siren pytorch

Pytorch 中的 SIRN

SIREN 的 Pytorch 實作 - 具有週期性活化函數的隱式神經表示

安裝

$ pip install siren-pytorch

用法

基於SIREN的多層神經網絡

 import torch
from torch import nn
from siren_pytorch import SirenNet

net = SirenNet (
    dim_in = 2 ,                        # input dimension, ex. 2d coor
    dim_hidden = 256 ,                  # hidden dimension
    dim_out = 3 ,                       # output dimension, ex. rgb value
    num_layers = 5 ,                    # number of layers
    final_activation = nn . Sigmoid (),   # activation of final layer (nn.Identity() for direct output)
    w0_initial = 30.                   # different signals may require different omega_0 in the first layer - this is a hyperparameter
)

coor = torch . randn ( 1 , 2 )
net ( coor ) # (1, 3) <- rgb value

一層 SIRN 層

 import torch
from siren_pytorch import Siren

neuron = Siren (
    dim_in = 3 ,
    dim_out = 256
)

coor = torch . randn ( 1 , 3 )
neuron ( coor ) # (1, 256)

正弦激活（只是torch.sin的包裝）

 import torch
from siren_pytorch import Sine

act = Sine ( 1. )
coor = torch . randn ( 1 , 2 )
act ( coor )

包裝器用於對來自給定SirenNet的指定高度和寬度的特定圖像進行訓練，然後進行生成。

 import torch
from torch import nn
from siren_pytorch import SirenNet , SirenWrapper

net = SirenNet (
    dim_in = 2 ,                        # input dimension, ex. 2d coor
    dim_hidden = 256 ,                  # hidden dimension
    dim_out = 3 ,                       # output dimension, ex. rgb value
    num_layers = 5 ,                    # number of layers
    w0_initial = 30.                   # different signals may require different omega_0 in the first layer - this is a hyperparameter
)

wrapper = SirenWrapper (
    net ,
    image_width = 256 ,
    image_height = 256
)

img = torch . randn ( 1 , 3 , 256 , 256 )
loss = wrapper ( img )
loss . backward ()

# after much training ...
# simply invoke the wrapper without passing in anything

pred_img = wrapper () # (1, 3, 256, 256) 

使用潛在代碼進行調製

一篇新論文提出，以潛在程式碼調節 Siren 的最佳方法是將潛在向量傳遞給調製器前饋網絡，其中每個層的隱藏狀態與 Siren 的相應層進行元素相乘。

您可以透過在SirenWrapper上設定額外的關鍵字latent_dim來簡單地使用它

 import torch
from torch import nn
from siren_pytorch import SirenNet , SirenWrapper

net = SirenNet (
    dim_in = 2 ,                        # input dimension, ex. 2d coor
    dim_hidden = 256 ,                  # hidden dimension
    dim_out = 3 ,                       # output dimension, ex. rgb value
    num_layers = 5 ,                    # number of layers
    w0_initial = 30.                   # different signals may require different omega_0 in the first layer - this is a hyperparameter
)

wrapper = SirenWrapper (
    net ,
    latent_dim = 512 ,
    image_width = 256 ,
    image_height = 256
)

latent = nn . Parameter ( torch . zeros ( 512 ). normal_ ( 0 , 1e-2 ))
img = torch . randn ( 1 , 3 , 256 , 256 )

loss = wrapper ( img , latent = latent )
loss . backward ()

# after much training ...
# simply invoke the wrapper without passing in anything

pred_img = wrapper ( latent = latent ) # (1, 3, 256, 256) 

引文

 @misc { sitzmann2020implicit ,
    title   = { Implicit Neural Representations with Periodic Activation Functions } ,
    author  = { Vincent Sitzmann and Julien N. P. Martel and Alexander W. Bergman and David B. Lindell and Gordon Wetzstein } ,
    year    = { 2020 } ,
    eprint  = { 2006.09661 } ,
    archivePrefix = { arXiv } ,
    primaryClass = { cs.CV }
}

 @misc { mehta2021modulated ,
    title   = { Modulated Periodic Activations for Generalizable Local Functional Representations } , 
    author  = { Ishit Mehta and Michaël Gharbi and Connelly Barnes and Eli Shechtman and Ravi Ramamoorthi and Manmohan Chandraker } ,
    year    = { 2021 } ,
    eprint  = { 2104.03960 } ,
    archivePrefix = { arXiv } ,
    primaryClass = { cs.CV }
}