add models

Author: clennan

models/vgg16.py

@@ -0,0 +1,137 @@
+
+from src.modules.sequential import Sequential
+from src.modules.linear import Linear
+from src.modules.convolution import Convolution
+from src.modules.convolution_ab import Convolution_ab
+from src.modules.reshape import Reshape
+from src.modules.proppool import PropPool
+from src.utils import load_json
+import numpy as np
+import cv2
+
+
+def vgg_addmean(images):
+    mean = [103.939, 116.779, 123.68]
+    img_trans = []
+    for img in images:
+        blue, green, red = np.split(img, 3, 2)
+        img_trans.append(np.concatenate(
+            [blue + mean[0], green + mean[1], red + mean[2]], 2),
+            )
+    return img_trans
+
+
+def bgr_to_rgb(images):
+    img_trans = []
+    for img in images:
+        blue, green, red = np.split(img, 3, 2)
+        img_trans.append(np.concatenate([red, green, blue], 2))
+    return img_trans
+
+
+class VGG16(object):
+    def __init__(self, weights_file, classes_file):
+        self.classes = load_json(classes_file)
+        self.weights = np.load(weights_file, encoding='latin1').item()
+
+    def vgg_addmean(self, images):
+        mean = [103.939, 116.779, 123.68]
+        img_trans = []
+        for img in images:
+            blue, green, red = np.split(img, 3, 2)
+            img_trans.append(np.concatenate(
+                [blue + mean[0], green + mean[1], red + mean[2]], 2),
+                )
+        return img_trans
+
+    def load_image(self, image_file):
+        image = cv2.imread(image_file)
+        image = cv2.resize(image, (224, 224))
+        image = np.expand_dims(image, axis=0)
+        return self.vgg_addmean(image)
+
+    def build_model(self, batch_size, alpha):
+        return Sequential([
+            Convolution(batch_size=batch_size,
+                        initializer=self.weights['conv1_1'],
+                        first=True,
+                        name='conv1_1_'),
+            Convolution_ab(
+                        batch_size=batch_size,
+                        initializer=self.weights['conv1_2'],
+                        alpha=alpha,
+                        name='conv1_2_'),
+            PropPool(name='PropPool1'),
+            Convolution_ab(
+                        batch_size=batch_size,
+                        initializer=self.weights['conv2_1'],
+                        alpha=alpha,
+                        name='conv2_1_'),
+            Convolution_ab(
+                        batch_size=batch_size,
+                        initializer=self.weights['conv2_2'],
+                        alpha=alpha,
+                        name='conv2_2_'),
+            PropPool(name='PropPool2'),
+            Convolution_ab(
+                        batch_size=batch_size,
+                        initializer=self.weights['conv3_1'],
+                        alpha=alpha,
+                        name='conv3_1_'),
+            Convolution_ab(
+                        batch_size=batch_size,
+                        initializer=self.weights['conv3_2'],
+                        alpha=alpha,
+                        name='conv3_2_'),
+            Convolution_ab(
+                        batch_size=batch_size,
+                        initializer=self.weights['conv3_3'],
+                        alpha=alpha,
+                        name='conv3_3_'),
+            PropPool(name='PropPool3'),
+            Convolution_ab(
+                        batch_size=batch_size,
+                        initializer=self.weights['conv4_1'],
+                        alpha=alpha,
+                        name='conv4_1_'),
+            Convolution_ab(
+                        batch_size=batch_size,
+                        initializer=self.weights['conv4_2'],
+                        alpha=alpha,
+                        name='conv4_2_'),
+            Convolution_ab(
+                        batch_size=batch_size,
+                        initializer=self.weights['conv4_3'],
+                        alpha=alpha,
+                        name='conv4_3_'),
+            PropPool(name='PropPool4'),
+            Convolution_ab(
+                        batch_size=batch_size,
+                        initializer=self.weights['conv5_1'],
+                        alpha=alpha,
+                        name='conv5_1_'),
+            Convolution_ab(
+                        batch_size=batch_size,
+                        initializer=self.weights['conv5_2'],
+                        alpha=alpha,
+                        name='conv5_2_'),
+            Convolution_ab(
+                        batch_size=batch_size,
+                        initializer=self.weights['conv5_3'],
+                        alpha=alpha,
+                        name='conv5_3_'),
+            PropPool(name='PropPool5'),
+            Reshape(name='flat1'),
+            Linear(batch_size=batch_size,
+                   initializer=self.weights['fc6'],
+                   alpha=alpha,
+                   name='fc6_'),
+            Linear(batch_size=batch_size,
+                   initializer=self.weights['fc7'],
+                   alpha=alpha,
+                   name='fc7_'),
+            Linear(batch_size=batch_size,
+                   initializer=self.weights['fc8'],
+                   alpha=alpha,
+                   name='fc8_'),
+                   ])