1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358
|
/*
* 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<SubStream, SubStream> new_state_1 =
add_lstm_cell(unstack_nid, 0, previous_state, previous_state, add_y, lstm_fc_weights, lstm_fc_bias);
std::pair<SubStream, SubStream> 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<SubStream, SubStream> 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<SubStream, SubStream> 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<SubStream, SubStream> 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<SubStream, SubStream> 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<SubStream, SubStream> 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<SubStream, SubStream> 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<SubStream, SubStream> 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<SubStream, SubStream> 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<SubStream, SubStream> 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<SubStream, SubStream> 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<SubStream, SubStream> 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<SubStream, SubStream> 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<SubStream, SubStream> 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<SubStream, SubStream> 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<SubStream, SubStream> 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<ConcatenateLayerNode>(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<ActivationLayerNode>(
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<ActivationLayerNode>(
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<EltwiseLayerNode>(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<ActivationLayerNode>(
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<SubStream, SubStream>(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<GraphDeepSpeechExample>(argc, argv);
}
|