fix pse_ctw1500 training error (#806)

* fix pse_ctw1500 training error

* fix pse_ctw1500 training error

* fix pse_ctw1500 training error

* ci fix

Author: alien-0119

mindocr/data/transforms/det_transforms.py

@@ -13,6 +13,8 @@
 import pyclipper
 from shapely.geometry import Polygon, box
 
+from ...utils.misc import is_uneven_nested_list
+
 __all__ = [
     "DetLabelEncode",
     "BorderMap",
@@ -595,9 +597,14 @@ def _shrink(self, text_polys, rate, max_shr=20):
             if not shrinked_bbox:
                 shrinked_text_polys.append(bbox)
                 continue
-
-            shrinked_bbox = np.array(shrinked_bbox)[0]
-            shrinked_bbox = np.array(shrinked_bbox)
+            if is_uneven_nested_list(shrinked_bbox):
+                shrinked_bbox = np.array(shrinked_bbox, dtype=object)[0]
+            else:
+                shrinked_bbox = np.array(shrinked_bbox)[0]
+            if is_uneven_nested_list(shrinked_bbox):
+                shrinked_bbox = np.array(shrinked_bbox, dtype=object)
+            else:
+                shrinked_bbox = np.array(shrinked_bbox)
             if shrinked_bbox.shape[0] <= 2:
                 shrinked_text_polys.append(bbox)
                 continue

mindocr/losses/det_loss.py

@@ -221,7 +221,7 @@ def __init__(self, alpha=0.7, ohem_ratio=3):
         self.zeros_like = ops.ZerosLike()
         self.add = ops.Add()
         self.gather = ops.Gather()
-        self.upsample = nn.ResizeBilinear()
+        self.upsample = ops.interpolate
 
     def ohem_batch(self, scores, gt_texts, training_masks):
         """
@@ -334,8 +334,10 @@ def construct(self, model_predict, gt_texts, gt_kernels, training_masks):
             Tensor: The computed loss value.
         """
         batch_size = model_predict.shape[0]
-        model_predict = self.upsample(model_predict, scale_factor=4)
-        h, w = model_predict.shape[2:]
+        scale_factor = 4
+        origin_h, origin_w = model_predict.shape[2:]
+        h, w = origin_h * scale_factor, origin_w * scale_factor
+        model_predict = self.upsample(model_predict, size=(h, w), mode="bilinear")
         texts = self.slice(model_predict, (0, 0, 0, 0), (batch_size, 1, h, w))
         texts = self.reshape(texts, (batch_size, h, w))
         selected_masks_text = self.ohem_batch(texts, gt_texts, training_masks)

mindocr/postprocess/det_db_postprocess.py

@@ -7,6 +7,7 @@
 from mindspore import Tensor
 
 from ..data.transforms.det_transforms import expand_poly
+from ..utils.misc import is_uneven_nested_list
 from .det_base_postprocess import DetBasePostprocess
 
 __all__ = ["DBPostprocess"]
@@ -111,7 +112,7 @@ def _extract_preds(self, pred: np.ndarray, bitmap: np.ndarray):
 
             poly = Polygon(points)
             poly_list = expand_poly(points, distance=poly.area * self._expand_ratio / poly.length)
-            if self._is_uneven_nested_list(poly_list):
+            if is_uneven_nested_list(poly_list):
                 poly = np.array(poly_list, dtype=object)
             else:
                 poly = np.array(poly_list)
@@ -138,18 +139,6 @@ def _extract_preds(self, pred: np.ndarray, bitmap: np.ndarray):
             return polys, scores
         return np.array(polys), np.array(scores).astype(np.float32)
 
-    def _is_uneven_nested_list(self, arr_list):
-        if not isinstance(arr_list, list):
-            return False
-
-        first_length = len(arr_list[0]) if isinstance(arr_list[0], list) else None
-
-        for sublist in arr_list:
-            if not isinstance(sublist, list) or len(sublist) != first_length:
-                return True
-
-        return False
-
     @staticmethod
     def _fit_box(contour):
         """

mindocr/utils/misc.py

@@ -73,3 +73,16 @@ def is_ms_version_2():
     make compatibilities in differenct Mindspore version
     """
     return version.parse(ms.__version__) >= version.parse("2.0.0rc")
+
+
+def is_uneven_nested_list(arr_list):
+    if not isinstance(arr_list, list):
+        return False
+
+    first_length = len(arr_list[0]) if isinstance(arr_list[0], list) else None
+
+    for sublist in arr_list:
+        if not isinstance(sublist, list) or len(sublist) != first_length:
+            return True
+
+    return False