In this notebook, we'll be building a neural network that recognizes handwritten numbers 0-9.
This kind of neural network is used in a variety of real-world applications including: recognizing phone numbers and sorting postal mail by address. To build the network, we'll be using the MNIST data set, which consists of images of handwritten numbers and their correct labels 0-9.
We'll be using TFLearn, a high-level library built on top of TensorFlow to build the neural network. We'll start off by importing all the modules we'll need, then load the data, and finally build the network.
In :# Import Numpy, TensorFlow, TFLearn, and MNIST data import numpy as np import tensorflow as tf import tflearn import tflearn.datasets.mnist as mnist
The MNIST data set already contains both training and test data. There are 55,000 data points of training data, and 10,000 points of test data.
Each MNIST data point has:
We'll call the images, which will be the input to our neural network, X and their corresponding labels Y.
We're going to want our labels as one-hot vectors, which are vectors that holds mostly 0's and one 1. It's easiest to see this in a example. As a one-hot vector, the number 0 is represented as [1, 0, 0, 0, 0, 0, 0, 0, 0, 0], and 4 is represented as [0, 0, 0, 0, 1, 0, 0, 0, 0, 0].
For this example, we'll be using flattened data or a representation of MNIST images in one dimension rather than two. So, each handwritten number image, which is 28x28 pixels, will be represented as a one dimensional array of 784 pixel values.
Flattening the data throws away information about the 2D structure of the image, but it simplifies our data so that all of the training data can be contained in one array whose shape is [55000, 784]; the first dimension is the number of training images and the second dimension is the number of pixels in each image. This is the kind of data that is easy to analyze using a simple neural network.
In :# Retrieve the training and test data trainX, trainY, testX, testY = mnist.load_data(one_hot=True)
Downloading MNIST... Succesfully downloaded train-images-idx3-ubyte.gz 9912422 bytes. Extracting mnist/train-images-idx3-ubyte.gz/opt/conda/envs/tflearn/lib/python3.5/gzip.py:274: VisibleDeprecationWarning: converting an array with ndim > 0 to an index will result in an error in the future return self._buffer.read(size) /opt/conda/envs/tflearn/lib/python3.5/site-packages/tflearn/datasets/mnist.py:52: VisibleDeprecationWarning: converting an array with ndim > 0 to an index will result in an error in the future data = data.reshape(num_images, rows, cols, 1)Downloading MNIST... Succesfully downloaded train-labels-idx1-ubyte.gz 28881 bytes. Extracting mnist/train-labels-idx1-ubyte.gz Downloading MNIST... Succesfully downloaded t10k-images-idx3-ubyte.gz 1648877 bytes. Extracting mnist/t10k-images-idx3-ubyte.gz Downloading MNIST... Succesfully downloaded t10k-labels-idx1-ubyte.gz 4542 bytes. Extracting mnist/t10k-labels-idx1-ubyte.gz
In :# Visualizing the data import matplotlib.pyplot as plt %matplotlib inline # Function for displaying a training image by it's index in the MNIST set def show_digit(index): label = trainY[index].argmax(axis=0) # Reshape 784 array into 28x28 image image = trainX[index].reshape([28,28]) plt.title('Training data, index: %d, Label: %d' % (index, label)) plt.imshow(image, cmap='gray_r') plt.show() # Display the first (index 0) training image show_digit(0)
TFLearn lets you build the network by defining the layers in that network.
For this example, you'll define:
Let's start with the input layer; to define the input layer, you'll define the type of data that the network expects. For example,
net = tflearn.input_data([None, 100])
would create a network with 100 inputs. The number of inputs to your network needs to match the size of your data. For this example, we're using 784 element long vectors to encode our input data, so we need 784 input units.
To add new hidden layers, you use
net = tflearn.fully_connected(net, n_units, activation='ReLU')
This adds a fully connected layer where every unit (or node) in the previous layer is connected to every unit in this layer. The first argument
net is the network you created in the
tflearn.input_data call, it designates the input to the hidden layer. You can set the number of units in the layer with
n_units, and set the activation function with the
activation keyword. You can keep adding layers to your network by repeated calling
Then, to set how you train the network, use:
net = tflearn.regression(net, optimizer='sgd', learning_rate=0.1, loss='categorical_crossentropy')
Again, this is passing in the network you've been building. The keywords:
optimizersets the training method, here stochastic gradient descent
learning_rateis the learning rate
lossdetermines how the network error is calculated. In this example, with categorical cross-entropy.
Finally, you put all this together to create the model with
Exercise: Below in the
build_model() function, you'll put together the network using TFLearn. You get to choose how many layers to use, how many hidden units, etc.
Hint: The final output layer must have 10 output nodes (one for each digit 0-9). It's also recommended to use a
softmax activation layer as your final output layer.
In :# Define the neural network def build_model(): # This resets all parameters and variables, leave this here tf.reset_default_graph() #### Your code #### # Include the input layer, hidden layer(s), and set how you want to train the model net = tflearn.input_data([None, 784]) net = tflearn.fully_connected(net, 300, activation='ReLU') net = tflearn.fully_connected(net, 150, activation='ReLU') net = tflearn.fully_connected(net, 10, activation='softmax') net = tflearn.regression(net, optimizer='sgd', learning_rate=0.1, loss='categorical_crossentropy') # This model assumes that your network is named "net" model = tflearn.DNN(net) return model
In :# Build the model model = build_model()
WARNING:tensorflow:From /opt/conda/envs/tflearn/lib/python3.5/site-packages/tflearn/summaries.py:46 in get_summary.: scalar_summary (from tensorflow.python.ops.logging_ops) is deprecated and will be removed after 2016-11-30. Instructions for updating: Please switch to tf.summary.scalar. Note that tf.summary.scalar uses the node name instead of the tag. This means that TensorFlow will automatically de-duplicate summary names based on the scope they are created in. Also, passing a tensor or list of tags to a scalar summary op is no longer supported. WARNING:tensorflow:From /opt/conda/envs/tflearn/lib/python3.5/site-packages/tflearn/summaries.py:46 in get_summary.: scalar_summary (from tensorflow.python.ops.logging_ops) is deprecated and will be removed after 2016-11-30. Instructions for updating: Please switch to tf.summary.scalar. Note that tf.summary.scalar uses the node name instead of the tag. This means that TensorFlow will automatically de-duplicate summary names based on the scope they are created in. Also, passing a tensor or list of tags to a scalar summary op is no longer supported. WARNING:tensorflow:From /opt/conda/envs/tflearn/lib/python3.5/site-packages/tflearn/helpers/trainer.py:766 in create_summaries.: merge_summary (from tensorflow.python.ops.logging_ops) is deprecated and will be removed after 2016-11-30. Instructions for updating: Please switch to tf.summary.merge. WARNING:tensorflow:VARIABLES collection name is deprecated, please use GLOBAL_VARIABLES instead; VARIABLES will be removed after 2017-03-02. WARNING:tensorflow:From /opt/conda/envs/tflearn/lib/python3.5/site-packages/tflearn/helpers/trainer.py:130 in __init__.: initialize_all_variables (from tensorflow.python.ops.variables) is deprecated and will be removed after 2017-03-02. Instructions for updating: Use `tf.global_variables_initializer` instead.
Now that we've constructed the network, saved as the variable
model, we can fit it to the data. Here we use the
model.fit method. You pass in the training features
trainX and the training targets
trainY. Below I set
validation_set=0.1 which reserves 10% of the data set as the validation set. You can also set the batch size and number of epochs with the
n_epoch keywords, respectively.
Too few epochs don't effectively train your network, and too many take a long time to execute. Choose wisely!
In :# Training model.fit(trainX, trainY, validation_set=0.1, show_metric=True, batch_size=100, n_epoch=10)
Training Step: 4950 | total loss: 0.05565 | SGD | epoch: 010 | loss: 0.05565 - acc: 0.9851 | val_loss: 0.08539 - val_acc: 0.9749 -- iter: 49500/49500 Training Step: 4950 | total loss: 0.05565 | SGD | epoch: 010 | loss: 0.05565 - acc: 0.9851 | val_loss: 0.08539 - val_acc: 0.9749 -- iter: 49500/49500 --
After you're satisified with the training output and accuracy, you can then run the network on the test data set to measure it's performance! Remember, only do this after you've done the training and are satisfied with the results.
A good result will be higher than 98% accuracy! Some simple models have been known to get up to 99.7% accuracy.
In :# Compare the labels that our model predicts with the actual labels predictions = (np.array(model.predict(testX))[:,0] >= 0.5).astype(np.int_) # Calculate the accuracy, which is the percentage of times the predicated labels matched the actual labels test_accuracy = np.mean(predictions == testY[:,0], axis=0) # Print out the result print("Test accuracy: ", test_accuracy)
Test accuracy: 0.9965
In [ ]: