Merge pull request #255 from yuedongli1/master

fix yolox yaml

configs/yolox/hyp.scratch.yaml

@@ -56,7 +56,6 @@ data:
       ],
       [
         { func_name: letterbox, scaleup: False },
-        { func_name: resample_segments },
         { func_name: hsv_augment, prob: 1.0, hgain: 0.015, sgain: 0.7, vgain: 0.4 },
         { func_name: fliplr, prob: 0.5 },
         { func_name: label_norm, xyxy2xywh_: True },

configs/yolox/yolox-nano.yaml

@@ -64,10 +64,13 @@ data:
     stage_epochs: [ 285, 15 ],
     trans_list: [
       [
-        { func_name: mosaic, prob: 0.5, mosaic9_prob: 0.0, translate: 0.1, scale: 0.5, degrees: 10.0, shear: 2.0 },
+        { func_name: mosaic, prob: 0.5, post_transform: [
+          { func_name: resample_segments },
+          { func_name: random_perspective, degrees: 10.0, translate: 0.1, scale: 0.5, shear: 2.0 },] },
+        { func_name: letterbox, scaleup: False,  },
         { func_name: hsv_augment, prob: 1.0, hgain: 0.015, sgain: 0.7, vgain: 0.4 },
-        { func_name: label_norm, xyxy2xywh_: True },
         { func_name: fliplr, prob: 0.5 },
+        { func_name: label_norm, xyxy2xywh_: True },
         { func_name: label_pad, padding_size: 160, padding_value: -1 },
         { func_name: image_norm, scale: 255. },
         { func_name: image_transpose, bgr2rgb: True, hwc2chw: True },

configs/yolox/yolox-tiny.yaml

@@ -24,10 +24,12 @@ data:
     stage_epochs: [ 285, 15 ],
     trans_list: [
       [
-        { func_name: mosaic, prob: 1.0, mosaic9_prob: 0.0, translate: 0.1, scale: 0.5, degrees: 10.0, shear: 2.0 },
+        { func_name: mosaic, prob: 1.0 },
+        { func_name: resample_segments },
+        { func_name: random_perspective, prob: 1.0, degrees: 10.0, translate: 0.1, scale: 0.5, shear: 2.0 },
         { func_name: hsv_augment, prob: 1.0, hgain: 0.015, sgain: 0.7, vgain: 0.4 },
-        { func_name: label_norm, xyxy2xywh_: True },
         { func_name: fliplr, prob: 0.5 },
+        { func_name: label_norm, xyxy2xywh_: True },
         { func_name: label_pad, padding_size: 160, padding_value: -1 },
         { func_name: image_norm, scale: 255. },
         { func_name: image_transpose, bgr2rgb: True, hwc2chw: True },

mindyolo/data/dataset.py

@@ -342,6 +342,7 @@ def mosaic(
         self,
         sample,
         mosaic9_prob=0.0,
+        post_transform=None,
     ):
         segment_format = sample['segment_format']
         bbox_format = sample['bbox_format']
@@ -350,9 +351,17 @@ def mosaic(
 
         mosaic9_prob = min(1.0, max(mosaic9_prob, 0.0))
         if random.random() < (1 - mosaic9_prob):
-            return self._mosaic4(sample)
+            sample = self._mosaic4(sample)
         else:
-            return self._mosaic9(sample)
+            sample = self._mosaic9(sample)
+
+        if post_transform:
+            for _i, ori_trans in enumerate(post_transform):
+                _trans = ori_trans.copy()
+                func_name, prob = _trans.pop("func_name"), _trans.pop("prob", 1.0)
+                sample = getattr(self, func_name)(sample, **_trans)
+
+        return sample
 
     def _mosaic4(self, sample):
         # loads images in a 4-mosaic
@@ -936,7 +945,7 @@ def fliplr(self, sample):
         sample['bboxes'] = bboxes
 
         # flip seg
-        if 'segments' in sample:
+        if self.return_segments:
             segment_format, segments = sample['segment_format'], sample['segments']
             assert segment_format == 'polygon', \
                 f'FlipLR: The segment format should be polygon, but got {segment_format}'
@@ -959,16 +968,21 @@ def letterbox(self, sample, new_shape=None, xywhn2xyxy_=True, scaleup=False, onl
 
         if not new_shape:
             new_shape = self.img_size
+
+        if isinstance(new_shape, int):
+            new_shape = (new_shape, new_shape)
+
         image = sample['img']
+        shape = image.shape[:2]  # current shape [height, width]
+        if shape == new_shape:
+            return sample
+
         bboxes = sample['bboxes']
         ori_shape = sample['ori_shape']
 
-        shape = image.shape[:2]  # current shape [height, width]
         h, w = shape[:]
         h0, w0 = ori_shape
         hw_scale = np.array([h / h0, w / w0])
-        if isinstance(new_shape, int):
-            new_shape = (new_shape, new_shape)
 
         # Scale ratio (new / old)
         r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])