This is an automated email from the ASF dual-hosted git repository.
taolv pushed a commit to branch v1.5.x
in repository https://gitbox.apache.org/repos/asf/incubator-mxnet.git
The following commit(s) were added to refs/heads/v1.5.x by this push:
new 964f288 [MXNET-1086] added sub and mul to ONNX->TensorRT conversion
(#15344) (#15875)
964f288 is described below
commit 964f2880c8c312a0e02579823e7c69875b604619
Author: Kellen Sunderland <[email protected]>
AuthorDate: Fri Aug 16 08:19:55 2019 -0700
[MXNET-1086] added sub and mul to ONNX->TensorRT conversion (#15344)
(#15875)
* added sub and mul to ONNX->TensorRT conversion
* add test for elementwise ops in TRT
---
CMakeLists.txt | 2 +-
src/operator/subgraph/tensorrt/nnvm_to_onnx-inl.h | 12 ++++
src/operator/subgraph/tensorrt/nnvm_to_onnx.cc | 12 ++++
tests/python/tensorrt/test_ops.py | 68 +++++++++++++++++++++++
4 files changed, 93 insertions(+), 1 deletion(-)
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 2142a09..0148ac3 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -47,7 +47,7 @@ mxnet_option(ENABLE_CUDA_RTC "Build with CUDA runtime
compilation support"
mxnet_option(BUILD_CPP_EXAMPLES "Build cpp examples" ON)
mxnet_option(INSTALL_EXAMPLES "Install the example source files." OFF)
mxnet_option(USE_SIGNAL_HANDLER "Print stack traces on segfaults." ON)
-mxnet_option(USE_TENSORRT "Enable infeference optimization with
TensorRT." OFF)
+mxnet_option(USE_TENSORRT "Enable inference optimization with
TensorRT." OFF)
mxnet_option(USE_ASAN "Enable Clang/GCC ASAN sanitizers." OFF)
mxnet_option(ENABLE_TESTCOVERAGE "Enable compilation with test coverage
metric output" OFF)
mxnet_option(USE_INT64_TENSOR_SIZE "Use int64_t to represent the total number
of elements in a tensor" OFF)
diff --git a/src/operator/subgraph/tensorrt/nnvm_to_onnx-inl.h
b/src/operator/subgraph/tensorrt/nnvm_to_onnx-inl.h
index f5bf8b7..55b3d93 100644
--- a/src/operator/subgraph/tensorrt/nnvm_to_onnx-inl.h
+++ b/src/operator/subgraph/tensorrt/nnvm_to_onnx-inl.h
@@ -125,6 +125,16 @@ void ConvertElementwiseAdd(NodeProto *node_proto,
const nnvm::IndexedGraph &ig,
const array_view<IndexedGraph::NodeEntry> &inputs);
+void ConvertElementwiseSub(NodeProto *node_proto,
+ const NodeAttrs &attrs,
+ const nnvm::IndexedGraph &ig,
+ const array_view<IndexedGraph::NodeEntry> &inputs);
+
+void ConvertElementwiseMul(NodeProto *node_proto,
+ const NodeAttrs &attrs,
+ const nnvm::IndexedGraph &ig,
+ const array_view<IndexedGraph::NodeEntry> &inputs);
+
void ConvertConcatenate(NodeProto *node_proto,
const NodeAttrs &attrs,
const nnvm::IndexedGraph &ig,
@@ -151,6 +161,8 @@ static const std::unordered_map<std::string,
ConverterFunction> converter_map =
{"Concat", ConvertConcatenate},
{"Dropout", ConvertDropout},
{"elemwise_add", ConvertElementwiseAdd},
+ {"elemwise_sub", ConvertElementwiseSub},
+ {"elemwise_mul", ConvertElementwiseMul},
{"Flatten", ConvertFlatten},
{"FullyConnected", ConvertFullyConnected},
{"Pad", ConvertPad},
diff --git a/src/operator/subgraph/tensorrt/nnvm_to_onnx.cc
b/src/operator/subgraph/tensorrt/nnvm_to_onnx.cc
index 111995d..6116f29 100644
--- a/src/operator/subgraph/tensorrt/nnvm_to_onnx.cc
+++ b/src/operator/subgraph/tensorrt/nnvm_to_onnx.cc
@@ -406,6 +406,18 @@ void ConvertElementwiseAdd(NodeProto* node_proto, const
NodeAttrs& /*attrs*/,
node_proto->set_op_type("Add");
}
+void ConvertElementwiseSub(NodeProto* node_proto, const NodeAttrs& /*attrs*/,
+ const nnvm::IndexedGraph& /*ig*/,
+ const array_view<IndexedGraph::NodeEntry>&
/*inputs*/) {
+ node_proto->set_op_type("Sub");
+}
+
+void ConvertElementwiseMul(NodeProto* node_proto, const NodeAttrs& /*attrs*/,
+ const nnvm::IndexedGraph& /*ig*/,
+ const array_view<IndexedGraph::NodeEntry>&
/*inputs*/) {
+ node_proto->set_op_type("Mul");
+}
+
void ConvertConcatenate(NodeProto* node_proto, const NodeAttrs& attrs,
const nnvm::IndexedGraph& /*ig*/,
const array_view<IndexedGraph::NodeEntry>& /*inputs*/)
{
diff --git a/tests/python/tensorrt/test_ops.py
b/tests/python/tensorrt/test_ops.py
new file mode 100644
index 0000000..2df9104
--- /dev/null
+++ b/tests/python/tensorrt/test_ops.py
@@ -0,0 +1,68 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+
+from mxnet.test_utils import assert_almost_equal
+import mxnet as mx
+import numpy as np
+import os
+
+def check_elementwise_random(op='sum', shape=(1, 3, 224, 224)):
+ """
+ Check elementwise operators with vanilla/TensorRT executors with uniform
random tensors
+ """
+ a = mx.sym.Variable('a')
+ b = mx.sym.Variable('b')
+ if op == 'sum':
+ sym = a + b
+ elif op == 'sub':
+ sym = a - b
+ elif op == 'mul':
+ sym = a * b
+
+ a_data = mx.ndarray.random.uniform(shape=shape, ctx=mx.gpu())
+ b_data = mx.ndarray.random.uniform(shape=shape, ctx=mx.gpu())
+
+ executor = sym.simple_bind(ctx=mx.gpu(), a=shape, b=shape,
+ grad_req='null', force_rebind=True)
+ y = executor.forward(is_train=False, a=a_data, b=b_data)
+ trt_sym = sym.get_backend_symbol('TensorRT')
+ original_precision_value = mx.contrib.tensorrt.get_use_fp16()
+ try:
+ mx.contrib.tensorrt.set_use_fp16(True)
+ executor = trt_sym.simple_bind(ctx=mx.gpu(), a=shape, b=shape,
+ grad_req='null', force_rebind=True)
+ y_trt = executor.forward(is_train=False, a=a_data, b=b_data)
+ mx.contrib.tensorrt.set_use_fp16(False)
+ executor = trt_sym.simple_bind(ctx=mx.gpu(), a=shape, b=shape,
+ grad_req='null', force_rebind=True)
+ y_trt_fp32 = executor.forward(is_train=False, a=a_data, b=b_data)
+ assert_almost_equal(y[0].asnumpy(), y_trt[0].asnumpy(), 1e-1, 1e-2)
+ assert_almost_equal(y[0].asnumpy(), y_trt_fp32[0].asnumpy(), 1e-4,
1e-4)
+ finally:
+ mx.contrib.tensorrt.set_use_fp16(original_precision_value)
+
+
+def test_elementwise():
+ for op in ['sum', 'sub', 'mul']:
+ for shape in [(20, 25), (3, 4, 20), (1, 3, 20, 25), (10, 10, 100,
100)]:
+ for itry in range(10):
+ check_elementwise_random(op, shape)
+
+
+if __name__ == '__main__':
+ import nose
+ nose.runmodule()
