solve the problem of sudden increases in losses for fasterrcnn model

model_zoo/official/cv/faster_rcnn/eval.py

@@ -19,6 +19,7 @@ import argparse
 import time
 import numpy as np
 from pycocotools.coco import COCO
+import mindspore.common.dtype as mstype
 from mindspore import context
 from mindspore.train.serialization import load_checkpoint, load_param_into_net
 from mindspore.common import set_seed, Parameter
@@ -51,7 +52,11 @@ def fasterrcnn_eval(dataset_path, ckpt_path, ann_file):
             tensor = value.asnumpy().astype(np.float32)
             param_dict[key] = Parameter(tensor, key)
     load_param_into_net(net, param_dict)
+
     net.set_train(False)
+    device_type = "Ascend" if context.get_context("device_target") == "Ascend" else "Others"
+    if device_type == "Ascend":
+        net.to_float(mstype.float16)
 
     eval_iter = 0
     total = ds.get_dataset_size()

model_zoo/official/cv/faster_rcnn/src/FasterRcnn/faster_rcnn_r50.py

@@ -16,6 +16,7 @@
 
 import numpy as np
 import mindspore.nn as nn
+from mindspore import context
 from mindspore.ops import operations as P
 from mindspore.common.tensor import Tensor
 import mindspore.common.dtype as mstype
@@ -144,6 +145,7 @@ class Faster_Rcnn_Resnet50(nn.Cell):
 
         # Init tensor
         self.init_tensor(config)
+        self.device_type = "Ascend" if context.get_context("device_target") == "Ascend" else "Others"
 
     def roi_init(self, config):
         self.roi_align = SingleRoIExtractor(config,
@@ -267,6 +269,8 @@ class Faster_Rcnn_Resnet50(nn.Cell):
                 bboxes_all = self.concat(bboxes_tuple)
             else:
                 bboxes_all = bboxes_tuple[0]
+            if self.device_type == "Ascend":
+                bboxes_all = self.cast(bboxes_all, mstype.float16)
             rois = self.concat_1((self.roi_align_index_test_tensor, bboxes_all))
 
         rois = self.cast(rois, mstype.float32)

model_zoo/official/cv/faster_rcnn/src/FasterRcnn/rcnn.py

@@ -40,7 +40,7 @@ class DenseNoTranpose(nn.Cell):
         if self.device_type == "Ascend":
             x = self.cast(x, mstype.float16)
             weight = self.cast(self.weight, mstype.float16)
-            output = self.bias_add(self.cast(self.matmul(x, weight), mstype.float32), self.bias)
+            output = self.bias_add(self.matmul(x, weight), self.bias)
         else:
             output = self.bias_add(self.matmul(x, self.weight), self.bias)
         return output

model_zoo/official/cv/faster_rcnn/src/FasterRcnn/rpn.py

@@ -16,7 +16,7 @@
 import numpy as np
 import mindspore.nn as nn
 import mindspore.common.dtype as mstype
-from mindspore import Tensor
+from mindspore import context, Tensor
 from mindspore.ops import operations as P
 from mindspore.ops import functional as F
 from mindspore.common.initializer import initializer
@@ -102,6 +102,7 @@ class RPN(nn.Cell):
         cfg_rpn = config
         self.dtype = np.float32
         self.ms_type = mstype.float32
+        self.device_type = "Ascend" if context.get_context("device_target") == "Ascend" else "Others"
         self.num_bboxes = cfg_rpn.num_bboxes
         self.slice_index = ()
         self.feature_anchor_shape = ()
@@ -180,9 +181,12 @@ class RPN(nn.Cell):
         bias_reg = initializer(0, shape=shp_bias_reg, dtype=self.ms_type).to_tensor()
 
         for i in range(num_layers):
-            rpn_layer.append(RpnRegClsBlock(in_channels, feat_channels, num_anchors, cls_out_channels, \
+            rpn_reg_cls_block = RpnRegClsBlock(in_channels, feat_channels, num_anchors, cls_out_channels, \
                                             weight_conv, bias_conv, weight_cls, \
-                                            bias_cls, weight_reg, bias_reg))
+                                            bias_cls, weight_reg, bias_reg)
+            if self.device_type == "Ascend":
+                rpn_reg_cls_block.to_float(mstype.float16)
+            rpn_layer.append(rpn_reg_cls_block)
 
         for i in range(1, num_layers):
             rpn_layer[i].rpn_conv.weight = rpn_layer[0].rpn_conv.weight
@@ -250,6 +254,7 @@ class RPN(nn.Cell):
                                                                                            mstype.bool_),
                                                                                  anchor_using_list, gt_valids_i)
 
+                bbox_target = self.cast(bbox_target, self.ms_type)
                 bbox_weight = self.cast(bbox_weight, self.ms_type)
                 label = self.cast(label, self.ms_type)
                 label_weight = self.cast(label_weight, self.ms_type)
@@ -286,8 +291,8 @@ class RPN(nn.Cell):
                 label_ = F.stop_gradient(label_with_batchsize)
                 label_weight_ = F.stop_gradient(label_weight_with_batchsize)
 
-                cls_score_i = rpn_cls_score[i]
-                reg_score_i = rpn_bbox_pred[i]
+                cls_score_i = self.cast(rpn_cls_score[i], self.ms_type)
+                reg_score_i = self.cast(rpn_bbox_pred[i], self.ms_type)
 
                 loss_cls = self.loss_cls(cls_score_i, label_)
                 loss_cls_item = loss_cls * label_weight_

model_zoo/official/cv/faster_rcnn/train.py

@@ -152,6 +152,10 @@ if __name__ == '__main__':
             param_dict[key] = Parameter(tensor, key)
         load_param_into_net(net, param_dict)
 
+    device_type = "Ascend" if context.get_context("device_target") == "Ascend" else "Others"
+    if device_type == "Ascend":
+        net.to_float(mstype.float16)
+
     loss = LossNet()
     lr = Tensor(dynamic_lr(config, dataset_size), mstype.float32)