#23: moidifed dataloader.py and dependency code

 1. move detection_collate function utils.py to dataloader.py
 2. class selective option in convert2Yolo/Format.py

pytorch_version/convertYolo/Format.py

@@ -216,7 +216,7 @@ def save(xml_list, path):
             return False, msg
 
     @staticmethod
-    def parse(path, cls_option=False, selective_cls=None):
+    def parse(path):
         try:
 
             (dir_path, dir_names, filenames) = next(os.walk(os.path.abspath(path)))
@@ -264,15 +264,9 @@ def parse(path, cls_option=False, selective_cls=None):
                     }
                     tmp["bndbox"] = bndbox
 
-                    if (cls_option==True) and (selective_cls != None):
-                        if _object.find("name").text != selective_cls:
-                            continue
-                        else:
-                            obj[str(obj_index)] = tmp
-                            obj_index += 1
-                    else:
-                        obj[str(obj_index)] = tmp
-                        obj_index += 1
+
+                    obj[str(obj_index)] = tmp
+                    obj_index += 1
 
                 annotation = {
                     "size": size,
@@ -281,11 +275,6 @@ def parse(path, cls_option=False, selective_cls=None):
 
                 if obj_index != 0:
 
-                    if cls_option==True:
-                        annotation["objects"]["num_obj"] = obj_index
-                    #else:
-                        #annotation["objects"]["num_obj"] = obj_index
-
                     data[filename.split(".")[0]] = annotation
 
                 printProgressBar(progress_cnt + 1, progress_length, prefix='VOC Parsing:'.ljust(15), suffix='Complete', length=40)

pytorch_version/utilities/dataloader.py

@@ -1,26 +1,66 @@
-import sys, os
-
-sys.path.insert(0, os.path.dirname(__file__))
+import sys
+import os
 
 import torch
-import torchvision
 import torch.utils.data as data
-from PIL import Image, ImageDraw
 import matplotlib.pyplot as plt
-
 import numpy as np
+from PIL import Image
 
 from convertYolo.Format import YOLO as cvtYOLO
 from convertYolo.Format import VOC as cvtVOC
-"""
-try:
-    from convertYolo.Format import YOLO as cvtYOLO
-    from convertYolo.Format import VOC as cvtVOC
-except Exception: # ImportError
-    import convertYolo.Format as Format
-    cvtYOLO = Format.YOLO
-    cvtVOC = Format.VOC
-"""
+
+sys.path.insert(0, os.path.dirname(__file__))
+
+def detection_collate(batch):
+    """ `Puts each data field into a tensor with outer dimension batch size`
+
+    Args:
+        batch : batch data ``batch[0]`` : image, ``batch[1]`` : label, ``batch[3]`` : size
+
+    Return:
+        image tensor, label tensor, sizes
+
+    Future work:
+        return value(torch.stack) change to Torch.FloatTensor()
+    """
+
+    targets = []
+    imgs = []
+    sizes = []
+
+    for sample in batch:
+        imgs.append(sample[0])
+
+        # for drawing box
+        # if using batch it should keep original image size.
+        sizes.append(sample[2])
+
+        np_label = np.zeros((7, 7, 6), dtype=np.float32)
+        for object in sample[1]:
+            objectness = 1
+            cls = object[0]
+            x_ratio = object[1]
+            y_ratio = object[2]
+            w_ratio = object[3]
+            h_ratio = object[4]
+
+            # can be acuqire grid (x,y) index when divide (1/S) of x_ratio
+            scale_factor = (1 / 7)
+            grid_x_index = int(x_ratio // scale_factor)
+            grid_y_index = int(y_ratio // scale_factor)
+            x_offset = (x_ratio / scale_factor) - grid_x_index
+            y_offset = (y_ratio / scale_factor) - grid_y_index
+
+            # insert object row in specific label tensor index as (x,y)
+            # object row follow as
+            # [objectness, class, x offset, y offset, width ratio, height ratio]
+            np_label[grid_x_index][grid_y_index] = np.array([objectness, cls, x_offset, y_offset, w_ratio, h_ratio])
+
+        label = torch.from_numpy(np_label)
+        targets.append(label)
+
+    return torch.stack(imgs, 0), torch.stack(targets, 0), sizes
 
 class VOC(data.Dataset):
     """ `VOC PASCAL Object Detection Challenge <http://host.robots.ox.ac.uk/pascal/VOC/voc2012/>_ Dataset `
@@ -39,23 +79,18 @@ class VOC(data.Dataset):
 
     """
 
-    #CLASSES = "./voc.names"
-    #CLASSES = "./person.names"
-    CLASSES = "./5class.names"
     IMAGE_FOLDER = "JPEGImages"
     LABEL_FOLDER = "Annotations"
     IMG_EXTENSIONS = '.jpg'
 
-    def __init__(self, root, train=True, transform=None, target_transform=None, resize=448, cls_option=False, selective_cls=None):
+    def __init__(self, root, train=True, transform=None, target_transform=None, resize=448, class_path='./voc.names'):
         self.root = root
         self.transform = transform
         self.target_transform = target_transform
         self.train = train
         self.resize_factor = resize
-        self.cls_option = cls_option
-        self.selective_cls = selective_cls
 
-        with open("./5class.names") as f:
+        with open(class_path) as f:
             self.classes = f.read().splitlines()
 
         if not self._check_exists():
@@ -75,10 +110,8 @@ def cvtData(self):
 
         result = []
         voc = cvtVOC()
-        yolo = cvtYOLO(os.path.abspath(self.CLASSES))
-        flag, self.dict_data =voc.parse(os.path.join(self.root, self.LABEL_FOLDER), cls_option=self.cls_option, selective_cls=self.selective_cls)
-        #print(flag, data)
-        #exit()
+        yolo = cvtYOLO(os.path.abspath(self.classes))
+        flag, self.dict_data =voc.parse(os.path.join(self.root, self.LABEL_FOLDER))
 
         try:
 
@@ -105,9 +138,11 @@ def cvtData(self):
         except Exception as e:
             raise RuntimeError("Error : {}".format(e))
 
+
     def __len__(self):
         return len(self.data)
 
+
     def __getitem__(self, index):
         """
         Args:
@@ -125,70 +160,17 @@ def __getitem__(self, index):
             ]
         """
 
-
         key = list(self.data[index].keys())[0]
-
         img = Image.open(key).convert('RGB')
-
         current_shape = img.size
-
         img = img.resize((self.resize_factor, self.resize_factor))
         img = np.array(img.getdata(), dtype=np.float).reshape(img.size[0], img.size[1], 3)
-
         target = self.data[index][key]
 
-        # for debug & visualization
-        """
-        _image = img.astype(dtype=np.uint8)
-        _image = Image.fromarray(_image, "RGB")
-        draw = ImageDraw.Draw(_image)
-
-
-        for obj in target:
-
-            cls = obj[0]
-            x_center = obj[1]
-            y_center = obj[2]
-            w_ratio = obj[3]
-            h_ratio = obj[4]
-
-            width = int(w_ratio * self.resize_factor)
-            height = int(h_ratio * self.resize_factor)
-
-            xmin = int(x_center * self.resize_factor - width/2)
-            ymin = int(y_center * self.resize_factor - height/2)
-            xmax = xmin + width
-            ymax = ymin + height
-
-            cls = self.classes[int(cls)]
-
-            draw.rectangle(((xmin, ymin), (xmax, ymax)), outline="blue")
-            draw.text((xmin, ymin), cls)
-
-
-        plt.imshow(_image)
-        plt.show()
-        """
-
-
         if self.transform is not None:
-            #img, target = self.transform([img, target])
-            #img = torchvision.transforms.ToTensor()(img)
-            img = img.astype(dtype=np.uint8)
-            print(img.shape)
-            print(img.dtype)
-            plt.imshow(img)
-            plt.show()
-            img = img.astype(dtype=np.float64)
-            img = self.transform(img)
-            print(img.shape)
-            image = img.type(torch.ByteTensor).view(448,448,3)
-            image = image.cpu().numpy()
-            plt.imshow(image)
-            plt.show()
+            print("It's not supported, transform parameter should be None")
             exit()
 
-
         else:
             img = torch.FloatTensor(img)
             img = torch.div(img, 255)
@@ -197,5 +179,4 @@ def __getitem__(self, index):
             # Future works
             pass
 
-
-        return img, target
+        return img, target, current_shape

pytorch_version/utilities/utils.py

@@ -21,53 +21,6 @@ def one_hot(output , label):
 
     return torch.from_numpy(dst)
 
-
-def detection_collate_with_size(batch):
-    r"""Puts each data field into a tensor with outer dimension batch size"""
-    targets = []
-    imgs = []
-    sizes = []
-    #print(len(batch[0]))
-    for sample in batch:
-        imgs.append(sample[0])
-        #sizes.append(sample[2])
-
-        #Add Size
-        #shape_np = sample[2]
-        #shape = torch.Tensor(shape_np)
-        sizes.append(sample[2])
-
-        np_label = np.zeros((7,7,6), dtype=np.float32)
-        for i in range(7):
-            for j in range(7):
-                np_label[i][j][1] = (num_classes-1)
-
-        for object in sample[1]:
-            objectness=1
-            #print(len(object))
-            cls = object[0]
-            x_ratio = object[1]
-            y_ratio = object[2]
-            w_ratio = object[3]
-            h_ratio = object[4]
-
-            # can be acuqire grid (x,y) index when divide (1/S) of x_ratio
-            grid_x_index = int(x_ratio // (1/7))
-            grid_y_index = int(y_ratio // (1/7))
-            x_offset = x_ratio - ((grid_x_index) * (1/7))
-            y_offset = y_ratio - ((grid_y_index) * (1/7))
-
-            # insert object row in specific label tensor index as (x,y)
-            # object row follow as
-            # [objectness, class, x offset, y offset, width ratio, height ratio]
-            np_label[grid_x_index-1][grid_y_index-1] = np.array([objectness, cls, x_offset, y_offset, w_ratio, h_ratio])
-
-        label = torch.from_numpy(np_label)
-        targets.append(label)
-
-    return torch.stack(imgs,0), torch.stack(targets, 0), sizes
-
-
 # visdom function
 
 def create_vis_plot(viz, _xlabel, _ylabel, _title, _legend):