This repository has been archived by the owner on Nov 17, 2023. It is now read-only.
Add unit tests for TensorRT integration and fix some bugs #15399
There are no files selected for viewing
Submodule onnx-tensorrt
updated
36 files
| +50 −16 | CMakeLists.txt | |
| +9 −9 | FancyActivation.cu | |
| +57 −13 | FancyActivation.hpp | |
| +3 −5 | ImporterContext.hpp | |
| +1 −1 | InstanceNormalization.cpp | |
| +48 −4 | InstanceNormalization.hpp | |
| +287 −32 | ModelImporter.cpp | |
| +6 −2 | ModelImporter.hpp | |
| +27 −6 | NvOnnxParser.h | |
| +30 −0 | NvOnnxParserTypedefs.h | |
| +6 −0 | OnnxAttrs.cpp | |
| +59 −39 | README.md | |
| +3 −2 | ResizeNearest.cu | |
| +47 −4 | ResizeNearest.hpp | |
| +66 −13 | ShapedWeights.cpp | |
| +2 −0 | ShapedWeights.hpp | |
| +4 −2 | Split.cu | |
| +47 −4 | Split.hpp | |
| +9 −0 | Status.hpp | |
| +627 −452 | builtin_op_importers.cpp | |
| +0 −43 | builtin_plugins.cpp | |
| +119 −0 | common.hpp | |
| +9 −0 | contributing.md | |
| +167 −0 | getSupportedAPITest.cpp | |
| +14 −94 | main.cpp | |
| +2 −2 | onnx2trt.hpp | |
| +5 −61 | onnx2trt_utils.cpp | |
| +161 −6 | onnx2trt_utils.hpp | |
| +109 −11 | onnx_backend_test.py | |
| +0 −2 | onnx_tensorrt/__init__.py | |
| +52 −12 | onnx_tensorrt/backend.py | |
| +57 −17 | onnx_tensorrt/tensorrt_engine.py | |
| +14 −53 | onnx_trt_backend.cpp | |
| +145 −0 | operators.md | |
| +78 −55 | plugin.cpp | |
| +63 −22 | plugin.hpp |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -34,6 +34,14 @@ | |
| #include <string> | ||
| #include <vector> | ||
|
|
||
| #include "../../nn/activation-inl.h" | ||
| #include "../../nn/batch_norm-inl.h" | ||
| #include "../../nn/concat-inl.h" | ||
| #include "../../nn/convolution-inl.h" | ||
| #include "../../nn/deconvolution-inl.h" | ||
| #include "../../nn/dropout-inl.h" | ||
| #include "../../nn/fully_connected-inl.h" | ||
| #include "../../nn/pooling-inl.h" | ||
| #include "../common.h" | ||
| #include "../subgraph_property.h" | ||
| #include "nnvm_to_onnx-inl.h" | ||
|
|
@@ -85,18 +93,12 @@ struct TRTEngineParam { | |
| class TensorrtSelector : public SubgraphSelector { | ||
| public: | ||
| const std::unordered_set<std::string> unconditionalTRTops = { | ||
| "_copy", | ||
| "clip", | ||
| "elemwise_add", | ||
| "elemwise_sub", | ||
| "elemwise_mul", | ||
| "Flatten", | ||
| "Pad", | ||
| "relu", | ||
| "rsqrt", | ||
|
|
@@ -112,13 +114,94 @@ class TensorrtSelector : public SubgraphSelector { | |
|
|
||
| bool isTRTCompatible(const nnvm::Node &n) { | ||
| const std::string op_name = n.op()->name; | ||
| if (op_name == "FullyConnected") { | ||
| const auto& param = nnvm::get<FullyConnectedParam>(n.attrs.parsed); | ||
| return !param.no_bias; | ||
| } | ||
|
|
||
| if (op_name == "Pooling") { | ||
| const auto& param = nnvm::get<PoolingParam>(n.attrs.parsed); | ||
| if (param.layout.has_value()) { | ||
| if (param.layout.value() == mshadow::kNHWC) { | ||
| LOG(INFO) << "Warning: NHWC layout (node: " << n.attrs.name | ||
| << ") is not supported by TensorRT"; | ||
| return false; | ||
| } else if (param.layout.value() == mshadow::kNDHWC) { | ||
| LOG(INFO) << "Warning: NDHWC layout (node: " << n.attrs.name | ||
| << ") is not supported by TensorRT"; | ||
| return false; | ||
| } | ||
| } | ||
| if (param.pooling_convention != pool_enum::kValid && !param.global_pool) | ||
| return false; | ||
| if (param.pool_type == pool_enum::kAvgPooling) { | ||
| if ((!param.global_pool) && | ||
| (!param.count_include_pad.has_value() || param.count_include_pad.value())) | ||
| return false; | ||
| return true; | ||
| } else if (param.pool_type == pool_enum::kMaxPooling) { | ||
| return true; | ||
| } else { | ||
| return false; | ||
| } | ||
| } | ||
|
|
||
| if (op_name == "Convolution") { | ||
| const auto& param = nnvm::get<ConvolutionParam>(n.attrs.parsed); | ||
| if (!param.layout.has_value()) | ||
| return true; | ||
| switch (param.layout.value()) { | ||
| case mshadow::kNCHW: | ||
| case mshadow::kNCW: | ||
| case mshadow::kNCDHW: | ||
| return true; | ||
| case mshadow::kNHWC: | ||
| LOG(INFO) << "Warning: NHWC layout (node: " << n.attrs.name | ||
| << ") is not supported by TensorRT"; | ||
| return false; | ||
| case mshadow::kNDHWC: | ||
| LOG(INFO) << "Warning: NDHWC layout (node: " << n.attrs.name | ||
| << ") is not supported by TensorRT"; | ||
| return false; | ||
| default: | ||
| LOG(INFO) << "Warning: Layout (node: " << n.attrs.name | ||
| << ") is unknown (so unsupported by TensorRT)"; | ||
| return false; | ||
| } | ||
| } | ||
|
|
||
| if (op_name == "Deconvolution") { | ||
| const auto& param = nnvm::get<DeconvolutionParam>(n.attrs.parsed); | ||
| if (!param.layout.has_value()) | ||
| return true; | ||
| switch (param.layout.value()) { | ||
| case mshadow::kNCHW: | ||
| case mshadow::kNCW: | ||
| case mshadow::kNCDHW: | ||
| return true; | ||
| case mshadow::kNHWC: | ||
| LOG(INFO) << "Warning: NHWC layout (node: " << n.attrs.name | ||
| << ") is no tsupported by TensorRT"; | ||
| return false; | ||
| case mshadow::kNDHWC: | ||
| LOG(INFO) << "Warning: NDHWC layout (node: " << n.attrs.name | ||
| << ") is not supported by TensorRT"; | ||
| return false; | ||
| default: | ||
| LOG(INFO) << "Warning: Layout (node: " << n.attrs.name | ||
| << ") is unknown (so unsupported by TensorRT)"; | ||
| return false; | ||
| } | ||
| } | ||
|
|
||
| if (op_name == "Concat") { | ||
| const auto& param = nnvm::get<ConcatParam>(n.attrs.parsed); | ||
| return (param.dim != 0); | ||
| } | ||
|
|
||
| if (op_name == "Dropout") { | ||
|
There was a problem hiding this comment. Again, will TensorRT optimize this out? We don't want it at inference time right? There was a problem hiding this comment. Dropout have always been seen as identity function in MXNet-TensorRT integration so I don't see any changement on this, regarding to whether or not identity is actually doing a copy or not I'm not quite sure, here is the onnx-tensorrt conversion: https://github.com/onnx/onnx-tensorrt/blob/0ab159579551cabfa05fd66f338357f116e96835/trt_utils.hpp#L169-L180 There was a problem hiding this comment. Ok, non-blocking comment for this PR. I'm just thinking about adding a warning in the future if people are using TRT with operations that don't make sense at inference time (Dropout, Ident, Empty Concats or Copies, etc.) |
||
| const auto& param = nnvm::get<DropoutParam>(n.attrs.parsed); | ||
| return param.mode == dropout::kTraining && param.axes.ndim() == 0; | ||
| } | ||
|
|
||
| if (op_name == "Activation") { | ||
|
|
@@ -127,6 +210,20 @@ class TensorrtSelector : public SubgraphSelector { | |
| n.attrs.dict.at("act_type") == "sigmoid"; | ||
| } | ||
|
|
||
| if (op_name == "BatchNorm") { | ||
| const auto& param = nnvm::get<BatchNormParam>(n.attrs.parsed); | ||
| if (param.axis != 1) { | ||
| LOG(INFO) << "Warning: Only Layout NC(D)(H)W are supported by TensorRT " | ||
| << "(node " << n.attrs.name << ")"; | ||
| return false; | ||
| } | ||
| return true; | ||
| } | ||
|
|
||
| if (unconditionalTRTops.count(op_name)) { | ||
| return true; | ||
| } | ||
|
|
||
| return false; | ||
| } | ||
|
|
||
|
|
||
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Any idea if TRT actually optimizes this out? I've seen this in a few prod services :-/
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I believe this should be optimized by ONNX-TRT