Create utils.py

ml/data_utils.py

@@ -0,0 +1,45 @@
+import numpy as np
+import os, csv, random
+from PIL import Image
+
+"""
+Code for extracting data and labels
+Katherine Mayo 4/15/2018
+"""
+
+def unpack_image(image):
+  #assumes all images come in as the same size
+  #otherwise use Image.resize((width, height))
+  im = Image.open(image, 'r')
+  width, height = im.size
+  pixels = list(im.getdata())
+  pixels = np.array(pixels).astype(np.float32)
+  return pixels, width, height
+
+#goes through a directory and gets the labels and images
+def load_data(directoryName):
+  files = os.listdir(directoryName)
+  labels = []
+  data = []
+  width = 0
+  height = 0
+  for f in files:
+    labels.append(f.split('-')[0])
+    im, width, height = unpack_image(directoryName + '/' + f)
+    data.append(im)
+  return data, labels, height, width
+
+def formatData(data, num, height, width):
+  data = np.array(data)
+  data = (data.reshape((num, height*width*3)).astype("float32"))
+  return data
+
+def sampleData(trainData, trainLabels, div=0.1):
+  x = np.random.choice(trainData.shape[0], size=int(len(trainData)*div))
+  val_data = trainData[x, :]
+  val_labels = []
+  for i in x:
+    val_labels.append(trainLabels[i])
+  train_data = trainData[-x,:]
+  train_labels = [i for i in trainLabels if i not in val_labels]
+  return train_data, train_labels, val_data, val_labels

ml/neuralNet.py

@@ -0,0 +1,68 @@
+from keras.models import Sequential
+from keras.layers.core import Dense, Activation
+from keras.optimizers import SGD, RMSprop, Adagrad, Adam
+
+"""
+Fully Connected Neural Network class
+--size adjustable [number of layers and neurons in each layer]
+Katherine Mayo 4/15/2018
+"""
+
+class neuralNet(object):
+
+  def __init__(self, numLayers=2, input_dim=734, num_classes=10, dropout=0, use_bn = False, reg=0.0):
+    self.use_batchnorm = use_bn
+    self.use_dropout = dropout > 0
+    self.reg = reg
+    self.numlayers = numLayers
+    self.num_classes = num_classes
+
+  """
+  default values batch norm used only if the network uses batch normalization
+  momentum and decay values are used only if the SGD variations are desired
+  """
+  def buildNet(self, neurons=4, optimizer='sgd', optParam={'lr': 0.001, 'momentum': 0, 'decay': 0}, bnorm={'momentum': 0.99, 'epsilon': 0.001}, dropout=0):
+    """
+    {affine - [batch norm] - relu - [dropout]} x (L - 1) - affine - softmax
+    """
+    model = Sequential()
+    model.add(Dense(neurons, input_shape=(784,), use_bias = True))
+    for i in range(self.numlayers-1):
+      model.add(Dense(neurons, use_bias=True))
+      if self.use_batchnorm:
+        model.add(BatchNormalization(momentum=bnorm['momentum'], epsilon=bnorm['epsilon']))
+      model.add(Activation('relu'))
+      if self.use_dropout:
+        model.add(Dropout(rate=dropout))
+    model.add(Dense(self.num_classes, activation='softmax'))
+
+    if optimizer == 'sgd':
+      opt = SGD(optParam['lr'], optParam['momentum'], optParam['decay'])
+    elif optimizer == 'rmsprop':
+      opt = RMSprop(optParam['lr'])
+    elif optimizer == 'adagrad':
+      opt = Adagrad(optParam['lr'])
+    else:
+      opt = Adam(optParam['lr'])
+    model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy'])
+    return model
+
+  def fitNet(self, model, data, batchSize=128, numEpochs=1, verbose=1):
+    trainData = data['train_data']
+    valData = data['val_data']
+    trainLab = data['train_labels']
+    valLab = data['val_labels']
+
+    networkHistory = model.fit(trainData, trainLab, batchSize, numEpochs, verbose, validation_data=(valData, valLab))
+    return networkHistory
+
+  """
+  ONLY USE THE FOLLOWING WHEN TESTING
+  """
+  def testModel(model, testData, testLabels):
+    loss, accuracy = model.evaluate(testData, testLabels, verbos=1)
+    return loss, accuracy
+
+  def predictTest(model, testData, testLabels):
+    predictions = model.predict(testData)
+    return predictions

ml/runNet.py

@@ -0,0 +1,26 @@
+from keras.datasets import mnist
+from neuralNet import *
+import utils as ut
+from data_utils import *
+
+"""
+Main code for running Neural Network
+"""
+
+if __name__ == '__main__':
+
+  trainDir = ""
+  testDir = ""
+  classes = 2
+
+  train_data, train_labels, height, width = load_data(trainDir)
+  test_data, test_labels, _, _ = load_data(testDir)
+  train_data = formatData(train_data, classes, height, width)
+  test_data = formatData(test_data, classes, height, width)
+  train_data, train_labels, val_data, val_labels = sampleData(train_data, train_labels)
+
+  data = {'train_data': train_data, 'val_data': val_data, 'train_labels': train_labels, 'val_labels': val_labels, 'test_data': test_data, 'test_labels': test_labels}
+
+  model = neuralNet(input_dim=height*width*3, num_classes=classes, dropout=0, use_bn=False, reg=0.0)
+  net = model.buildNet(neurons=100, optimizer='sgd', optParam={'lr': 0.001, 'momentum': 0, 'decay': 0}, bnorm={'momentum': 0.99, 'epsilon': 0.001}, dropout=0)
+  hist = model.fitNet(net, data, batchSize=128, numEpochs=2, verbose=1)

ml/utils.py

@@ -0,0 +1,26 @@
+from keras.utils import np_utils
+import matplotlib.pyplot as plt
+
+"""
+Utility methods for fixing categories and plotting graphs
+Katherine Mayo 4/15/2018
+"""
+
+def setLabels(labels, numCategories):
+  return np_utils.to_categorical(labels, numCategories)
+
+def plotNetHistory(networkHistory):
+  plot.figure()
+  plt.xlabel('Epochs')
+  plt.ylabel('Loss')
+  plt.plot(networkHistory.history['loss'])
+  plt.plot(networkHistory.history['val_loss'])
+  plt.legend(['Training', 'Validation'])
+  plt.figure()
+
+  plt.xlabel('Epochs')
+  plt.ylabel('Accuracy')
+  plt.plot(networkHistory.history['acc'])
+  plt.plot(networkHistory.history['val_acc'])
+  plt.legend(['Training', 'Validation'], loc='lowerright')
+  plt.show()