The CIFAR-10 dataset (Canadian Institute For Advanced Research) is a collection of images that are commonly used to train machine learning and computer vision algorithms. It is one of the most widely used datasets for machine learning research. The CIFAR-10 dataset contains 60,000 32x32 color images in 10 different classes. The 10 different classes represent airplanes, cars, birds, cats, deer, dogs, frogs, horses, ships, and trucks. There are 6,000 images of each class.
import numpy as np
import pandas as pd
from packaging import version
from sklearn.metrics import confusion_matrix, classification_report
from sklearn.metrics import accuracy_score
from sklearn.metrics import mean_squared_error as MSE
from sklearn.model_selection import train_test_split
import matplotlib.pyplot as plt
import seaborn as sns
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import models, layers
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, MaxPool2D, BatchNormalization, Dropout, Flatten, Dense
from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping
from tensorflow.keras.preprocessing import image
from tensorflow.keras.utils import to_categorical
%matplotlib inline
np.set_printoptions(precision=3, suppress=True)
print("This notebook requires TensorFlow 2.0 or above")
print("TensorFlow version: ", tf.__version__)
assert version.parse(tf.__version__).release[0] >=2
This notebook requires TensorFlow 2.0 or above TensorFlow version: 2.10.0
print("Keras version: ", keras.__version__)
Keras version: 2.10.0
# from google.colab import drive
# drive.mount('/content/gdrive')
def get_three_classes(x, y):
def indices_of(class_id):
indices, _ = np.where(y == float(class_id))
return indices
indices = np.concatenate([indices_of(0), indices_of(1), indices_of(2)], axis=0)
x = x[indices]
y = y[indices]
count = x.shape[0]
indices = np.random.choice(range(count), count, replace=False)
x = x[indices]
y = y[indices]
y = tf.keras.utils.to_categorical(y)
return x, y
def show_random_examples(x, y, p):
indices = np.random.choice(range(x.shape[0]), 10, replace=False)
x = x[indices]
y = y[indices]
p = p[indices]
plt.figure(figsize=(10, 5))
for i in range(10):
plt.subplot(2, 5, i + 1)
plt.imshow(x[i])
plt.xticks([])
plt.yticks([])
col = 'green' if np.argmax(y[i]) == np.argmax(p[i]) else 'red'
plt.xlabel(class_names_preview[np.argmax(p[i])], color=col)
plt.show()
def plot_history(history):
losses = history.history['loss']
accs = history.history['accuracy']
val_losses = history.history['val_loss']
val_accs = history.history['val_accuracy']
epochs = len(losses)
plt.figure(figsize=(16, 4))
for i, metrics in enumerate(zip([losses, accs], [val_losses, val_accs], ['Loss', 'Accuracy'])):
plt.subplot(1, 2, i + 1)
plt.plot(range(epochs), metrics[0], label='Training {}'.format(metrics[2]))
plt.plot(range(epochs), metrics[1], label='Validation {}'.format(metrics[2]))
plt.legend()
plt.show()
def print_validation_report(y_test, predictions):
print("Classification Report")
print(classification_report(y_test, predictions))
print('Accuracy Score: {}'.format(accuracy_score(y_test, predictions)))
print('Root Mean Square Error: {}'.format(np.sqrt(MSE(y_test, predictions))))
def plot_confusion_matrix(y_true, y_pred):
mtx = confusion_matrix(y_true, y_pred)
fig, ax = plt.subplots(figsize=(8,8))
sns.heatmap(mtx, annot=True, fmt='d', linewidths=.75, cbar=False, ax=ax,cmap='Blues',linecolor='white')
# square=True,
plt.ylabel('true label')
plt.xlabel('predicted label')
The CIFAR-10 dataset consists of 60000 32x32 colour images in 10 classes, with 6000 images per class. There are 50000 training images and 10000 test images.
The dataset is divided into five training batches and one test batch, each with 10000 images. The test batch contains exactly 1000 randomly-selected images from each class. The training batches contain the remaining images in random order, but some training batches may contain more images from one class than another. Between them, the training batches contain exactly 5000 images from each class.
(x_train, y_train), (x_test, y_test) = keras.datasets.cifar10.load_data()
print('train_images:\t{}'.format(x_train.shape))
print('train_labels:\t{}'.format(y_train.shape))
print('test_images:\t\t{}'.format(x_test.shape))
print('test_labels:\t\t{}'.format(y_test.shape))
train_images: (50000, 32, 32, 3) train_labels: (50000, 1) test_images: (10000, 32, 32, 3) test_labels: (10000, 1)
print("First ten labels training dataset:\n {}\n".format(y_train[0:10]))
print("This output the numeric label, need to convert to item description")
First ten labels training dataset: [[6] [9] [9] [4] [1] [1] [2] [7] [8] [3]] This output the numeric label, need to convert to item description
(train_images, train_labels),(test_images, test_labels)= tf.keras.datasets.cifar10.load_data()
x_preview, y_preview = get_three_classes(train_images, train_labels)
x_preview, y_preview = get_three_classes(test_images, test_labels)
class_names_preview = ['aeroplane', 'car', 'bird']
show_random_examples(x_preview, y_preview, y_preview)
The labels are an array of integers, ranging from 0 to 9. These correspond to the class of clothing the image represents:
| Label | Class_ |
|---|---|
| 0 | airplane |
| 1 | automobile |
| 2 | bird |
| 3 | cat |
| 4 | deer |
| 5 | dog |
| 6 | frog |
| 7 | horse |
| 8 | ship |
| 9 | truck |
class_names = ['airplane'
,'automobile'
,'bird'
,'cat'
,'deer'
,'dog'
,'frog'
,'horse'
,'ship'
,'truck']
x_train_split, x_valid_split, y_train_split, y_valid_split = train_test_split(x_train
,y_train
,test_size=.1
,random_state=42
,shuffle=True)
print(x_train_split.shape, x_valid_split.shape, x_test.shape)
(45000, 32, 32, 3) (5000, 32, 32, 3) (10000, 32, 32, 3)
The images are 28x28 NumPy arrays, with pixel values ranging from 0 to 255
x_train_norm = x_train_split/255
x_valid_norm = x_valid_split/255
x_test_norm = x_test/255
We use a Sequential class defined in Keras to create our model. The first 9 layers Conv2D MaxPooling handle feature learning. The last 3 layers, handle classification
model = Sequential([
Conv2D(filters=128, kernel_size=(3, 3), strides=(1, 1), activation=tf.nn.relu,input_shape=x_train_norm.shape[1:]),
MaxPool2D((2, 2),strides=2),
Dropout(0.3),
Conv2D(filters=256, kernel_size=(3, 3), strides=(1, 1), activation=tf.nn.relu),
MaxPool2D((2, 2),strides=2),
Dropout(0.3),
Flatten(),
Dense(units=10, activation=tf.nn.softmax)
])
model.summary()
Model: "sequential_1"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_2 (Conv2D) (None, 30, 30, 128) 3584
max_pooling2d_2 (MaxPooling (None, 15, 15, 128) 0
2D)
dropout_2 (Dropout) (None, 15, 15, 128) 0
conv2d_3 (Conv2D) (None, 13, 13, 256) 295168
max_pooling2d_3 (MaxPooling (None, 6, 6, 256) 0
2D)
dropout_3 (Dropout) (None, 6, 6, 256) 0
flatten_1 (Flatten) (None, 9216) 0
dense_1 (Dense) (None, 10) 92170
=================================================================
Total params: 390,922
Trainable params: 390,922
Non-trainable params: 0
_________________________________________________________________
keras.utils.plot_model(model, "CIFAR10.png", show_shapes=True)
You must install pydot (`pip install pydot`) and install graphviz (see instructions at https://graphviz.gitlab.io/download/) for plot_model to work.
In addition to setting up our model architecture, we also need to define which algorithm should the model use in order to optimize the weights and biases as per the given data. We will use stochastic gradient descent.
We also need to define a loss function. Think of this function as the difference between the predicted outputs and the actual outputs given in the dataset. This loss needs to be minimised in order to have a higher model accuracy. That's what the optimization algorithm essentially does - it minimises the loss during model training. For our multi-class classification problem, categorical cross entropy is commonly used.
Finally, we will use the accuracy during training as a metric to keep track of as the model trains.
model.compile(optimizer='adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False),
metrics=['accuracy'])
history = model.fit(x_train_norm
,y_train_split
,epochs=200
,batch_size=64
,validation_data=(x_valid_norm, y_valid_split)
,callbacks=[
tf.keras.callbacks.ModelCheckpoint("CNN_model.h5",save_best_only=True,save_weights_only=False)
,tf.keras.callbacks.EarlyStopping(monitor='val_accuracy', patience=3),
]
)
Epoch 1/200 704/704 [==============================] - 133s 188ms/step - loss: 1.4931 - accuracy: 0.4670 - val_loss: 1.3009 - val_accuracy: 0.5442 Epoch 2/200 704/704 [==============================] - 133s 189ms/step - loss: 1.1343 - accuracy: 0.6057 - val_loss: 1.1113 - val_accuracy: 0.6064 Epoch 3/200 704/704 [==============================] - 133s 188ms/step - loss: 1.0079 - accuracy: 0.6508 - val_loss: 0.9516 - val_accuracy: 0.6696 Epoch 4/200 704/704 [==============================] - 137s 194ms/step - loss: 0.9285 - accuracy: 0.6790 - val_loss: 0.8927 - val_accuracy: 0.6888 Epoch 5/200 704/704 [==============================] - 220s 313ms/step - loss: 0.8535 - accuracy: 0.7064 - val_loss: 0.8738 - val_accuracy: 0.6986 Epoch 6/200 704/704 [==============================] - 349s 496ms/step - loss: 0.8148 - accuracy: 0.7207 - val_loss: 0.8661 - val_accuracy: 0.7006 Epoch 7/200 704/704 [==============================] - 359s 510ms/step - loss: 0.7819 - accuracy: 0.7332 - val_loss: 0.8422 - val_accuracy: 0.7066 Epoch 8/200 704/704 [==============================] - 281s 399ms/step - loss: 0.7449 - accuracy: 0.7425 - val_loss: 0.7893 - val_accuracy: 0.7270 Epoch 9/200 704/704 [==============================] - 233s 331ms/step - loss: 0.7095 - accuracy: 0.7550 - val_loss: 0.7980 - val_accuracy: 0.7170 Epoch 10/200 704/704 [==============================] - 146s 207ms/step - loss: 0.6957 - accuracy: 0.7602 - val_loss: 0.8033 - val_accuracy: 0.7162 Epoch 11/200 704/704 [==============================] - 132s 188ms/step - loss: 0.6702 - accuracy: 0.7676 - val_loss: 0.7936 - val_accuracy: 0.7242
In order to ensure that this is not a simple "memorization" by the machine, we should evaluate the performance on the test set. This is easy to do, we simply use the evaluate method on our model.
model = tf.keras.models.load_model("CNN_model.h5")
print(f"Test acc: {model.evaluate(x_test_norm, y_test)[1]:.3f}")
313/313 [==============================] - 9s 26ms/step - loss: 0.7502 - accuracy: 0.7411 Test acc: 0.741
preds = model.predict(x_test_norm)
print('shape of preds: ', preds.shape)
313/313 [==============================] - 9s 27ms/step shape of preds: (10000, 10)
We use Matplotlib to create 2 plots--displaying the training and validation loss (resp. accuracy) for each (training) epoch side by side.
history_dict = history.history
history_dict.keys()
dict_keys(['loss', 'accuracy', 'val_loss', 'val_accuracy'])
history_df=pd.DataFrame(history_dict)
history_df.tail().round(3)
| loss | accuracy | val_loss | val_accuracy | |
|---|---|---|---|---|
| 6 | 0.782 | 0.733 | 0.842 | 0.707 |
| 7 | 0.745 | 0.743 | 0.789 | 0.727 |
| 8 | 0.710 | 0.755 | 0.798 | 0.717 |
| 9 | 0.696 | 0.760 | 0.803 | 0.716 |
| 10 | 0.670 | 0.768 | 0.794 | 0.724 |
plot_history(history)
Using both sklearn.metrics. Then we visualize the confusion matrix and see what that tells us.
pred1= model.predict(x_test_norm)
pred1=np.argmax(pred1, axis=1)
313/313 [==============================] - 9s 29ms/step
print_validation_report(y_test, pred1)
Classification Report
precision recall f1-score support
0 0.72 0.83 0.77 1000
1 0.80 0.92 0.85 1000
2 0.71 0.57 0.63 1000
3 0.62 0.47 0.54 1000
4 0.72 0.67 0.69 1000
5 0.69 0.65 0.67 1000
6 0.66 0.91 0.76 1000
7 0.79 0.78 0.79 1000
8 0.81 0.87 0.84 1000
9 0.90 0.74 0.81 1000
accuracy 0.74 10000
macro avg 0.74 0.74 0.74 10000
weighted avg 0.74 0.74 0.74 10000
Accuracy Score: 0.7411
Root Mean Square Error: 2.092008604188807
plot_confusion_matrix(y_test,pred1)
model = tf.keras.models.load_model('CNN_model.h5')
preds = model.predict(x_test_norm)
313/313 [==============================] - 9s 28ms/step
preds.shape
(10000, 10)
cm = sns.light_palette((260, 75, 60), input="husl", as_cmap=True)
df = pd.DataFrame(preds[0:20], columns = ['airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck'])
df.style.format("{:.2%}").background_gradient(cmap=cm)
| airplane | automobile | bird | cat | deer | dog | frog | horse | ship | truck | |
|---|---|---|---|---|---|---|---|---|---|---|
| 0 | 7.17% | 0.88% | 0.67% | 76.30% | 0.63% | 6.43% | 2.86% | 0.19% | 4.44% | 0.42% |
| 1 | 0.51% | 12.79% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 86.66% | 0.03% |
| 2 | 4.13% | 11.31% | 0.11% | 0.10% | 0.03% | 0.02% | 0.04% | 0.07% | 82.28% | 1.91% |
| 3 | 84.84% | 10.86% | 0.37% | 0.03% | 0.04% | 0.00% | 0.01% | 0.00% | 3.76% | 0.09% |
| 4 | 0.00% | 0.00% | 1.32% | 1.37% | 9.50% | 0.20% | 87.59% | 0.00% | 0.01% | 0.00% |
| 5 | 0.05% | 0.48% | 0.15% | 0.39% | 0.28% | 0.11% | 98.40% | 0.02% | 0.04% | 0.08% |
| 6 | 0.09% | 82.34% | 0.01% | 0.08% | 0.00% | 0.02% | 0.01% | 0.01% | 0.01% | 17.43% |
| 7 | 0.11% | 0.02% | 5.90% | 2.57% | 5.02% | 0.23% | 86.11% | 0.05% | 0.00% | 0.00% |
| 8 | 0.05% | 0.01% | 3.33% | 76.03% | 3.97% | 11.18% | 4.11% | 1.30% | 0.01% | 0.01% |
| 9 | 0.27% | 97.43% | 0.00% | 0.00% | 0.00% | 0.00% | 0.02% | 0.00% | 0.06% | 2.22% |
| 10 | 33.02% | 0.20% | 11.38% | 6.33% | 8.10% | 6.32% | 0.71% | 5.76% | 26.29% | 1.88% |
| 11 | 0.00% | 0.34% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 99.65% |
| 12 | 0.03% | 0.17% | 5.43% | 19.39% | 13.67% | 52.17% | 7.58% | 1.41% | 0.11% | 0.05% |
| 13 | 0.01% | 0.01% | 0.08% | 0.04% | 0.12% | 1.67% | 0.00% | 98.05% | 0.00% | 0.02% |
| 14 | 0.03% | 0.31% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.02% | 99.64% |
| 15 | 4.74% | 1.29% | 1.29% | 0.42% | 0.44% | 0.02% | 13.35% | 0.00% | 78.42% | 0.02% |
| 16 | 0.02% | 0.03% | 0.27% | 9.21% | 0.02% | 89.36% | 0.24% | 0.75% | 0.02% | 0.08% |
| 17 | 1.82% | 0.87% | 2.65% | 11.01% | 7.24% | 6.20% | 1.26% | 65.09% | 0.89% | 2.97% |
| 18 | 0.95% | 0.99% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 97.41% | 0.65% |
| 19 | 0.00% | 0.04% | 0.24% | 0.52% | 0.38% | 0.06% | 98.72% | 0.02% | 0.00% | 0.00% |
(_,_), (test_images, test_labels) = tf.keras.datasets.cifar10.load_data()
img = test_images[2000]
img_tensor = image.img_to_array(img)
img_tensor = np.expand_dims(img_tensor, axis=0)
class_names = ['airplane'
,'automobile'
,'bird'
,'cat'
,'deer'
,'dog'
,'frog'
,'horse'
,'ship'
,'truck']
plt.imshow(img, cmap='viridis')
plt.axis('off')
plt.show()
# Extracts the outputs of the top 8 layers:
layer_outputs = [layer.output for layer in model.layers[:8]]
# Creates a model that will return these outputs, given the model input:
activation_model = models.Model(inputs=model.input, outputs=layer_outputs)
activations = activation_model.predict(img_tensor)
len(activations)
1/1 [==============================] - 0s 215ms/step
8
layer_names = []
for layer in model.layers:
layer_names.append(layer.name)
layer_names
['conv2d_3', 'max_pooling2d_3', 'dropout_3', 'conv2d_4', 'max_pooling2d_4', 'dropout_4', 'conv2d_5', 'max_pooling2d_5', 'dropout_5', 'flatten_1', 'dense_1']
# These are the names of the layers, so can have them as part of our plot
layer_names = []
for layer in model.layers[:3]:
layer_names.append(layer.name)
images_per_row = 16
# Now let's display our feature maps
for layer_name, layer_activation in zip(layer_names, activations):
# This is the number of features in the feature map
n_features = layer_activation.shape[-1]
# The feature map has shape (1, size, size, n_features)
size = layer_activation.shape[1]
# We will tile the activation channels in this matrix
n_cols = n_features // images_per_row
display_grid = np.zeros((size * n_cols, images_per_row * size))
# We'll tile each filter into this big horizontal grid
for col in range(n_cols):
for row in range(images_per_row):
channel_image = layer_activation[0,
:, :,
col * images_per_row + row]
# Post-process the feature to make it visually palatable
channel_image -= channel_image.mean()
channel_image /= channel_image.std()
channel_image *= 64
channel_image += 128
channel_image = np.clip(channel_image, 0, 255).astype('uint8')
display_grid[col * size : (col + 1) * size,
row * size : (row + 1) * size] = channel_image
# Display the grid
scale = 1. / size
plt.figure(figsize=(scale * display_grid.shape[1],
scale * display_grid.shape[0]))
plt.title(layer_name)
plt.grid(False)
plt.imshow(display_grid, aspect='auto', cmap='viridis')
plt.show();
/var/folders/zz/wfk650fx7lx3mmyd6_y3yxgr0000gr/T/ipykernel_32438/1872159762.py:28: RuntimeWarning: invalid value encountered in true_divide channel_image /= channel_image.std()
