In the previous labs in the "Learn Tensorflow" series, you used the Fashion MNIST dataset to train an image classifier. In this case you had images that were 28x28 where the subject was centered. In this lab you'll take this to the next level, training to recognize features in an image where the subject can be anywhere in the image!

You'll be using TensorFlow in this lab to create a Convolutional Neural Network that is trained to recognize images of horses and humans and classify them.

If you've never done this before, I'd strongly recommend you use Google Colaboratory, a browser-based environment that contains all the required dependencies, including TensorFlow and the required libraries. You can find the code for the rest of this lab running in a colab here.

Otherwise, the main language you will use for training models is Python, so you will need to have that installed. In addition to that you'll also need TensorFlow. Details on installing it are here. You'll also need the numpy library.

You'll do this by building a horses-or-humans classifier that will tell you if a given image contains a horse or a human, where the network is trained to recognize features that determine which is which.

In the case of Fashion MNIST, the data was built into TensorFlow via Keras. In this case the data isn't so you'll have to do some processing of it before you can train.

First, let's download the data:

!wget --no-check-certificate  -O /tmp/

The following python code will use the OS library to use Operating System libraries, giving you access to the file system, and the zipfile library allowing you to unzip the data.

import os
import zipfile
local_zip = '/tmp/'
zip_ref = zipfile.ZipFile(local_zip, 'r')

The contents of the .zip are extracted to the base directory /tmp/horse-or-human, which in turn each contain horses and humans subdirectories.

In short: The training set is the data that is used to tell the neural network model that 'this is what a horse looks like', 'this is what a human looks like' etc.

One thing to pay attention to in this sample: We do not explicitly label the images as horses or humans. If you remember with the fashion example earlier, we had labelled 'this is a 1', 'this is a 7' etc.

Later you'll see something called an ImageGenerator being used -- and that reads images from subdirectories, and automatically label them from the name of that subdirectory. So, for example, you will have a 'training' directory containing a 'horses' directory and a 'humans' one. ImageGenerator will label the images appropriately for you, reducing a coding step.

Let's define each of these directories:

# Directory with our training horse pictures
train_horse_dir = os.path.join('/tmp/horse-or-human/horses')
# Directory with our training human pictures
train_human_dir = os.path.join('/tmp/horse-or-human/humans')

Now, let's see what the filenames look like in the horses and humans training directories:

train_horse_names = os.listdir(train_horse_dir)
train_human_names = os.listdir(train_human_dir)

Let's find out the total number of horse and human images in the directories:

print('total training horse images:', len(os.listdir(train_horse_dir)))
print('total training human images:', len(os.listdir(train_human_dir)))

Now let's take a look at a few pictures to get a better sense of what they look like. First, configure the matplot parameters:

%matplotlib inline
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
# Parameters for our graph; we'll output images in a 4x4 configuration
nrows = 4
ncols = 4
# Index for iterating over images
pic_index = 0

Now, display a batch of 8 horse and 8 human pictures. You can rerun the cell to see a fresh batch each time:

# Set up matplotlib fig, and size it to fit 4x4 pics
fig = plt.gcf()
fig.set_size_inches(ncols * 4, nrows * 4)
pic_index += 8
next_horse_pix = [os.path.join(train_horse_dir, fname) 
                for fname in train_horse_names[pic_index-8:pic_index]]
next_human_pix = [os.path.join(train_human_dir, fname) 
                for fname in train_human_names[pic_index-8:pic_index]]
for i, img_path in enumerate(next_horse_pix+next_human_pix):
  # Set up subplot; subplot indices start at 1
  sp = plt.subplot(nrows, ncols, i + 1)
  sp.axis('Off') # Don't show axes (or gridlines)
  img = mpimg.imread(img_path)

Here are some example images showing horses and humans in different poses and orientations.

Let's start defining the model:

Step 1 will be to import tensorflow.

import tensorflow as tf

We then add convolutional layers as in the previous example, and flatten the final result to feed into the densely connected layers. Finally we add the densely connected layers.

Note that because we are facing a two-class classification problem, i.e. a binary classification problem, we will end our network with a sigmoid activation, so that the output of our network will be a single scalar between 0 and 1, encoding the probability that the current image is class 1 (as opposed to class 0).

model = tf.keras.models.Sequential([
    # Note the input shape is the desired size of the image 300x300 with 3 bytes color
    # This is the first convolution
    tf.keras.layers.Conv2D(16, (3,3), activation='relu', input_shape=(300, 300, 3)),
    tf.keras.layers.MaxPooling2D(2, 2),
    # The second convolution
    tf.keras.layers.Conv2D(32, (3,3), activation='relu'),
    # The third convolution
    tf.keras.layers.Conv2D(64, (3,3), activation='relu'),
    # The fourth convolution
    tf.keras.layers.Conv2D(64, (3,3), activation='relu'),
    # The fifth convolution
    tf.keras.layers.Conv2D(64, (3,3), activation='relu'),
    # Flatten the results to feed into a DNN
    # 512 neuron hidden layer
    tf.keras.layers.Dense(512, activation='relu'),
    # Only 1 output neuron. It will contain a value from 0-1 where 0 for 1 class ('horses') and 1 for the other ('humans')
    tf.keras.layers.Dense(1, activation='sigmoid')

The model.summary() method call prints a summary of the NN


You can see the results here:

Layer (type)                 Output Shape              Param #   
conv2d (Conv2D)              (None, 298, 298, 16)      448       
max_pooling2d (MaxPooling2D) (None, 149, 149, 16)      0         
conv2d_1 (Conv2D)            (None, 147, 147, 32)      4640      
max_pooling2d_1 (MaxPooling2 (None, 73, 73, 32)        0         
conv2d_2 (Conv2D)            (None, 71, 71, 64)        18496     
max_pooling2d_2 (MaxPooling2 (None, 35, 35, 64)        0         
conv2d_3 (Conv2D)            (None, 33, 33, 64)        36928     
max_pooling2d_3 (MaxPooling2 (None, 16, 16, 64)        0         
conv2d_4 (Conv2D)            (None, 14, 14, 64)        36928     
max_pooling2d_4 (MaxPooling2 (None, 7, 7, 64)          0         
flatten (Flatten)            (None, 3136)              0         
dense (Dense)                (None, 512)               1606144   
dense_1 (Dense)              (None, 1)                 513       
Total params: 1,704,097
Trainable params: 1,704,097
Non-trainable params: 0

The "output shape" column shows how the size of your feature map evolves in each successive layer. The convolution layers reduce the size of the feature maps by a bit due to padding, and each pooling layer halves the dimensions.

Next, we'll configure the specifications for model training. We will train our model with the binary_crossentropy loss, because it's a binary classification problem and our final activation is a sigmoid. (For a refresher on loss metrics, see the Machine Learning Crash Course.) We will use the rmsprop optimizer with a learning rate of 0.001. During training, we will want to monitor classification accuracy.

NOTE: In this case, using the RMSprop optimization algorithm is preferable to stochastic gradient descent (SGD), because RMSprop automates learning-rate tuning for us. (Other optimizers, such as Adam and Adagrad, also automatically adapt the learning rate during training, and would work equally well here.)

from tensorflow.keras.optimizers import RMSprop

Let's set up data generators that will read pictures in our source folders, convert them to float32 tensors, and feed them (with their labels) to our network. We'll have one generator for the training images and one for the validation images. Our generators will yield batches of images of size 300x300 and their labels (binary).

As you may already know, data that goes into neural networks should usually be normalized in some way to make it more amenable to processing by the network. (It is uncommon to feed raw pixels into a convnet.) In our case, we will preprocess our images by normalizing the pixel values to be in the [0, 1] range (originally all values are in the [0, 255] range).

In Keras this can be done via the keras.preprocessing.image.ImageDataGenerator class using the rescale parameter. This ImageDataGenerator class allows you to instantiate generators of augmented image batches (and their labels) via .flow(data, labels) or .flow_from_directory(directory). These generators can then be used with the Keras model methods that accept data generators as inputs: fit_generator, evaluate_generator, and predict_generator.

from tensorflow.keras.preprocessing.image import ImageDataGenerator
# All images will be rescaled by 1./255
train_datagen = ImageDataGenerator(rescale=1./255)
# Flow training images in batches of 128 using train_datagen generator
train_generator = train_datagen.flow_from_directory(
        '/tmp/horse-or-human/',  # This is the source directory for training images
        target_size=(300, 300),  # All images will be resized to 150x150
        # Since we use binary_crossentropy loss, we need binary labels

Let's train for 15 epochs -- this may take a few minutes to run.

history = model.fit_generator(

Do note the values per epoch.

The Loss and Accuracy are a great indication of progress of training. It's making a guess as to the classification of the training data, and then measuring it against the known label, calculating the result. Accuracy is the portion of correct guesses.

Epoch 1/15
9/9 [==============================] - 9s 1s/step - loss: 0.8662 - acc: 0.5151
Epoch 2/15
9/9 [==============================] - 8s 927ms/step - loss: 0.7212 - acc: 0.5969
Epoch 3/15
9/9 [==============================] - 8s 921ms/step - loss: 0.6612 - acc: 0.6592
Epoch 4/15
9/9 [==============================] - 8s 925ms/step - loss: 0.3135 - acc: 0.8481
Epoch 5/15
9/9 [==============================] - 8s 919ms/step - loss: 0.4640 - acc: 0.8530
Epoch 6/15
9/9 [==============================] - 8s 896ms/step - loss: 0.2306 - acc: 0.9231
Epoch 7/15
9/9 [==============================] - 8s 915ms/step - loss: 0.1464 - acc: 0.9396
Epoch 8/15
9/9 [==============================] - 8s 935ms/step - loss: 0.2663 - acc: 0.8919
Epoch 9/15
9/9 [==============================] - 8s 883ms/step - loss: 0.0772 - acc: 0.9698
Epoch 10/15
9/9 [==============================] - 9s 951ms/step - loss: 0.0403 - acc: 0.9805
Epoch 11/15
9/9 [==============================] - 8s 891ms/step - loss: 0.2618 - acc: 0.9075
Epoch 12/15
9/9 [==============================] - 8s 902ms/step - loss: 0.0434 - acc: 0.9873
Epoch 13/15
9/9 [==============================] - 8s 904ms/step - loss: 0.0187 - acc: 0.9932
Epoch 14/15
9/9 [==============================] - 9s 951ms/step - loss: 0.0974 - acc: 0.9649
Epoch 15/15
9/9 [==============================] - 8s 877ms/step - loss: 0.2859 - acc: 0.9338

Let's now take a look at actually running a prediction using the model. This code will allow you to choose 1 or more files from your file system, it will then upload them, and run them through the model, giving an indication of whether the object is a horse or a human.

You can download images from the internet to your file system to try them out!

Note that you might see that the network makes a LOT of mistakes, despite the fact that the training accuracy is above 99%.

This is due to something called overfitting, which means that the neural network is trained with very limited data -- there are only 500ish images of each class. So it's very good at recognizing images that look like those in the training set, but it can fail a lot at images that are not in the training set.

This is a data point proving that the more data you train on, the better your final network will be!

There are many techniques that can be used to make your training better, despite limited data, including something called Image Augmentation. That's beyond the scope of this lab!

import numpy as np
from google.colab import files
from keras.preprocessing import image
uploaded = files.upload()
for fn in uploaded.keys():
  # predicting images
  path = '/content/' + fn
  img = image.load_img(path, target_size=(300, 300))
  x = image.img_to_array(img)
  x = np.expand_dims(x, axis=0)
  images = np.vstack([x])
  classes = model.predict(images, batch_size=10)
  if classes[0]>0.5:
    print(fn + " is a human")
    print(fn + " is a horse")

So, for example to test with this image:

Here's what the colab produces:

..we can see that despite it being a cartoon graphic, it still classifies correctly.

This image also classifies correctly:

Try some images yourself and explore!

To get a feel for what kind of features our convnet has learned, one fun thing to do is to visualize how an input gets transformed as it goes through the convnet.

Let's pick a random image from the training set, and then generate a figure where each row is the output of a layer, and each image in the row is a specific filter in that output feature map. Rerun this cell to generate intermediate representations for a variety of training images.

import numpy as np
import random
from tensorflow.keras.preprocessing.image import img_to_array, load_img
# Let's define a new Model that will take an image as input, and will output
# intermediate representations for all layers in the previous model after
# the first.
successive_outputs = [layer.output for layer in model.layers[1:]]
#visualization_model = Model(img_input, successive_outputs)
visualization_model = tf.keras.models.Model(inputs = model.input, outputs = successive_outputs)
# Let's prepare a random input image from the training set.
horse_img_files = [os.path.join(train_horse_dir, f) for f in train_horse_names]
human_img_files = [os.path.join(train_human_dir, f) for f in train_human_names]
img_path = random.choice(horse_img_files + human_img_files)
img = load_img(img_path, target_size=(300, 300))  # this is a PIL image
x = img_to_array(img)  # Numpy array with shape (150, 150, 3)
x = x.reshape((1,) + x.shape)  # Numpy array with shape (1, 150, 150, 3)
# Rescale by 1/255
x /= 255
# Let's run our image through our network, thus obtaining all
# intermediate representations for this image.
successive_feature_maps = visualization_model.predict(x)
# These are the names of the layers, so can have them as part of our plot
layer_names = [ for layer in model.layers]
# Now let's display our representations
for layer_name, feature_map in zip(layer_names, successive_feature_maps):
  if len(feature_map.shape) == 4:
    # Just do this for the conv / maxpool layers, not the fully-connected layers
    n_features = feature_map.shape[-1]  # number of features in feature map
    # The feature map has shape (1, size, size, n_features)
    size = feature_map.shape[1]
    # We will tile our images in this matrix
    display_grid = np.zeros((size, size * n_features))
    for i in range(n_features):
      # Postprocess the feature to make it visually palatable
      x = feature_map[0, :, :, i]
      x -= x.mean()
      x /= x.std()
      x *= 64
      x += 128
      x = np.clip(x, 0, 255).astype('uint8')
      # We'll tile each filter into this big horizontal grid
      display_grid[:, i * size : (i + 1) * size] = x
    # Display the grid
    scale = 20. / n_features
    plt.figure(figsize=(scale * n_features, scale))
    plt.imshow(display_grid, aspect='auto', cmap='viridis')

Here are example results:

As you can see we go from the raw pixels of the images to increasingly abstract and compact representations. The representations downstream start highlighting what the network pays attention to, and they show fewer and fewer features being "activated"; most are set to zero. This is called "sparsity." Representation sparsity is a key feature of deep learning.

These representations carry increasingly less information about the original pixels of the image, but increasingly refined information about the class of the image. You can think of a convnet (or a deep network in general) as an information distillation pipeline.