Neural Tangent Kernels

The neural tangent kernel (NTK) is a kernel that describes how a neural network evolves during training. There has been a lot of research around it in recent years. This tutorial, inspired by the implementation of NTKs in JAX (see Fast Finite Width Neural Tangent Kernel for details), demonstrates how to easily compute this quantity using torch.func, composable function transforms for PyTorch.

참고

This tutorial requires PyTorch 2.0.0 or later.

Setup

First, some setup. Let’s define a simple CNN that we wish to compute the NTK of.

import torch
import torch.nn as nn
from torch.func import functional_call, vmap, vjp, jvp, jacrev
device = 'cuda' if torch.cuda.device_count() > 0 else 'cpu'

class CNN(nn.Module):
    def __init__(self):
        super(CNN, self).__init__()
        self.conv1 = nn.Conv2d(3, 32, (3, 3))
        self.conv2 = nn.Conv2d(32, 32, (3, 3))
        self.conv3 = nn.Conv2d(32, 32, (3, 3))
        self.fc = nn.Linear(21632, 10)

    def forward(self, x):
        x = self.conv1(x)
        x = x.relu()
        x = self.conv2(x)
        x = x.relu()
        x = self.conv3(x)
        x = x.flatten(1)
        x = self.fc(x)
        return x

And let’s generate some random data

x_train = torch.randn(20, 3, 32, 32, device=device)
x_test = torch.randn(5, 3, 32, 32, device=device)

Create a function version of the model

torch.func transforms operate on functions. In particular, to compute the NTK, we will need a function that accepts the parameters of the model and a single input (as opposed to a batch of inputs!) and returns a single output.

We’ll use torch.func.functional_call, which allows us to call an nn.Module using different parameters/buffers, to help accomplish the first step.

Keep in mind that the model was originally written to accept a batch of input data points. In our CNN example, there are no inter-batch operations. That is, each data point in the batch is independent of other data points. With this assumption in mind, we can easily generate a function that evaluates the model on a single data point:

net = CNN().to(device)

# Detaching the parameters because we won't be calling Tensor.backward().
params = {k: v.detach() for k, v in net.named_parameters()}

def fnet_single(params, x):
    return functional_call(net, params, (x.unsqueeze(0),)).squeeze(0)

Compute the NTK: method 1 (Jacobian contraction)

We’re ready to compute the empirical NTK. The empirical NTK for two data points \(x_1\) and \(x_2\) is defined as the matrix product between the Jacobian of the model evaluated at \(x_1\) and the Jacobian of the model evaluated at \(x_2\):

\[J_{net}(x_1) J_{net}^T(x_2)\]

In the batched case where \(x_1\) is a batch of data points and \(x_2\) is a batch of data points, then we want the matrix product between the Jacobians of all combinations of data points from \(x_1\) and \(x_2\).

The first method consists of doing just that - computing the two Jacobians, and contracting them. Here’s how to compute the NTK in the batched case:

def empirical_ntk_jacobian_contraction(fnet_single, params, x1, x2):
    # Compute J(x1)
    jac1 = vmap(jacrev(fnet_single), (None, 0))(params, x1)
    jac1 = jac1.values()
    jac1 = [j.flatten(2) for j in jac1]

    # Compute J(x2)
    jac2 = vmap(jacrev(fnet_single), (None, 0))(params, x2)
    jac2 = jac2.values()
    jac2 = [j.flatten(2) for j in jac2]

    # Compute J(x1) @ J(x2).T
    result = torch.stack([torch.einsum('Naf,Mbf->NMab', j1, j2) for j1, j2 in zip(jac1, jac2)])
    result = result.sum(0)
    return result

result = empirical_ntk_jacobian_contraction(fnet_single, params, x_train, x_test)
print(result.shape)

torch.Size([20, 5, 10, 10])

In some cases, you may only want the diagonal or the trace of this quantity, especially if you know beforehand that the network architecture results in an NTK where the non-diagonal elements can be approximated by zero. It’s easy to adjust the above function to do that:

def empirical_ntk_jacobian_contraction(fnet_single, params, x1, x2, compute='full'):
    # Compute J(x1)
    jac1 = vmap(jacrev(fnet_single), (None, 0))(params, x1)
    jac1 = jac1.values()
    jac1 = [j.flatten(2) for j in jac1]

    # Compute J(x2)
    jac2 = vmap(jacrev(fnet_single), (None, 0))(params, x2)
    jac2 = jac2.values()
    jac2 = [j.flatten(2) for j in jac2]

    # Compute J(x1) @ J(x2).T
    einsum_expr = None
    if compute == 'full':
        einsum_expr = 'Naf,Mbf->NMab'
    elif compute == 'trace':
        einsum_expr = 'Naf,Maf->NM'
    elif compute == 'diagonal':
        einsum_expr = 'Naf,Maf->NMa'
    else:
        assert False

    result = torch.stack([torch.einsum(einsum_expr, j1, j2) for j1, j2 in zip(jac1, jac2)])
    result = result.sum(0)
    return result

result = empirical_ntk_jacobian_contraction(fnet_single, params, x_train, x_test, 'trace')
print(result.shape)

torch.Size([20, 5])

The asymptotic time complexity of this method is \(N O [FP]\) (time to compute the Jacobians) + \(N^2 O^2 P\) (time to contract the Jacobians), where \(N\) is the batch size of \(x_1\) and \(x_2\), \(O\) is the model’s output size, \(P\) is the total number of parameters, and \([FP]\) is the cost of a single forward pass through the model. See section 3.2 in Fast Finite Width Neural Tangent Kernel for details.

Compute the NTK: method 2 (NTK-vector products)

The next method we will discuss is a way to compute the NTK using NTK-vector products.

This method reformulates NTK as a stack of NTK-vector products applied to columns of an identity matrix \(I_O\) of size \(O\times O\) (where \(O\) is the output size of the model):

\[J_{net}(x_1) J_{net}^T(x_2) = J_{net}(x_1) J_{net}^T(x_2) I_{O} = \left[J_{net}(x_1) \left[J_{net}^T(x_2) e_o\right]\right]_{o=1}^{O},\]

where \(e_o\in \mathbb{R}^O\) are column vectors of the identity matrix \(I_O\).

• Let \(\textrm{vjp}_o = J_{net}^T(x_2) e_o\). We can use a vector-Jacobian product to compute this.

• Now, consider \(J_{net}(x_1) \textrm{vjp}_o\). This is a Jacobian-vector product!

• Finally, we can run the above computation in parallel over all columns \(e_o\) of \(I_O\) using vmap.

This suggests that we can use a combination of reverse-mode AD (to compute the vector-Jacobian product) and forward-mode AD (to compute the Jacobian-vector product) to compute the NTK.

Let’s code that up:

def empirical_ntk_ntk_vps(func, params, x1, x2, compute='full'):
    def get_ntk(x1, x2):
        def func_x1(params):
            return func(params, x1)

        def func_x2(params):
            return func(params, x2)

        output, vjp_fn = vjp(func_x1, params)

        def get_ntk_slice(vec):
            # This computes ``vec @ J(x2).T``
            # `vec` is some unit vector (a single slice of the Identity matrix)
            vjps = vjp_fn(vec)
            # This computes ``J(X1) @ vjps``
            _, jvps = jvp(func_x2, (params,), vjps)
            return jvps

        # Here's our identity matrix
        basis = torch.eye(output.numel(), dtype=output.dtype, device=output.device).view(output.numel(), -1)
        return vmap(get_ntk_slice)(basis)

    # ``get_ntk(x1, x2)`` computes the NTK for a single data point x1, x2
    # Since the x1, x2 inputs to ``empirical_ntk_ntk_vps`` are batched,
    # we actually wish to compute the NTK between every pair of data points
    # between {x1} and {x2}. That's what the ``vmaps`` here do.
    result = vmap(vmap(get_ntk, (None, 0)), (0, None))(x1, x2)

    if compute == 'full':
        return result
    if compute == 'trace':
        return torch.einsum('NMKK->NM', result)
    if compute == 'diagonal':
        return torch.einsum('NMKK->NMK', result)

# Disable TensorFloat-32 for convolutions on Ampere+ GPUs to sacrifice performance in favor of accuracy
with torch.backends.cudnn.flags(allow_tf32=False):
    result_from_jacobian_contraction = empirical_ntk_jacobian_contraction(fnet_single, params, x_test, x_train)
    result_from_ntk_vps = empirical_ntk_ntk_vps(fnet_single, params, x_test, x_train)

assert torch.allclose(result_from_jacobian_contraction, result_from_ntk_vps, atol=1e-5)

Our code for empirical_ntk_ntk_vps looks like a direct translation from the math above! This showcases the power of function transforms: good luck trying to write an efficient version of the above by only using torch.autograd.grad.

The asymptotic time complexity of this method is \(N^2 O [FP]\), where \(N\) is the batch size of \(x_1\) and \(x_2\), \(O\) is the model’s output size, and \([FP]\) is the cost of a single forward pass through the model. Hence this method performs more forward passes through the network than method 1, Jacobian contraction (\(N^2 O\) instead of \(N O\)), but avoids the contraction cost altogether (no \(N^2 O^2 P\) term, where \(P\) is the total number of model’s parameters). Therefore, this method is preferable when \(O P\) is large relative to \([FP]\), such as fully-connected (not convolutional) models with many outputs \(O\). Memory-wise, both methods should be comparable. See section 3.3 in Fast Finite Width Neural Tangent Kernel for details.