Merge pull request pytorch#82 from jma127/master

Python 3.7 compatibility

CMakeLists.txt

@@ -43,7 +43,7 @@ endif()
 set(CMAKE_CXX_STANDARD 17)
 set(PYBIND11_CPP_STANDARD -std=c++17)
 set(CMAKE_CXX_FLAGS
-    "${CMAKE_CXX_FLAGS} -Wall -Werror -Wextra -Wno-register -fPIC -march=native")
+    "${CMAKE_CXX_FLAGS} -Wall -Werror -Wextra -Wno-register -Wno-deprecated-declarations -fPIC -march=native")
 set(CMAKE_CXX_FLAGS_RELEASE "-O3")
 if(NOT CMAKE_BUILD_TYPE)
   set(CMAKE_BUILD_TYPE Release)

src_py/elf/utils_elf.py

@@ -245,28 +245,28 @@ def half(self):
                            for k, v in self.batch.items()}
         return new_batch
 
-    def cpu2gpu(self, gpu, async=True):
+    def cpu2gpu(self, gpu, non_blocking=True):
         ''' transfer batch data to gpu '''
         # For each time step
         new_batch = self.empty_copy()
-        new_batch.batch = {k: v.cuda(gpu, async=async)
+        new_batch.batch = {k: v.cuda(gpu, non_blocking=non_blocking)
                            for k, v in self.batch.items()}
         return new_batch
 
-    def cpu2cpu(self, gpu, async=True):
+    def cpu2cpu(self, gpu, non_blocking=True):
         ''' transfer batch data to gpu '''
         # For each time step
         new_batch = self.empty_copy()
         new_batch.batch = {k: v.clone() for k, v in self.batch.items()}
         return new_batch
 
-    def transfer_cpu2gpu(self, batch_gpu, async=True):
+    def transfer_cpu2gpu(self, batch_gpu, non_blocking=True):
         ''' transfer batch data to gpu '''
         # For each time step
         for k, v in self.batch.items():
-            batch_gpu[k].copy_(v, async=async)
+            batch_gpu[k].copy_(v, non_blocking=non_blocking)
 
-    def transfer_cpu2cpu(self, batch_dst, async=True):
+    def transfer_cpu2cpu(self, batch_dst, non_blocking=True):
         ''' transfer batch data to cpu '''
 
         # For each time step

src_py/rlpytorch/runner/parameter_server.py

@@ -245,7 +245,7 @@ def process_main(self, i, gpu_id):
 
         while True:
             self.cvs_recv[i].wait()
-            utils_elf.transfer_cpu2gpu(batch, batch_gpu, async=True)
+            utils_elf.transfer_cpu2gpu(batch, batch_gpu, non_blocking=True)
             self.cvs_send[i].notify()
             self.cb_remote_batch_process(context, batch_gpu)