如何在pytorch中改变NN的权重

Question

Im playing around with genetic algorithms with pytorch and I'm looking for a more efficient way of mutating the weights of the network (applying a small modification to them)

Right now I have a suboptimal solution where I loop through the parameters and apply a random modification.

child_agent = network()
for param in child_agent.parameters():
        if len(param.shape) == 4:  # weights of Conv2D
            for i0 in range(param.shape[0]):
                for i1 in range(param.shape[1]):
                    for i2 in range(param.shape[2]):
                        for i3 in range(param.shape[3]):
                            param[i0][i1][i2][i3] += mutation_power * np.random.randn()
        elif len(param.shape) == 2:  # weights of linear layer
            for i0 in range(param.shape[0]):
                for i1 in range(param.shape[1]):
                    param[i0][i1] += mutation_power * np.random.randn()
        elif len(param.shape) == 1:  # biases of linear layer or conv layer
            for i0 in range(param.shape[0]):
                param[i0] += mutation_power * np.random.randn()

This solution is bound to my architecture and needs recoding If I decide to add more layers. Is there any way to do this more efficiently and clean? Preferable that it works regardless of how the architecture of my network looks like.

Thanks

Answer 1

pytorch and numpy are tensor oriented, eg you make operations on multiple items contained within multidimensional array-like object.

You can change your whole code to this single line:

import torch

child_agent = network()
for param in child_agent.parameters():
    param.data += mutation.power * torch.randn_like(param)

randn_like (docs here ) creates random normal tensor with the same shape as param .

Also, if this parameter requires grad (which it probably does), you should modify it's data field.

MCVE :

import torch

mutation_power = 0.4

child_agent = torch.nn.Sequential(
    torch.nn.Conv2d(1, 3, 3, padding=1), torch.nn.Linear(10, 20)
)

for param in child_agent.parameters():
    param.data += mutation_power * torch.randn_like(param)