How would I increase my accuracy in the cifar-100 dataset? I have a 10% accuracy at the moment

Question

I am doing a small project for fun with the cifar-100 dataset. I'm not sure why I have a low 10% accuracy. Here is my code, could anyone help? thanks all (fyi, cifar-100 dataset is a keras dataset with 100 types of images.)

import tensorflow as tf
from tensorflow import keras
import numpy as np
import matplotlib.pyplot as plt

#print(train_images[0])
#print("Network Accuracy: " + str(test_acc))
# plt.imshow(train_images[0], cmap=plt.cm.binary) #greyscale
# plt.imshow(train_images[0]) #neon
# plt.show()

cifar100_mnist = keras.datasets.cifar100

(train_images, train_labels), (test_images, test_labels) = cifar100_mnist.load_data()
print(train_images)
print("-")
print(train_labels)
train_images = train_images/255
test_images = test_images/255
classes = [
    'apple', 'aquarium_fish', 'baby', 'bear', 'beaver', 'bed', 'bee', 'beetle',
    'bicycle', 'bottle', 'bowl', 'boy', 'bridge', 'bus', 'butterfly', 'camel',
    'can', 'castle', 'caterpillar', 'cattle', 'chair', 'chimpanzee', 'clock',
    'cloud', 'cockroach', 'couch', 'crab', 'crocodile', 'cup', 'dinosaur',
    'dolphin', 'elephant', 'flatfish', 'forest', 'fox', 'girl', 'hamster',
    'house', 'kangaroo', 'keyboard', 'lamp', 'lawn_mower', 'leopard', 'lion',
    'lizard', 'lobster', 'man', 'maple_tree', 'motorcycle', 'mountain', 'mouse',
    'mushroom', 'oak_tree', 'orange', 'orchid', 'otter', 'palm_tree', 'pear',
    'pickup_truck', 'pine_tree', 'plain', 'plate', 'poppy', 'porcupine',
    'possum', 'rabbit', 'raccoon', 'ray', 'road', 'rocket', 'rose',
    'sea', 'seal', 'shark', 'shrew', 'skunk', 'skyscraper', 'snail', 'snake',
    'spider', 'squirrel', 'streetcar', 'sunflower', 'sweet_pepper', 'table',
    'tank', 'telephone', 'television', 'tiger', 'tractor', 'train', 'trout',
    'tulip', 'turtle', 'wardrobe', 'whale', 'willow_tree', 'wolf', 'woman',
    'worm'
]
model = keras.Sequential([
    keras.layers.Flatten(input_shape=(32, 32, 3)),
    keras.layers.Dense(500, activation="relu"),
    keras.layers.Dense(100, activation="softmax")
])

model.compile(optimizer="adam", loss="sparse_categorical_crossentropy", metrics=["accuracy"])
model.fit(train_images, train_labels, epochs=5)

test_loss, test_acc = model.evaluate(test_images, test_labels)
print(test_acc)
# print(test_images)
prediction = model.predict(test_images)
answer = np.argmax(prediction[0])
print(classes[answer])
# print(train_images[0])
plt.imshow(train_images[0])
plt.show()

Hope that I can get an answer, thanks guys, I really appreciate the help.

Answer 1

An example of Implementing Convolutions regarding your task:

model = tf.keras.models.Sequential([
    tf.keras.layers.Conv2D(64, (3,3), activation='relu', input_shape=(32, 32, 3)),
    tf.keras.layers.MaxPooling2D(2, 2),
    tf.keras.layers.Conv2D(64, (3, 3), activation='relu'),
    tf.keras.layers.MaxPooling2D(2, 2),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(512, activation='relu'),
    tf.keras.layers.Dense(100, activation='softmax')
])

model.summary()

Try more epochs (e.g. 25):

model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])

history = model.fit(train_images, train_labels, epochs=25)

You gain high Accuracy on the Trainingset:

Epoch 25/25 50000/50000 [==============================] - 12s 248us/sample - loss: 0.2362 - acc: 0.9243

Let's have a look on the Test Accuracy:

test_loss, test_acc = model.evaluate(test_images, test_labels)
print(test_acc)
prediction = model.predict(test_images)

10000/10000 [==============================] - 1s 139us/sample - loss: 7.7701 - acc: 0.3272 0.3272

Which means your model overfits the trainingset.

Try it yourself:

image_number = 5
answer = np.argmax(prediction[image_number])
print(classes[answer])
plt.imshow(train_images[image_number])

lizard

obviously no lizard :)

Here begins the Task of Machine Learning. To improve your model to reduce bias and variance and gain high accuracy. I am still learning too, so i hope this points you in the right direction.

But be aware of this dataset. You can read on the CIFAR-100 Page:

There are 100 classes containing 600 images each. There are 500 training images and 100 testing images per class.

500 images is way too less to train a CNN and this will lead to overfitting. I think the CIFAR 10 is more likely for beginners? They mention on the same Page:

[...] 18% test error without data augmentation [...]

Which would be much more fun to experiment with :)