leleamol commented on a change in pull request #13680: [MXNET-1121] Example to demonstrate the inference workflow using RNN URL: https://github.com/apache/incubator-mxnet/pull/13680#discussion_r254424142
########## File path: cpp-package/example/inference/sentiment_analysis_rnn.cpp ########## @@ -0,0 +1,495 @@ +/* + * 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. + */ + +/* + * This example demonstrates sentiment prediction workflow with pre-trained RNN model using MXNet C++ API. + * The example performs following tasks. + * 1. Load the pre-trained RNN model, + * 2. Load the dictionary file that contains word to index mapping. + * 3. Create executors for pre-determined input lengths. + * 4. Convert each line in the input to the vector of indices. + * 5. Predictor finds the right executor for each line. + * 4. Run the forward pass for each line and predicts the sentiment scores. + * The example uses a pre-trained RNN model that is trained with the IMDB dataset. + */ + +#include <sys/stat.h> +#include <iostream> +#include <fstream> +#include <cstdlib> +#include <map> +#include <string> +#include <algorithm> +#include <vector> +#include <sstream> +#include "mxnet-cpp/MxNetCpp.h" + +using namespace mxnet::cpp; + +static const int DEFAULT_BUCKET_KEYS[] = {30, 25, 20, 15, 10, 5}; +static const char DEFAULT_S3_URL[] = "https://s3.amazonaws.com/mxnet-cpp/RNN_model/"; + + +/* + * class Predictor + * + * This class encapsulates the functionality to load the model, process input image and run the forward pass. + */ + +class Predictor { + public: + Predictor() {} + Predictor(const std::string& model_json, + const std::string& model_params, + const std::string& input_dictionary, + const std::vector<int>& bucket_keys, + bool use_gpu = false); + float PredictSentiment(const std::string &input_review); + ~Predictor(); + + private: + void LoadModel(const std::string& model_json_file); + void LoadParameters(const std::string& model_parameters_file); + void LoadDictionary(const std::string &input_dictionary); + inline bool FileExists(const std::string& name) { + struct stat buffer; + return (stat(name.c_str(), &buffer) == 0); + } + float PredictSentimentForOneLine(const std::string &input_line); + int ConvertToIndexVector(const std::string& input, + std::vector<float> *input_vector); + int GetIndexForOutputSymbolName(const std::string& output_symbol_name); + float GetIndexForWord(const std::string& word); + int GetClosestBucketKey(int num_words); + + std::map<std::string, NDArray> args_map; + std::map<std::string, NDArray> aux_map; + std::map<std::string, int> wordToIndex; + Symbol net; + std::map<int, Executor*> executor_buckets; + Context global_ctx = Context::cpu(); +}; + + +/* + * The constructor takes the following parameters as input: + * 1. model_json: The RNN model in json formatted file. + * 2. model_params: File containing model parameters + * 3. input_dictionary: File containing the word and associated index. + * 4. bucket_keys: A vector of bucket keys for creating executors. + * + * The constructor: + * 1. Loads the model and parameter files. + * 2. Loads the dictionary file to create index to word and word to index maps. + * 3. For each bucket key in the input vector of bucket keys, it creates an executor. + * The executors share the memory. The bucket key determines the length of input data + * required for that executor. + * 4. Creates a map of bucket key to corresponding executor. + * 5. The model is loaded only once. The executors share the memory for the parameters. + */ +Predictor::Predictor(const std::string& model_json, + const std::string& model_params, + const std::string& input_dictionary, + const std::vector<int>& bucket_keys, + bool use_gpu) { + if (use_gpu) { + global_ctx = Context::gpu(); + } + + /* + * Load the dictionary file that contains the word and its index. + * The function creates word to index and index to word map. The maps are used to create index + * vector for the input sentence. + */ + LoadDictionary(input_dictionary); + + // Load the model + LoadModel(model_json); + + // Load the model parameters. + LoadParameters(model_params); + + /* + * Create the executors for each bucket key. The bucket key represents the shape of input data. + * The executors will share the memory by using following technique: + * 1. Infer the executor arrays and bind the first executor with the first bucket key. + * 2. Then for creating the next bucket key, adjust the shape of input argument to match that key. + * 3. Create the executor for the next bucket key by passing the inferred executor arrays and + * pointer to the executor created for the first key. + */ + std::vector<NDArray> arg_arrays; + std::vector<NDArray> grad_arrays; + std::vector<OpReqType> grad_reqs; + std::vector<NDArray> aux_arrays; + + /* + * Create master executor with highest bucket key for optimizing the shared memory between the + * executors for the remaining bucket keys. + */ + int highest_bucket_key = *(std::max_element(bucket_keys.begin(), bucket_keys.end())); + args_map["data0"] = NDArray(Shape(highest_bucket_key, 1), global_ctx, false); + args_map["data1"] = NDArray(Shape(1), global_ctx, false); + + net.InferExecutorArrays(global_ctx, &arg_arrays, &grad_arrays, &grad_reqs, + &aux_arrays, args_map, std::map<std::string, NDArray>(), + std::map<std::string, OpReqType>(), aux_map); + Executor *master_executor = net.Bind(global_ctx, arg_arrays, grad_arrays, grad_reqs, aux_arrays, + std::map<std::string, Context>(), nullptr); + executor_buckets[highest_bucket_key] = master_executor; + + for (int bucket : bucket_keys) { + if (executor_buckets.find(bucket) == executor_buckets.end()) { + arg_arrays[0] = NDArray(Shape(bucket, 1), global_ctx, false); + Executor *executor = net.Bind(global_ctx, arg_arrays, grad_arrays, grad_reqs, aux_arrays, + std::map<std::string, Context>(), master_executor); + executor_buckets[bucket] = executor; + } + } +} + + +/* + * The following function loads the model from json file. + */ +void Predictor::LoadModel(const std::string& model_json_file) { + if (!FileExists(model_json_file)) { + LG << "Model file " << model_json_file << " does not exist"; + throw std::runtime_error("Model file does not exist"); + } + LG << "Loading the model from " << model_json_file << std::endl; + net = Symbol::Load(model_json_file); +} + + +/* + * The following function loads the model parameters. + */ +void Predictor::LoadParameters(const std::string& model_parameters_file) { + if (!FileExists(model_parameters_file)) { + LG << "Parameter file " << model_parameters_file << " does not exist"; + throw std::runtime_error("Model parameters does not exist"); + } + LG << "Loading the model parameters from " << model_parameters_file << std::endl; + std::map<std::string, NDArray> parameters; + NDArray::Load(model_parameters_file, 0, ¶meters); + for (const auto &k : parameters) { + if (k.first.substr(0, 4) == "aux:") { + auto name = k.first.substr(4, k.first.size() - 4); + aux_map[name] = k.second.Copy(global_ctx); + } + if (k.first.substr(0, 4) == "arg:") { + auto name = k.first.substr(4, k.first.size() - 4); + args_map[name] = k.second.Copy(global_ctx); + } + } + /*WaitAll is need when we copy data between GPU and the main memory*/ + NDArray::WaitAll(); +} + + +/* + * The following function loads the dictionary file. + * The function constructs the word to index and index to word maps. + * These maps will be used to represent words in the input sentence to their indices. + * Ensure to use the same dictionary file that was used for training the network. + */ +void Predictor::LoadDictionary(const std::string& input_dictionary) { + if (!FileExists(input_dictionary)) { + LG << "Dictionary file " << input_dictionary << " does not exist"; + throw std::runtime_error("Dictionary file does not exist"); + } + LG << "Loading the dictionary file."; + std::ifstream fi(input_dictionary.c_str()); + if (!fi.is_open()) { + std::cerr << "Error opening dictionary file " << input_dictionary << std::endl; + assert(false); + } + + std::string line; + std::string word; + int index; + while (std::getline(fi, line)) { + std::istringstream stringline(line); + stringline >> word >> index; + wordToIndex[word] = index; + } + fi.close(); +} + + +/* + * The function returns the index associated with the word in the dictionary. + * If the word is not present, the index representing "<unk>" is returned. + * If the "<unk>" is not present then 0 is returned. + */ +float Predictor::GetIndexForWord(const std::string& word) { + if (wordToIndex.find(word) == wordToIndex.end()) { + if (wordToIndex.find("<unk>") == wordToIndex.end()) + return 0; + else + return static_cast<float>(wordToIndex["<unk>"]); + } + return static_cast<float>(wordToIndex[word]); +} + +/* + * The function populates the input vector with indices from the dictionary that + * correspond to the words in the input string. + * The function returns the number of words in the input line. + */ +int Predictor::ConvertToIndexVector(const std::string& input, std::vector<float> *input_vector) { + std::istringstream input_string(input); + input_vector->clear(); + const char delimiter = ' '; + std::string token; + size_t words = 0; + while (std::getline(input_string, token, delimiter) && (words <= input_vector->size())) { + input_vector->push_back(GetIndexForWord(token)); + words++; + } + return words; +} + + +/* + * The function returns the index at which the given symbol name will appear + * in the output vector of NDArrays obtained after running the forward pass on the executor. + */ +int Predictor::GetIndexForOutputSymbolName(const std::string& output_symbol_name) { + int index = 0; + for (const std::string op : net.ListOutputs()) { + if (op == output_symbol_name) { + return index; + } else { + index++; + } + } + throw std::runtime_error("The output symbol name can not be found"); +} + + +/* + * The function finds the closest bucket for the given num_words in the input line. + * If the exact bucket key exists, function returns that bucket key. + * If the matching bucket key does not exist, function looks for the next bucket key + * that is greater than given num_words. + * If the next larger bucket does not exist, function looks for the previous bucket key + * that is lesser than given num_words. + */ +int Predictor::GetClosestBucketKey(int num_words) { Review comment: @szha I agree, with lower_bounds we will get matching or a smaller bucket key. My concern is that, we will always find such bucket and we will trim the input sequence. In that process, we would lose some of the information. Will it be acceptable? ---------------------------------------------------------------- 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