/* * Copyright (c) 2019-2021 Arm Limited. * * SPDX-License-Identifier: MIT * * Permission is hereby granted, free of charge, to any person obtaining a copy * of this software and associated documentation files (the "Software"), to * deal in the Software without restriction, including without limitation the * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or * sell copies of the Software, and to permit persons to whom the Software is * furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice shall be included in all * copies or substantial portions of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ #include "arm_compute/graph.h" #include "arm_compute/graph/Types.h" #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" #include "utils/Utils.h" using namespace arm_compute::utils; using namespace arm_compute::graph; using namespace arm_compute::graph::frontend; using namespace arm_compute::graph_utils; /** Example demonstrating how to implement DeepSpeech v0.4.1's network using the Compute Library's graph API */ class GraphDeepSpeechExample : public Example { public: GraphDeepSpeechExample() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "DeepSpeech v0.4.1") { } bool do_setup(int argc, char **argv) override { // Parse arguments cmd_parser.parse(argc, argv); cmd_parser.validate(); // Consume common parameters common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested if(common_params.help) { cmd_parser.print_help(argv[0]); return false; } // Print parameter values std::cout << common_params << std::endl; // Get trainable parameters data path std::string data_path = common_params.data_path; const std::string model_path = "/cnn_data/deepspeech_model/"; if(!data_path.empty()) { data_path += model_path; } // How many timesteps to process at once, higher values mean more latency // Notice that this corresponds to the number of LSTM cells that will be instantiated const unsigned int n_steps = 16; // ReLU clipping value for non-recurrent layers const float cell_clip = 20.f; // Create input descriptor const TensorShape tensor_shape = permute_shape(TensorShape(26U, 19U, n_steps, 1U), DataLayout::NHWC, common_params.data_layout); TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NHWC; graph << common_params.target << common_params.fast_math_hint << InputLayer(input_descriptor, get_weights_accessor(data_path, "input_values_x" + std::to_string(n_steps) + ".npy", weights_layout)) .set_name("input_node"); if(common_params.data_layout == DataLayout::NCHW) { graph << PermuteLayer(PermutationVector(2U, 0U, 1U), common_params.data_layout).set_name("permute_to_nhwc"); } graph << ReshapeLayer(TensorShape(494U, n_steps)).set_name("Reshape_input") // Layer 1 << FullyConnectedLayer( 2048U, get_weights_accessor(data_path, "h1_transpose.npy", weights_layout), get_weights_accessor(data_path, "MatMul_bias.npy")) .set_name("fc0") << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, cell_clip)) .set_name("Relu") // Layer 2 << FullyConnectedLayer( 2048U, get_weights_accessor(data_path, "h2_transpose.npy", weights_layout), get_weights_accessor(data_path, "MatMul_1_bias.npy")) .set_name("fc1") << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, cell_clip)) .set_name("Relu_1") // Layer 3 << FullyConnectedLayer( 2048U, get_weights_accessor(data_path, "h3_transpose.npy", weights_layout), get_weights_accessor(data_path, "MatMul_2_bias.npy")) .set_name("fc2") << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, cell_clip)) .set_name("Relu_2") // Layer 4 << ReshapeLayer(TensorShape(2048U, 1U, n_steps)).set_name("Reshape_1"); // Unstack Layer (using SplitLayerNode) NodeParams unstack_params = { "unstack", graph.hints().target_hint }; NodeID unstack_nid = GraphBuilder::add_split_node(graph.graph(), unstack_params, { graph.tail_node(), 0 }, n_steps, 2); // Create input state descriptor TensorDescriptor state_descriptor = TensorDescriptor(TensorShape(2048U), common_params.data_type).set_layout(common_params.data_layout); SubStream previous_state(graph); SubStream add_y(graph); // Initial state for LSTM is all zeroes for both state_h and state_c, therefore only one input is created previous_state << InputLayer(state_descriptor, get_weights_accessor(data_path, "zeros.npy")) .set_name("previous_state_c_h"); add_y << InputLayer(state_descriptor, get_weights_accessor(data_path, "ones.npy")) .set_name("add_y"); // Create LSTM Fully Connected weights and bias descriptors TensorDescriptor lstm_weights_descriptor = TensorDescriptor(TensorShape(4096U, 8192U), common_params.data_type).set_layout(common_params.data_layout); TensorDescriptor lstm_bias_descriptor = TensorDescriptor(TensorShape(8192U), common_params.data_type).set_layout(common_params.data_layout); SubStream lstm_fc_weights(graph); SubStream lstm_fc_bias(graph); lstm_fc_weights << ConstantLayer(lstm_weights_descriptor, get_weights_accessor(data_path, "rnn_lstm_cell_kernel_transpose.npy", weights_layout)) .set_name("h5/transpose"); lstm_fc_bias << ConstantLayer(lstm_bias_descriptor, get_weights_accessor(data_path, "rnn_lstm_cell_MatMul_bias.npy")) .set_name("MatMul_3_bias"); // LSTM Block std::pair new_state_1 = add_lstm_cell(unstack_nid, 0, previous_state, previous_state, add_y, lstm_fc_weights, lstm_fc_bias); std::pair new_state_2 = add_lstm_cell(unstack_nid, 1, new_state_1.first, new_state_1.second, add_y, lstm_fc_weights, lstm_fc_bias); std::pair new_state_3 = add_lstm_cell(unstack_nid, 2, new_state_2.first, new_state_2.second, add_y, lstm_fc_weights, lstm_fc_bias); std::pair new_state_4 = add_lstm_cell(unstack_nid, 3, new_state_3.first, new_state_3.second, add_y, lstm_fc_weights, lstm_fc_bias); std::pair new_state_5 = add_lstm_cell(unstack_nid, 4, new_state_4.first, new_state_4.second, add_y, lstm_fc_weights, lstm_fc_bias); std::pair new_state_6 = add_lstm_cell(unstack_nid, 5, new_state_5.first, new_state_5.second, add_y, lstm_fc_weights, lstm_fc_bias); std::pair new_state_7 = add_lstm_cell(unstack_nid, 6, new_state_6.first, new_state_6.second, add_y, lstm_fc_weights, lstm_fc_bias); std::pair new_state_8 = add_lstm_cell(unstack_nid, 7, new_state_7.first, new_state_7.second, add_y, lstm_fc_weights, lstm_fc_bias); std::pair new_state_9 = add_lstm_cell(unstack_nid, 8, new_state_8.first, new_state_8.second, add_y, lstm_fc_weights, lstm_fc_bias); std::pair new_state_10 = add_lstm_cell(unstack_nid, 9, new_state_9.first, new_state_9.second, add_y, lstm_fc_weights, lstm_fc_bias); std::pair new_state_11 = add_lstm_cell(unstack_nid, 10, new_state_10.first, new_state_10.second, add_y, lstm_fc_weights, lstm_fc_bias); std::pair new_state_12 = add_lstm_cell(unstack_nid, 11, new_state_11.first, new_state_11.second, add_y, lstm_fc_weights, lstm_fc_bias); std::pair new_state_13 = add_lstm_cell(unstack_nid, 12, new_state_12.first, new_state_12.second, add_y, lstm_fc_weights, lstm_fc_bias); std::pair new_state_14 = add_lstm_cell(unstack_nid, 13, new_state_13.first, new_state_13.second, add_y, lstm_fc_weights, lstm_fc_bias); std::pair new_state_15 = add_lstm_cell(unstack_nid, 14, new_state_14.first, new_state_14.second, add_y, lstm_fc_weights, lstm_fc_bias); std::pair new_state_16 = add_lstm_cell(unstack_nid, 15, new_state_15.first, new_state_15.second, add_y, lstm_fc_weights, lstm_fc_bias); // Concatenate new states on height const int axis = 1; graph << StackLayer(axis, std::move(new_state_1.second), std::move(new_state_2.second), std::move(new_state_3.second), std::move(new_state_4.second), std::move(new_state_5.second), std::move(new_state_6.second), std::move(new_state_7.second), std::move(new_state_8.second), std::move(new_state_9.second), std::move(new_state_10.second), std::move(new_state_11.second), std::move(new_state_12.second), std::move(new_state_13.second), std::move(new_state_14.second), std::move(new_state_15.second), std::move(new_state_16.second)) .set_name("concat"); graph << FullyConnectedLayer( 2048U, get_weights_accessor(data_path, "h5_transpose.npy", weights_layout), get_weights_accessor(data_path, "MatMul_3_bias.npy")) .set_name("fc3") << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, cell_clip)) .set_name("Relu3") << FullyConnectedLayer( 29U, get_weights_accessor(data_path, "h6_transpose.npy", weights_layout), get_weights_accessor(data_path, "MatMul_4_bias.npy")) .set_name("fc3") << SoftmaxLayer().set_name("logits"); graph << OutputLayer(get_output_accessor(common_params, 5)); // Finalize graph GraphConfig config; config.num_threads = common_params.threads; config.use_tuner = common_params.enable_tuner; config.tuner_file = common_params.tuner_file; config.mlgo_file = common_params.mlgo_file; config.use_synthetic_type = arm_compute::is_data_type_quantized(common_params.data_type); config.synthetic_type = common_params.data_type; graph.finalize(common_params.target, config); return true; } void do_run() override { // Run graph graph.run(); } private: CommandLineParser cmd_parser; CommonGraphOptions common_opts; CommonGraphParams common_params; Stream graph; Status set_node_params(Graph &g, NodeID nid, NodeParams ¶ms) { INode *node = g.node(nid); ARM_COMPUTE_RETURN_ERROR_ON(!node); node->set_common_node_parameters(params); return Status{}; } std::pair add_lstm_cell(NodeID unstack_nid, unsigned int unstack_idx, SubStream previous_state_c, SubStream previous_state_h, SubStream add_y, SubStream lstm_fc_weights, SubStream lstm_fc_bias) { const std::string cell_name("rnn/lstm_cell_" + std::to_string(unstack_idx)); const DataLayoutDimension concat_dim = (common_params.data_layout == DataLayout::NHWC) ? DataLayoutDimension::CHANNEL : DataLayoutDimension::WIDTH; // Concatenate result of Unstack with previous_state_h NodeParams concat_params = { cell_name + "/concat", graph.hints().target_hint }; NodeID concat_nid = graph.graph().add_node(2, concat_dim); graph.graph().add_connection(unstack_nid, unstack_idx, concat_nid, 0); graph.graph().add_connection(previous_state_h.tail_node(), 0, concat_nid, 1); set_node_params(graph.graph(), concat_nid, concat_params); graph.forward_tail(concat_nid); graph << FullyConnectedLayer( 8192U, lstm_fc_weights, lstm_fc_bias) .set_name(cell_name + "/BiasAdd"); // Split Layer const unsigned int num_splits = 4; const unsigned int split_axis = 0; NodeParams split_params = { cell_name + "/split", graph.hints().target_hint }; NodeID split_nid = GraphBuilder::add_split_node(graph.graph(), split_params, { graph.tail_node(), 0 }, num_splits, split_axis); NodeParams sigmoid_1_params = { cell_name + "/Sigmoid_1", graph.hints().target_hint }; NodeParams add_params = { cell_name + "/add", graph.hints().target_hint }; NodeParams sigmoid_2_params = { cell_name + "/Sigmoid_2", graph.hints().target_hint }; NodeParams tanh_params = { cell_name + "/Tanh", graph.hints().target_hint }; // Sigmoid 1 (first split) NodeID sigmoid_1_nid = graph.graph().add_node(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); graph.graph().add_connection(split_nid, 0, sigmoid_1_nid, 0); set_node_params(graph.graph(), sigmoid_1_nid, sigmoid_1_params); // Tanh (second split) NodeID tanh_nid = graph.graph().add_node(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f)); graph.graph().add_connection(split_nid, 1, tanh_nid, 0); set_node_params(graph.graph(), tanh_nid, tanh_params); SubStream tanh_ss(graph); tanh_ss.forward_tail(tanh_nid); // Add (third split) NodeID add_nid = graph.graph().add_node(descriptors::EltwiseLayerDescriptor{ EltwiseOperation::Add }); graph.graph().add_connection(split_nid, 2, add_nid, 0); graph.graph().add_connection(add_y.tail_node(), 0, add_nid, 1); set_node_params(graph.graph(), add_nid, add_params); // Sigmoid 2 (fourth split) NodeID sigmoid_2_nid = graph.graph().add_node(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); graph.graph().add_connection(split_nid, 3, sigmoid_2_nid, 0); set_node_params(graph.graph(), sigmoid_2_nid, sigmoid_2_params); SubStream sigmoid_1_ss(graph); sigmoid_1_ss.forward_tail(sigmoid_1_nid); SubStream mul_1_ss(sigmoid_1_ss); mul_1_ss << EltwiseLayer(std::move(sigmoid_1_ss), std::move(tanh_ss), EltwiseOperation::Mul) .set_name(cell_name + "/mul_1"); SubStream tanh_1_ss_tmp(graph); tanh_1_ss_tmp.forward_tail(add_nid); tanh_1_ss_tmp << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)) .set_name(cell_name + "/Sigmoid"); SubStream tanh_1_ss_tmp2(tanh_1_ss_tmp); tanh_1_ss_tmp2 << EltwiseLayer(std::move(tanh_1_ss_tmp), std::move(previous_state_c), EltwiseOperation::Mul) .set_name(cell_name + "/mul"); SubStream tanh_1_ss(tanh_1_ss_tmp2); tanh_1_ss << EltwiseLayer(std::move(tanh_1_ss_tmp2), std::move(mul_1_ss), EltwiseOperation::Add) .set_name(cell_name + "/new_state_c"); SubStream new_state_c(tanh_1_ss); tanh_1_ss << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f)) .set_name(cell_name + "/Tanh_1"); SubStream sigmoid_2_ss(graph); sigmoid_2_ss.forward_tail(sigmoid_2_nid); graph << EltwiseLayer(std::move(sigmoid_2_ss), std::move(tanh_1_ss), EltwiseOperation::Mul) .set_name(cell_name + "/new_state_h"); SubStream new_state_h(graph); return std::pair(new_state_c, new_state_h); } }; /** Main program for DeepSpeech v0.4.1 * * Model is based on: * https://arxiv.org/abs/1412.5567 * "Deep Speech: Scaling up end-to-end speech recognition" * Awni Hannun, Carl Case, Jared Casper, Bryan Catanzaro, Greg Diamos, Erich Elsen, Ryan Prenger, Sanjeev Satheesh, Shubho Sengupta, Adam Coates, Andrew Y. Ng * * Provenance: https://github.com/mozilla/DeepSpeech * * @note To list all the possible arguments execute the binary appended with the --help option * * @param[in] argc Number of arguments * @param[in] argv Arguments * * @return Return code */ int main(int argc, char **argv) { return arm_compute::utils::run_example(argc, argv); }