diff --git a/paddle/function/RowConvOpGpu.cu b/paddle/function/RowConvOpGpu.cu
index c0b947e224313abaf4fadfb8293dc78ca085ff84..d9dcc7d59d1e3c222f5a7ce448daa8d7edb6c978 100644
--- a/paddle/function/RowConvOpGpu.cu
+++ b/paddle/function/RowConvOpGpu.cu
@@ -32,7 +32,7 @@ __global__ void KeRowConv(real* y, const real* x,  const real* w,
   for (int i = tidy; i < context; i += blky) {
     sw[i][tidx] = gidx + tidx < width ? w[i*width + gidx + tidx] : 0.0;
   }
-  
+
   __syncthreads();
 
   for (int i = 0; i < numSeq; ++i) {
@@ -144,12 +144,15 @@ __global__ void KeRowConvBwWeight(real* dw, const real* x, const real* dy,
       int yoff = start + j;
 
       // transpose
-      sh_x[tidx][tidy] = (xoff < width && yoff < end) ? x[yoff * width + xoff] : 0.0;
-      sh_dy[tidx][tidy + context - 1] = (xoff < width && yoff < end) ? dy[yoff * width + xoff] : 0.0;
+      sh_x[tidx][tidy] = (xoff < width && yoff < end) ?
+      x[yoff * width + xoff] : 0.0;
+      sh_dy[tidx][tidy + context - 1] = (xoff < width && yoff < end) ?
+      dy[yoff * width + xoff] : 0.0;
       __syncthreads();
       if (tidy < (context - 1)) {
         yoff = yoff - context + 1;
-        sh_dy[tidx][tidy] = (xoff < width && yoff >= start) ? dy[yoff * width + xoff] : 0.0;
+        sh_dy[tidx][tidy] = (xoff < width && yoff >= start) ?
+        dy[yoff * width + xoff] : 0.0;
       }
       __syncthreads();
 
@@ -199,11 +202,13 @@ __global__ void KeRowConvBwWeight2(real* dw, const real* x, const real* dy,
       int yoff = start + j;
 
       // transpose
-      sh_x[tidx][tidy] = (xoff < width && yoff < end) ? x[yoff * width + xoff] : 0.0;
+      sh_x[tidx][tidy] = (xoff < width && yoff < end) ?
+      x[yoff * width + xoff] : 0.0;
       __syncthreads();
 
       for (int t = 0; t < context; t++) {
-        sh_dy[tidx][tidy] = (xoff < width && (yoff - t) >= start && yoff - t < end) ? dy[(yoff - t) * width + xoff] : 0.0;
+        sh_dy[tidx][tidy] = (xoff < width && (yoff - t) >= start &&
+        yoff - t < end) ? dy[(yoff - t) * width + xoff] : 0.0;
         __syncthreads();
 
         real val = sh_x[tidy][tidx] * sh_dy[tidy][tidx];
@@ -239,7 +244,7 @@ __global__ void KeRowConvBwData(real* dx, const real* w, const real* dy,
   for (int i = tidy; i < context; i += blky) {
     sw[i][tidx] = gidx + tidx < width ? w[i*width + gidx + tidx] : 0.0;
   }
-  
+
   __syncthreads();
 
   for (int i = 0; i < numSeq; ++i) {
@@ -312,7 +317,7 @@ void RowConvGrad<DEVICE_TYPE_GPU>(const GpuMatrix& outG,
     dim3 dimBlock(32, 32);
     dim3 dimGrid(DIVUP(width, dimBlock.x), 1);
     real* dw = filterG.getData();
-    if (contextLength <= 32) { 
+    if (contextLength <= 32) {
       KeRowConvBwWeight<32, 32, 32>
         <<<dimGrid, dimBlock, 0, STREAM_DEFAULT>>>
         (dw, x, dy, starts, height, width, numSeq, contextLength);
diff --git a/python/paddle/v2/optimizer.py b/python/paddle/v2/optimizer.py
index 755b1e09d7f07dac3c0a81a0ed4a1f204932604c..7e8a3bece9acb93ffba634e7b3656f2b580a48f2 100644
--- a/python/paddle/v2/optimizer.py
+++ b/python/paddle/v2/optimizer.py
@@ -1,3 +1,4 @@
+import py_paddle.swig_paddle as swig_api
 import paddle.trainer_config_helpers.config_parser_utils as config_parser_utils
 import paddle.trainer_config_helpers.optimizers as v1_optimizers
 """
@@ -16,7 +17,6 @@ __all__ = [
 
 class Optimizer(object):
     def __init__(self, **kwargs):
-        import py_paddle.swig_paddle as swig_api
         if 'batch_size' in kwargs:
             del kwargs['batch_size']  # not important for python library.