I don't understand how images are actually fed into a CNN? If I have a directory containing a few thousand images, what steps do I need to take in order to feed them to a neural network (for instance resizing, grey scale, labeling, etc) I don't understand how even the labeling of an image works. What would this dataset actually look like? Or can you not look at it at all (in a summarized form, I'm thinking something like a table)?
This is a very packed question. Let's try to go through it and I will try to provide some example for image processing using a CNN.
Pre-processing the data
Pre-processing the data such as resizing, and grey scale is the first step of your machine learning pipeline. Most deep learning frameworks will require your training data to all have the same shape. So it is best to resize your images to some standard.
Whenever training any kind of machine learning model it is important to remember the bias variance trade-off. The more complex the model the harder it will be to train it. That means it is best to limit the number of model parameters in your model. You can lower the number of inputs to your model by downsampling the images. Greyscaling is often used for the same reason. If the colors in the images do not contain any distinguishing information then you can reduce the number of inputs by a third by greyscaling.
There are a number of other pre-processing methods which can be used depending on your data. It is also a good idea to do some data augmentation, this is altering your input data slightly without changing the resulting label to increase the number of instances you have to train your model.
How to structure the data?
The shape of the variable which you will use as the input for your CNN will depend on the package you choose. I prefer using tensorflow, which is developed by Google. If you are planning on using a pretty standard architecture, then there is a very useful wrapper library named Keras which will help make designing and training a CNN very easy.
When using tensorflow you will want to get your set of images into a numpy matrix. The first dimension is your instances, then your image dimensions and finally the last dimension is for
So for example if you are using MNIST data as shown below, then you are working with greyscale images which each have dimensions 28 by 28. Then the numpy matrix shape that you would feed into your deep learning model would be (n, 28, 28, 1), where $n$ is the number of images you have in your dataset.
How to label images?
For most data the labeling would need to be done manually. This is often named data collection and is the hardest and most expensive part of any machine learning solution. It is often best to either use readily available data, or to use less complex models and more pre-processing if the data is just unavailable.
Here is an example of the use of a CNN for the MNIST dataset
First we load the data
from keras.datasets import mnist import numpy as np (x_train, y_train), (x_test, y_test) = mnist.load_data() x_train = x_train.astype('float32') / 255. x_test = x_test.astype('float32') / 255. print('Training data shape: ', x_train.shape) print('Testing data shape : ', x_test.shape)
Training data shape: (60000, 28, 28)
Testing data shape : (10000, 28, 28)
from __future__ import print_function import keras from keras.datasets import mnist from keras.models import Sequential from keras.layers import Dense, Dropout, Flatten from keras.layers import Conv2D, MaxPooling2D from keras.callbacks import ModelCheckpoint from keras.models import model_from_json from keras import backend as K
Then we need to reshape our data to add the channel dimension at the end of our numpy matrix. Furthermore, we will one-hot encode the labels. So you will have 10 output neurons, where each represent a different class.
# The known number of output classes. num_classes = 10 # Input image dimensions img_rows, img_cols = 28, 28 # Channels go last for TensorFlow backend x_train_reshaped = x_train.reshape(x_train.shape, img_rows, img_cols, 1) x_test_reshaped = x_test.reshape(x_test.shape, img_rows, img_cols, 1) input_shape = (img_rows, img_cols, 1) # Convert class vectors to binary class matrices. This uses 1 hot encoding. y_train_binary = keras.utils.to_categorical(y_train, num_classes) y_test_binary = keras.utils.to_categorical(y_test, num_classes)
Now we design our model
model = Sequential() model.add(Conv2D(32, kernel_size=(3, 3), activation='relu', input_shape=input_shape)) model.add(Conv2D(64, (3, 3), activation='relu')) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(Dropout(0.25)) model.add(Flatten()) model.add(Dense(128, activation='relu')) model.add(Dropout(0.5)) model.add(Dense(num_classes, activation='softmax')) model.compile(loss=keras.losses.categorical_crossentropy, optimizer=keras.optimizers.Adadelta(), metrics=['accuracy'])
Finally we can train the model
epochs = 4 batch_size = 128 # Fit the model weights. model.fit(x_train_reshaped, y_train_binary, batch_size=batch_size, epochs=epochs, verbose=1, validation_data=(x_test_reshaped, y_test_binary))
Dataset just consists of Features and Labels. Here features are your images and labels are the classes.
There is a fit() method for every CNN model, which will take in Features and Labels, and performs training.
for the first layer, you need to mention the input dimension of image, and the output layer should be a softmax (if you're doing classification) with dimension as the number of classes you have.
model = Sequential() model.add(Conv2D(32, (3, 3), padding='same', input_shape=(64, 64, 3))) model.add(Activation('relu')) model.add(Conv2D(32, (3, 3))) model.add(Activation('relu')) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(Dropout(0.25)) model.add(Flatten()) model.add(Dense(512)) model.add(Activation('relu')) model.add(Dropout(0.5)) model.add(Dense(num_classes)) model.add(Activation('softmax')) model.compile(loss='categorical_crossentropy', optimizer='Adam', metrics=['accuracy']) model.fit(x_train, y_train, epochs=10)
The above is the code for training a Keras sequenctioal model.
- input_shape should be the dimension of X_train.
- You need to get this shape when you do X_train.shape (numpy)
- Convolutions are then applied with respective Activations
- Dropout and Pooling layers are optional.
- After the convolution layers, the data is flattened. using Flatten()
- Then it is sent to few Fully Connected layers
- The last but one layer should have the dimensions of number of classes
- Last layer will be softmax.
- Now, compile the model with the loss, optimizer and metric
- Then fit()
Vote up ;) if you like it.