!1692 Fix topk bug for fasterrcnn

Merge pull request !1692 from meixiaowei/master

mindspore/ccsrc/pre_activate/ascend/ir_fission/topk_split.cc

@@ -35,7 +35,7 @@ tensor::TensorPtr CreateTensor(const AnfNodePtr &node) {
   // 1 create tensor
   auto shape = AnfAlgo::GetPrevNodeOutputInferShape(node, 0);
   auto last_dim = shape[shape.size() - 1];
-  std::vector<int> indices_shape = {SizeToInt(last_dim)};
+  std::vector<int> indices_shape = {SizeToInt(last_dim * 2)};
   TensorTypePtr tensor_type = std::make_shared<TensorType>(kFloat16);
   MS_EXCEPTION_IF_NULL(tensor_type);
   tensor::DeviceInfo device_info{kOpFormat_DEFAULT, tensor_type};
@@ -50,7 +50,11 @@ tensor::TensorPtr CreateTensor(const AnfNodePtr &node) {
   for (size_t i = 0; i < last_dim; ++i) {
     half_data.emplace_back(Eigen::half(static_cast<float>(i)));
   }
-  auto elem_num = last_dim * kFloat16Len;
+  for (size_t i = 0; i < last_dim; ++i) {
+    auto gap = static_cast<int>(i) - static_cast<int>(Eigen::half(static_cast<float>(i)));
+    half_data.emplace_back(Eigen::half(static_cast<float>(gap)));
+  }
+  auto elem_num = last_dim * kFloat16Len * 2;
   auto ret_code = memcpy_s(data_ptr, static_cast<size_t>(indices_tensor->data().nbytes()), half_data.data(), elem_num);
   if (ret_code != 0) {
     MS_LOG(ERROR) << "Failed to copy data into Tensor.";
@@ -108,6 +112,13 @@ const AnfNodePtr TopKSplit::Process(const FuncGraphPtr &func_graph, const AnfNod
     MS_LOG(INFO) << "The input k of topk has been converted to attr";
     return nullptr;
   }
+  auto shape = AnfAlgo::GetPrevNodeOutputInferShape(node, 0);
+  auto last_dim = shape[shape.size() - 1];
+  const size_t kMaxFloat16 = 65500;
+  if (last_dim > kMaxFloat16) {
+    MS_LOG(INFO) << "The last dim is more than 65500, switch to aicpu ops.";
+    return nullptr;
+  }
   // Copy a new node to check supported.
   std::vector<AnfNodePtr> new_inputs{NewValueNode(std::make_shared<Primitive>(kTopKOpName))};
   new_inputs.insert(new_inputs.end(), cnode->inputs().begin() + 1, cnode->inputs().end());

model_zoo/faster_rcnn/scripts/run_distribute_train.sh

@@ -59,6 +59,7 @@ do
     mkdir ./train_parallel$i
     cp ../*.py ./train_parallel$i
     cp *.sh ./train_parallel$i
+    cp -r ../src ./train_parallel$i
     cd ./train_parallel$i || exit
     echo "start training for rank $RANK_ID, device $DEVICE_ID"
     env > env.log

model_zoo/faster_rcnn/scripts/run_eval.sh

@@ -57,6 +57,7 @@ fi
 mkdir ./eval
 cp ../*.py ./eval
 cp *.sh ./eval
+cp -r ../src ./eval
 cd ./eval || exit
 env > env.log
 echo "start eval for device $DEVICE_ID"

model_zoo/faster_rcnn/scripts/run_standalone_train.sh

@@ -49,6 +49,7 @@ fi
 mkdir ./train
 cp ../*.py ./train
 cp *.sh ./train
+cp -r ../src ./train
 cd ./train || exit
 echo "start training for device $DEVICE_ID"
 env > env.log

model_zoo/faster_rcnn/src/config.py

@@ -134,7 +134,7 @@ config = ed({
     "keep_checkpoint_max": 10,
     "save_checkpoint_path": "./checkpoint",
 
-    "mindrecord_dir": "../MindRecoid_COCO_TRAIN",
+    "mindrecord_dir": "../MindRecord_COCO_TRAIN",
     "coco_root": "./cocodataset/",
     "train_data_type": "train2017",
     "val_data_type": "val2017",

model_zoo/faster_rcnn/src/dataset.py

@@ -24,7 +24,7 @@ import mmcv
 import mindspore.dataset as de
 import mindspore.dataset.transforms.vision.c_transforms as C
 from mindspore.mindrecord import FileWriter
-from config import config
+from src.config import config
 
 
 def bbox_overlaps(bboxes1, bboxes2, mode='iou'):

tests/ut/cpp/pre_activate/ascend/ir_fission/topk_split_test.cc

@@ -90,7 +90,7 @@ TEST_F(TestHWTopKSplit, test_topk_split) {
   EXPECT_TRUE(value_node->value()->isa<tensor::Tensor>());
   auto tensor = value_node->value()->cast<tensor::TensorPtr>();
   EXPECT_EQ(tensor->shape().size(), 1);
-  EXPECT_EQ(tensor->shape()[0], 4);
+  EXPECT_EQ(tensor->shape()[0], 8);
 }
 
 TEST_F(TestHWTopKSplit, test_topk_no_split) {