diff --git a/include/data/tensor.hpp b/include/data/tensor.hpp
index f712c398..e4f816d7 100644
--- a/include/data/tensor.hpp
+++ b/include/data/tensor.hpp
@@ -258,6 +258,9 @@ class Tensor {
    */
   T* matrix_raw_ptr(uint32_t index);
 
+ private:
+  void Review(const std::vector<uint32_t>& shapes);
+
  private:
   std::vector<uint32_t> raw_shapes_;  // 张量数据的实际尺寸大小
   arma::Cube<T> data_;                // 张量数据
diff --git a/source/data/tensor.cpp b/source/data/tensor.cpp
index 5c1abe88..ad5d829d 100644
--- a/source/data/tensor.cpp
+++ b/source/data/tensor.cpp
@@ -321,23 +321,28 @@ void Tensor<T>::Reshape(const std::vector<uint32_t>& shapes, bool row_major) {
   CHECK(shapes.size() <= 3);
   CHECK(current_size == origin_size);
 
-  std::vector<T> values;
-  if (row_major) {
-    values = this->values(true);
-  }
-  if (shapes.size() == 3) {
-    this->data_.reshape(shapes.at(1), shapes.at(2), shapes.at(0));
-    this->raw_shapes_ = {shapes.at(0), shapes.at(1), shapes.at(2)};
-  } else if (shapes.size() == 2) {
-    this->data_.reshape(shapes.at(0), shapes.at(1), 1);
-    this->raw_shapes_ = {shapes.at(0), shapes.at(1)};
+  if (!row_major) {
+    if (shapes.size() == 3) {
+      this->data_.reshape(shapes.at(1), shapes.at(2), shapes.at(0));
+      this->raw_shapes_ = {shapes.at(0), shapes.at(1), shapes.at(2)};
+    } else if (shapes.size() == 2) {
+      this->data_.reshape(shapes.at(0), shapes.at(1), 1);
+      this->raw_shapes_ = {shapes.at(0), shapes.at(1)};
+    } else {
+      this->data_.reshape(1, shapes.at(0), 1);
+      this->raw_shapes_ = {shapes.at(0)};
+    }
   } else {
-    this->data_.reshape(1, shapes.at(0), 1);
-    this->raw_shapes_ = {shapes.at(0)};
-  }
-
-  if (row_major) {
-    this->Fill(values, true);
+    if (shapes.size() == 3) {
+      this->Review({shapes.at(0), shapes.at(1), shapes.at(2)});
+      this->raw_shapes_ = {shapes.at(0), shapes.at(1), shapes.at(2)};
+    } else if (shapes.size() == 2) {
+      this->Review({1, shapes.at(0), shapes.at(1)});
+      this->raw_shapes_ = {shapes.at(0), shapes.at(1)};
+    } else {
+      this->Review({1, 1, shapes.at(0)});
+      this->raw_shapes_ = {shapes.at(0)};
+    }
   }
 }
 
@@ -395,6 +400,40 @@ void Tensor<T>::set_data(arma::Cube<T>&& data) {
   this->data_ = std::move(data);
 }
 
+template <typename T>
+void Tensor<T>::Review(const std::vector<uint32_t>& shapes) {
+  CHECK(!this->data_.empty());
+  CHECK_EQ(shapes.size(), 3);
+  const uint32_t target_channels = shapes.at(0);
+  const uint32_t target_rows = shapes.at(1);
+  const uint32_t target_cols = shapes.at(2);
+
+  CHECK_EQ(this->data_.size(), target_channels * target_cols * target_rows);
+  arma::Cube<T> new_data(target_rows, target_cols, target_channels);
+  const uint32_t plane_size = target_rows * target_cols;
+#pragma omp parallel for
+  for (uint32_t channel = 0; channel < this->data_.n_slices; ++channel) {
+    const arma::Mat<T>& channel_data = this->data_.slice(channel);
+    const uint32_t plane_start = channel * data_.n_rows * data_.n_cols;
+    for (uint32_t src_col = 0; src_col < this->data_.n_cols; ++src_col) {
+      const T* col_ptr = channel_data.colptr(src_col);
+      for (uint32_t src_row = 0; src_row < this->data_.n_rows; ++src_row) {
+        const uint32_t pos_index =
+            plane_start + src_row * data_.n_cols + src_col;
+        const uint32_t dest_channel = pos_index / plane_size;
+        const uint32_t dest_row =
+            (pos_index - dest_channel * plane_size) / target_cols;
+        const uint32_t dest_col =
+            (pos_index - dest_channel * plane_size - dest_row * target_cols);
+        CHECK(dest_channel < new_data.n_slices && dest_col < new_data.n_cols &&
+              dest_row < new_data.n_rows);
+        new_data.at(dest_row, dest_col, dest_channel) = *(col_ptr + src_row);
+      }
+    }
+  }
+  this->data_ = std::move(new_data);
+}
+
 template class Tensor<float>;
 template class Tensor<int>;
 template class Tensor<uint8_t>;