[PYTHON] Effects of image rotation, enlargement, color, etc. on convolutional neural networks (CNN)

For the purpose of understanding the characteristics of convolutional neural networks (CNN) and other machine learning methods (gradient boosting, multi-layer perceptron), the classification performance is improved when the image is rotated, enlarged, or changed in color. I tried how it would change.

The image was generated by the method of Automatically generate images of koala and bear. Can you distinguish between koalas and bears by silhouette?

code

Image data reading

from PIL import Image

koalas = []
for i in range(num_data):
    koala = Image.open("koala_or_bear/koala_{}.jpg ".format(i))
    koalas.append(koala)

bears = []
for i in range(num_data):
    bear = Image.open("koala_or_bear/bear_{}.jpg ".format(i))
    bears.append(bear)

First data display

%matplotlib inline
import matplotlib.pyplot as plt

fig = plt.figure(figsize=(10,10))
for i in range(16):
    ax = fig.add_subplot(4, 4, i+1)
    ax.axis('off')
    if i < 8:
        ax.set_title('koala_{}'.format(i))
        ax.imshow(koalas[i],cmap=plt.cm.gray, interpolation='none')
    else:
        ax.set_title('bear_{}'.format(i - 8))
        ax.imshow(bears[i - 8],cmap=plt.cm.gray, interpolation='none')
plt.show()

Explanatory variable / objective variable

import numpy as np

X = [] #Explanatory variable
Y = [] #Objective variable

index = 0
for koala in koalas:
    resize_img = koala.resize((128, 128))
    r, g, b = resize_img.split()
    r_resize_img = np.asarray(np.float32(r)/255.0)
    g_resize_img = np.asarray(np.float32(g)/255.0)
    b_resize_img = np.asarray(np.float32(b)/255.0)
    rgb_resize_img = np.asarray([r_resize_img, g_resize_img, b_resize_img])
    X.append(rgb_resize_img)
    Y.append(0)
    index += 1
    if index >= num_data:
        break

index = 0
for bear in bears:
    resize_img = bear.resize((128, 128))
    r, g, b = resize_img.split()
    r_resize_img = np.asarray(np.float32(r)/255.0)
    g_resize_img = np.asarray(np.float32(g)/255.0)
    b_resize_img = np.asarray(np.float32(b)/255.0)
    rgb_resize_img = np.asarray([r_resize_img, g_resize_img, b_resize_img])
    X.append(rgb_resize_img)
    Y.append(1)
    index += 1
    if index >= num_data:
        break

X = np.array(X, dtype='float32')
Y = np.array(Y, dtype='int64')

Separated into teacher set and test set

from sklearn import model_selection
X_train, X_test, Y_train, Y_test = model_selection.train_test_split(
    X, Y, test_size=0.1
)

Data type conversion for scikit-learn

d1, d2, d3, d4 = X_train.shape
X_train_a = X_train.reshape((d1, d2 * d3 * d4))
Y_train_onehot = np.identity(2)[Y_train]
d1, d2, d3, d4 = X_test.shape
X_test_a = X_test.reshape((d1, d2 * d3 * d4))
Y_test_onehot = np.identity(2)[Y_test]

Data type conversion for PyTorch

import torch
from torch.utils.data import TensorDataset
from torch.utils.data import DataLoader

X_train_t = torch.from_numpy(X_train).float()
Y_train_t = torch.from_numpy(Y_train).long()

X_train_v = torch.autograd.Variable(X_train_t)
Y_train_v = torch.autograd.Variable(Y_train_t)

X_test_t = torch.from_numpy(X_test).float()
Y_test_t = torch.from_numpy(Y_test).long()

X_test_v = torch.autograd.Variable(X_test_t)
Y_test_v = torch.autograd.Variable(Y_test_t)

train = TensorDataset(X_train_t, Y_train_t)
train_loader = DataLoader(train, batch_size=32, shuffle=True)

Gradient boosting

%%time
from sklearn.ensemble import GradientBoostingClassifier

classifier = GradientBoostingClassifier()
classifier.fit(X_train_a, Y_train)
print("Accuracy score (train): ", classifier.score(X_train_a, Y_train))
print("Accuracy score (test): ", classifier.score(X_test_a, Y_test))

Multilayer perceptron (1 intermediate layer)

%%time
from sklearn.neural_network import MLPClassifier

classifier = MLPClassifier(max_iter=10000, early_stopping=True)
classifier.fit(X_train_a, Y_train)
print("Accuracy score (train): ", classifier.score(X_train_a, Y_train))
print("Accuracy score (test): ", classifier.score(X_test_a, Y_test))

Multilayer perceptron (2 intermediate layers)

%%time
from sklearn.neural_network import MLPClassifier
classifier = MLPClassifier(max_iter=10000, early_stopping=True,
                           hidden_layer_sizes=(100, 100))
classifier.fit(X_train_a, Y_train)
print("Accuracy score (train): ", classifier.score(X_train_a, Y_train))
print("Accuracy score (test): ", classifier.score(X_test_a, Y_test))

Convolutional Neural Network (CNN)

Network definition

class CNN(torch.nn.Module):
    def __init__(self):
        super(CNN, self).__init__()
        self.conv1 = torch.nn.Conv2d(3, 10, 5)
        self.conv2 = torch.nn.Conv2d(10, 20, 5)
        self.fc1 = torch.nn.Linear(20 * 29 * 29, 50)
        self.fc2 = torch.nn.Linear(50, 2)
    
    def forward(self, x):
        x = torch.nn.functional.relu(self.conv1(x))
        x = torch.nn.functional.max_pool2d(x, 2)
        x = torch.nn.functional.relu(self.conv2(x))
        x = torch.nn.functional.max_pool2d(x, 2)
        x = x.view(-1, 20 * 29 * 29)
        x = torch.nn.functional.relu(self.fc1(x))
        x = torch.nn.functional.log_softmax(self.fc2(x), 1)
        return x

Checking the network structure

from torchsummary import summary
model = CNN()
summary(model, X[0].shape)

----------------------------------------------------------------
        Layer (type)               Output Shape         Param #
================================================================
            Conv2d-1         [-1, 10, 124, 124]             760
            Conv2d-2           [-1, 20, 58, 58]           5,020
            Linear-3                   [-1, 50]         841,050
            Linear-4                    [-1, 2]             102
================================================================
Total params: 846,932
Trainable params: 846,932
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.19
Forward/backward pass size (MB): 1.69
Params size (MB): 3.23
Estimated Total Size (MB): 5.11
----------------------------------------------------------------

Function for execution

def learn(model, criterion, optimizer, n_iteration):
    for epoch in range(n_iteration):
        total_loss = np.array(0, dtype='float64')
        for x, y in train_loader:
            x = torch.autograd.Variable(x)
            y = torch.autograd.Variable(y)
            optimizer.zero_grad()
            y_pred = model(x)
            loss = criterion(y_pred, y)
            loss.backward()
            optimizer.step()
            total_loss += loss.data.numpy()
        
        if (epoch + 1) % 10 == 0:
            print(epoch + 1, total_loss)
        
        if total_loss == np.array(0, dtype='float64'):
            break
        
        loss_history.append(total_loss)
    
    return model

Execution code

#%%time
model = CNN()
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
loss_history = []
model = learn(model, criterion, optimizer, 300)

Loss history display

ax = plt.subplot(2, 1, 1)
ax.plot(loss_history)
ax.grid()
ax = plt.subplot(2, 1, 2)
ax.plot(loss_history)
ax.set_yscale('log')
ax.grid()

Correct answer rate

Correct answer rate (teacher set)

Y_pred = torch.max(model(X_train_v).data, 1)[1]
accuracy = sum(Y_train == Y_pred.numpy()) / len(Y_train)
print(accuracy)

Correct answer rate (test set)

Y_pred = torch.max(model(X_test_v).data, 1)[1]
accuracy = sum(Y_test == Y_pred.numpy()) / len(Y_test)
print(accuracy)

result

The first is the simplest question to distinguish between koalas and bears.

All prediction methods were able to answer perfectly.

Expansion

Let's magnify Koala and Kuma at random magnifications. At this time, I will try to shift it slightly vertically and horizontally.

The percentage of correct answers for gradient boosting has dropped slightly. Convolutional Neural Networks (CNN) have the perfect percentage of correct answers.

There is an obstacle

Draw something that might get in the way in the background.

It seems to have almost no effect.

Color only the background

Let's make the background colorful.

This also seems to have little effect.

Color only animals

Let's make koalas and bears colorful.

The prediction accuracy has dropped considerably. The performance of the convolutional neural network (CNN) has dropped. that? Surprisingly, the multi-layer perceptron (middle two layers) is doing well. And the performance of gradient boosting has hardly deteriorated. It's amazing.

Enlarged / colored / with obstacles

The difficulty has increased, but the gradient boosting is pretty good. Convolutional neural networks (CNNs) are completely useless.

Enlargement / coloring / animals only black / with obstacles

All prediction performance was restored just by unifying the colors of Koala and Kuma to black. After all, that is the criterion.

rotation

Koalas spin, bears spin

The multi-layer perceptron (middle 1 layer) didn't work, but everyone else seems to have done it about rotation. Gradient boosting seems to be a little weak in rotation.

Rotate / enlarge

Let's expand while rotating.

It's okay if it's just rotation, it's okay if it's just expansion. However, it seems that everyone gets confused if it expands while rotating. Still, gradient boosting, I'm doing my best.

Rotation / enlargement / coloring / black only for animals / with obstacles

If various factors (rotation, enlargement, obstacles, colors) are attacked individually, even if it is not a big problem, it seems that it will be difficult if those factors are mixed and attacked.

All the various factors

I was able to ride all the various factors. This seems to be a mess in any way.

Summary

I wanted to get a feel for the convolutional neural network (CNN), so I wanted to get a "CNN wins alone !!!" </ b> result, but the conclusion is "gradient boosting". Great yeah yeah! "</ B>. (ヽ ´ω`)