Feature/tensor array lod pack (#5007)

Author: Superjomn

paddle/framework/lod_tensor.cc

@@ -106,6 +106,15 @@ size_t LoDTensor::NumElements(size_t level, size_t idx) const {
   return lod_[level][idx + 1] - lod_[level][idx];
 }
 
+size_t LoDTensor::NumInstancesInElement(size_t level, size_t idx) const {
+  PADDLE_ENFORCE_LT(level, NumLevels());
+  PADDLE_ENFORCE_LT(idx, NumElements(level));
+  auto abs_lod = ToAbsOffset(lod());
+  size_t begin = abs_lod[level][idx];
+  size_t end = abs_lod[level][idx + 1];
+  return end - begin;
+}
+
 void LoDTensor::ShrinkLevels(size_t level_begin, size_t level_end) {
   auto new_lod = framework::SliceLevels(lod_, level_begin, level_end);
   lod_ = new_lod;
@@ -117,8 +126,15 @@ void LoDTensor::ShrinkInLevel(size_t level, size_t elem_begin,
   PADDLE_ENFORCE_LT(elem_begin, NumElements(level));
   PADDLE_ENFORCE_LT(elem_end, NumElements(level) + 1);
 
+  auto abs_lod = framework::ToAbsOffset(lod());
   auto new_lod = framework::SliceInLevel(lod_, level, elem_begin, elem_end);
   lod_ = new_lod;
+
+  // slice the underlying tensor
+  size_t begin = abs_lod[level][elem_begin];
+  size_t end = abs_lod[level][elem_end];
+  PADDLE_ENFORCE_LT(begin, end, "Cannot shrink, the result tensor is empty.");
+  ShareDataWith(Slice(begin, end));
 }
 
 std::string LoDTensor::SerializeToString() const {

paddle/framework/lod_tensor.h

@@ -122,6 +122,12 @@ class LoDTensor : public Tensor {
    */
   size_t NumElements(size_t level, size_t idx) const;
 
+  /*
+   * Get the number of instances in the underlying tensor in the `idx`-th
+   * element.
+   */
+  size_t NumInstancesInElement(size_t level, size_t idx) const;
+
   /*
    * Shrink levels[level_begin:level_end]
    */
@@ -157,5 +163,42 @@ class LoDTensor : public Tensor {
  private:
   LoD lod_;
 };
+
+/*
+ * Expand the `source` to fit the LoD of `lod`. For example, a `source`
+ * LoDTensor is
+ *  - LoD: [0, 2]
+ *  - tensor: [a0, a1]
+ * a `lod` is
+ *  - LoD: [0 3 5]
+ * returns a new LoDTensor
+ *  - [a0 a0 a0 a1 a1]
+ */
+template <typename T>
+LoDTensor LodExpand(const LoDTensor& source, const LoD& lod, size_t level,
+                    const platform::Place& place) {
+  LoD abs_lod = ToAbsOffset(lod);
+  const auto& lod_level = lod[level];
+  size_t num_instances = source.dims()[0];
+
+  // new tensor
+  LoDTensor tensor;
+  tensor.set_lod(lod);
+  auto dims = source.dims();
+  dims[0] = lod_level.back();
+  tensor.Resize(dims);
+  tensor.mutable_data<T>(place);
+
+  PADDLE_ENFORCE_EQ(num_instances, lod_level.size() - 1);
+  for (size_t ins = 0; ins < num_instances; ins++) {
+    for (size_t elem = lod_level[ins]; elem < lod_level[ins + 1]; elem++) {
+      tensor.Slice(elem, elem + 1)
+          .CopyFrom(source.Slice(ins, ins + 1), platform::CPUPlace(),
+                    platform::CPUDeviceContext());
+    }
+  }
+  return tensor;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/lod_tensor_test.cc

@@ -92,19 +92,56 @@ TEST_F(LoDTensorTester, ShrinkInLevel) {
   size_t level = 0;
   LoDTensor new_lod_tensor = lod_tensor_;
   new_lod_tensor.ShrinkInLevel(level, 0, 1);
-  EXPECT_EQ(new_lod_tensor.NumLevels(), 3UL);
-  EXPECT_EQ(new_lod_tensor.NumElements(0), 1UL);
-  EXPECT_EQ(new_lod_tensor.NumElements(1), 2UL);
-  EXPECT_EQ(new_lod_tensor.NumElements(2), 5UL);
-  ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
+  ASSERT_EQ(new_lod_tensor.NumLevels(), 3UL);
+  ASSERT_EQ(new_lod_tensor.NumElements(0), 1UL);
+  ASSERT_EQ(new_lod_tensor.NumElements(1), 2UL);
+  ASSERT_EQ(new_lod_tensor.NumElements(2), 5UL);
+  ASSERT_EQ(new_lod_tensor.dims()[0], 12);
+  for (int i = 0; i < 12 * 128; i++) {
+    ASSERT_EQ(new_lod_tensor.data<float>()[i], i);
+  }
 
   level = 1;
   new_lod_tensor = lod_tensor_;
   new_lod_tensor.ShrinkInLevel(level, 1, 2);
   ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
   ASSERT_EQ(new_lod_tensor.NumElements(0), 1UL);
   ASSERT_EQ(new_lod_tensor.NumElements(1), 3UL);
-  ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
+  ASSERT_EQ(new_lod_tensor.dims()[0], 7);
+  for (int i = 5 * 128; i < 12 * 128; i++) {
+    ASSERT_EQ(new_lod_tensor.data<float>()[i - 5 * 128], i);
+  }
+
+  LoDTensor t1;
+  t1.set_lod(lod_tensor_.lod());
+  t1.ShareDataWith(lod_tensor_);
+
+  LoDTensor t2;
+  t2.set_lod(lod_tensor_.lod());
+  t2.ShareDataWith(lod_tensor_);
+
+  t1.ShrinkInLevel(0, 1, 2);
+  t2.ShrinkInLevel(0, 0, 1);
+  EXPECT_NE(t1.data<float>(), t2.data<float>());
+  EXPECT_NE(t1.data<float>(), lod_tensor_.data<float>());
+}
+
+TEST(LodExpand, test) {
+  LoD lod{{0, 2}};
+  LoDTensor tensor;
+  tensor.set_lod(lod);
+  tensor.Resize({2, 1});
+  tensor.mutable_data<float>(platform::CPUPlace());
+  tensor.data<float>()[0] = 0;
+  tensor.data<float>()[1] = 1;
+
+  LoD target;
+  target.emplace_back(std::vector<size_t>{0, 3, 5});
+  auto new_tensor = LodExpand<float>(tensor, target, 0UL, platform::CPUPlace());
+  std::vector<int> result{{0, 0, 0, 1, 1}};
+  for (size_t i = 0; i < 5; i++) {
+    ASSERT_EQ(new_tensor.data<float>()[i], result[i]);
+  }
 }
 
 TEST_F(LoDTensorTester, SerializeDeserialize) {

paddle/framework/tensor_array.cc

@@ -20,6 +20,8 @@
 #include <algorithm>
 #include <limits>
 
+#include "paddle/framework/eigen.h"
+
 namespace paddle {
 namespace framework {
 
@@ -104,10 +106,10 @@ void TensorArray::Write(size_t index, const LoDTensor& value) {
     values_.resize(index + 1);
   }
 
+  values_[index].set_lod(value.lod());
   values_[index].Resize(value.dims());
-  values_[index].mutable_data<value_type>(platform::CPUPlace());
-  values_[index].CopyFrom(value, platform::CPUPlace(),
-                          platform::CPUDeviceContext());
+  values_[index].mutable_data<value_type>(value.place());
+  values_[index].CopyFrom(value, value.place(), platform::CPUDeviceContext());
 }
 
 void TensorArray::WriteShared(size_t index, const LoDTensor& value) {
@@ -116,6 +118,7 @@ void TensorArray::WriteShared(size_t index, const LoDTensor& value) {
     values_.resize(index + 1);
   }
 
+  values_[index].set_lod(value.lod());
   values_[index].ShareDataWith(value);
 }
 
@@ -144,6 +147,156 @@ DySeqMetaBatch TensorArray::Unpack(const LoDTensor& source, int level,
   return unpacker.meta;
 }
 
+LoDTensor TensorArray::LodPack(size_t level) const {
+  PADDLE_ENFORCE_GT(size(), 0UL, "no time step exists");
+  // the levels should be no less than 2
+  LoDTensor merged;
+  const LoDTensor *pre, *cur;
+  pre = &Read(0);
+
+  for (size_t step = 1; step < size(); step++) {
+    cur = &Read(step);
+    PADDLE_ENFORCE_GT(cur->NumLevels(), 0);
+    PADDLE_ENFORCE_GT(pre->NumLevels(), 0);
+    PADDLE_ENFORCE_EQ(pre->NumLevels(), cur->NumLevels());
+    PADDLE_ENFORCE_EQ(pre->NumElements(level), cur->NumElements(level));
+
+    merged = LodPackTwo(*pre, *cur, level);
+    pre = &merged;
+  }
+  return merged;
+}
+
+/*
+ * NOTE currently, only the lowest level supports packing.
+ * The lowest LoD will be changed, while the relative offsets in levels above
+ * stay unchanged.
+ *
+ * previous step : [0] [1] [3]
+ * current step: [0 1 2] [2 3] []
+ * packed to
+ *   [0 0] [0 1] [0 2] [1 2] [1 3] [3]
+ */
+LoDTensor TensorArray::LodPackTwo(const LoDTensor& pre, const LoDTensor& cur,
+                                  size_t level) const {
+  PADDLE_ENFORCE_EQ(pre.NumLevels(), cur.NumLevels());
+  PADDLE_ENFORCE_EQ(pre.NumLevels(), level + 1,
+                    "Only the lowest LoD level supports pack temporarily.");
+  // calculate the result tensor's shape first
+  size_t num_instances = 0;
+  for (size_t elem = 0; elem < pre.NumElements(level); elem++) {
+    size_t prefix_size = pre.NumElements(level, elem);
+    size_t num_candidates = cur.NumElements(level, elem);
+    if (num_candidates > 0) {
+      num_instances += num_candidates * (prefix_size + 1);
+    } else {
+      num_instances += prefix_size;
+    }
+  }
+
+  auto res_dims = pre.dims();
+  res_dims[0] = num_instances;
+  LoDTensor result;
+  result.Resize(res_dims);
+  result.mutable_data<value_type>(cur.place());
+
+  Vector<size_t> last_lod_level;
+  // copy data
+  size_t index = 0;
+  last_lod_level.push_back(index);
+  for (size_t elem = 0; elem < pre.NumElements(level); elem++) {
+    size_t prefix_size = pre.NumElements(level, elem);
+    size_t num_candidates = cur.NumElements(level, elem);
+
+    // slice the prefix Tensor
+    LoDTensor prefix = pre;
+    prefix.ShrinkInLevel(level, elem, elem + 1);
+    LoDTensor candidate = cur;
+    if (num_candidates > 0) {
+      candidate.ShrinkInLevel(level, elem, elem + 1);
+    } else {  // just push prefix
+      result.Slice(index, index + prefix_size)
+          .CopyFrom(prefix, result.place(), platform::CPUDeviceContext());
+      index += prefix_size;
+      last_lod_level.push_back(index);
+    }
+    for (size_t candi = 0; candi < num_candidates; candi++) {
+      // TODO(superjom) support GPU
+      result.Slice(index, index + prefix_size)
+          .CopyFrom(prefix, result.place(), platform::CPUDeviceContext());
+      index += prefix_size;
+      // copy candidate record
+      result.Slice(index, index + 1)
+          .CopyFrom(candidate.Slice(candi, candi + 1), result.place(),
+                    platform::CPUDeviceContext());
+      index++;
+      last_lod_level.push_back(index);
+    }
+  }
+
+  // update lod
+  auto lod = cur.lod();
+  lod.back() = last_lod_level;
+  result.set_lod(lod);
+  return result;
+}
+
+/*
+ * source [0 1 2] [3 4] [5 6 7] will be transformd to a list of LoDTensors such
+ * as
+ * [0 3 5] [1 4 6] [2 7] with 1-level LoDs:
+ * - [0 1 2 3]
+ * - [0 1 2 3]
+ * - [0 1 1 2], the [1,1) here means the second sequence is empty
+ *
+ * NOTE Unpack a LoDTensor in this approach may result in a big LoD.
+ */
+void TensorArray::LodUnpack(const LoDTensor& source, size_t level) {
+  PADDLE_ENFORCE_EQ(level, source.NumLevels() - 1,
+                    "only the lowest LoD level supports unpack.");
+  int non_empty_instances = -1;
+  size_t index = 0;
+  Vector<size_t> lowest_lod_level;
+  lowest_lod_level.push_back(index);
+
+  for (size_t step = 0; non_empty_instances > 0 || non_empty_instances == -1;
+       step++) {
+    size_t num_instances = 0;
+    for (size_t id = 0; id < source.NumElements(level); id++) {
+      auto instance = source;
+      instance.ShrinkInLevel(level, id, id + 1);
+      if (static_cast<size_t>(instance.dims()[0]) > step) {
+        num_instances++;
+        index++;
+      }
+      lowest_lod_level.push_back(index);
+    }
+
+    // create tensor for this time step
+    LoDTensor tensor;
+    auto dims = source.dims();
+    dims[0] = num_instances;
+    // set lod
+    auto lod = source.lod();
+    lod.back() = lowest_lod_level;
+    tensor.set_lod(lod);
+
+    index = 0;
+    for (size_t id = 0; id < source.NumElements(level); id++) {
+      auto instance = source;
+      instance.ShrinkInLevel(level, id, id + 1);
+      if (static_cast<size_t>(instance.dims()[0]) > step) {
+        // copy this instance
+        tensor.Slice(index, index + 1)
+            .CopyFrom(instance.Slice(step, step + 1), tensor.place(),
+                      platform::CPUDeviceContext());
+        index++;
+      }
+    }
+    Write(step, tensor);
+  }
+}
+
 LoDTensor TensorArray::Stack() const {
   LoDTensor result;
   if (size() == 0) return result;

paddle/framework/tensor_array.h

@@ -86,6 +86,16 @@ class TensorArray {
    */
   DySeqMetaBatch Unpack(const LoDTensor &source, int level, bool length_desend);
 
+  /*
+   * Pack an array of LoDTensors to a LoDTensor.
+   */
+  LoDTensor LodPack(size_t level) const;
+
+  /*
+   * Unpack a LoDTensor to an array of LoDTensors.
+   */
+  void LodUnpack(const LoDTensor &source, size_t level);
+
   /*
    * Pack the values into a tensor with rank one higher than each tensor in
    * values.
@@ -111,6 +121,9 @@ class TensorArray {
  protected:
   void Unstack(const LoDTensor &source, bool data_shared) const;
 
+  LoDTensor LodPackTwo(const LoDTensor &pre, const LoDTensor &cur,
+                       size_t level) const;
+
  private:
   mutable std::vector<LoDTensor> values_;
 };  // class TensorArray