Below is a CNN trained on Alien and Predator images obtained form: https://www.kaggle.com/pmigdal/alien-vs-predator-images
- Dataset contains 347 test samples of class Alien and 347 test samples of class Predator
In [2]:
import os
import matplotlib.pyplot as plt
import numpy as np
import cv2
import random
from sklearn.utils import shuffle
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Flatten, Conv2D, Dropout, MaxPooling2D
In [3]:
class_names_label = {'alien':0, 'predator':1}
In [4]:
# Preparing data
def load_data():
datasets = ['alien-vs-predator-images/data/train', 'alien-vs-predator-images/data/test']
size = (150,150)
output = []
for dataset in datasets:
images = []
labels = []
for class_name in os.listdir(dataset):
curr_label = class_names_label[class_name]
for img in os.listdir(dataset + '/' + '/' + class_name):
img_path = dataset + '/' + '/' + class_name + '/' + img
curr_img = cv2.imread(img_path)
curr_img = cv2.resize(curr_img, size)
images.append(curr_img)
labels.append(curr_label)
images = np.asarray(images, dtype='float32')
labels = np.asarray(labels, dtype='float32')
images, labels = shuffle(images, labels)
output.append((images, labels))
return output
In [5]:
(x_train, y_train), (x_test, y_test) = load_data()
x_train /= 255
x_test /= 255
In [23]:
# Exploring Data
sample = [0, 10, 40, 193, 333, 412, 252, 14, 600]
fig = plt.figure()
plt.subplots(3,3, figsize=(16,8))
for i in range(9):
plt.subplot(3,3,i+1)
if y_train[sample[i]] == 0:
title = 'Alien'
else:
title = 'Predator'
plt.title(title)
plt.tight_layout()
plt.imshow(x_train[sample[i]])
In [7]:
# Building Model
model = Sequential()
model.add(Conv2D(32,(3,3), activation='relu', input_shape = (150, 150, 3)))
model.add(MaxPooling2D(pool_size=(2,2)))
model.add(Conv2D(32,(3,3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2,2)))
model.add(Conv2D(32,(3,3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2,2)))
model.add(Flatten())
model.add(Dense(128,activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(1,activation='sigmoid'))
model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])
In [29]:
model.summary()
In [30]:
history = model.fit(x_train, y_train, epochs=20)
In [31]:
model.evaluate(x_test, y_test)
Out[31]:
In [ ]:
model.save('AVPmodel_1.h5')
In [108]:
#Looking at misclassified images
predicted = np.round(model.predict(x_test),0).reshape(200,)
misclassified = predicted != y_test
plt.figure()
plt.subplots(1,4, figsize=(16,8))
misclassified_samples = np.arange(1,5,1)
for i in misclassified_samples:
if y_test[i] == 0:
title = 'Alien'
else:
title = 'Predator'
plt.subplot(1,4,i)
plt.imshow(x_test[misclassified][i])
plt.tight_layout()
plt.title(title)
RESULTS
- Accuracy on training set: 98.4%
- Accuracy on test set: 77.5%
DISCUSSION
The model overfit on training data and is not able to generalize as well to the test set. The model architecture chosen has 3 convolutional layers making it deep with a total of 1.2M trainable parameters. Such a deep CNN is susceptible to overfitting with only 694 samples.
However, a dropout layer was introduced to reduce overfitting. Also, images were downscaled to 150 x 150 to help combat overfitting.
Improvements
- Model capacity could be reduced by removing layers (This was attempted with minimal success)
- Training set can be augmented using a data generated to distort and transform images allowing better generalization
- 694 training samples may be a small dataset for a CNN of this depth. More data would help with generalization
- Upon analysis of dataset quality; data could be better represented with more diversity of colour images and types of images (sketch vs CGI)
- Add weight regularizers