Merge branch 'develop' of upstream into add_lstmp_doc

paddle/operators/pool_op.h

@@ -139,10 +139,8 @@ class PoolGradKernel : public framework::OpKernel<T> {
     auto& dev_ctx = context.template device_context<DeviceContext>();
     if (in_x_grad) {
       in_x_grad->mutable_data<T>(context.GetPlace());
-      auto temp = framework::EigenVector<T>::Flatten(*in_x_grad);
-      temp.device(
-          *context.template device_context<DeviceContext>().eigen_device()) =
-          temp.constant(static_cast<T>(0));
+      paddle::operators::math::SetConstant<DeviceContext, T> set_constant;
+      set_constant(dev_ctx, in_x_grad, 0.0);
 
       switch (ksize.size()) {
         case 2: {

python/paddle/v2/fluid/distribute_transpiler.py

@@ -225,7 +225,7 @@ class DistributeTranspiler:
             if len(splited_vars) <= 1:
                 continue
             orig_var = program.global_block().vars[varname]
-            if orig_var == core.VarDesc.VarType.SELECTED_ROWS:
+            if orig_var.type == core.VarDesc.VarType.SELECTED_ROWS:
                 height_sections = []
                 for v in splited_vars:
                     height_sections.append(v.shape[0])
@@ -234,7 +234,7 @@ class DistributeTranspiler:
                     inputs={"X": orig_var},
                     outputs={"Out": splited_vars},
                     attrs={"height_sections": height_sections})
-            elif orig_var == core.VarDesc.VarType.LOD_TENSOR:
+            elif orig_var.type == core.VarDesc.VarType.LOD_TENSOR:
                 sections = []
                 for v in splited_vars:
                     sections.append(v.shape[0])

python/paddle/v2/fluid/tests/book/test_rnn_encoder_decoder.py

@@ -118,12 +118,13 @@ def seq_to_seq_net():
     src_forward, src_backward = bi_lstm_encoder(
         input_seq=src_embedding, hidden_size=encoder_size)
 
-    encoded_vector = fluid.layers.concat(
-        input=[src_forward, src_backward], axis=1)
+    src_forward_last = fluid.layers.sequence_last_step(input=src_forward)
+    src_backward_first = fluid.layers.sequence_first_step(input=src_backward)
 
-    enc_vec_last = fluid.layers.sequence_last_step(input=encoded_vector)
+    encoded_vector = fluid.layers.concat(
+        input=[src_forward_last, src_backward_first], axis=1)
 
-    decoder_boot = fluid.layers.fc(input=enc_vec_last,
+    decoder_boot = fluid.layers.fc(input=encoded_vector,
                                    size=decoder_size,
                                    bias_attr=False,
                                    act='tanh')
@@ -137,7 +138,7 @@ def seq_to_seq_net():
         dtype='float32')
 
     prediction = lstm_decoder_without_attention(trg_embedding, decoder_boot,
-                                                enc_vec_last, decoder_size)
+                                                encoded_vector, decoder_size)
     label = fluid.layers.data(
         name='label_sequence', shape=[1], dtype='int64', lod_level=1)
     cost = fluid.layers.cross_entropy(input=prediction, label=label)