maxpool2d_stride backward")  # TODO

    def _bw():
        g = out.grad
        mxk = win.max(axis=(2, 3), keepdims=True)
        mask = (win == mxk).astype(np.float64)
        counts = mask.sum(axis=(2, 3), keepdims=True)
        gg = g[:, :, None, None, :, :]
        contrib = mask * gg / counts
        dxp = np.zeros_like(xp)
        for i in range(k):
            for j in range(k):
                dxp[:, :, i:i + st * OH:st, j:j + st * OW:st] += contrib[:, :, i, j]
        if pad:
            dx = dxp[:, :, pad:pad + H, pad:pad + W]
        else:
            dx = dxp
        x._accum(dx)

    def _bw():
        raise NotImplementedError("avgpool2d backward")  # TODO

    def _bw():
        g = out.grad
        gg = g[:, :, :, None, :, None] / (k * k)
        dx = np.broadcast_to(gg, (N, C, H // k, k, W // k, k)).reshape(N, C, H, W)
        x._accum(dx.copy())

    def _bw():
        raise NotImplementedError("global_avg_pool backward")  # TODO

    def _bw():
        g = out.grad
        dx = np.broadcast_to(g[:, :, None, None], (N, C, H, W)) / (H * W)
        x._accum(dx.copy())

    def _bw():
        raise NotImplementedError("embedding backward")  # TODO

    def _bw():
        gz = np.zeros_like(weight.data)
        np.add.at(gz, idx, out.grad)
        weight._accum(gz)

        def _bw_id():
            raise NotImplementedError("dropout backward")  # TODO

        def _bw_id():
            a._accum(out.grad)

    def _bw():
        raise NotImplementedError("dropout backward")  # TODO

    def _bw():
        a._accum(out.grad * keep)

    def forward(self, x: Tensor) -> Tensor:
        # TODO: implement (compose ag.* ops). See the docstring/contract.
        raise NotImplementedError("forward")
    def named_params(self):
        d = {"weight": self.weight}

    def forward(self, x: Tensor) -> Tensor:
        out = ag.matmul(x, self.weight)
        if self.bias is not None:
            out = ag.add(out, self.bias)
        return out
    def named_params(self):
        d = {"weight": self.weight}

    def forward(self, x: Tensor) -> Tensor:
        # TODO: implement (compose ag.* ops). See the docstring/contract.
        raise NotImplementedError("forward")
    def named_params(self):
        return {"weight": self.weight, "bias": self.bias}

    def forward(self, x: Tensor) -> Tensor:
        return ag.conv2d(x, self.weight, self.bias, pad=self.pad, stride=1)
    def named_params(self):
        return {"weight": self.weight, "bias": self.bias}

    def forward(self, x: Tensor, training: bool = True) -> Tensor:
        # TODO: implement (compose ag.* ops). See the docstring/contract.
        raise NotImplementedError("forward")
    def named_params(self):
        return {"gamma": self.gamma, "beta": self.beta}

    def buffers(self):

    def forward(self, x: Tensor, training: bool = True) -> Tensor:
        return ag.batchnorm2d(x, self.gamma, self.beta, self.running_mean, self.running_var,
                              training=training, momentum=self.momentum, eps=self.eps)
    def named_params(self):
        return {"gamma": self.gamma, "beta": self.beta}

    def buffers(self):

    def forward(self, x: Tensor) -> Tensor:
        # TODO: implement (compose ag.* ops). See the docstring/contract.
        raise NotImplementedError("forward")
    def named_params(self):
        return {"gamma": self.gamma, "beta": self.beta}


class LayerNorm(Module):

    def forward(self, x: Tensor) -> Tensor:
        return ag.batchnorm(x, self.gamma, self.beta, eps=self.eps)
    def named_params(self):
        return {"gamma": self.gamma, "beta": self.beta}


class LayerNorm(Module):

    def forward(self, x: Tensor) -> Tensor:
        # TODO: implement (compose ag.* ops). See the docstring/contract.
        raise NotImplementedError("forward")
    def named_params(self):
        return {"gamma": self.gamma, "beta": self.beta}


class Dropout(Module):

    def forward(self, x: Tensor) -> Tensor:
        return ag.layernorm(x, self.gamma, self.beta, eps=self.eps)
    def named_params(self):
        return {"gamma": self.gamma, "beta": self.beta}


class Dropout(Module):

    def forward(self, x: Tensor, training: bool = False, rng=None) -> Tensor:
        # TODO: implement (compose ag.* ops). See the docstring/contract.
        raise NotImplementedError("forward")

    def forward(self, x: Tensor, training: bool = False, rng=None) -> Tensor:
        return ag.dropout(x, self.p, training, rng)

    def forward(self, x: Tensor) -> Tensor:
        # TODO: implement (compose ag.* ops). See the docstring/contract.
        raise NotImplementedError("forward")
    def named_params(self):
        return {}


class MaxPool2d(Module):

    def forward(self, x: Tensor) -> Tensor:
        return ag.avgpool2d(x, self.k)
    def named_params(self):
        return {}


class MaxPool2d(Module):

    def forward(self, x: Tensor) -> Tensor:
        # TODO: implement (compose ag.* ops). See the docstring/contract.
        raise NotImplementedError("forward")
    def named_params(self):
        return {}


class RMSNorm(Module):

    def forward(self, x: Tensor) -> Tensor:
        return ag.maxpool2d(x, self.k)
    def named_params(self):
        return {}


class RMSNorm(Module):

    def forward(self, x: Tensor) -> Tensor:
        # TODO: implement (compose ag.rms_norm). See the docstring/contract.
        raise NotImplementedError("forward")

    def forward(self, x: Tensor) -> Tensor:
        return ag.rms_norm(x, self.gamma, eps=self.eps)

    def forward(self, x: Tensor) -> Tensor:
        # TODO: implement (compose ag.groupnorm2d). See the docstring/contract.
        raise NotImplementedError("forward")

    def forward(self, x: Tensor) -> Tensor:
        return ag.groupnorm2d(x, self.gamma, self.beta, self.num_groups, eps=self.eps)

    def forward(self, X, training: bool = True) -> Tensor:
        # TODO: implement (stem->gn0->silu; then 2 residual blocks
        #   r=h; h=silu(b*n1(b*c1(h))); h=b*n2(b*c2(h)); h=silu(h + r);
        #   then strided downsample via ag.conv2d(h, down.weight, down.bias, pad=1, stride=2) + silu;
        #   then global-avg-pool over axes (2,3); then the linear head). Compose ag.* ops.
        raise NotImplementedError("forward")

    def forward(self, X, training: bool = True) -> Tensor:
        if not isinstance(X, Tensor):
            X = Tensor(X)
        h = ag.silu(self.gn0(self.stem(X)))
        # block 1
        r = h
        h = ag.silu(self.b1n1(self.b1c1(h)))
        h = self.b1n2(self.b1c2(h))
        h = ag.silu(ag.add(h, r))
        # block 2
        r = h
        h = ag.silu(self.b2n1(self.b2c1(h)))
        h = self.b2n2(self.b2c2(h))
        h = ag.silu(ag.add(h, r))
        # strided downsample
        h = ag.silu(ag.conv2d(h, self.down.weight, self.down.bias, pad=1, stride=2))
        h = ag.global_avg_pool(h)
        return self.head(h)

    def featurize(self, X, training: bool = True) -> Tensor:
        # TODO: implement (compose ag.* ops). See the docstring/contract.
        raise NotImplementedError("featurize")
    def forward(self, X, training: bool = True) -> Tensor:
        # TODO: implement (compose ag.* ops). See the docstring/contract.
        raise NotImplementedError("forward")

    def featurize(self, X, training: bool = True) -> Tensor:
        if not isinstance(X, Tensor):
            X = Tensor(X)
        h = ag.maxpool2d(ag.relu(self.bn1(self.conv1(X), training=training)), 2)
        h = ag.maxpool2d(ag.relu(self.bn2(self.conv2(h), training=training)), 2)
        h = ag.relu(self.bn3(self.conv3(h), training=training))
        h = ag.global_avg_pool(h)
        h = ag.relu(self.embed(h))
        return h

    def forward(self, X, training: bool = True) -> Tensor:
        feat = self.featurize(X, training=training)
        return self.classifier(feat)

    def forward(self, x) -> Tensor:
        # TODO: implement (compose ag.* ops). See the docstring/contract.
        raise NotImplementedError("forward")
    def named_params(self) -> Dict[str, Tensor]:
        d: Dict[str, Tensor] = {}
        for i, lin in enumerate(self.layers):

    def forward(self, x) -> Tensor:
        if not isinstance(x, Tensor):
            x = Tensor(x)
        for i, lin in enumerate(self.layers):
            x = lin(x)
            if i < len(self.layers) - 1:
                x = ag.relu(x)
        return x
    def named_params(self) -> Dict[str, Tensor]:
        d: Dict[str, Tensor] = {}
        for i, lin in enumerate(self.layers):

    def forward(self, x) -> Tensor:
        # TODO: implement (compose ag.* ops). See the docstring/contract.
        raise NotImplementedError("forward")
    def named_params(self) -> Dict[str, Tensor]:
        d: Dict[str, Tensor] = {}
        for k, v in self.proj_in.named_params().items():

    def forward(self, x) -> Tensor:
        if not isinstance(x, Tensor):
            x = Tensor(x)
        x = self.proj_in(x)
        for blk in self.blocks:
            h = blk["ln"](x)
            h = blk["fc1"](h)
            h = ag.gelu(h)
            h = blk["fc2"](h)
            x = ag.add(x, h)
        return self.proj_out(x)
    def named_params(self) -> Dict[str, Tensor]:
        d: Dict[str, Tensor] = {}
        for k, v in self.proj_in.named_params().items():

    def forward(self, X, training: bool = True) -> Tensor:
        # TODO: implement (compose ag.* ops). See the docstring/contract.
        raise NotImplementedError("forward")
    def named_params(self) -> Dict[str, Tensor]:
        d: Dict[str, Tensor] = {}
        for name in ("conv1", "bn1", "conv2", "bn2", "head"):

    def forward(self, X, training: bool = True) -> Tensor:
        if not isinstance(X, Tensor):
            X = Tensor(X)
        h = ag.maxpool2d(ag.relu(self.bn1(self.conv1(X), training=training)), 2)
        h = ag.avgpool2d(ag.relu(self.bn2(self.conv2(h), training=training)), 2)
        N = h.shape[0]
        h = ag.reshape(h, (N, 2 * self.W * self.side * self.side))
        return self.head(h)
    def named_params(self) -> Dict[str, Tensor]:
        d: Dict[str, Tensor] = {}
        for name in ("conv1", "bn1", "conv2", "bn2", "head"):

    # TODO: implement (numpy only). See docstring.
    raise NotImplementedError("balanced_group_weights")
def class_balanced_weights(targets, n_classes=None, beta=0.0):

    gids = np.asarray(group_ids).astype(np.int64).reshape(-1)
    if n_groups is None:
        n_groups = int(gids.max()) + 1
    present = np.unique(gids)
    w = np.zeros(n_groups, dtype=np.float64)
    w[present] = 1.0 / len(present)
    return w
def class_balanced_weights(targets, n_classes=None, beta=0.0):

    # TODO: implement (numpy only). See docstring.
    raise NotImplementedError("class_balanced_weights")
def log_class_prior(targets, n_classes=None, eps=1e-6):

    t = np.asarray(targets).astype(np.int64).reshape(-1)
    if n_classes is None:
        n_classes = int(t.max()) + 1
    counts = np.bincount(t, minlength=n_classes).astype(np.float64)
    if beta == 0.0:
        eff = counts
    else:
        eff = (1.0 - beta ** counts) / (1.0 - beta)
    cw = 1.0 / np.maximum(eff, 1e-12)
    sw = cw[t]
    sw = sw / sw.mean()
    return sw
def log_class_prior(targets, n_classes=None, eps=1e-6):

    # TODO: implement (numpy only). See docstring.
    raise NotImplementedError("log_class_prior")
def ldam_margins(targets, n_classes=None, max_margin=0.5):

    t = np.asarray(targets).astype(np.int64).reshape(-1)
    if n_classes is None:
        n_classes = int(t.max()) + 1
    counts = np.bincount(t, minlength=n_classes).astype(np.float64)
    N = t.size
    prior = counts / N
    return np.where(prior > 0, np.log(np.maximum(prior, eps)), np.log(eps))
def ldam_margins(targets, n_classes=None, max_margin=0.5):

    # TODO: implement (numpy only). See docstring.
    raise NotImplementedError("ldam_margins")
def effective_number_weights(targets, n_classes=None, beta=0.999):

    t = np.asarray(targets).astype(np.int64).reshape(-1)
    if n_classes is None:
        n_classes = int(t.max()) + 1
    counts = np.bincount(t, minlength=n_classes).astype(np.float64)
    m = np.zeros(n_classes, dtype=np.float64)
    pres = counts > 0
    m[pres] = counts[pres] ** (-0.25)
    if m.max() > 0:
        m = m / m.max() * max_margin
    return m
def effective_number_weights(targets, n_classes=None, beta=0.999):

    # TODO: implement (numpy only). See docstring.
    raise NotImplementedError("effective_number_weights")
class EMAGroupLoss:

    t = np.asarray(targets).astype(np.int64).reshape(-1)
    if n_classes is None:
        n_classes = int(t.max()) + 1
    counts = np.bincount(t, minlength=n_classes).astype(np.float64)
    pres = counts > 0
    eff = (1.0 - beta ** counts) / (1.0 - beta)
    w = np.zeros(n_classes, dtype=np.float64)
    w[pres] = 1.0 / eff[pres]
    w[pres] = w[pres] / w[pres].mean()
    return w
class EMAGroupLoss:

    def update(self, group_losses, present=None):
        # TODO: implement the masked EMA update (see docstring). Return self.l.copy().
        raise NotImplementedError("EMAGroupLoss.update")

    def worst_group(self):
        # TODO: implement (argmax of the current EMA). See docstring.
        raise NotImplementedError("EMAGroupLoss.worst_group")

    def update(self, group_losses, present=None):
        gl = np.asarray(group_losses, dtype=np.float64).reshape(-1)
        if present is None:
            present = np.ones(self.n_groups, dtype=bool)
        else:
            present = np.asarray(present, dtype=bool).reshape(-1)
        for k in range(self.n_groups):
            if not present[k]:
                continue
            if not self._init[k]:
                self.l[k] = gl[k]
                self._init[k] = True
            else:
                self.l[k] = self.beta * self.l[k] + (1.0 - self.beta) * gl[k]
        return self.l.copy()

    def worst_group(self):
        return int(np.argmax(self.l))

    def update(self, group_losses):
        # TODO: exponentiated-gradient update of self.q on group_losses; renormalize.
        raise NotImplementedError("GroupDROState.update")

    def update(self, group_losses):
        gl = np.asarray(group_losses, dtype=np.float64).reshape(-1)
        self.q = self.q * np.exp(self.eta_q * gl)
        self.q = self.q / self.q.sum()
        return self.q.copy()

    def step(self):
        # TODO: implement the SGD update (see module docstring). Update p.data in place.
        raise NotImplementedError("SGD.step")

    def step(self):
        for i, p in enumerate(self.params):
            if p.grad is None:
                continue
            g = p.grad + self.weight_decay * p.data
            self._v[i] = self.momentum * self._v[i] + g
            upd = (g + self.momentum * self._v[i]) if self.nesterov else self._v[i]
            p.data -= self.lr * upd

    def step(self):
        # TODO: implement the Adagrad update (see module docstring). Update p.data in place.
        raise NotImplementedError("Adagrad.step")

    def step(self):
        for i, p in enumerate(self.params):
            if p.grad is None:
                continue
            g = p.grad + self.weight_decay * p.data
            self._s[i] = self._s[i] + g * g
            p.data -= self.lr * g / (np.sqrt(self._s[i]) + self.eps)

    def step(self):
        # TODO: implement the RMSprop update (see module docstring). Update p.data in place.
        raise NotImplementedError("RMSprop.step")

    def step(self):
        for i, p in enumerate(self.params):
            if p.grad is None:
                continue
            g = p.grad + self.weight_decay * p.data
            self._s[i] = self.alpha * self._s[i] + (1.0 - self.alpha) * (g * g)
            p.data -= self.lr * g / (np.sqrt(self._s[i]) + self.eps)

    def step(self):
        # TODO: implement the Adam update (see module docstring). Update p.data in place.
        raise NotImplementedError("Adam.step")

    def step(self):
        self.t += 1
        for i, p in enumerate(self.params):
            if p.grad is None:
                continue
            g = p.grad + self.weight_decay * p.data
            self._m[i] = self.b1 * self._m[i] + (1.0 - self.b1) * g
            self._v[i] = self.b2 * self._v[i] + (1.0 - self.b2) * (g * g)
            mhat = self._m[i] / (1.0 - self.b1 ** self.t)
            vhat = self._v[i] / (1.0 - self.b2 ** self.t)
            p.data -= self.lr * mhat / (np.sqrt(vhat) + self.eps)

    def step(self):
        # TODO: implement the AdamW update (see module docstring). Update p.data in place.
        raise NotImplementedError("AdamW.step")

    def step(self):
        self.t += 1
        for i, p in enumerate(self.params):
            if p.grad is None:
                continue
            g = p.grad
            self._m[i] = self.b1 * self._m[i] + (1.0 - self.b1) * g
            self._v[i] = self.b2 * self._v[i] + (1.0 - self.b2) * (g * g)
            mhat = self._m[i] / (1.0 - self.b1 ** self.t)
            vhat = self._v[i] / (1.0 - self.b2 ** self.t)
            p.data -= self.lr * (mhat / (np.sqrt(vhat) + self.eps) + self.weight_decay * p.data)

    def step(self):
        # TODO: implement the Adamax update (see module docstring). Update p.data in place.
        raise NotImplementedError("Adamax.step")

    def step(self):
        self.t += 1
        for i, p in enumerate(self.params):
            if p.grad is None:
                continue
            g = p.grad + self.weight_decay * p.data
            self._m[i] = self.b1 * self._m[i] + (1.0 - self.b1) * g
            self._u[i] = np.maximum(self.b2 * self._u[i], np.abs(g))
            p.data -= (self.lr / (1.0 - self.b1 ** self.t)) * self._m[i] / (self._u[i] + self.eps)

    def step(self):
        # TODO: implement the NAdam update (see module docstring). Update p.data in place.
        raise NotImplementedError("NAdam.step")

    def step(self):
        self.t += 1
        for i, p in enumerate(self.params):
            if p.grad is None:
                continue
            g = p.grad + self.weight_decay * p.data
            self._m[i] = self.b1 * self._m[i] + (1.0 - self.b1) * g
            self._v[i] = self.b2 * self._v[i] + (1.0 - self.b2) * (g * g)
            mhat = self._m[i] / (1.0 - self.b1 ** self.t)
            vhat = self._v[i] / (1.0 - self.b2 ** self.t)
            mnes = self.b1 * mhat + (1.0 - self.b1) * g / (1.0 - self.b1 ** self.t)
            p.data -= self.lr * mnes / (np.sqrt(vhat) + self.eps)

    def step(self):
        # TODO: implement the Lion update (see module docstring). Update p.data in place.
        raise NotImplementedError("Lion.step")

    def step(self):
        for i, p in enumerate(self.params):
            if p.grad is None:
                continue
            g = p.grad
            upd = np.sign(self.b1 * self._m[i] + (1.0 - self.b1) * g)
            p.data -= self.lr * (upd + self.weight_decay * p.data)
            self._m[i] = self.b2 * self._m[i] + (1.0 - self.b2) * g

    def step(self):
        # TODO: implement the RAdam update (see module docstring). Update p.data in place.
        raise NotImplementedError("RAdam.step")

    def step(self):
        self.t += 1
        t = self.t
        b1, b2 = self.b1, self.b2
        rinf = 2.0 / (1.0 - b2) - 1.0
        rt = rinf - 2.0 * t * (b2 ** t) / (1.0 - b2 ** t)
        for i, p in enumerate(self.params):
            if p.grad is None:
                continue
            g = p.grad + self.weight_decay * p.data
            self._m[i] = b1 * self._m[i] + (1.0 - b1) * g
            self._v[i] = b2 * self._v[i] + (1.0 - b2) * (g * g)
            mhat = self._m[i] / (1.0 - b1 ** t)
            if rt > 4.0:
                l = np.sqrt((1.0 - b2 ** t) / (self._v[i] + self.eps))
                r = math.sqrt(((rt - 4.0) * (rt - 2.0) * rinf) / ((rinf - 4.0) * (rinf - 2.0) * rt))
                p.data -= self.lr * mhat * r * l
            else:
                p.data -= self.lr * mhat

    def step(self):
        # TODO: implement the AdaBelief update (see module docstring). Update p.data in place.
        raise NotImplementedError("AdaBelief.step")

    def step(self):
        self.t += 1
        for i, p in enumerate(self.params):
            if p.grad is None:
                continue
            g = p.grad
            self._m[i] = self.b1 * self._m[i] + (1.0 - self.b1) * g
            d = g - self._m[i]
            self._s[i] = self.b2 * self._s[i] + (1.0 - self.b2) * (d * d) + self.eps
            mhat = self._m[i] / (1.0 - self.b1 ** self.t)
            shat = self._s[i] / (1.0 - self.b2 ** self.t)
            p.data -= self.lr * (mhat / (np.sqrt(shat) + self.eps) + self.weight_decay * p.data)

    # TODO: implement (see docstring).
    raise NotImplementedError("clip_grad_norm")
def clip_grad_value(params: List[Tensor], clip_value: float) -> None:
    """Element-wise gradient clamp to [-clip_value, clip_value] (in place)."""
    # TODO: implement (see docstring).
    raise NotImplementedError("clip_grad_value")

    total = 0.0
    for p in params:
        if p.grad is not None:
            total += float((p.grad ** 2).sum())
    total = math.sqrt(total)
    if total > max_norm:
        scale = max_norm / (total + 1e-6)
        for p in params:
            if p.grad is not None:
                p.grad = p.grad * scale
    return total
def clip_grad_value(params: List[Tensor], clip_value: float) -> None:
    """Element-wise gradient clamp to [-clip_value, clip_value] (in place)."""
    for p in params:
        if p.grad is not None:
            p.grad = np.clip(p.grad, -clip_value, clip_value)

    # TODO: implement (see docstring).
    raise NotImplementedError("cosine_lr")

    if warmup > 0 and step < warmup:
        return base_lr * (step + 1) / warmup
    prog = min(1.0, max(0.0, (step - warmup) / max(1, total - warmup)))
    return min_lr + 0.5 * (base_lr - min_lr) * (1.0 + math.cos(math.pi * prog))

    # TODO: implement (see docstring).
    raise NotImplementedError("warmup_lr")

    if step < warmup:
        return base_lr * (step + 1) / warmup
    return base_lr

    # TODO: implement (see docstring).
    raise NotImplementedError("step_lr")

    return base_lr * (gamma ** (step // step_size))

    # TODO: implement (see docstring).
    raise NotImplementedError("exponential_lr")

    return base_lr * (gamma ** step)

    # TODO: implement (see docstring).
    raise NotImplementedError("polynomial_lr")

    if step >= total:
        return min_lr
    return (base_lr - min_lr) * ((1.0 - step / total) ** power) + min_lr

    # TODO: implement (see docstring).
    raise NotImplementedError("cosine_restart_lr")

    s = step
    T = period
    while s >= T:
        s -= T
        T *= t_mult
    return min_lr + 0.5 * (base_lr - min_lr) * (1.0 + math.cos(math.pi * s / max(1, T)))

    """Flatten (y, g) to a single group id y*n_classes + g (int64)."""
    # TODO
    raise NotImplementedError("group_ids")

    """Flatten (y, g) to a single group id y*n_classes + g (int64)."""
    return (np.asarray(y).astype(np.int64) * n_classes + np.asarray(g).astype(np.int64)).astype(np.int64)

    """Yield (Xb, yb, gb) minibatches over a random permutation of the dataset."""
    # TODO
    raise NotImplementedError("iterate")

    """Yield (Xb, yb, gb) minibatches over a random permutation of the dataset."""
    n = len(y)
    perm = rng.permutation(n)
    for s in range(0, n, batch_size):
        idx = perm[s:s + batch_size]
        yield X[idx], y[idx], g[idx]

    """Return (min over (y,g) groups of within-group accuracy, average accuracy, per-group dict)."""
    # TODO
    raise NotImplementedError("worst_group_acc")

    """Return (min over (y,g) groups of within-group accuracy, average accuracy, per-group dict)."""
    preds = np.asarray(preds).reshape(-1)
    y = np.asarray(y).reshape(-1)
    g = np.asarray(g).reshape(-1)
    gid = group_ids(y, g, n_classes)
    correct = (preds == y)
    avg = float(correct.mean())
    per = {}
    for k in np.unique(gid):
        mask = gid == k
        per[int(k)] = float(correct[mask].mean())
    wga = min(per.values()) if per else 0.0
    return wga, avg, per

    """Boolean mask selecting ~frac of EACH (y,g) group (keeps both slices group-balanced)."""
    # TODO
    raise NotImplementedError("stratified_group_split")

    """Boolean mask selecting ~frac of EACH (y,g) group (keeps both slices group-balanced)."""
    y = np.asarray(y).reshape(-1)
    g = np.asarray(g).reshape(-1)
    gid = group_ids(y, g, n_classes)
    mask = np.zeros(len(y), dtype=bool)
    for k in np.unique(gid):
        idx = np.where(gid == k)[0]
        rng.shuffle(idx)
        ntake = int(round(frac * len(idx)))
        ntake = max(1, min(len(idx), ntake)) if len(idx) > 0 else 0
        mask[idx[:ntake]] = True
    return mask

def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--train", default="/app/data/train.npz")
    ap.add_argument("--reweight", default="/app/data/reweight.npz")
    ap.add_argument("--out", default="/app/submission/model.npz")
    ap.add_argument("--seed", type=int, default=0)
    args = ap.parse_args()

    # TODO: build a SmallCNN(n_classes, img_size, width=WIDTH, embed_dim=EMBED_DIM), train a
    #       worst-group-robust recipe with your framework (forward -> loss -> backward -> clip ->
    #       optimizer step), then save:
    #           save_checkpoint(args.out, model.state_dict(), n_classes, img_size)
    raise NotImplementedError("implement training and call save_checkpoint")


if __name__ == "__main__":
    main()

def _featurize_all(model, X, batch=256):
    feats = []
    for s in range(0, len(X), batch):
        f = model.featurize(X[s:s + batch], training=False).data
        feats.append(np.asarray(f, dtype=np.float64))
    return np.concatenate(feats, axis=0)


def _predict(model, X, batch=256):
    preds = []
    for s in range(0, len(X), batch):
        logits = model(X[s:s + batch], training=False).data
        preds.append(logits.argmax(axis=-1))
    return np.concatenate(preds, axis=0)


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--train", default="/app/data/train.npz")
    ap.add_argument("--reweight", default="/app/data/reweight.npz")
    ap.add_argument("--out", default="/app/submission/model.npz")
    ap.add_argument("--seed", type=int, default=0)
    ap.add_argument("--epochs", type=int, default=24)
    args = ap.parse_args()

    rng = np.random.default_rng(args.seed)

    Xtr, ytr, gtr = D.load_npz(args.train)
    Xrw, yrw, grw = D.load_npz(args.reweight)
    n_classes = int(max(ytr.max(), yrw.max())) + 1
    img_size = Xtr.shape[-1]

    model = SmallCNN(n_classes=n_classes, img_size=img_size,
                     width=WIDTH, embed_dim=EMBED_DIM, seed=args.seed)

    # ---- Phase 1: ERM training of the full network on the correlated split -------------
    bs = 128
    nsteps_per = (len(Xtr) + bs - 1) // bs
    total_steps = args.epochs * nsteps_per
    base_lr = 2e-3
    opt = O.AdamW(model.params(), lr=base_lr, weight_decay=1e-4)
    step = 
…[truncated 2732 chars]