Merge pull request #2072 from heavengate/fix_yolo_param_pick

[cherry-pick] refine yolov3 param name

PaddleCV/yolov3/README.md

@@ -142,6 +142,8 @@ Evalutaion result is shown as below:
 | 416x416 | 36.5 | 58.2 | 39.1 |
 | 320x320 | 34.1 | 55.4 | 36.3 |
 
+- **NOTE:** evaluations based on `pycocotools` evaluator, predict bounding boxes with `score < 0.05` were not filtered out. Some frameworks which filtered out predict bounding boxes with `score < 0.05` will cause a drop in accuracy.
+
 ## Inference and Visualization
 
 Inference is used to get prediction score or image features based on trained models. `infer.py`  is the main executor for inference, one can start infer step by:

PaddleCV/yolov3/README_cn.md

@@ -142,6 +142,8 @@ Train Loss
 | 416x416 | 36.5 | 58.2 | 39.1 |
 | 320x320 | 34.1 | 55.4 | 36.3 |
 
+- **注意: **评估结果基于`pycocotools`评估器，没有滤除`score < 0.05`的预测框，其他框架有此滤除操作会导致精度下降。
+
 
 
 ## 模型推断及可视化

PaddleCV/yolov3/image_utils.py

@@ -51,7 +51,14 @@ def random_distort(img):
     return img
 
 
-def random_crop(img, boxes, labels, scores, scales=[0.3, 1.0], max_ratio=2.0, constraints=None, max_trial=50):
+def random_crop(img, 
+                boxes, 
+                labels, 
+                scores, 
+                scales=[0.3, 1.0], 
+                max_ratio=2.0, 
+                constraints=None, 
+                max_trial=50):
     if len(boxes) == 0:
         return img, boxes
 
@@ -90,10 +97,12 @@ def random_crop(img, boxes, labels, scores, scales=[0.3, 1.0], max_ratio=2.0, co
 
     while crops:
         crop = crops.pop(np.random.randint(0, len(crops)))
-        crop_boxes, crop_labels, crop_scores, box_num = box_utils.box_crop(boxes, labels, scores, crop, (w, h))
+        crop_boxes, crop_labels, crop_scores, box_num = \
+            box_utils.box_crop(boxes, labels, scores, crop, (w, h))
         if box_num < 1:
             continue
-        img = img.crop((crop[0], crop[1], crop[0] + crop[2], crop[1] + crop[3])).resize(img.size, Image.LANCZOS)
+        img = img.crop((crop[0], crop[1], crop[0] + crop[2], 
+                        crop[1] + crop[3])).resize(img.size, Image.LANCZOS)
         img = np.asarray(img)
         return img, crop_boxes, crop_labels, crop_scores
     img = np.asarray(img)
@@ -118,10 +127,16 @@ def random_interp(img, size, interp=None):
     h, w, _ = img.shape
     im_scale_x = size / float(w)
     im_scale_y = size / float(h)
-    img = cv2.resize(img, None, None, fx=im_scale_x, fy=im_scale_y, interpolation=interp)
+    img = cv2.resize(img, None, None, fx=im_scale_x, fy=im_scale_y, 
+                     interpolation=interp)
     return img
 
-def random_expand(img, gtboxes, max_ratio=4., fill=None, keep_ratio=True, thresh=0.5):
+def random_expand(img, 
+                  gtboxes, 
+                  max_ratio=4., 
+                  fill=None, 
+                  keep_ratio=True, 
+                  thresh=0.5):
     if random.random() > thresh:
         return img, gtboxes
 
@@ -153,13 +168,21 @@ def random_expand(img, gtboxes, max_ratio=4., fill=None, keep_ratio=True, thresh
     return out_img.astype('uint8'), gtboxes
 
 def shuffle_gtbox(gtbox, gtlabel, gtscore):
-    gt = np.concatenate([gtbox, gtlabel[:, np.newaxis], gtscore[:, np.newaxis]], axis=1)
+    gt = np.concatenate([gtbox, gtlabel[:, np.newaxis], 
+                         gtscore[:, np.newaxis]], axis=1)
     idx = np.arange(gt.shape[0])
     np.random.shuffle(idx)
     gt = gt[idx, :]
     return gt[:, :4], gt[:, 4], gt[:, 5]
 
-def image_mixup(img1, gtboxes1, gtlabels1, gtscores1, img2, gtboxes2, gtlabels2, gtscores2):
+def image_mixup(img1, 
+                gtboxes1, 
+                gtlabels1, 
+                gtscores1, 
+                img2, 
+                gtboxes2, 
+                gtlabels2, 
+                gtscores2):
     factor = np.random.beta(1.5, 1.5)
     factor = max(0.0, min(1.0, factor))
     if factor >= 1.0:
@@ -173,7 +196,8 @@ def image_mixup(img1, gtboxes1, gtlabels1, gtscores1, img2, gtboxes2, gtlabels2,
     w = max(img1.shape[1], img2.shape[1])
     img = np.zeros((h, w, img1.shape[2]), 'float32')
     img[:img1.shape[0], :img1.shape[1], :] = img1.astype('float32') * factor
-    img[:img2.shape[0], :img2.shape[1], :] += img2.astype('float32') * (1.0 - factor)
+    img[:img2.shape[0], :img2.shape[1], :] += \
+            img2.astype('float32') * (1.0 - factor)
     gtboxes = np.zeros_like(gtboxes1)
     gtlabels = np.zeros_like(gtlabels1)
     gtscores = np.zeros_like(gtscores1)
@@ -208,7 +232,8 @@ def image_mixup(img1, gtboxes1, gtlabels1, gtscores1, img2, gtboxes2, gtlabels2,
 def image_augment(img, gtboxes, gtlabels, gtscores,  size, means=None):
     img = random_distort(img)
     img, gtboxes = random_expand(img, gtboxes, fill=means)
-    img, gtboxes, gtlabels, gtscores = random_crop(img, gtboxes, gtlabels, gtscores)
+    img, gtboxes, gtlabels, gtscores = \
+            random_crop(img, gtboxes, gtlabels, gtscores)
     img = random_interp(img, size)
     img, gtboxes = random_flip(img, gtboxes)
     gtboxes, gtlabels, gtscores = shuffle_gtbox(gtboxes, gtlabels, gtscores)

PaddleCV/yolov3/models/darknet.py

@@ -55,7 +55,13 @@ def conv_bn_layer(input,
         out = fluid.layers.leaky_relu(x=out, alpha=0.1)
     return out
 
-def downsample(input, ch_out, filter_size=3, stride=2, padding=1, is_test=True, name=None):
+def downsample(input, 
+               ch_out, 
+               filter_size=3, 
+               stride=2, 
+               padding=1, 
+               is_test=True, 
+               name=None):
     return conv_bn_layer(input, 
             ch_out=ch_out, 
             filter_size=filter_size, 
@@ -65,15 +71,19 @@ def downsample(input, ch_out, filter_size=3, stride=2, padding=1, is_test=True,
             name=name)
 
 def basicblock(input, ch_out, is_test=True, name=None):
-    conv1 = conv_bn_layer(input, ch_out, 1, 1, 0, is_test=is_test, name=name+".0")
-    conv2 = conv_bn_layer(conv1, ch_out*2, 3, 1, 1, is_test=is_test, name=name+".1")
+    conv1 = conv_bn_layer(input, ch_out, 1, 1, 0, 
+                          is_test=is_test, name=name+".0")
+    conv2 = conv_bn_layer(conv1, ch_out*2, 3, 1, 1, 
+                          is_test=is_test, name=name+".1")
     out = fluid.layers.elementwise_add(x=input, y=conv2, act=None)
     return out
 
 def layer_warp(block_func, input, ch_out, count, is_test=True, name=None):
-    res_out = block_func(input, ch_out, is_test=is_test, name='{}.0'.format(name))
+    res_out = block_func(input, ch_out, is_test=is_test, 
+                         name='{}.0'.format(name))
     for j in range(1, count):
-        res_out = block_func(res_out, ch_out, is_test=is_test, name='{}.{}'.format(name, j))
+        res_out = block_func(res_out, ch_out, is_test=is_test, 
+                             name='{}.{}'.format(name, j))
     return res_out
 
 DarkNet_cfg = {
@@ -83,14 +93,21 @@ DarkNet_cfg = {
 def add_DarkNet53_conv_body(body_input, is_test=True):
     stages, block_func = DarkNet_cfg[53]
     stages = stages[0:5]
-    conv1 = conv_bn_layer(
-            body_input, ch_out=32, filter_size=3, stride=1, padding=1, is_test=is_test, name="yolo_input")
-    downsample_ = downsample(conv1, ch_out=conv1.shape[1]*2, is_test=is_test, name="yolo_input.downsample")
+    conv1 = conv_bn_layer(body_input, ch_out=32, filter_size=3, 
+                          stride=1, padding=1, is_test=is_test, 
+                          name="yolo_input")
+    downsample_ = downsample(conv1, ch_out=conv1.shape[1]*2, 
+                             is_test=is_test, 
+                             name="yolo_input.downsample")
     blocks = []
     for i, stage in enumerate(stages):
-        block = layer_warp(block_func, downsample_, 32 *(2**i), stage, is_test=is_test, name="stage.{}".format(i))
+        block = layer_warp(block_func, downsample_, 32 *(2**i), 
+                           stage, is_test=is_test, 
+                           name="stage.{}".format(i))
         blocks.append(block)
         if i < len(stages) - 1: # do not downsaple in the last stage
-            downsample_ = downsample(block, ch_out=block.shape[1]*2, is_test=is_test, name="stage.{}.downsample".format(i))
+            downsample_ = downsample(block, ch_out=block.shape[1]*2, 
+                                     is_test=is_test, 
+                                     name="stage.{}.downsample".format(i))
     return blocks[-1:-4:-1]
 

PaddleCV/yolov3/models/yolov3.py

@@ -27,13 +27,22 @@ from .darknet import add_DarkNet53_conv_body
 from .darknet import conv_bn_layer
 
 def yolo_detection_block(input, channel, is_test=True, name=None):
-    assert channel % 2 == 0, "channel {} cannot be divided by 2".format(channel)
+    assert channel % 2 == 0, \
+            "channel {} cannot be divided by 2".format(channel)
     conv = input
     for j in range(2):
-        conv = conv_bn_layer(conv, channel, filter_size=1, stride=1, padding=0, is_test=is_test, name='{}.{}.0'.format(name, j))
-        conv = conv_bn_layer(conv, channel*2, filter_size=3, stride=1, padding=1, is_test=is_test, name='{}.{}.1'.format(name, j))
-    route = conv_bn_layer(conv, channel, filter_size=1, stride=1, padding=0, is_test=is_test, name='{}.2'.format(name))
-    tip = conv_bn_layer(route,channel*2, filter_size=3, stride=1, padding=1, is_test=is_test, name='{}.tip'.format(name))
+        conv = conv_bn_layer(conv, channel, filter_size=1, 
+                             stride=1, padding=0, is_test=is_test, 
+                             name='{}.{}.0'.format(name, j))
+        conv = conv_bn_layer(conv, channel*2, filter_size=3, 
+                             stride=1, padding=1, is_test=is_test, 
+                             name='{}.{}.1'.format(name, j))
+    route = conv_bn_layer(conv, channel, filter_size=1, stride=1, 
+                          padding=0, is_test=is_test, 
+                          name='{}.2'.format(name))
+    tip = conv_bn_layer(route,channel*2, filter_size=3, stride=1, 
+                        padding=1, is_test=is_test, 
+                        name='{}.tip'.format(name))
     return route, tip
 
 def upsample(input, scale=2,name=None):
@@ -68,11 +77,15 @@ class YOLOv3(object):
         if self.is_train:
             self.py_reader = fluid.layers.py_reader(
                 capacity=64,
-                shapes = [[-1] + self.image_shape, [-1, cfg.max_box_num, 4], [-1, cfg.max_box_num], [-1, cfg.max_box_num]],
+                shapes = [[-1] + self.image_shape, 
+                          [-1, cfg.max_box_num, 4], 
+                          [-1, cfg.max_box_num], 
+                          [-1, cfg.max_box_num]],
                 lod_levels=[0, 0, 0, 0],
                 dtypes=['float32'] * 2 + ['int32'] + ['float32'],
                 use_double_buffer=True)
-            self.image, self.gtbox, self.gtlabel, self.gtscore = fluid.layers.read_file(self.py_reader)
+            self.image, self.gtbox, self.gtlabel, self.gtscore = \
+                    fluid.layers.read_file(self.py_reader)
         else:
             self.image = fluid.layers.data(
                     name='image', shape=self.image_shape, dtype='float32'
@@ -139,9 +152,9 @@ class YOLOv3(object):
             if self.is_train:
                 loss = fluid.layers.yolov3_loss(
                         x=out,
-                        gtbox=self.gtbox,
-                        gtlabel=self.gtlabel,
-                        gtscore=self.gtscore,
+                        gt_box=self.gtbox,
+                        gt_label=self.gtlabel,
+                        gt_score=self.gtscore,
                         anchors=cfg.anchors,
                         anchor_mask=anchor_mask,
                         class_num=cfg.class_num,

PaddleCV/yolov3/reader.py

@@ -53,13 +53,17 @@ class DataSetReader(object):
                 cfg.dataset))
 
         if mode == 'train':
-            cfg.train_file_list = os.path.join(cfg.data_dir, cfg.train_file_list)
-            cfg.train_data_dir = os.path.join(cfg.data_dir, cfg.train_data_dir)
+            cfg.train_file_list = os.path.join(cfg.data_dir, 
+                                               cfg.train_file_list)
+            cfg.train_data_dir = os.path.join(cfg.data_dir, 
+                                              cfg.train_data_dir)
             self.COCO = COCO(cfg.train_file_list)
             self.img_dir = cfg.train_data_dir
         elif mode == 'test' or mode == 'infer':
-            cfg.val_file_list = os.path.join(cfg.data_dir, cfg.val_file_list)
-            cfg.val_data_dir = os.path.join(cfg.data_dir, cfg.val_data_dir)
+            cfg.val_file_list = os.path.join(cfg.data_dir, 
+                                             cfg.val_file_list)
+            cfg.val_data_dir = os.path.join(cfg.data_dir, 
+                                            cfg.val_data_dir)
             self.COCO = COCO(cfg.val_file_list)
             self.img_dir = cfg.val_data_dir
 
@@ -88,7 +92,8 @@ class DataSetReader(object):
     def _parse_gt_annotations(self, img):
         img_height = img['height']
         img_width = img['width']
-        anno = self.COCO.loadAnns(self.COCO.getAnnIds(imgIds=img['id'], iscrowd=None))
+        anno = self.COCO.loadAnns(
+                self.COCO.getAnnIds(imgIds=img['id'], iscrowd=None))
         gt_index = 0
         for target in anno:
             if target['area'] < cfg.gt_min_area:
@@ -96,13 +101,15 @@ class DataSetReader(object):
             if 'ignore' in target and target['ignore']:
                 continue
 
-            box = box_utils.coco_anno_box_to_center_relative(target['bbox'], img_height, img_width)
+            box = box_utils.coco_anno_box_to_center_relative(
+                            target['bbox'], img_height, img_width)
             if box[2] <= 0 and box[3] <= 0:
                 continue
 
             img['gt_id'][gt_index] = np.int32(target['id'])
             img['gt_boxes'][gt_index] = box
-            img['gt_labels'][gt_index] = self.category_to_id_map[target['category_id']]
+            img['gt_labels'][gt_index] = \
+                self.category_to_id_map[target['category_id']]
             gt_index += 1
             if gt_index >= cfg.max_box_num:
                 break
@@ -136,10 +143,18 @@ class DataSetReader(object):
         else:
             return self._parse_images(is_train=(mode=='train'))
 
-    def get_reader(self, mode, size=416, batch_size=None, shuffle=False, mixup_iter=0, random_sizes=[], image=None):
+    def get_reader(self, 
+                   mode, 
+                   size=416, 
+                   batch_size=None, 
+                   shuffle=False, 
+                   mixup_iter=0, 
+                   random_sizes=[], 
+                   image=None):
         assert mode in ['train', 'test', 'infer'], "Unknow mode type!"
         if mode != 'infer':
-            assert batch_size is not None, "batch size connot be None in mode {}".format(mode)
+            assert batch_size is not None, \
+                "batch size connot be None in mode {}".format(mode)
             self._parse_dataset_dir(mode)
             self._parse_dataset_catagory()
 
@@ -151,7 +166,9 @@ class DataSetReader(object):
             h, w, _ = im.shape
             im_scale_x = size / float(w)
             im_scale_y = size / float(h)
-            out_img = cv2.resize(im, None, None, fx=im_scale_x, fy=im_scale_y, interpolation=cv2.INTER_CUBIC)
+            out_img = cv2.resize(im, None, None, 
+                                 fx=im_scale_x, fy=im_scale_y, 
+                                 interpolation=cv2.INTER_CUBIC)
             mean = np.array(mean).reshape((1, 1, -1))
             std = np.array(std).reshape((1, 1, -1))
             out_img = (out_img / 255.0 - mean) / std
@@ -173,11 +190,14 @@ class DataSetReader(object):
                 mixup_gt_boxes = np.array(mixup_img['gt_boxes']).copy()
                 mixup_gt_labels = np.array(mixup_img['gt_labels']).copy()
                 mixup_gt_scores = np.ones_like(mixup_gt_labels)
-                im, gt_boxes, gt_labels, gt_scores = image_utils.image_mixup(im, gt_boxes, \
-                        gt_labels, gt_scores, mixup_im, mixup_gt_boxes, mixup_gt_labels, \
-                        mixup_gt_scores)
+                im, gt_boxes, gt_labels, gt_scores = \
+                    image_utils.image_mixup(im, gt_boxes, gt_labels, 
+                                            gt_scores, mixup_im, mixup_gt_boxes, 
+                                            mixup_gt_labels, mixup_gt_scores)
 
-            im, gt_boxes, gt_labels, gt_scores = image_utils.image_augment(im, gt_boxes, gt_labels, gt_scores, size, mean)
+            im, gt_boxes, gt_labels, gt_scores = \
+                image_utils.image_augment(im, gt_boxes, gt_labels, 
+                                          gt_scores, size, mean)
 
             mean = np.array(mean).reshape((1, 1, -1))
             std = np.array(std).reshape((1, 1, -1))
@@ -214,7 +234,9 @@ class DataSetReader(object):
                     read_cnt += 1
                     if read_cnt % len(imgs) == 0 and shuffle:
                         np.random.shuffle(imgs)
-                    im, gt_boxes, gt_labels, gt_scores = img_reader_with_augment(img, img_size, cfg.pixel_means, cfg.pixel_stds, mixup_img)
+                    im, gt_boxes, gt_labels, gt_scores = \
+                        img_reader_with_augment(img, img_size, cfg.pixel_means, 
+                                                cfg.pixel_stds, mixup_img)
                     batch_out.append([im, gt_boxes, gt_labels, gt_scores])
 
                     if len(batch_out) == batch_size:
@@ -227,7 +249,9 @@ class DataSetReader(object):
                 imgs = self._parse_images_by_mode(mode)
                 batch_out = []
                 for img in imgs:
-                    im, im_id, im_shape = img_reader(img, size, cfg.pixel_means, cfg.pixel_stds)
+                    im, im_id, im_shape = img_reader(img, size, 
+                                                     cfg.pixel_means, 
+                                                     cfg.pixel_stds)
                     batch_out.append((im, im_id, im_shape))
                     if len(batch_out) == batch_size:
                         yield batch_out
@@ -238,7 +262,9 @@ class DataSetReader(object):
                 img = {}
                 img['image'] = image
                 img['id'] = 0
-                im, im_id, im_shape = img_reader(img, size, cfg.pixel_means, cfg.pixel_stds)
+                im, im_id, im_shape = img_reader(img, size, 
+                                                 cfg.pixel_means, 
+                                                 cfg.pixel_stds)
                 batch_out = [(im, im_id, im_shape)]
                 yield batch_out
 
@@ -256,7 +282,8 @@ def train(size=416,
           num_workers=8,
           max_queue=32,
           use_multiprocessing=True):
-    generator = dsr.get_reader('train', size, batch_size, shuffle, int(mixup_iter/num_workers), random_sizes)
+    generator = dsr.get_reader('train', size, batch_size, shuffle, 
+                               int(mixup_iter/num_workers), random_sizes)
 
     if not use_multiprocessing:
         return generator

PaddleCV/yolov3/train.py

@@ -90,7 +90,13 @@ def train():
 
     total_iter = cfg.max_iter - cfg.start_iter
     mixup_iter = total_iter - cfg.no_mixup_iter
-    train_reader = reader.train(input_size, batch_size=cfg.batch_size, shuffle=True, total_iter=total_iter*devices_num, mixup_iter=mixup_iter*devices_num, random_sizes=random_sizes, use_multiprocessing=cfg.use_multiprocess)
+    train_reader = reader.train(input_size, 
+                                batch_size=cfg.batch_size, 
+                                shuffle=True, 
+                                total_iter=total_iter*devices_num, 
+                                mixup_iter=mixup_iter*devices_num, 
+                                random_sizes=random_sizes, 
+                                use_multiprocessing=cfg.use_multiprocess)
     py_reader = model.py_reader
     py_reader.decorate_paddle_reader(train_reader)
 
@@ -112,7 +118,8 @@ def train():
         for iter_id in range(cfg.start_iter, cfg.max_iter):
             prev_start_time = start_time
             start_time = time.time()
-            losses = exe.run(compile_program, fetch_list=[v.name for v in fetch_list])
+            losses = exe.run(compile_program, 
+                             fetch_list=[v.name for v in fetch_list])
             smoothed_loss.add_value(np.mean(np.array(losses[0])))
             snapshot_loss += np.mean(np.array(losses[0]))
             snapshot_time += start_time - prev_start_time
@@ -120,12 +127,15 @@ def train():
                           .get_tensor())
             print("Iter {:d}, lr {:.6f}, loss {:.6f}, time {:.5f}".format(
                   iter_id, lr[0],
-                smoothed_loss.get_mean_value(), start_time - prev_start_time))
+                  smoothed_loss.get_mean_value(), 
+                  start_time - prev_start_time))
             sys.stdout.flush()
             if (iter_id + 1) % cfg.snapshot_iter == 0:
                 save_model("model_iter{}".format(iter_id))
-                print("Snapshot {} saved, average loss: {}, average time: {}".format(
-                    iter_id + 1, snapshot_loss / float(cfg.snapshot_iter), 
+                print("Snapshot {} saved, average loss: {}, \
+                      average time: {}".format(
+                      iter_id + 1, 
+                      snapshot_loss / float(cfg.snapshot_iter), 
                       snapshot_time / float(cfg.snapshot_iter)))
                 snapshot_loss = 0
                 snapshot_time = 0

PaddleCV/yolov3/utility.py

@@ -119,9 +119,12 @@ def parse_args():
     add_arg('nms_posk',         int,    100,    "The number of boxes of NMS output.")
     add_arg('debug',            bool,   False,  "Debug mode")
     # SINGLE EVAL AND DRAW
-    add_arg('image_path',       str,   'image', "The image path used to inference and visualize.")
-    add_arg('image_name',       str,    None,   "The single image used to inference and visualize. None to inference all images in image_path")
-    add_arg('draw_thresh',      float,  0.5,    "Confidence score threshold to draw prediction box in image in debug mode")
+    add_arg('image_path',       str,   'image', 
+            "The image path used to inference and visualize.")
+    add_arg('image_name',       str,    None,   
+            "The single image used to inference and visualize. None to inference all images in image_path")
+    add_arg('draw_thresh',      float,  0.5,    
+            "Confidence score threshold to draw prediction box in image in debug mode")
     # yapf: enable
     args = parser.parse_args()
     file_name = sys.argv[0]