diff options
Diffstat (limited to 'examples/graph_lenet.cpp')
-rw-r--r-- | examples/graph_lenet.cpp | 72 |
1 files changed, 39 insertions, 33 deletions
diff --git a/examples/graph_lenet.cpp b/examples/graph_lenet.cpp index 7b475c2c03..7d6dce7b17 100644 --- a/examples/graph_lenet.cpp +++ b/examples/graph_lenet.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 ARM Limited. + * Copyright (c) 2017-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/graph.h" + #include "support/ToolchainSupport.h" #include "utils/CommonGraphOptions.h" #include "utils/GraphUtils.h" @@ -35,8 +36,7 @@ using namespace arm_compute::graph_utils; class GraphLenetExample : public Example { public: - GraphLenetExample() - : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "LeNet") + GraphLenetExample() : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "LeNet") { } bool do_setup(int argc, char **argv) override @@ -49,14 +49,15 @@ public: common_params = consume_common_graph_parameters(common_opts); // Return when help menu is requested - if(common_params.help) + if (common_params.help) { cmd_parser.print_help(argv[0]); return false; } // Checks - ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), "QASYMM8 not supported for this graph"); + ARM_COMPUTE_EXIT_ON_MSG(arm_compute::is_data_type_quantized_asymmetric(common_params.data_type), + "QASYMM8 not supported for this graph"); // Print parameter values std::cout << common_params << std::endl; @@ -67,43 +68,39 @@ public: // Create input descriptor const auto operation_layout = common_params.data_layout; - const TensorShape tensor_shape = permute_shape(TensorShape(28U, 28U, 1U, batches), DataLayout::NCHW, operation_layout); - TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); + const TensorShape tensor_shape = + permute_shape(TensorShape(28U, 28U, 1U, batches), DataLayout::NCHW, operation_layout); + TensorDescriptor input_descriptor = + TensorDescriptor(tensor_shape, common_params.data_type).set_layout(operation_layout); // Set weights trained layout const DataLayout weights_layout = DataLayout::NCHW; //conv1 << pool1 << conv2 << pool2 << fc1 << act1 << fc2 << smx - graph << common_params.target - << common_params.fast_math_hint + graph << common_params.target << common_params.fast_math_hint << InputLayer(input_descriptor, get_input_accessor(common_params)) << ConvolutionLayer( - 5U, 5U, 20U, - get_weights_accessor(data_path, "/cnn_data/lenet_model/conv1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/lenet_model/conv1_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv1") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool1") + 5U, 5U, 20U, get_weights_accessor(data_path, "/cnn_data/lenet_model/conv1_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/lenet_model/conv1_b.npy"), PadStrideInfo(1, 1, 0, 0)) + .set_name("conv1") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))) + .set_name("pool1") << ConvolutionLayer( - 5U, 5U, 50U, - get_weights_accessor(data_path, "/cnn_data/lenet_model/conv2_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/lenet_model/conv2_b.npy"), - PadStrideInfo(1, 1, 0, 0)) - .set_name("conv2") - << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))).set_name("pool2") - << FullyConnectedLayer( - 500U, - get_weights_accessor(data_path, "/cnn_data/lenet_model/ip1_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/lenet_model/ip1_b.npy")) - .set_name("ip1") + 5U, 5U, 50U, get_weights_accessor(data_path, "/cnn_data/lenet_model/conv2_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/lenet_model/conv2_b.npy"), PadStrideInfo(1, 1, 0, 0)) + .set_name("conv2") + << PoolingLayer(PoolingLayerInfo(PoolingType::MAX, 2, operation_layout, PadStrideInfo(2, 2, 0, 0))) + .set_name("pool2") + << FullyConnectedLayer(500U, + get_weights_accessor(data_path, "/cnn_data/lenet_model/ip1_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/lenet_model/ip1_b.npy")) + .set_name("ip1") << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)).set_name("relu") - << FullyConnectedLayer( - 10U, - get_weights_accessor(data_path, "/cnn_data/lenet_model/ip2_w.npy", weights_layout), - get_weights_accessor(data_path, "/cnn_data/lenet_model/ip2_b.npy")) - .set_name("ip2") - << SoftmaxLayer().set_name("prob") - << OutputLayer(get_output_accessor(common_params)); + << FullyConnectedLayer(10U, + get_weights_accessor(data_path, "/cnn_data/lenet_model/ip2_w.npy", weights_layout), + get_weights_accessor(data_path, "/cnn_data/lenet_model/ip2_b.npy")) + .set_name("ip2") + << SoftmaxLayer().set_name("prob") << OutputLayer(get_output_accessor(common_params)); // Finalize graph GraphConfig config; @@ -111,6 +108,7 @@ public: config.use_tuner = common_params.enable_tuner; config.tuner_mode = common_params.tuner_mode; config.tuner_file = common_params.tuner_file; + config.mlgo_file = common_params.mlgo_file; graph.finalize(common_params.target, config); @@ -131,6 +129,14 @@ private: /** Main program for LeNet * + * Model is based on: + * http://yann.lecun.com/exdb/publis/pdf/lecun-98.pdf + * "Gradient-Based Learning Applied to Document Recognition" + * Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner + * + * The original model uses tanh instead of relu activations. However the use of relu activations in lenet has been + * widely adopted to improve accuracy.* + * * @note To list all the possible arguments execute the binary appended with the --help option * * @param[in] argc Number of arguments |