Convolution layer using numpy in python

Question

I'm trying to make a CNN in python using numpy. I have a finished product but it seems that it can be improved. On testing the convolutional layer is the biggest bottleneck

    def forward(self, a_prev, training):
        batch_size = a_prev.shape[0]
        a_prev_padded = Conv.zero_pad(a_prev, self.pad)
        out = np.zeros((batch_size, self.n_h, self.n_w, self.n_c))

        # Convolve
        for i in range(self.n_h):
            v_start = i * self.stride
            v_end = v_start + self.kernel_size

            for j in range(self.n_w):
                h_start = j * self.stride
                h_end = h_start + self.kernel_size

                out[:, i, j, :] = np.sum(
                    a_prev_padded[:, v_start:v_end, h_start:h_end, :, np.newaxis] * self.w[np.newaxis, :, :, :],
                    axis=(1, 2, 3),
                )

        z = out + self.b
        a = self.activation.f(z)

        if training:
            # Cache for backward pass
            self.cache.update({"a_prev": a_prev, "z": z, "a": a})

        return a

This is the code I have written. The nested for loops are one of my concerns but not sure. Is there a way to speed this up or should I use another algorithm?

Thanks in advance