!14982 [MS][LITE]support the anlysis to inserted 3d transpose op

From: @probiotics_53
Reviewed-by: @zhang_xue_tong
Signed-off-by: @zhang_xue_tong

mindspore/ccsrc/backend/kernel_compiler/cpu/nnacl/infer/transpose_infer.c

@@ -60,32 +60,36 @@ int TransposeInferShape(const TensorC *const *inputs, size_t inputs_size, Tensor
   for (size_t i = 0; i < perms_num; i++) {
     ShapePush(perm, &perm_size, perm_data[i]);
   }
-  int out_shape[MAX_TRANSPOSE_DIM_SIZE] = {0};
-  if (input->shape_size_ != 4 && perms_num == 4) {
-    for (size_t i = 0; i < input->shape_size_; ++i) {
-      out_shape[i] = input->shape_[i];
-    }
-    SetShapeArray(output, out_shape, input->shape_size_);
-    return NNACL_OK;
-  }
   const int nchw2nhwc[4] = {0, 2, 3, 1};
   const int nhwc2nchw[4] = {0, 3, 1, 2};
+  const int trans3d[3] = {0, 2, 1};
   if (perms_num == 4) {
     if (input->format_ == Format_NCHW && CheckPermTransFormat(perm, nchw2nhwc, perms_num)) {
       output->format_ = Format_NHWC;
     } else if (input->format_ == Format_NHWC && CheckPermTransFormat(perm, nhwc2nchw, perms_num)) {
       output->format_ = Format_NCHW;
     }
+    // though the perm is 4d in default, the input can be a 3d tensor. The op implementation should be adapted to this.
+    if (input->shape_size_ == 3) {
+      ShapeSet(perm, &perm_size, trans3d, 3);
+    }
   }
-  output->shape_size_ = perm_size;
-  for (size_t i = 0; i < perm_size; ++i) {
-    out_shape[i] = input->shape_[perm[i]];
-  }
+  // set output shape
+  int out_shape[MAX_TRANSPOSE_DIM_SIZE] = {0};
+  size_t in_shape_size = input->shape_size_;
+  output->shape_size_ = in_shape_size;
   if (perm_size == 0) {
-    size_t shape_size = input->shape_size_;
-    output->shape_size_ = shape_size;
-    for (size_t i = 0; i < shape_size; ++i) {
-      out_shape[shape_size - i - 1] = input->shape_[i];
+    for (size_t i = 0; i < in_shape_size; ++i) {
+      out_shape[in_shape_size - i - 1] = input->shape_[i];
+    }
+  } else if (perm_size != in_shape_size) {
+    for (size_t i = 0; i < in_shape_size; ++i) {
+      out_shape[i] = input->shape_[i];
+    }
+  } else {
+    output->shape_size_ = perm_size;
+    for (size_t i = 0; i < perm_size; ++i) {
+      out_shape[i] = input->shape_[perm[i]];
     }
   }
   SetShapeArray(output, out_shape, output->shape_size_);

mindspore/lite/src/runtime/kernel/arm/fp16/transpose_fp16.cc

@@ -41,19 +41,6 @@ int TransposeFp16CPUKernel::Run() {
   MS_ASSERT(in_tensors_.size() == 1 || in_tensors_.size() == 2);
   TransposeParameter *param = reinterpret_cast<TransposeParameter *>(this->op_parameter_);
   param->data_size_ = in_tensors_[0]->Size();
-  if (in_tensors_.size() == 2) {
-    auto input_perm = in_tensors_.at(1);
-    MS_ASSERT(input_perm != nullptr);
-    MS_ASSERT(input_perm->data_c() != nullptr);
-    int *perm_data = reinterpret_cast<int *>(input_perm->data_c());
-    for (int i = 0; i < input_perm->ElementsNum(); ++i) {
-      param->perm_[i] = perm_data[i];
-    }
-    for (int i = input_perm->ElementsNum(); i < MAX_TRANSPOSE_DIM_SIZE; ++i) {
-      param->perm_[i] = 0;
-    }
-    param->num_axes_ = input_perm->ElementsNum();
-  }
   MS_ASSERT(out_tensors_.size() == 1);
   auto &in_tensor = in_tensors_.front();
   auto &out_tensor = out_tensors_.front();

mindspore/lite/src/runtime/kernel/arm/fp32/transpose_fp32.cc

@@ -40,13 +40,22 @@ int TransposeCPUKernel::ReSize() {
   if (in_tensors_.size() == 2) {
     param->num_axes_ = in_tensors_.at(1)->ElementsNum();
   }
-  if (in_tensors_.at(kInputIndex)->shape().size() != static_cast<size_t>(param->num_axes_)) {
-    return RET_OK;
+  int trans3d[3] = {0, 2, 1};
+  int *perm_data = nullptr;
+  auto input_tensor = in_tensors_.at(kInputIndex);
+  if (input_tensor->shape().size() != static_cast<size_t>(param->num_axes_)) {
+    if (input_tensor->shape().size() == 3 && param->num_axes_ == 4) {
+      param->num_axes_ = 3;
+      perm_data = trans3d;
+    } else {
+      return RET_OK;
+    }
+  } else {
+    MS_ASSERT(in_tensors_.size() == 2);
+    auto perm_tensor = in_tensors_.at(1);
+    perm_data = reinterpret_cast<int *>(perm_tensor->data_c());
   }
-  // get perm data
-  MS_ASSERT(in_tensors_.size() == 2);
-  auto perm_tensor = in_tensors_.at(1);
-  int *perm_data = reinterpret_cast<int *>(perm_tensor->data_c());
+  // set perm data
   MS_ASSERT(perm_data != nullptr);
   for (int i = 0; i < param->num_axes_; ++i) {
     param->perm_[i] = perm_data[i];
@@ -144,18 +153,6 @@ int TransposeCPUKernel::Run() {
     memcpy(out_data_, in_data_, in_tensor->ElementsNum() * sizeof(float));
     return RET_OK;
   }
-  if (in_tensors_.size() == 2) {
-    auto input_perm = in_tensors_.at(1);
-    MS_ASSERT(input_perm != nullptr);
-    MS_ASSERT(input_perm->data_c() != nullptr);
-    int *perm_data = reinterpret_cast<int *>(input_perm->data_c());
-    for (int i = 0; i < input_perm->ElementsNum(); ++i) {
-      param_->perm_[i] = perm_data[i];
-    }
-    for (int i = input_perm->ElementsNum(); i < MAX_TRANSPOSE_DIM_SIZE; ++i) {
-      param_->perm_[i] = 0;
-    }
-  }
   thread_count_ = op_parameter_->thread_num_;
   GetNHNCTransposeFunc(in_tensor, out_tensor, param_);
   if (NHNCTransposeFunc_ != nullptr) {

mindspore/lite/test/models_onnx.cfg

@@ -79,3 +79,4 @@ Q_face_recognition.onnx
 Q888_face_recognition.onnx
 Q888_iris_detect.onnx
 simple_IPS_model_4D_input.onnx
+Harmony_Voiceprint.onnx;1,200,40,1

mindspore/lite/test/models_onnx_fp16.cfg

@@ -86,3 +86,6 @@ ml_video_edit_enhance_update_tmp.onnx 0.5
 Q888_face_recognition.onnx 3.5
 Q888_iris_detect.onnx 0.5
 ssd_mobilenet_v1_10.onnx;1,383,640,3 0.5
+# The output from a conv in the later part contains many minus values, the following leakyRelu makes them become very
+# close to 0 (-e^-4). The fp16 precision lost a lot in this case and it affects the following computation.
+Harmony_Voiceprint.onnx;1,200,40,1 5.5