多値分類

 

参考文献: PyTorch & 深層学習プログラミング 赤石雅典著 日経BP社

pip3.11 install torchviz
pip3.11 install torchinfo

python3.11

import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim

# データ準備
x_data = torch.randn(150,2)

rng = np.random.default_rng()
y_data = rng.integers(3, size=(150) )
y_data = torch.tensor(y_data)

print(x_data[:5,:])
print(y_data[:5])

# 訓練データ、検証データに分割 (シャフルも同時に実施)
from sklearn.model_selection import train_test_split
x_train, x_test, y_train, y_test = train_test_split(
  x_data, y_data, train_size=75, test_size=75,
  random_state=123)
print(x_train.shape, x_test.shape, y_train.shape, y_test.shape)

# モデル定義
n_input= 2
n_output = 3

class Net(nn.Module):
  def __init__(self, n_input, n_output):
    super().__init__()
    self.l1 = nn.Linear(n_input, n_output)
  
  def forward(self, x):
    x1 = self.l1(x)
    return x1

 

# インスタンス生成

net = Net(n_input, n_output)

for parameter in net.named_parameters():
  print(parameter)

for parameter in net.parameters():
  print(parameter)

print(net)

from torchinfo import summary
summary(net, (2,))

# データのテンソル化
inputs = torch.tensor(x_train).float()
labels = torch.tensor(y_train).long()

inputs_test = torch.tensor(x_test).float()
labels_test = torch.tensor(y_test).long()

# 損失関数
criterion = nn.CrossEntropyLoss()

# 学習率
lr = 0.01

# 最適化関数
optimizer = optim.SGD(net.parameters(), lr=lr)

# 繰り返し回数
num_epochs = 100

# 繰り返し計算

for epoch in range(num_epochs):
  # 訓練フェーズ
  #勾配の初期化★
  optimizer.zero_grad()
  # 予測計算
  outputs = net(inputs)
  # 損失計算
  loss = criterion(outputs, labels)
  # 勾配計算★
  loss.backward()
  # パラメータ修正★
  optimizer.step()
  # 予測ラベル算出
  predicted = torch.max(outputs, 1)[1]
  # 損失と精度の計算
  train_loss = loss.item()
  train_acc = (predicted == labels).sum()  / len(labels)
  
  #予測フェーズ
  # 予測計算
  outputs_test = net(inputs_test)
  # 損失計算
  loss_test = criterion(outputs_test, labels_test)
  # 予測ラベル算出
  predicted_test = torch.max(outputs_test, 1)[1]
  # 損失と精度の計算
  val_loss =  loss_test.item()
  val_acc =  (predicted_test == labels_test).sum() / len(labels_test)
  if ((epoch) % 10 == 0):
    print (f'Epoch [{epoch}/{num_epochs}], loss: {train_loss:.5f} acc: {train_acc:.5f} val_loss: {val_loss:.5f}, val_acc: {val_acc:.5f}')

 

# 重み行列
print(net.l1.weight.data)

# バイアス
print(net.l1.bias.data)