You write custom CUDA kernels to replace the pytorch operators in the given architecture to get speedups.   
  
You have complete freedom to choose the set of operators you want to replace. You may make the decision to replace some operators with custom CUDA kernels and leave others unchanged. You may replace multiple operators with custom implementations, consider operator fusion opportunities (combining multiple operators into a single kernel, for example, combining matmul+relu), or algorithmic changes (such as online softmax). You are only limited by your imagination.  
  
Here's an example to show you the syntax of inline embedding custom CUDA operators in torch: The example given architecture is:   
  
```python
# switchablenorm_torch.py 
import torch
import torch.nn as nn
import torch.nn.functional as F

N, C, H, W = 16, 64, 32, 32
EPS = 1e-5

class SN(nn.Module):

    def __init__(self, num_channels, eps):
        super().__init__()
        self.eps = eps
        
        self.weight = nn.Parameter(torch.ones(1, num_channels, 1, 1)) # gamma
        self.bias = nn.Parameter(torch.zeros(1, num_channels, 1, 1))   # beta

        self.w_in = nn.Parameter(torch.ones(num_channels))
        self.w_ln = nn.Parameter(torch.ones(num_channels))
        self.w_bn = nn.Parameter(torch.ones(num_channels))
        
        self.register_buffer('running_mean', torch.zeros(num_channels))
        self.register_buffer('running_var', torch.ones(num_channels))

    def forward(self, x: torch.Tensor) -> torch.Tensor:
    
        ln_mean = x.mean(dim=[1, 2, 3], keepdim=True)
        ln_var = x.var(dim=[1, 2, 3], keepdim=True)  
        

        in_mean = x.mean(dim=[2, 3], keepdim=True) 
        in_var = x.var(dim=[2, 3], keepdim=True)  
        

        bn_mean = self.running_mean.view(1, C, 1, 1)
        bn_var = self.running_var.view(1, C, 1, 1)

        w_sum = self.w_in.abs() + self.w_ln.abs() + self.w_bn.abs()
        w_in_norm = (self.w_in.abs() / w_sum).view(1, C, 1, 1)
        w_ln_norm = (self.w_ln.abs() / w_sum).view(1, C, 1, 1)
        w_bn_norm = (self.w_bn.abs() / w_sum).view(1, C, 1, 1)

        mean = w_in_norm * in_mean + w_ln_norm * ln_mean + w_bn_norm * bn_mean
        

        var_in_M2 = in_var + in_mean.pow(2)
        var_ln_M2 = ln_var + ln_mean.pow(2)
        var_bn_M2 = bn_var + bn_mean.pow(2)
        
        aggregated_var_M2 = w_in_norm * var_in_M2 + w_ln_norm * var_ln_M2 + w_bn_norm * var_bn_M2
        var = aggregated_var_M2 - mean.pow(2)
        

        x_norm = (x - mean) / torch.sqrt(var + self.eps)
        return x_norm * self.weight + self.bias


class Model(nn.Module):
    def __init__(self, weight, bias, w_in, w_ln, w_bn):
        super().__init__()
        self.sn = SN(C, EPS)
        
        with torch.no_grad():
            self.sn.weight.data.copy_(weight)
            self.sn.bias.data.copy_(bias)
            self.sn.w_in.data.copy_(w_in.squeeze())
            self.sn.w_ln.data.copy_(w_ln.squeeze())
            self.sn.w_bn.data.copy_(w_bn.squeeze())
            
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.sn(x)


def get_inputs():

    x = torch.randn(N, C, H, W, dtype=torch.float32)
    return [x]


def get_init_inputs():

    w_in = torch.ones(C)
    w_ln = torch.ones(C)
    w_bn = torch.ones(C)
    weight = torch.ones(1, C, 1, 1)
    bias = torch.zeros(1, C, 1, 1)
    return [weight, bias, w_in, w_ln, w_bn]