fix the bug

paddlex/cv/nets/detection/loss/yolo_loss.py

@@ -19,6 +19,7 @@ from __future__ import print_function
 from paddle import fluid
 from .iou_loss import IouLoss
 from .iou_aware_loss import IouAwareLoss
+import paddlex
 
     
 class YOLOv3Loss(object):
@@ -157,7 +158,7 @@ class YOLOv3Loss(object):
 
             loss_cls = fluid.layers.sigmoid_cross_entropy_with_logits(cls, tcls)
             loss_cls = fluid.layers.elementwise_mul(loss_cls, tobj, axis=0)
-            loss_cls = fluid.layers.reduce_sum(loss_cls)
+            loss_cls = fluid.layers.reduce_sum(loss_cls, dim=[1, 2, 3, 4])
 
             loss_xys.append(fluid.layers.reduce_mean(loss_x + loss_y))
             loss_whs.append(fluid.layers.reduce_mean(loss_w + loss_h))
@@ -270,6 +271,7 @@ class YOLOv3Loss(object):
 
         # 1. get pred bbox, which is same with YOLOv3 infer mode, use yolo_box here
         # NOTE: img_size is set as 1.0 to get noramlized pred bbox
+        batch_size = int(batch_size / paddlex.env_info['num'])
         bbox, prob = fluid.layers.yolo_box(
             x=output,
             img_size=fluid.layers.ones(

paddlex/cv/nets/detection/yolo_v3.py

@@ -205,7 +205,7 @@ class YOLOv3:
         p = fluid.layers.expand_as(gamma, input)
         input_shape_tmp = fluid.layers.cast(input_shape, dtype="int64")
         random_matrix = fluid.layers.uniform_random(
-            input_shape_tmp, dtype='float32', min=0.0, max=1.0, seed=1)
+            input_shape_tmp, dtype='float32', min=0.0, max=1.0)
         one_zero_m = fluid.layers.less_than(random_matrix, p)
         one_zero_m.stop_gradient = True
         one_zero_m = fluid.layers.cast(one_zero_m, dtype="float32")

paddlex/cv/transforms/det_transforms.py

@@ -1299,7 +1299,6 @@ class BatchRandomShape(DetTransform):
             im = np.swapaxes(im, 1, 0)
             data_list[0] = im
             batch_data[data_id] = tuple(data_list)
-            np.save('im.npy', im)
         return batch_data