Hi. I created a VAE which takes in a 720x1 vector of numbers as the input. I wanted to know how I could initialize the weights of my VAE using kaiming_uniform, so I could improve its performance. I tried searching online but was unable to put this in place.
Here is the code:
import os
import numpy as np
import matplotlib.pyplot as plt
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torchvision
from torchvision import datasets, transforms
from torch.autograd import Variable
from torchvision.utils import save_image
import os
from PIL import Image as PImage
device=“cpu”
trainloss=0
testloss=0
class VAE(nn.Module):
def init(self, z_dim):
super(VAE, self).init()
# Encoder
self.encode=nn.Sequential(
nn.Linear(720,540),
nn.ReLU(),
nn.Linear(540,400),
nn.ReLU(),
nn.Linear(400,300),
nn.ReLU(),
nn.Linear(300,225),
nn.ReLU(),
nn.Linear(225,150),
nn.ReLU(),
nn.Linear(150,100),
nn.ReLU(),
nn.Linear(100,60),
nn.ReLU(),
nn.Linear(60,45)
)
self.fc_mu = nn.Linear(45, z_dim)
self.fc_logvar = nn.Linear(45, z_dim)
# Decoder
self.decode=nn.Sequential(
nn.Linear(z_dim, 45),
nn.ReLU(),
nn.Linear(45, 60),
nn.ReLU(),
nn.Linear(60,100),
nn.ReLU(),
nn.Linear(100,150),
nn.ReLU(),
nn.Linear(150,225),
nn.ReLU(),
nn.Linear(225,300),
nn.ReLU(),
nn.Linear(300,400),
nn.ReLU(),
nn.Linear(400,540),
nn.ReLU(),
nn.Linear(540,720)
)
def encoder(self, x):
a=self.encode(x)
return self.fc_mu(a),self.fc_logvar(a)
def sampling(self, mu, logvar):
std = torch.exp(0.5 * logvar)
eps = torch.randn_like(std)
return eps.mul(std).add_(mu) # return z sample
def decoder(self, z):
return self.decode(z)
def forward(self, x):
mu, logvar = self.encoder(x)
z = self.sampling(mu, logvar)
return self.decoder(z), mu, logvar
vae = VAE(z_dim=20).to(device)
optimizer = optim.Adam(vae.parameters(), lr=0.001)
def loss_function(recon_x, x, mu, logvar):
myloss=nn.MSELoss()
MSE=myloss(recon_x,x)
KLD = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp())
return MSE + KLD
def trainortest(myvector):
optimizer.zero_grad()
recon_batch, mu, logvar = vae(myvector)
loss = loss_function(recon_batch, myvector, mu, logvar)
if(epoch<34000):
loss.backward()
myloss=loss.item()
optimizer.step()
else:
myloss=loss.item()
print(myloss)
if(epoch==100):
print(input_list)
print(recon_batch)
return myloss
with open(‘x_y (1).txt’, ‘r’) as f:
lines = f.readlines()
epoch=0
trainloss=0
testloss=0
for line in lines:
epoch+=1
x, y = line.split(‘=’)[0], line.split(‘=’)[1]
x, y = x.split(’ ‘), y.split(’ ')
x = [i for i in x if i]
y = [i for i in y if i]
x[0] = x[0][1:]
y[0] = y[0][1:]
x[-1] = x[-1][:-1]
y[-1] = y[-1][:-1]
x = [float(i) for i in x if i]
y = [float(i) for i in y if i]
S=np.zeros(360, dtype='complex_')
for i in range(0,360):
for k in range(360):
a=x[k]
b=y[k]
tmp = ((-2j*np.pi*i*k)) /360
S[i] += (complex(a,b)) * np.exp(tmp)
S[i]=S[i]/360
input_list=[]
for i in range(0,360):
input_list.append(np.real(S[i]))
input_list.append(np.imag(S[i]))
input_list=torch.FloatTensor(input_list)
answer=trainortest(input_list)
if(epoch<34000):
trainloss+=answer
else:
testloss+=answer
I would appreciate any help as to how I could initialize the weights of my VAE in the above code. Thank you!