KellenSunderland commented on a change in pull request #11325: Added TensorRT runtime integration URL: https://github.com/apache/incubator-mxnet/pull/11325#discussion_r204207228
########## File path: src/operator/contrib/nnvm_to_onnx.cc ########## @@ -0,0 +1,545 @@ +/* + * 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. + */ + +/*! + * Copyright (c) 2018 by Contributors + * \file trt.cc + * \brief TensorRT operation registration + * \author Marek Kolodziej, Clement Fuji Tsang +*/ + +#if MXNET_USE_TENSORRT + +#include "./nnvm_to_onnx-inl.h" + +#include <mxnet/base.h> +#include <nnvm/graph.h> +#include <nnvm/pass_functions.h> + +#include <algorithm> +#include <fstream> +#include <iostream> +#include <unordered_map> +#include <vector> + +#include "./tensorrt-inl.h" +#include "../../common/serialization.h" +#include "../../common/utils.h" +#include "../../ndarray/ndarray_function.h" +#include "../../operator/nn/activation-inl.h" +#include "../../operator/nn/batch_norm-inl.h" +#include "../../operator/nn/convolution-inl.h" +#include "../../operator/nn/fully_connected-inl.h" +#include "../../operator/nn/pooling-inl.h" +#include "../../operator/softmax_output-inl.h" + +// #include <onnx/checker.h> + +namespace mxnet { +namespace op { +namespace nnvm_to_onnx { + +op::TRTParam ConvertNnvmGraphToOnnx( + const nnvm::Graph& g, + std::unordered_map<std::string, NDArray>* const shared_buffer) { + op::TRTParam trt_param; + op::tensorrt::NameToIdx_t trt_input_map; + op::tensorrt::InferenceMap_t trt_output_map; + + const nnvm::IndexedGraph& ig = g.indexed_graph(); + const StorageTypeVector& storage_types = + g.GetAttr<StorageTypeVector>("storage_type"); + const DTypeVector& dtypes = g.GetAttr<DTypeVector>("dtype"); + const ShapeVector& shape_inputs = g.GetAttr<ShapeVector>("shape_inputs"); + + for (auto& e : storage_types) { + if (e != mshadow::kFloat32) { + LOG(FATAL) << "ONNX converter does not support types other than float32 " + "right now."; + } + } + + ModelProto model_proto; + // Need to determine IR versions and features to support + model_proto.set_ir_version(static_cast<int64>(2)); + GraphProto* graph_proto = model_proto.mutable_graph(); + // graph_proto->set_name(graph_name); + + std::unordered_map<std::string, TShape> placeholder_shapes = + GetPlaceholderShapes(shape_inputs, ig); + std::unordered_map<std::string, uint32_t> output_lookup = GetOutputLookup(ig); + uint32_t current_input = 0; + + // can't do a foreach over IndexedGraph since it doesn't + // implement begin(), etc. + for (uint32_t node_idx = 0; node_idx < ig.num_nodes(); ++node_idx) { + const IndexedGraph::Node& node = ig[node_idx]; + const nnvm::Node* source = node.source; + const NodeAttrs& attrs = source->attrs; + const Op* op = source->op(); + + std::string node_name = attrs.name; + // Here, "variable" actually means anything that's not an op, + // i.e. a constant (weights) or a placeholder + if (source->is_variable()) { + // Is this a placeholder? + if (shared_buffer->count(node_name) == 0) { + // This fixes the problem with a SoftmaxOutput node during inference, + // but it's hacky. + // Need to figure out how to properly fix it. + if (node_name.find("label") != std::string::npos) { + current_input++; + continue; + } + trt_input_map.emplace(node_name, current_input++); + ConvertPlaceholder(node_name, placeholder_shapes, graph_proto); + } else { + // If it's not a placeholder, then by exclusion it's a constant. + ConvertConstant(graph_proto, node_name, shared_buffer); + } // is_placeholder + } else { + // It's an op, rather than a "variable" (constant or placeholder) + NodeProto* node_proto = graph_proto->add_node(); + node_proto->set_name(node_name); + if (converter_map.count(op->name) == 0) { + LOG(FATAL) << "Conversion for node of type " << op->name << " (node " + << node_name << ") " + << " is not supported yet."; + } + // Find function ptr to a converter based on the op name, and + // invoke the converter. This looks unsafe because find may not + // succeed, but it does because we're in the operator logic after + // testing that this node name does not represent a variable. + converter_map.find(op->name)->second(node_proto, attrs, ig, node.inputs); + // Add all inputs to the current node (i.e. add graph edges) + for (const nnvm::IndexedGraph::NodeEntry& entry : node.inputs) { + std::string in_node_name = ig[entry.node_id].source->attrs.name; + // As before, we're not adding labels e.g. for SoftmaxOutput, + // but I wish there was a less hacky way to do it than name matching. + if (in_node_name.find("label") != std::string::npos) { + continue; + } + node_proto->add_input(in_node_name); + } + // The node's output will have the same name as the node name. + node_proto->add_output(node_name); + // See if the current node is an output node + auto out_iter = output_lookup.find(node_name); + // We found an output + if (out_iter != output_lookup.end()) { + ConvertOutput(&trt_output_map, graph_proto, out_iter, node_name, g, + storage_types, dtypes); + } // output found + } // conversion function exists + } // loop over i from 0 to num_nodes + + model_proto.SerializeToString(&trt_param.serialized_onnx_graph); + common::Serialize<op::tensorrt::NameToIdx_t>(trt_input_map, + &trt_param.serialized_input_map); + common::Serialize<op::tensorrt::InferenceMap_t>(trt_output_map, + &trt_param.serialized_output_map); + // onnx::checker::check_model(model_proto); + return trt_param; +} + +void ConvertConvolution(NodeProto* node_proto, const NodeAttrs& attrs, + const nnvm::IndexedGraph& ig, + const array_view<IndexedGraph::NodeEntry>& inputs) { + const op::ConvolutionParam& conv_param = + nnvm::get<op::ConvolutionParam>(attrs.parsed); + + node_proto->set_op_type("Conv"); + + const TShape kernel = conv_param.kernel; + const TShape stride = conv_param.stride; + const TShape dilate = conv_param.dilate; + const TShape pad = conv_param.pad; + // const uint32_t num_filter = conv_param.num_filter; Review comment: Let's remove this comment if the code is not needed. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
