more augmentation

helpers/DataLoader.py

@@ -20,10 +20,10 @@ def flip(x, label, size):
     return x, label
 
 def color(x: tf.Tensor, label, size):
-    x = tf.image.random_hue(x, 0.08)
-    x = tf.image.random_saturation(x, 0.6, 1.6)
-    x = tf.image.random_brightness(x, 0.05)
-    x = tf.image.random_contrast(x, 0.7, 1.3)
+    x = tf.image.random_hue(x, 0.1)
+    x = tf.image.random_saturation(x, 0.5, 1.7)
+    x = tf.image.random_brightness(x, 0.2)
+    x = tf.image.random_contrast(x, 0.5, 1.5)
     return x, label
 
 def rotate(x: tf.Tensor, label, size):
@@ -31,7 +31,7 @@ def rotate(x: tf.Tensor, label, size):
 
 def zoom(x: tf.Tensor, label, size):
     # Generate 20 crop settings, ranging from a 1% to 30% crop.
-    scales = list(np.arange(0.7, 1.0, 0.01))
+    scales = list(np.arange(0.6, 1.0, 0.03))
     boxes = np.zeros((len(scales), 4))
 
     for i, scale in enumerate(scales):
@@ -48,8 +48,8 @@ def zoom(x: tf.Tensor, label, size):
 
     choice = tf.random.uniform(shape=[], minval=0., maxval=1., dtype=tf.float32)
 
-    # Only apply cropping 70% of the time
-    return tf.cond(choice < 0.7, lambda: x, lambda: random_crop(x)), label
+    # Only apply cropping 80% of the time
+    return tf.cond(choice < 0.8, lambda: x, lambda: random_crop(x)), label
 
 class DataLoader(object):
 
@@ -63,7 +63,7 @@ class DataLoader(object):
         self.IMG_HEIGHT = img_height
         self.CHANNELS = channels
         self.AUGMENT = augment
-        self.AUGMENTATIONS = [flip, color, zoom, rotate]
+        self.AUGMENTATIONS = [flip, color, zoom]
 
         self.classes = [item.name for item in data_path.glob('*') if item.name != '.DS_Store']
         self.n_classes = len(self.classes)
@@ -180,14 +180,14 @@ class DataLoader(object):
         self.dataset_1 = self.dataset_1.cache()
         if self.AUGMENT:
             for f in self.AUGMENTATIONS:
-                self.dataset_1 = self.dataset_1.map(lambda x,y: tf.cond(tf.random.uniform([], 0, 1) > 0.6, lambda: f(x,y, (self.IMG_WIDTH, self.IMG_HEIGHT)), lambda: (x,y)),
+                self.dataset_1 = self.dataset_1.map(lambda x,y: tf.cond(tf.random.uniform([], 0, 1) > 0.1, lambda: f(x,y, (self.IMG_WIDTH, self.IMG_HEIGHT)), lambda: (x,y)),
                                       num_parallel_calls=self.NR_THREADS)
             self.dataset_1 = self.dataset_1.map(lambda x,y: (tf.clip_by_value(x, 0, 1),y))
             if self.dataset_2:
                 self.dataset_2 = self.dataset_2.map(self.process_path, num_parallel_calls=self.NR_THREADS)
                 self.dataset_2 = self.dataset_2.cache()
                 for f in self.AUGMENTATIONS:
-                    self.dataset_2 = self.dataset_2.map(lambda x, y: tf.cond(tf.random.uniform([], 0, 1) > 0.6, lambda: f(x, y, (self.IMG_WIDTH, self.IMG_HEIGHT)),
+                    self.dataset_2 = self.dataset_2.map(lambda x, y: tf.cond(tf.random.uniform([], 0, 1) > 0.1, lambda: f(x, y, (self.IMG_WIDTH, self.IMG_HEIGHT)),
                                                                              lambda: (x, y)), num_parallel_calls=self.NR_THREADS)
                 self.dataset_2 = self.dataset_2.map(lambda x, y: (tf.clip_by_value(x, 0, 1), y))
 

network/VGG_16.py

@@ -2,7 +2,7 @@ from tensorflow.keras import regularizers
 
 from network.NetworkBase import NetworkBase
 from tensorflow.keras.models import Model
-from tensorflow.keras.layers import Input, Dense, GlobalAveragePooling2D, Dropout, Flatten, AveragePooling2D
+from tensorflow.keras.layers import Input, Dense, GlobalAveragePooling2D, Dropout, Flatten, AveragePooling2D, GaussianNoise
 from tensorflow.keras.applications.vgg16 import VGG16
 
 import pathlib
@@ -25,19 +25,20 @@ class VGG_16(NetworkBase):
                 print('Given weigths-file for base ResNet not found. (Missing: {})'.format(self.WEIGHTS_PATH))
 
         input_tensor = Input(shape=(self.IMG_WIDTH, self.IMG_HEIGHT, self.CHANNELS))
+        input_tensor = GaussianNoise(stddev=0.01)(input_tensor)
 
         base_model = VGG16(weights=weights, input_tensor=input_tensor, include_top=False)
 
         if shouldSave:
             base_model.save('input/vgg_base_model.h5')
 
-        x = GlobalAveragePooling2D()(base_model.output)
-        x = Dense(256, activation='relu')(x)
+        x = Flatten()(base_model.output)
+        x = Dense(1024, activation='relu')(x)
         x = Dense(256, activation='relu')(x)
         out = Dense(self.NUM_CLASSES, activation='softmax', name='probs')(x)
 
         model = Model(base_model.input, out)
 
-        for layer in model.layers[:15]:
+        for layer in model.layers[:16]:
             layer.trainable = False
         return model

test_augmentation.py

 import pathlib
 import numpy as np
 import matplotlib.pylab as plt
-from classifier.DataLoader import DataLoader
+from helpers.DataLoader import DataLoader
 
+'''
+Tool to visually inspect the (augmented) output of the DataLoader
 
-train_dir = pathlib.Path('data/training_data_one')
+This class should be run in a local environment and not in a docker container. 
+The purpose is to assure the plt.show will actually show the created images.  
+'''
+
+class_index_map = {'caries': 0, 'no_caries': 1}
+index_class_map = {}
+for element in class_index_map:
+    index_class_map[class_index_map[element]] = element
+
+
+train_dir = pathlib.Path('input/training_data_mini')
 batch_size = 5
-img_width = 1200
-img_height = 1200
+img_width = 500
+img_height = 500
 channels = 3
-batch_count_dataset1 = 5
+batch_count_dataset1 = 9
 
 train_loader = DataLoader(data_path=train_dir,
                           batch_size=batch_size,
-                          batch_count_dataset1=batch_count_dataset1,
+                          should_size_dataset1=batch_count_dataset1,
                           img_width=img_width,
                           img_height=img_height,
                           channels=channels,
@@ -34,14 +46,17 @@ def plot_images(dataset, n_images, samples_per_image):
     plt.imshow(output)
     plt.show()
 
-#plot_images(train_dataset, 1, dataset1_size)
 
 count = 0
-for (img, label) in train_dataset:
-    for batch_img in img:
+for batch_img, batch_label in train_dataset.as_numpy_iterator():
+    for i in range(len(batch_label)):
+        img = batch_img[i]
+        label = batch_label[i]
         count += 1
-        print(batch_img.shape)
-        plt.imshow(batch_img)
+        print(img.shape)
+        label = index_class_map[np.argmax(label)]
+        plt.title(label)
+        plt.imshow(img)
         plt.show()
     print("new batch")
 print("{} image(s) processed".format(count))