add gpu training for Executor.train_from_dataset

test=develop

paddle/fluid/framework/data_feed.cc

@@ -466,6 +466,17 @@ void MultiSlotDataFeed::Init(
     if (slot.is_used()) {
       use_slots_.push_back(all_slots_[i]);
       use_slots_is_dense_.push_back(slot.is_dense());
+      std::vector<int> local_shape;
+      if (slot.is_dense()) {
+        // for batch size holder if is_dense
+        if (slot.shape(0) > 0) {
+          local_shape.push_back(0);
+        }
+      }
+      for (size_t i = 0; i < slot.shape_size(); ++i) {
+        local_shape.push_back(slot.shape(i));
+      }
+      use_slots_shape_.push_back(local_shape);
     }
   }
   feed_vec_.resize(use_slots_.size());
@@ -752,8 +763,8 @@ void MultiSlotDataFeed::PutToFeedVec(
     LoD data_lod{offset};
     feed_vec_[i]->set_lod(data_lod);
     if (use_slots_is_dense_[i]) {
-      int dim = total_instance / batch_size_;
+      use_slots_shape_[i][0] = batch_size_;
-      feed_vec_[i]->Resize({batch_size_, dim});
+      feed_vec_[i]->Resize(framework::make_ddim(use_slots_shape_[i]));
     }
   }
 #endif
@@ -785,6 +796,16 @@ void MultiSlotInMemoryDataFeed::Init(
     if (slot.is_used()) {
       use_slots_.push_back(all_slots_[i]);
       use_slots_is_dense_.push_back(slot.is_dense());
+      std::vector<int> local_shape;
+      if (slot.is_dense()) {
+        if (slot.shape(0) > 0) {
+          local_shape.push_back(0);
+        }
+      }
+      for (size_t i = 0; i < slot.shape_size(); ++i) {
+        local_shape.push_back(slot.shape(i));
+      }
+      use_slots_shape_.push_back(local_shape);
     }
   }
   feed_vec_.resize(use_slots_.size());
@@ -940,8 +961,8 @@ void MultiSlotInMemoryDataFeed::PutToFeedVec(
     LoD data_lod{offset};
     feed_vec_[i]->set_lod(data_lod);
     if (use_slots_is_dense_[i]) {
-      int dim = total_instance / batch_size_;
+      use_slots_shape_[i][0] = batch_size_;
-      feed_vec_[i]->Resize({batch_size_, dim});
+      feed_vec_[i]->Resize(framework::make_ddim(use_slots_shape_[i]));
     }
   }
 #endif

paddle/fluid/framework/data_feed.h

@@ -142,6 +142,7 @@ class DataFeed {
   // object)
   std::vector<std::string> all_slots_;
   std::vector<std::string> all_slots_type_;
+  std::vector<std::vector<int>> use_slots_shape_;
   std::vector<int>
       use_slots_index_;  // -1: not used; >=0: the index of use_slots_
 

paddle/fluid/framework/data_feed.proto

@@ -19,6 +19,7 @@ message Slot {
   required string type = 2;
   optional bool is_dense = 3 [ default = false ];
   optional bool is_used = 4 [ default = false ];
+  repeated int32 shape = 5; // we can define N-D Tensor
 }
 
 message MultiSlotDesc { repeated Slot slots = 1; }

paddle/fluid/imperative/nccl_context.h

@@ -16,6 +16,7 @@
 // network header files
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include <arpa/inet.h>
+#include <asm-generic/socket.h>
 #include <netinet/in.h>
 #include <stdlib.h>
 #include <sys/socket.h>

python/paddle/fluid/dataset.py

@@ -136,6 +136,8 @@ class DatasetBase(object):
             slot_var.name = var.name
             if var.lod_level == 0:
                 slot_var.is_dense = True
+                print(var.shape)
+                slot_var.shape.extend(var.shape)
             if var.dtype == core.VarDesc.VarType.FP32:
                 slot_var.type = "float"
             elif var.dtype == core.VarDesc.VarType.INT64:

python/paddle/fluid/executor.py

@@ -712,10 +712,6 @@ class Executor(object):
         if dataset == None:
             raise RuntimeError("dataset is needed and should be initialized")
 
-        if not isinstance(self.place, core.CPUPlace):
-            raise RuntimeError("infer_from_dataset is verified on CPUPlace"
-                               "We will open CUDAPlace in the future")
-
         scope, trainer = self._prepare_trainer(
             program=program,
             dataset=dataset,
@@ -796,10 +792,6 @@ class Executor(object):
         if dataset == None:
             raise RuntimeError("dataset is need and should be initialized")
 
-        if not isinstance(self.place, core.CPUPlace):
-            raise RuntimeError("train_from_dataset is verified on CPUPlace"
-                               "We will open CUDAPlace in the future")
-
         scope, trainer = self._prepare_trainer(
             program=program,
             dataset=dataset,