fix data generate.

add faceswap
change paper name

cnn_visualization/generate_class_specific_samples.py

@@ -43,7 +43,7 @@ def generate(self, iterations=150):
             np.ndarray -- Final maximally activated class image
         """
         print("bat dau generate xong ... ")
-        initial_learning_rate = 6
+        initial_learning_rate = 200
         for i in range(1, iterations):
             print(i)
             # Process image and return variable
@@ -57,9 +57,9 @@ def generate(self, iterations=150):
             print(output)
             class_loss = -output[0, self.target_class]
 
-            if i % 10 == 0 or i == iterations-1:
+            if i % 1 == 0 or i == iterations-1:
                 print('Iteration:', str(i), 'Loss',
-                      "{0:.2f}".format(class_loss.data.numpy()))
+                      "{0:.2f}".format(class_loss.cpu().data.numpy()))
             # Zero grads
             self.model.zero_grad()
             # Backward
@@ -68,16 +68,18 @@ def generate(self, iterations=150):
             optimizer.step()
             # Recreate image
             self.created_image = recreate_image(self.processed_image)
-            if i % 10 == 0 or i == iterations-1:
+            print(self.created_image.size)
+            if i % 1 == 0 or i == iterations-1:
                 # Save image
+                initial_learning_rate /=2
                 im_path = 'generated/class_'+str(self.target_class)+'/c_'+str(self.target_class)+'_'+'iter_'+str(i)+'.png'
                 save_image(self.created_image, im_path)
 
         return self.processed_image
 
 def parse_args():
     parser = argparse.ArgumentParser(description="Deepfake detection")
-    parser.add_argument('--model_path', default="../../model/xception/model_pytorch_4.pt", help='path to model ')
+    parser.add_argument('--model_path', default="../../../model/xception/model_pytorch_4.pt", help='path to model ')
     parser.add_argument('--gpu_id',type=int, default=-1, help='path to model ')
     parser.add_argument('--image_size',type=int, default=256, help='path to model ')
     parser.add_argument('--iterations',type=int, default=256, help='iterations random number')
@@ -171,13 +173,13 @@ def parse_args():
         pass
 
 
-    # from pytorch_model.xception import xception
+    from pytorch_model.xception import xception
 
-    # model = xception(pretrained=False)
+    model = xception(pretrained=False)
     device = torch.device("cuda" if torch.cuda.is_available()
                           else "cpu")
     model = model.to(device)
-    model.load_state_dict(torch.load(args.model_path))
+    model.load_state_dict(torch.load(args.model_path,map_location=torch.device('cpu')))
     print("Load xong ... ")
     model.eval()
     csig = ClassSpecificImageGeneration(model, target_class,image_size)

cnn_visualization/generate_regularized_class_specific_samples.py

@@ -76,7 +76,7 @@ def generate(self, iterations=150, blur_freq=4, blur_rad=1, wd=0.0001, clipping_
             optimizer = SGD([self.processed_image],
                             lr=initial_learning_rate, weight_decay=wd)
             # Forward
-            output = self.model(self.processed_image)
+            output = self.model(self.processed_image).cpu()
             # Target specific class
             class_loss = -output[0, self.target_class]
 

cnn_visualization/misc_functions.py

@@ -45,6 +45,7 @@ def save_gradient_images(gradient, file_name):
     # Normalize
     gradient = gradient - gradient.min()
     gradient /= gradient.max()
+    print(gradient.max())
     # Save image
     path_to_file = os.path.join('../results', file_name + '.jpg')
     save_image(gradient, path_to_file)
@@ -108,6 +109,7 @@ def format_np_output(np_arr):
     """
     # Phase/Case 1: The np arr only has 2 dimensions
     # Result: Add a dimension at the beginning
+    print("format_np_output  :",np_arr.shape)
     if len(np_arr.shape) == 2:
         np_arr = np.expand_dims(np_arr, axis=0)
     # Phase/Case 2: Np arr has only 1 channel (assuming first dim is channel)
@@ -116,12 +118,16 @@ def format_np_output(np_arr):
         np_arr = np.repeat(np_arr, 3, axis=0)
     # Phase/Case 3: Np arr is of shape 3xWxH
     # Result: Convert it to WxHx3 in order to make it saveable by PIL
+    if np_arr.shape[0] > 3:
+        np_arr = np_arr[6:9]
     if np_arr.shape[0] == 3:
         np_arr = np_arr.transpose(1, 2, 0)
     # Phase/Case 4: NP arr is normalized between 0-1
     # Result: Multiply with 255 and change type to make it saveable by PIL
     if np.max(np_arr) <= 1:
         np_arr = (np_arr*255).astype(np.uint8)
+    else:
+        np_arr = (np_arr * 255).astype(np.uint8)
     return np_arr
 
 
@@ -134,11 +140,12 @@ def save_image(im, path):
     """
     if isinstance(im, (np.ndarray, np.generic)):
         im = format_np_output(im)
+        # print(im)
         im = Image.fromarray(im)
     im.save(path)
 
 
-def preprocess_image(pil_im, resize_im=True,image_size=256):
+def preprocess_image(pil_im, resize_im=True,image_size=128):
     """
         Processes image for CNNs
 
@@ -210,6 +217,8 @@ def get_positive_negative_saliency(gradient):
     returns:
         pos_saliency ( )
     """
+    print(gradient.min())
+    print(gradient.max())
     pos_saliency = (np.maximum(0, gradient) / gradient.max())
     neg_saliency = (np.maximum(0, -gradient) / -gradient.min())
     return pos_saliency, neg_saliency
@@ -230,9 +239,15 @@ def get_example_params(example_index):
         pretrained_model(Pytorch model): Model to use for the operations
     """
     # Pick one of the examples
-    example_list = (('../input_images/snake.jpg', 56),
-                    ('../input_images/cat_dog.png', 243),
-                    ('../input_images/spider.png', 72))
+    # example_list = (("../../../data/extract_raw_img_test/real/aansscoqsl.mp4_8.jpg", 0),
+    #                 ("../../../data/extract_raw_img_test/real/aansscoqsl.mp4_10.jpg", 0),
+    #                 ("../../../data/extract_raw_img_test/real/aansscoqsl.mp4_17.jpg", 0))
+    # example_list = (("G:/fake_test/1/fake/1/376.png", 0),
+    #                 ("G:/fake_test/1/fake/1/377.png", 0),
+    #                 ("G:/fake_test/1/fake/1/378.png", 0))
+    example_list = (("G:/real_test/3/real/59/1214.png", 0),
+                    ("G:/real_test/3/real/59/1169.png", 0),
+                    ("G:/real_test/3/real/59/1169.png", 0))
     img_path = example_list[example_index][0]
     target_class = example_list[example_index][1]
     file_name_to_export = img_path[img_path.rfind('/')+1:img_path.rfind('.')]
@@ -241,9 +256,9 @@ def get_example_params(example_index):
     # Process image
     prep_img = preprocess_image(original_image)
     # Define model
-    pretrained_model = models.alexnet(pretrained=True)
+    # pretrained_model = models.alexnet(pretrained=True)
     return (original_image,
             prep_img,
             target_class,
-            file_name_to_export,
-            pretrained_model)
+            file_name_to_export)
+            # pretrained_model)

paper/README.md

@@ -6,23 +6,23 @@ https://github.com/NVlabs/stylegan2
 
 [paper v1](1812.04948.pdf)
 
-[paper v1](1912.04958.pdf)
+[paper v1](stylegan2.pdf)
 
 
 ## starGAN
 https://github.com/clovaai/stargan-v2
 
-[paper v1](1711.09020.pdf)
+[paper v1](stargan.pdf)
 
-[paper v2](1912.01865.pdf)
+[paper v2](stargan2.pdf)
 
 
 
 
 # Phương pháp phát hiện ảnh giả mạo
 ## Classic ML
 ### Head pose 
-[paper](1811.00661.pdf)
+[paper](headpose.pdf)
 
 Exposing deep fakes using inconsistent head poses
 
@@ -33,7 +33,7 @@ có thể xác phân biệt được ảnh thật giả.
 
 
 ### Visual Arrtifact
-[paper](Exploiting visual artifacts to expose deepfakes and face manipulations-annotated.pdf)
+[paper](visual_artifacts.pdf)
 
 Exploiting Visual Artifacts to Expose Deepfakes and Face Manipulations
 
@@ -45,7 +45,7 @@ Tính toán những đặc điểm trên khuôn mặt:
 
 
 ### frequency domain 
-[paper](1911.00686.pdf)
+[paper](frequency_domain.pdf)
 
 Unmasking DeepFakes with simple Features
 
@@ -58,7 +58,7 @@ toán này tính trung bình các giá trị có cùng khoảng cách tới tâm
 
 ## Deep learning
 ### Mesonet
-[paper](1809.00888.pdf)
+[paper](mesonet.pdf)
 
 MesoNet: a compact facial video forgery detection network
 
@@ -68,7 +68,7 @@ tầng cao có thể dẫn đến không phân biệt được ảnh khuôn mặ
  thuận lợi cho việc phân biệt các ảnh thật và giả. 
 
 ### Capsule
-[paper](Capsule-forensics Using Capsule Networks to Detect Forged Images and Videos-annotated.pdfs)
+[paper](capsule.pdf)
 
 Capsule-forensics: Using capsule networks to detect forged images and videos.
 
@@ -79,7 +79,7 @@ Một lớp capsule network được thiết kế phía sau mạng VGG-19 để
 
 
 ### xception
-[paper](1901.08971-annotated.pdf)
+[paper](facefornsic++.pdf)
 
 Faceforensics++: Learning to detect manipulated facial images
 
@@ -93,6 +93,6 @@ tính toán so với lớp tích chập thông thường.
 
 ## Figerprint
 
-[paper](1811.08180.pdf)
+[paper](fingerprint.pdf)