From 12f2b8eb07b034a24e3a0e0538a757e389fb4c45 Mon Sep 17 00:00:00 2001
From: Liu Yiqun <liuyiqun01@baidu.com>
Date: Mon, 25 Sep 2017 04:12:04 +0000
Subject: [PATCH] Correct the forward of sequence_softmax_op.

---
 paddle/operators/reshape_op.cc          |  3 +--
 paddle/operators/sequence_softmax_op.cc | 10 ++++++----
 paddle/operators/sequence_softmax_op.h  |  6 +++++-
 3 files changed, 12 insertions(+), 7 deletions(-)

diff --git a/paddle/operators/reshape_op.cc b/paddle/operators/reshape_op.cc
index 0d05e344148..d5d04dac275 100644
--- a/paddle/operators/reshape_op.cc
+++ b/paddle/operators/reshape_op.cc
@@ -42,8 +42,7 @@ class ReshapeOp : public framework::OperatorWithKernel {
     int64_t capacity =
         std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int>());
     auto *in = ctx.Input<framework::Tensor>("X");
-    int64_t in_size = framework::product(in->dims());
-    PADDLE_ENFORCE_EQ(capacity, in_size,
+    PADDLE_ENFORCE_EQ(capacity, in->numel(),
                       "The size of Input(X) mismatches with Attr(shape).");
     // resize output
     std::vector<int64_t> shape_int64(shape.size(), 0);
diff --git a/paddle/operators/sequence_softmax_op.cc b/paddle/operators/sequence_softmax_op.cc
index 0a99717440e..58ef77b1a3a 100644
--- a/paddle/operators/sequence_softmax_op.cc
+++ b/paddle/operators/sequence_softmax_op.cc
@@ -30,18 +30,20 @@ class SequenceSoftmaxOp : public framework::OperatorWithKernel {
         "Output(Out) of SequenceSoftmaxOp should not be null.");
 
     auto *x = ctx.Input<framework::LoDTensor>("X");
-    auto dims = x->dims();
     auto lod = x->lod();
-    PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now.");
+    auto dims = x->dims();
     PADDLE_ENFORCE_GE(
         dims[0],
         /* batch_size */ static_cast<int64_t>(lod[0].size() - 1),
         "The first dimension of Input(X) should be larger than batch size.");
-    PADDLE_ENFORCE_EQ(x->numel(), static_cast<int64_t>(lod[0].size() - 1),
+
+    const size_t level = lod.size() - 1;
+    PADDLE_ENFORCE_EQ(x->numel(), static_cast<int64_t>(lod[level].back()),
                       "The width of each timestep in Input(X) of "
                       "SequenceSoftmaxOp should be 1.");
 
-    dims[0] = lod[0].size() - 1;
+    std::cout << DebugString() << std::endl;
+
     ctx.Output<framework::LoDTensor>("Out")->Resize({dims});
   }
 };
diff --git a/paddle/operators/sequence_softmax_op.h b/paddle/operators/sequence_softmax_op.h
index 54d82652711..f39c2ec6c31 100644
--- a/paddle/operators/sequence_softmax_op.h
+++ b/paddle/operators/sequence_softmax_op.h
@@ -38,7 +38,7 @@ class SequenceSoftmaxKernel : public framework::OpKernel {
     auto* out = ctx.Output<LoDTensor>("Out");
 
     auto lod = x->lod();
-    const size_t level = lod.size();
+    const size_t level = lod.size() - 1;
 
     out->mutable_data<T>(ctx.GetPlace());
     for (int i = 0; i < static_cast<int>(lod[level].size()) - 1; ++i) {
@@ -47,6 +47,10 @@ class SequenceSoftmaxKernel : public framework::OpKernel {
       Tensor x_i = x->Slice<T>(start_pos, end_pos);
       Tensor out_i = out->Slice<T>(start_pos, end_pos);
 
+      // Reshape from (end_pos - start_pos) x 1UL to 1UL x (end_pos - start_pos)
+      framework::DDim dims = framework::make_ddim({1UL, end_pos - start_pos});
+      x_i.Resize(dims);
+      out_i.Resize(dims);
       math::SoftmaxFunctor<Place, T>()(&x_i, &out_i, ctx);
     }
   }
-- 
GitLab