Add fp16 support for python backend.

Author: Arynthon

onnx_tensorrt/backend.py

@@ -58,7 +58,7 @@ def count_trailing_ones(vals):
 
 
 class TensorRTBackendRep(BackendRep):
-    def __init__(self, model, device, max_batch_size=32,
+    def __init__(self, model, device, max_batch_size=32, fp16_mode=False,
                  max_workspace_size=None, serialize_engine=False, **kwargs):
         if not isinstance(device, Device):
             device = Device(device)
@@ -89,6 +89,7 @@ def __init__(self, model, device, max_batch_size=32,
 
         self.builder.max_batch_size = max_batch_size
         self.builder.max_workspace_size = max_workspace_size
+        self.builder.fp16_mode = fp16_mode
 
         for layer in self.network:
             print(layer.name)
@@ -231,4 +232,4 @@ def supports_device(cls, device_str):
 prepare         = TensorRTBackend.prepare
 run_node        = TensorRTBackend.run_node
 run_model       = TensorRTBackend.run_model
-supports_device = TensorRTBackend.supports_device
+supports_device = TensorRTBackend.supports_device