piiswrong commented on a change in pull request #10451: [MXNET-432] Add Foreach URL: https://github.com/apache/incubator-mxnet/pull/10451#discussion_r189704747
########## File path: src/operator/nn/control_flow.cc ########## @@ -0,0 +1,532 @@ +/* + * 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. + */ + +#include <mxnet/io.h> +#include <mxnet/base.h> +#include <mxnet/ndarray.h> +#include <mxnet/operator.h> +#include <mxnet/operator_util.h> +#include <dmlc/logging.h> +#include <dmlc/optional.h> +#include "../operator_common.h" +#include "../elemwise_op_common.h" +#include "../../imperative/imperative_utils.h" +#include "./subgraph_op_common.h" + +namespace mxnet { +namespace op { + +struct ForeachParam : public dmlc::Parameter<ForeachParam> { + int num_args; + int dim; + int num_outputs; + int num_out_data; + nnvm::Tuple<dim_t> in_state_locs; + nnvm::Tuple<dim_t> in_data_locs; + DMLC_DECLARE_PARAMETER(ForeachParam) { + DMLC_DECLARE_FIELD(num_args).set_lower_bound(1) + .describe("Number of inputs."); + DMLC_DECLARE_FIELD(dim).set_default(1) + .describe("the dimension of the input array to iterate."); + DMLC_DECLARE_FIELD(num_outputs) + .describe("The number of outputs of the subgraph."); + DMLC_DECLARE_FIELD(num_out_data) + .describe("The number of output data of the subgraph."); + DMLC_DECLARE_FIELD(in_state_locs) + .describe("The locations of loop states among the inputs."); + DMLC_DECLARE_FIELD(in_data_locs) + .describe("The locations of input data among the inputs."); + } +}; // struct ForeachParam + +DMLC_REGISTER_PARAMETER(ForeachParam); + +class ForeachState { + // These are output arrays from all iterations. + // They also contain the Op state for each CachedOp. + std::vector<std::vector<NDArray> > all_outputs; + std::vector<std::vector<NDArray> > all_inputs; + std::vector<std::vector<NDArray> > all_gradients; + std::vector<CachedOpPtr> iter_ops; + + public: + Symbol subgraph_sym; + nnvm::Graph subgraph; + ForeachParam params; + + ForeachState(const Symbol &g, const ForeachParam ¶ms) { + this->subgraph_sym = g; + this->subgraph.outputs = g.outputs; + this->params = params; + } + + void Forward(std::vector<NDArray> cinputs, + const std::vector<OpReqType>& req, + std::vector<NDArray> coutputs, bool is_recording); + void Backward(int iter_no, std::vector<NDArray> ograds, + const std::vector<OpReqType> &req, + std::vector<NDArray> igrads); + void Cleanup() { + all_outputs.clear(); + all_inputs.clear(); + all_gradients.clear(); + iter_ops.clear(); + } +}; + +void ForeachState::Forward(const std::vector<NDArray> &cinputs, + const std::vector<OpReqType>& req, + const std::vector<NDArray> &coutputs, bool is_recording) { + using namespace nnvm; + using namespace imperative; + + bool orig_is_record; + if (is_recording) + orig_is_record = Imperative::Get()->set_is_recording(true); + else + orig_is_record = Imperative::Get()->is_recording(); + + std::vector<NDArray *> inputs(cinputs.size()); + std::vector<NDArray *> outputs(coutputs.size()); + for (size_t i = 0; i < inputs.size(); i++) + inputs[i] = &cinputs[i]; + for (size_t i = 0; i < outputs.size(); i++) + outputs[i] = &coutputs[i]; + + if (is_recording) { + all_inputs.push_back(cinputs); + std::vector<NDArray> gradients(cinputs.size()); + std::vector<NDArray *> input_ptrs(cinputs.size()); + std::vector<NDArray *> gradient_ptrs(cinputs.size()); + std::vector<mx_uint> grad_reqs(cinputs.size()); + for (size_t i = 0; i < gradients.size(); i++) { + gradients[i] = NDArray(cinputs[i].shape(), cinputs[i].ctx(), + true, cinputs[i].dtype()); + input_ptrs[i] = &cinputs[i]; + gradient_ptrs[i] = &gradients[i]; + grad_reqs[i] = kWriteTo; + } + Imperative::Get()->MarkVariables(input_ptrs, grad_reqs, gradient_ptrs);; + } + + std::vector<std::pair<std::string, std::string> > kwargs; + kwargs.push_back(std::pair<std::string, std::string>("inline_limit", "0")); + // Get input names. + const auto& idx = subgraph.indexed_graph(); + std::vector<std::string> arg_names(idx.input_nodes().size()); + for (size_t i = 0; i < idx.input_nodes().size(); ++i) + arg_names[i] = idx[idx.input_nodes()[i]].source->attrs.name; + // We don't have parameters for the cached op. + std::unordered_map<std::string, std::vector<NDArray> > params; + CachedOpPtr op = std::make_shared<Imperative::CachedOp>(subgraph_sym, kwargs, + arg_names, params); + // TODO(zhengda) we need to avoid shape inference and memory plan whenever the op is Review comment: why not allocate memory? ---------------------------------------------------------------- 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
