Pytorch 05) - Image Recognition

Pytorch - 05) Image Recognition

마지막으루다가 손글씨 분류해보는 실습을 해보겠습니다.

Dataset

import torch
import matplotlib.pyplot as plt
import numpy as np

from torch import nn
from torch.nn import functional as F
from torchvision import datasets, transforms


# 리사이즈
# 텐서화
# channel, height, width를 정규화 한거임
# 처음 세개는 mean
# 그 다음 세개는 sd
transformer = transforms.Compose([transforms.Resize((28, 28)),
                                  transforms.ToTensor(),
                                  transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
                                 ])


training_dataset = datasets.MNIST(root='./data', train=True, download=True, transform=transformer)
validation_dataset = datasets.MNIST(root='./data', train=False, download=True, transform=transformer)

# 6000개를 100개 배치로 나눠서 돌림
training_loader = torch.utils.data.DataLoader(dataset=training_dataset, batch_size=100, shuffle=True)
validation_loader = torch.utils.data.DataLoader(dataset=validation_dataset, batch_size=100, shuffle=False)

텐서를 받아서 image로 보여주는 메소드를 만들자.

def im_convert(tensor):
  # 복제하고, 자동미분 끄고, numpy로
  image = tensor.clone().detach().numpy()
  # 데이터 형태는 color channel 1 28 px 28 px , 즉 1, 28, 28로 되어있음
  # 이거를 28, 28, 1 로 변경
  image = image.transpose(1, 2, 0)
  # denormalize
  image = image * np.array([0.5, 0.5, 0.5] + np.array([0.5, 0.5, 0.5]))
  # 데이터를 0과 1사이로만 있도록 보정
  image = image.clip(0, 1)
  return image

맛보기로 이미지를 한번 보자.

dataiter = iter(training_loader)
images, labels = dataiter.next()

fig = plt.figure(figsize=(25, 4))

for i in np.arange(20):
  # row 2 column 10
  ax = fig.add_subplot(2, 10, i+1, xticks=[], yticks=[])
  plt.imshow(im_convert(images[i]))
  ax.set_title(labels[i].item())

Model

class Classifier(nn.Module):

  def __init__(self, n_input, H1, H2, n_output):
    super().__init__()
    self.linear1 = nn.Linear(n_input, H1)
    self.linear2 = nn.Linear(H1, H2)
    self.linear3 = nn.Linear(H2, n_output)

  def forward(self, x):
    x = F.relu(self.linear1(x))
    x = F.relu(self.linear2(x))
    x = self.linear3(x)
    return x

hidden layer2개에, linear와 활성화 함수로 relu를 사용했다. sigmoid와 다르게 relu는 값 그자체가 확률을 나타내지 않으므로, 마지막에는 ReLU를 쓰지 않았다.

model = Classifier(784, 125, 65, 10)
model

Classifier(
  (linear1): Linear(in_features=784, out_features=125, bias=True)
  (linear2): Linear(in_features=125, out_features=65, bias=True)
  (linear3): Linear(in_features=65, out_features=10, bias=True)
)

이미지가 28*28이므로 784개의 input, 총 10개 (0~9)로 구별해야 하므로 output은 10개로 나타넀다. 가운데 125와 65는 그냥 임의로 넣어봤다. 해보면서 조절해보면 된다.

criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.005)

classification 문제라서 CrossEntroypyLoss를, 그리고 optimizer로는 Adam을 사용하였다.

training

epochs = 12
running_loss_history = []
running_correct_history = []
validation_running_loss_history = []
validation_running_correct_history = []

for e in range(epochs):

  running_loss = 0.0
  running_correct = 0.0
  validation_running_loss = 0.0
  validation_running_correct = 0.0

  for inputs, labels in training_loader:

    inputs = inputs.view(inputs.shape[0], -1)
    outputs = model(inputs)
    loss = criterion(outputs, labels)

    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

    _, preds = torch.max(outputs, 1)

    running_correct += torch.sum(preds == labels.data)
    running_loss += loss.item()



  else:
    # 훈련팔 필요가 없으므로 메모리 절약
    with torch.no_grad():

      for val_input, val_label in validation_loader:
        val_input = val_input.view(val_input.shape[0], -1)
        val_outputs = model(val_input)
        val_loss = criterion(val_outputs, val_label)

        _, val_preds = torch.max(val_outputs, 1)
        validation_running_loss += val_loss.item()
        validation_running_correct += torch.sum(val_preds == val_label.data)


    epoch_loss = running_loss / len(training_loader)
    epoch_acc = running_correct.float() / len(training_loader)
    running_loss_history.append(epoch_loss)
    running_correct_history.append(epoch_acc)

    val_epoch_loss = validation_running_loss / len(validation_loader)
    val_epoch_acc = validation_running_correct.float() / len(validation_loader)
    validation_running_loss_history.append(val_epoch_loss)
    validation_running_correct_history.append(val_epoch_acc)

    print("===================================================")
    print("epoch: ", e + 1)
    print("training loss: {:.5f}, {:5f}".format(epoch_loss, epoch_acc))
    print("validation loss: {:.5f}, {:5f}".format(val_epoch_loss, val_epoch_acc))

===================================================
epoch:  1
training loss: 0.34698, 89.241669
validation loss: 0.18939, 94.029999
===================================================
epoch:  2
training loss: 0.17725, 94.583336
validation loss: 0.16674, 94.790001
===================================================
epoch:  3
training loss: 0.15004, 95.341667
validation loss: 0.16987, 94.989998
===================================================
epoch:  4
training loss: 0.12949, 95.991669
validation loss: 0.13101, 96.040001
===================================================
epoch:  5
training loss: 0.11537, 96.478333
validation loss: 0.14699, 95.970001
===================================================
epoch:  6
training loss: 0.10967, 96.628334
validation loss: 0.11986, 96.410004
===================================================
epoch:  7
training loss: 0.10788, 96.681664
validation loss: 0.16216, 95.400002
===================================================
epoch:  8
training loss: 0.10198, 96.928337
validation loss: 0.13875, 96.029999
===================================================
epoch:  9
training loss: 0.09768, 97.044998
validation loss: 0.13843, 96.129997
===================================================
epoch:  10
training loss: 0.09128, 97.230003
validation loss: 0.14281, 96.349998
===================================================
epoch:  11
training loss: 0.09073, 97.208336
validation loss: 0.15617, 95.629997
===================================================
epoch:  12
training loss: 0.09084, 97.209999
validation loss: 0.13772, 96.470001

Loss

plt.plot(running_loss_history, label='training loss')
plt.plot(validation_running_loss_history, label='validation loss')
plt.legend()

Accuracy

plt.plot(running_correct_history, label='training accuracy')
plt.plot(validation_running_correct_history, label='validation accuracy')

정확도는 95% 안팎으로 나왔다. 그러나 validation loss가 training loss보다 커지는 것을 보아, 훈련하는 과정에서 overfitting이 일어났다고 봐야할 것이다. hyperparameter를 조금씩 조절할 필요가 있어 보인다.

Test

import requests
from PIL import Image
import PIL

url = 'https://images.homedepot-static.com/productImages/007164ea-d47e-4f66-8d8c-fd9f621984a2/svn/architectural-mailboxes-house-letters-numbers-3585b-5-64_1000.jpg'
response = requests.get(url, stream=True)
img = Image.open(response.raw)
plt.imshow(img)

# 반전
img = PIL.ImageOps.invert(img)
# 흑백으로 좀더 또렷하게
img = img.convert('1')
img = transformer(img)

plt.imshow(im_convert(img))

img = img.view(img.shape[0], -1)
outputs = model(img)
_, pred = torch.max(outputs, 1)
print(pred.item())

5

맞췄다. 한번 validation dataset에 다시 해보자.

dataiter = iter(validation_loader)
images, labels = dataiter.next()
images_ = images.view(images.shape[0], -1)
output = model(images_)
_, preds = torch.max(output, 1)

fig = plt.figure(figsize=(25, 4))

for i in np.arange(20):
  # row 2 column 10
  ax = fig.add_subplot(2, 10, i+1, xticks=[], yticks=[])
  plt.imshow(im_convert(images[i]))
  ax.set_title("{} ({})".format(str(preds[i].item()), str(labels[i].item())), color=('green' if preds[i] == labels[i] else 'red'))

하나 빼고 다 맞췄는데, 저건 내가봐도 6으로 볼 수도 있을 것 같다. 착한 분류기 ㅇㅈ 합니다.