diff --git a/paddle/function/ContextProjectionOp.cpp b/paddle/function/ContextProjectionOp.cpp
index 8803ea78963a00c4724bcfd5cb02381fd2d1a89e..f1e42cad7259e3fb9c862bb53a6da254dabc5c1e 100644
--- a/paddle/function/ContextProjectionOp.cpp
+++ b/paddle/function/ContextProjectionOp.cpp
@@ -232,7 +232,7 @@ public:
     /// input grad and output grad have the same batch_size
     CHECK_EQ(inouts[0].dims_[0], inputs[1].dims_[0]);
     /// dim of output = dim of input * context_length
-    CHECK_EQ(inputs[1].dims_[1], inputs[0].dims_[1] * context_length_);
+    CHECK_EQ(inputs[1].dims_[1], inouts[0].dims_[1] * context_length_);
 
     typename SequenceT<Device>::type seq_vec(
         inputs[0].dims_[0], reinterpret_cast<int*>(inputs[0].getData()));
diff --git a/paddle/function/ContextProjectionOpGpu.cu b/paddle/function/ContextProjectionOpGpu.cu
index 6194ad8e74a37edb287d11a631a968d7eef6e7d9..c5a636dce8ea31721bfec7daef7b815054d680ac 100644
--- a/paddle/function/ContextProjectionOpGpu.cu
+++ b/paddle/function/ContextProjectionOpGpu.cu
@@ -256,7 +256,7 @@ __global__ void KeContextProjectionBackwardWeight(const real* out_grad,
     for (int seqId = idy; seqId < num_sequences; seqId += THREADS_Y) {
       int seq_start = sequence[seqId];
       int seq_end = sequence[seqId+1];
-      output_r = const_cast<real*>(out_grad) 
+      output_r = const_cast<real*>(out_grad)
                     + seq_start * w_dim * context_length;
 
       if (context_start < 0) {