Feat: major functions

1. Build Siamese NIMA model network
2. Train and predict NIMA model network

Author: ryanfwy

demo_predict.py

@@ -0,0 +1,17 @@
+'''Predict with NIMA model network.'''
+
+from model.siamese_nima import SiameseNIMA
+
+
+if __name__ == '__main__':
+    # dirs and paths to load data
+    predict_image_dir = './assets/demo/predict_images'
+    predict_data_path = './assets/demo/predict_data.csv'
+
+    # load data and train model
+    siamese = SiameseNIMA()
+    predict_raw = siamese.load_data(predict_image_dir, predict_data_path,
+                                    columns=['file_name'])
+    results = siamese.predict(predict_raw,
+                              nima_weight_path='./assets/weights/nima_weights_pre_trained.h5')
+    print(results)

demo_train.py

@@ -0,0 +1,17 @@
+'''Train Siamese NIMA model networks.'''
+
+from model.siamese_nima import SiameseNIMA
+
+
+if __name__ == '__main__':
+    # dirs and paths to load data
+    train_image_dir = './assets/demo/train_images'
+    train_data_path = './assets/demo/train_data.csv'
+
+    # load data and train model
+    siamese = SiameseNIMA(output_dir='./assets')
+    train_raw = siamese.load_data(train_image_dir, train_data_path)
+    siamese.train(train_raw,
+                  epochs=5,
+                  batch_size=16,
+                  nima_weight_path='./assets/weights/nima_weights_pre_trained.h5')

installation.sh

@@ -0,0 +1,35 @@
+# 0. Check python and pip3
+(python3 -V && pip3 -V) &>/dev/null;
+if [[ $? -ne 0 ]]; then
+    echo "Error: please check whether 'python3' and 'pip3' is valid";
+    exit 1;
+fi
+
+# 1. Create virtual environment
+conda -V &>/dev/null;
+if [[ $? -eq 0 ]]; then
+    echo "1. Creating conda environment: image_quality";
+    conda create --name image_quality python=3.6;
+    source activate image_quality;
+else
+    echo "1. Creating virtual environment: .env"
+    python3 -m venv .env
+    source .env/bin/activate
+fi
+
+# 2. Install tensorflow backend
+nvidia-smi &>/dev/null;
+if [[ $? -eq 0 ]]; then
+    echo "2. Installing backend: tensorflow-gpu==1.12.0";
+    pip3 install numpy==1.15.4 Keras==2.2.4 tensorflow-gpu==1.12.0;
+else
+    echo "2. Installing backend: tensorflow==1.12.0";
+    pip3 install numpy==1.15.4 Keras==2.2.4 tensorflow==1.12.0;
+fi
+
+# 3. Install dependencies
+echo "3. Installing the rest dependencies";
+pip3 install -r requirements.txt;
+
+# 4. Deactivate
+deactivate;

model/__init__.py

@@ -0,0 +1,86 @@
+'''DataSequence for model training.
+Inherit from keras.utils.Sequence,
+see https://keras.io/utils/#sequence for more details.
+'''
+
+import random
+import numpy as np
+
+from keras.utils import Sequence
+from keras.applications.inception_resnet_v2 import preprocess_input
+from keras.preprocessing.image import load_img, img_to_array
+
+
+class DataSequence(Sequence):
+    def __init__(self, x_raw, y_raw, batch_size, num_classes, target_size=(224, 224)):
+        '''Create train / validate generator.
+
+        Args:
+            x_raw (np.ndarray): an array of image paths, not image data.
+            y_raw (np.ndarray): an array of image classes.
+            batch_size (int): the sample size of each batch.
+            num_classes (int): the number of classes that dataset have.
+            target_size (tuple, optional): the size (width, height) of image data. Defaults to (224, 224).
+        '''
+        self.batch_size = batch_size
+        self.num_classes = num_classes
+        self.target_size = target_size
+        self.x, self.y = self.__create_pairs(x_raw, y_raw)
+
+    def __create_pairs(self, x_raw, y_raw):
+        '''Create positive and negative pairs.
+        Wont load image data itself but create a list of sample pairs,
+        so that we can train model with generator and load image data asynchronous.
+
+        Args:
+            x_raw (np.ndarray): an array of image paths, not image data.
+            y_raw (np.ndarray): an array of image classes.
+
+        Returns:
+            pairs (list): the samples of image paths combined into pairs, known as `x`.
+            labels (list): the samples of image labels combined into pairs, known as `y`.
+        '''
+        pairs = []
+        labels = []
+        digit_indices = [np.where(y_raw == i)[0] for i in range(self.num_classes)]
+        n = min([len(digit_indices[d]) for d in range(self.num_classes)]) - 1
+        for d in range(self.num_classes):
+            for i in range(n):
+                z1, z2 = digit_indices[d][i], digit_indices[d][i + 1]
+                pairs += [[x_raw[z1], x_raw[z2]]]
+                inc = random.randrange(1, self.num_classes)
+                dn = (d + inc) % self.num_classes
+                z1, z2 = digit_indices[d][i], digit_indices[dn][i]
+                pairs += [[x_raw[z1], x_raw[z2]]]
+                labels += [1, 0]
+        return pairs, labels
+
+    def __load_image_data(self, image_path):
+        '''Load image data from local image path.
+
+        Args: 
+            image_path (str): the image path of an image.
+
+        Returns:
+            x (np.ndarray): the image data, known as `x`.
+        '''
+        img = load_img(image_path, target_size=self.target_size)
+        x = img_to_array(img)
+        x = preprocess_input(x) # normalization
+        return x
+
+    def __len__(self):
+        return int(np.ceil(len(self.x) / float(self.batch_size)))
+
+    def __getitem__(self, idx):
+        batch_x_raw = self.x[idx * self.batch_size:(idx + 1) * self.batch_size]
+        batch_y_raw = self.y[idx * self.batch_size:(idx + 1) * self.batch_size]
+        batch_x1, batch_x2 = [], []
+        for x1, x2 in batch_x_raw:
+            batch_x1.append(self.__load_image_data(x1))
+            batch_x2.append(self.__load_image_data(x2))
+        batch_x1 = np.array(batch_x1)
+        batch_x2 = np.array(batch_x2)
+        batch_x = [batch_x1, batch_x2]
+        batch_y = np.array(batch_y_raw)
+        return batch_x, batch_y

model/data_sequence.py

@@ -0,0 +1,47 @@
+'''DataSequence for model predicting.
+Inherit from keras.utils.Sequence,
+see https://keras.io/utils/#sequence for more details.
+'''
+
+import random
+import numpy as np
+
+from keras.utils import Sequence
+from keras.applications.inception_resnet_v2 import preprocess_input
+from keras.preprocessing.image import load_img, img_to_array
+
+
+class DataSequencePred(Sequence):
+    def __init__(self, x_raw, batch_size=1, target_size=None):
+        '''Create predict generator.
+
+        Args:
+            x_raw (np.ndarray | list): an array of image paths, not image data.
+            batch_size (int, optional): the sample size of each batch. Defaults to 1.
+            target_size (tuple, optional): the size (width, height) of image data. Defaults to None.
+        '''
+        self.batch_size = batch_size
+        self.target_size = target_size
+        self.x = x_raw
+
+    def __load_image_data(self, image_path):
+        '''Load image data from local image path.
+
+        Args: 
+            image_path (str): the image path of an image.
+
+        Returns:
+            x (np.ndarray): the image data, known as `x`.
+        '''
+        img = load_img(image_path, target_size=self.target_size)
+        x = img_to_array(img)
+        x = preprocess_input(x) # normalization
+        return x
+
+    def __len__(self):
+        return int(np.ceil(len(self.x) / float(self.batch_size)))
+
+    def __getitem__(self, idx):
+        batch_x_raw = self.x[idx * self.batch_size:(idx + 1) * self.batch_size]
+        batch_x = [self.__load_image_data(x) for x in batch_x_raw]
+        return np.array(batch_x)