torch.nn.Module 신경망에서 초기가중치 변경하기

torch.nn.Module에서 초기가중치 변경

# Import Numpy & PyTorch

import numpy as np

import torch

 

# Imports

import torch.nn as nn

 

# Import tensor dataset & data loader

from torch.utils.data import TensorDataset, DataLoader

 

# Import nn.functional

import torch.nn.functional as F

 

# to view module

import inspect

 
# 입력

inputs = np.array([[3.0, 13], [4, 14], [5, 15]], dtype='float32') # 입력변수가 1개

targets = np.array([[6.0, 26], [8, 28], [10, 30]], dtype='float32') # 타깃이 1개

 

inputs = torch.from_numpy(inputs)

targets = torch.from_numpy(targets)

 

# Define dataset

train_ds = TensorDataset(inputs, targets)

print(train_ds[0:1])

 

# Define data loader

batch_size = 3

train_dl = DataLoader(train_ds, batch_size, shuffle=False)

 

# Draw data loader

for xb, yb in train_dl:

print("xb =", xb)

print("xb =", xb)

--------------------------------------

(tensor([[ 3., 13.]]), tensor([[ 6., 26.]]))

xb = tensor([[ 3., 13.],

[ 4., 14.],

[ 5., 15.]])

xb = tensor([[ 3., 13.],

[ 4., 14.],

[ 5., 15.]])

---------------------------------------

# 가중치 세팅

torch.manual_seed(777)

a = torch.randn(2,4).t()

a

---------------------------------------

tensor([[-0.4015, -0.9433],

[-0.5934, -0.6758],

[ 1.6885, 1.4425],

[ 0.5554, -1.4711]])

-----------------------------------

class SimpleNet(nn.Module):

    

    # Initialize the layers

    def __init__(self):

        super().__init__()  # 

        self.linear1 = nn.Linear(2, 4, bias=False)

        self.linear2 = nn.Linear(4, 2, bias=False)

        

        # 가중치를 변경할 경우 important !!! 이거 알려고 하루가 걸림 !!!

        with torch.no_grad():

            torch.manual_seed(777)

            a = torch.randn(2,4, requires_grad=True).t()

            b = torch.randn(4,2, requires_grad=True).t()

            self.linear1.weight.copy_(a)

            self.linear2.weight.copy_(b)

            

    # Perform the computation

    def forward(self, x):

        x = self.linear1(x)

        x = self.linear2(x)

        return x

model2 = SimpleNet()

model2(inputs)

 

# model2(inputs)와 동일한 기능

model2.forward(inputs)

 
# 모델, 옵티마이저, 비용함수

model4 = SimpleNet()

opt = torch.optim.SGD(model4.parameters(), 1e-3)

loss_fn = F.mse_loss

 

# Define a utility function to train the model

def fit(num_epochs, model, loss_fn, opt):

   for epoch in range(num_epochs):

   for xb, yb in train_dl:

   # Generate predictions

   pred = model(xb)

   loss = loss_fn(pred, yb)

   print(epoch, "loss =", loss)

 

   # Perform gradient descent

   loss.backward()

   opt.step() # 가중치를 update하는 것 (update rule에 의거)

   opt.zero_grad()  # 미분결과 축적을 방지하기 위해 초기화

 

fit(1, model4, loss_fn, opt)

-----------------------------------------------------

0 loss = tensor(2957.1760, grad_fn=<MseLossBackward>)