As I was going through the chapter on convnets in Chollet’s book, Deep Learning with Python, one of the things that I found interesting was the ability to extend an existing trained model. When you think about image recognition, I can’t imagine being able to collect enough images to be able to build a decent model. I thought about an application that could track soccer players on the field and detect how often they were engaged with the ball. Engagement would be defined as the amount of time they were spotted within 1 to 2 meters of the ball.
I think this would be an interesting project, but collecting images of all of the angles that players might find themselves in would be a challenge if not impossible. I wonder how many images of each player would it take to train the model. One of the things that might work is using data augmentation or more specifically taking the same images and mutate them into new images. The mutation could be a translation or an offset in the frame to make the image different, at least to the machine. These images would add to the existing pool and improve the model since there is now more training data. Keras takes care of those mutations with their ImageDataGenerator class.
Convolution using a windowed view of the image and moves that image around to find local features. In contrast, Dense layers look at the whole of the features to train. I am an not an artist so I won’t try to show an example of a convolution window, but think of reading the news paper through a magnifying glass and moving it around until eventually you have covered the whole page.
Chollet gives an explanation of strides and padding, which seem straightforward. I think the best explanation from another well known site MachineLearningMastery. The purpose of the padding is really to give each pixel the change the be in the center of the window. Since the window move 1 pixel at a time from left to right, unless the size of this image is large enough, it is impossible to center the border pixels. For a 5 X 5 image and a 3 X 3 window, it will be impossible to center each pixel, but it you add that image such that it is a 7 X 7 dimensioned image, you can center the pixels.
I am going to post the code from the book since it is more concise, but you can get the true source from Chollet’s GitHub. This code assumes that you have downloaded the cats vs dogs dataset from Kaggle and you have loaded the data and separated them out. I have posted my version on GitHub, but again it is derived from the authors code.
#We need to setup the environment and some paths to our images:
import os
import shutil
from keras import layers
from keras import models
from keras import optimizers
from keras.preprocessing.image import ImageDataGenerator
os.environ['KMP_DUPLICATE_LIB_OK']='True'
base_dir = '/Users/heathivie/Downloads/cats_and_dogs_small'
train_dir = os.path.join(base_dir, 'train')
validation_dir = os.path.join(base_dir, 'validation')
test_dir = os.path.join(base_dir, 'test')
I needed to add this piece os.environ[‘KMP_DUPLICATE_LIB_OK’]=’True’ because it was failing with this error:
OMP: Error #15: Initializing libiomp5.dylib, but found libiomp5.dylib already initialized.
OMP: Hint: This means that multiple copies of the OpenMP runtime have been linked into the program. That is dangerous, since it can degrade performance or cause incorrect results. The best thing to do is to ensure that only a single OpenMP runtime is linked into the process, e.g. by avoiding static linking of the OpenMP runtime in any library. As an unsafe, unsupported, undocumented workaround you can set the environment variable KMP_DUPLICATE_LIB_OK=TRUE to allow the program to continue to execute, but that may cause crashes or silently produce incorrect results. For more information, please see http://www.intel.com/software/products/support/
There are a ton of images that need to be processed and used for training, so we will need to use Keras’ ImageDataGenerator. It is a python generator that looks through the files and yields the image as it is available. Here we will load the data for training.
train_datagen = ImageDataGenerator(rescale=1./255,
rotation_range=40,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True)
test_datagen = ImageDataGenerator(rescale=1./255)
train_generator = train_datagen.flow_from_directory(
train_dir,
target_size=(150, 150),
batch_size=20,
class_mode='binary')
validation_generator = test_datagen.flow_from_directory(
validation_dir,
target_size=(150, 150),
batch_size=20,
class_mode='binary')
The first train_datagen is filled with parameters to support the data augmentation. The next to pieces there simple setup the path, target (image size) and batch size. It also specifies what the class model is and since we are doing a classification of two types (cats & dogs) we will use binary.
Like the earlier posts we will still be using a Sequential model, but we will start with the ConvD layers.
model = models.Sequential()
model.add(layers.Conv2D(32, (3, 3), activation='relu',
input_shape=(150, 150, 3)))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(64, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(128, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(128, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
Above we have specified that we want to have a 3 X 3 window, 32 filters (channels), relu as our activation and the image shape of 150 X 150 X 3. One thing to note, we need to do a classification which requires a Dense layer to process, so how do we translate a 3D tensor to fit the dense layer. Keras gives a Flatten method to do this. It final shape is a 1D tensor (X * Y * Channel).
model.add(layers.Flatten())
model.add(layers.Dense(512, activation='relu'))
model.add(layers.Dense(1, activation='sigmoid'))
We finalize it with a single Dense layer with the sigmoid activation. The last piece is to compile the model. For this we will use the loss function binary_crossentropy since this is a classification problem with 2 possible outcomes.We will again use the optimizer RMS prop, but here we will specify a learning rate or the rate at which it moves when doing the gradient. Lastly, we configure it to return the accuracy metrics.
model.compile(loss='binary_crossentropy',
optimizer=optimizers.RMSprop(lr=1e-4),
metrics=['acc'])
Now we can run the fit method, supplying it with the training and validation generators that we created above. The step* parameters are there to make sure that our generators don’t run forever. This is configured to run 30 epochs at 100 steps each, so on my machine this takes about 10 minutes. Make sure you save your model.
history = model.fit(
train_generator,
steps_per_epoch=100,
epochs=30,
validation_data=validation_generator,
validation_steps=50)
model.save('ch_5_cat_dogs.h5')
After running through all of the epoch’s, I achieved a 0.75 accuracy. This is what it looked like:
After you have saved your model, you can go a take picture of your cat or dog (or grab one of the internet) and use it to predict whether it is a cat or a dog.
import os
import numpy
from keras.models import load_model
from keras.preprocessing import image
base_dir = '/Users/heathivie/Downloads/cats_and_dogs_small'
# load model
model = load_model('ch_5_cat_dogs.h5')
# summarize model.
model.summary()
file = test_dir = os.path.join(base_dir, 'test/cats/download.jpeg')
f = image.load_img(file, target_size=(150, 150, 3))
x = image.img_to_array(f)
# the first param is the batch size
y = x.reshape((1, 150, 150, 3)).astype('float32')
classes = model.predict_classes(y)
print(classes)
I used this image of my amazing dog Fergus and the prediction was correct, he was indeed a dog.
The next post I will do is use a pre-training convnet, which I think it awesome. I am going to continue talking about the goal of a model that can detect someone and their proximity to a ball.
It's just me 