complete version

ResNet.py

@@ -1,22 +1,19 @@
 """
-A Trainable ResNet Class is defined in this file
+A Trainable ResNet-50 Class is defined in this file
 Author: Kaihua Tang
 """
 import math
 import numpy as np
 import tensorflow as tf
 from functools import reduce
 
-VGG_MEAN = [104.7546, 124.328, 167.1754]
-_BATCH_NORM_DECAY = 0.99
-_BATCH_NORM_EPSILON = 1e-12
-
 class ResNet:
-    # some properties
-    """
-    Initialize function
-    """
-    def __init__(self, ResNet_npy_path=None, trainable=True, open_tensorboard=False, dropout=0.8):
+    def __init__(self, ResNet_npy_path=None, trainable=True, open_tensorboard=False):
+        """
+        Initialize function
+        ResNet_npy_path: If path is not none, loading the model. Otherwise, initialize all parameters at random.
+        open_tensorboard: Is Open Tensorboard or not. 
+        """
         if ResNet_npy_path is not None:
             self.data_dict = np.load(ResNet_npy_path, encoding='latin1').item()
         else:
@@ -25,26 +22,29 @@ def __init__(self, ResNet_npy_path=None, trainable=True, open_tensorboard=False,
         self.var_dict = {}
         self.trainable = trainable
         self.open_tensorboard = open_tensorboard
-        self.dropout = dropout
         self.is_training = True
 
     def set_is_training(self, isTrain):
+        """
+        Set is training bool.
+        """
         self.is_training = isTrain
 
-    def build(self, rgb, label_num, train_mode=None, last_layer_type = "softmax"):
+    def build(self, rgb, label_num, last_layer_type = "softmax"):
         """
         load variable from npy to build the Resnet or Generate a new one
         :param rgb: rgb image [batch, height, width, 3] values scaled [0, 1]
-        :param train_mode: a bool tensor, usually a placeholder: if True, dropout will be turned on
         """
+        # Preprocessing: Turning RGB to BGR - Mean.
+        BGR_MEAN = [104.7546, 124.328, 167.1754]
         red, green, blue = tf.split(axis=3, num_or_size_splits=3, value=rgb)
         assert red.get_shape().as_list()[1:] == [224, 224, 1]
         assert green.get_shape().as_list()[1:] == [224, 224, 1]
         assert blue.get_shape().as_list()[1:] == [224, 224, 1]
         bgr = tf.concat(axis=3, values=[
-            blue - VGG_MEAN[0],
-            green - VGG_MEAN[1],
-            red - VGG_MEAN[2],
+            blue - BGR_MEAN[0],
+            green - BGR_MEAN[1],
+            red - BGR_MEAN[2],
         ])
         print(bgr.get_shape().as_list())
         assert bgr.get_shape().as_list()[1:] == [224, 224, 3]
@@ -114,6 +114,11 @@ def res_block_3_layers(self, bottom, channel_list, name, change_dimension = Fals
         return relu
 
     def batch_norm(self, inputsTensor):
+        """
+        Batchnorm
+        """
+        _BATCH_NORM_DECAY = 0.99
+        _BATCH_NORM_EPSILON = 1e-12
         return tf.layers.batch_normalization(inputs=inputsTensor, axis = 3, momentum=_BATCH_NORM_DECAY, epsilon=_BATCH_NORM_EPSILON, center=True, scale=True, training=self.is_training)
 
     def avg_pool(self, bottom, kernal_size = 2, stride = 2, name = "avg"):
@@ -241,7 +246,7 @@ def get_var(self, initial_value, name, idx, var_name):
         return var
 
 
-    def save_npy(self, sess, npy_path="./Resnet-save.npy"):
+    def save_npy(self, sess, npy_path="./model/Resnet-save.npy"):
         """
         Save this model into a npy file
         """

TestResNet.py

@@ -0,0 +1,69 @@
+"""
+Extract ResNet feature
+Author: Kaihua Tang
+"""
+
+import math
+import time
+import tensorflow as tf
+import ResNet as resnet
+import numpy as np
+import scipy.io as scio
+from scipy import misc
+from utils import *
+
+# image size
+WIDTH = 224
+HEIGHT = 224
+CHANNELS = 3
+#Number of output labels
+LABELSNUM = 1200
+#"Path of Label.npy"
+label_path = "./label/label.npy"
+#"Path of image file names"
+image_name_path = "./label/name.npy"
+# image path
+parentPath = "F:\\CACD2000_Crop\\"
+
+
+def CalculateFeature():
+    """
+    EXtract ResNet Feature by trained model
+    model_path: The model we use
+    feature_path: The path to save feature
+    """
+    model_path = "./model/11.npy"
+    feature_path = "./resnet_feature.mat"
+
+    #Lists that store name of image and its label
+    testNameList = np.load(image_name_path)
+    testLabelList = np.load(label_path)
+
+    #num of total training image
+    num_test_image = testLabelList.shape[0]
+    #load all image data
+    allImageData = load_all_image(testNameList, HEIGHT, WIDTH, CHANNELS, parentPath)
+    #container for ResNet Feature
+    res_feature = np.zeros((num_test_image, 2048))
+
+    with tf.Session() as sess:
+        images = tf.placeholder(tf.float32, shape = [None, WIDTH, HEIGHT, CHANNELS])
+
+        # build resnet model
+        resnet_model = resnet.ResNet(ResNet_npy_path = model_path)
+        resnet_model.build(images, LABELSNUM, "softmax")
+
+        sess.run(tf.global_variables_initializer())
+        resnet_model.set_is_training(False)
+
+        for i in range(num_test_image):
+           if(i%1000 == 0):
+                print(i)
+           (minibatch_X, minibatch_Y) = get_minibatch([i], testLabelList, HEIGHT, WIDTH, CHANNELS, LABELSNUM, allImageData, True)
+           pool2 = sess.run([resnet_model.pool2], feed_dict={images: minibatch_X})
+           res_feature[i][:] = pool2[0][:]
+
+        scio.savemat(feature_path,{'feature' : res_feature})
+
+if __name__ == '__main__':
+    CalculateFeature()