handgesture_cnn_(81_).py

from keras.models import Sequential
from keras.layers import Dense, Activation, Flatten, Conv2D, MaxPooling2D
import numpy as np
import random
import os
import cv2

# Variables:
img_size = 100
iterations = 2

training_data = []

def printMsg(msg):
    print()
    print()
    print(msg)
    print()
    print()

def img_to_array():
    X = []
    Y = []
    categories = []
    images_done = 0

    for i in range(10):
        path = os.path.join('../leapGestRecog','0' + str(i))

        for cat in os.listdir(path):
            new_path = os.path.join(path, cat)

            if cat not in categories:
                categories.append(cat)

            for img in os.listdir(new_path):
                new_new_path = os.path.join(new_path, img)
                label = categories.index(cat)

                try:
                    img_array = cv2.imread(new_new_path)
                    resized_image = cv2.resize(img_array, (img_size, img_size)) # Resize image to make training less expensive
                    training_data.append([resized_image, label])
                    images_done = images_done + 1
                    print(new_new_path, "Done!", str(images_done)+'/20000')
                except Exception as e:
                    print(e)

    random.shuffle(training_data) # Make sure to shuffle data in order to train on heterogenous data

    for samples, labels in training_data:
        X.append(samples)
        Y.append(labels)

    X = np.array(X)
    Y = np.array(Y)

    # Save arrays so that we don't have to preprocess images each time
    np.save('samples', X)
    np.save('labels', Y)
    printMsg("img_to_array Done!")

# One hot encode our labels for each sample
# Ex: 5 encodes to -> [0, 0, 0, 0, 0, 1, 0, 0, 0, 0]
def to_categorical(array):
    one_hot = np.zeros((len(array), 10)) # There are 10 categories
    printMsg("to_categorical Started!")
    for idx, num in enumerate(array):
        one_hot[idx][num] = 1
    return one_hot

def train():
    printMsg("Training Started")
    samples = np.load('samples.npy')
    printMsg("Loading Complete!")
    labels = to_categorical(np.load('labels.npy'))
    printMsg("Category labels complete!")
    samples = samples.astype('float32') / 255 # Image data so we divide by 255 to normalize data.

    network = Sequential()

    network.add(Conv2D(64, (3,3), input_shape=(100, 100, 3)))
    network.add(Activation('relu'))
    network.add(MaxPooling2D(pool_size=(2,2)))

    network.add(Conv2D(128, (3,3)))
    network.add(Activation('relu'))
    network.add(MaxPooling2D(pool_size=(2,2)))

    network.add(Conv2D(128, (3,3)))
    network.add(Activation('relu'))
    network.add(MaxPooling2D(pool_size=(2,2)))

    network.add(Flatten())

    network.add(Dense(10))
    network.add(Activation('softmax')) # Using softmax instead of sigmoid because there are more than 2 classes.

    network.compile(loss="categorical_crossentropy", optimizer="rmsprop", metrics=['accuracy'])

    printMsg("Network created. Fitting Started.")

    network.fit(samples, labels, batch_size=32, validation_split=0.3, epochs=iterations, verbose=1) # Train network and take 30% of samples as validation data

    # Save Model
    network.save('Network.model')

print("Started!")
from keras import backend as K
K.tensorflow_backend._get_available_gpus()
if (os.path.isfile('samples.npy') and os.path.isfile('labels.npy')):
  print("Option 1")
  train()
else:
  print("Option 2")
  img_to_array()
  train()