Commit 39a33e5c authored by Raphael Zumer's avatar Raphael Zumer

Add libaom ML framework

parent 8ac57e6b
/*
* Copyright (c) 2016, Alliance for Open Media. All rights reserved
*
* This source code is subject to the terms of the BSD 2 Clause License and
* the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
* was not distributed with this source code in the LICENSE file, you can
* obtain it at www.aomedia.org/license/software. If the Alliance for Open
* Media Patent License 1.0 was not distributed with this source code in the
* PATENTS file, you can obtain it at www.aomedia.org/license/patent.
*/
#include <assert.h>
#include "av1/encoder/ml.h"
void av1_nn_predict(const float *features, const NN_CONFIG *nn_config,
float *output) {
int num_input_nodes = nn_config->num_inputs;
int buf_index = 0;
float buf[2][NN_MAX_NODES_PER_LAYER];
const float *input_nodes = features;
// Propagate hidden layers.
const int num_layers = nn_config->num_hidden_layers;
assert(num_layers <= NN_MAX_HIDDEN_LAYERS);
for (int layer = 0; layer < num_layers; ++layer) {
const float *weights = nn_config->weights[layer];
const float *bias = nn_config->bias[layer];
float *output_nodes = buf[buf_index];
const int num_output_nodes = nn_config->num_hidden_nodes[layer];
assert(num_output_nodes < NN_MAX_NODES_PER_LAYER);
for (int node = 0; node < num_output_nodes; ++node) {
float val = 0.0f;
for (int i = 0; i < num_input_nodes; ++i)
val += weights[i] * input_nodes[i];
val += bias[node];
// ReLU as activation function.
val = val > 0.0f ? val : 0.0f; // Could use AOMMAX().
output_nodes[node] = val;
weights += num_input_nodes;
}
num_input_nodes = num_output_nodes;
input_nodes = output_nodes;
buf_index = 1 - buf_index;
}
// Final output layer.
const float *weights = nn_config->weights[num_layers];
for (int node = 0; node < nn_config->num_outputs; ++node) {
const float *bias = nn_config->bias[num_layers];
float val = 0.0f;
for (int i = 0; i < num_input_nodes; ++i)
val += weights[i] * input_nodes[i];
output[node] = val + bias[node];
weights += num_input_nodes;
}
}
/*
* Copyright (c) 2016, Alliance for Open Media. All rights reserved
*
* This source code is subject to the terms of the BSD 2 Clause License and
* the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
* was not distributed with this source code in the LICENSE file, you can
* obtain it at www.aomedia.org/license/software. If the Alliance for Open
* Media Patent License 1.0 was not distributed with this source code in the
* PATENTS file, you can obtain it at www.aomedia.org/license/patent.
*/
#ifndef AV1_ENCODER_ML_H_
#define AV1_ENCODER_ML_H_
#ifdef __cplusplus
extern "C" {
#endif
#define NN_MAX_HIDDEN_LAYERS 10
#define NN_MAX_NODES_PER_LAYER 128
typedef struct {
int num_inputs; // Number of input nodes, i.e. features.
int num_outputs; // Number of output nodes.
int num_hidden_layers; // Number of hidden layers, maximum 10.
// Number of nodes for each hidden layer.
int num_hidden_nodes[NN_MAX_HIDDEN_LAYERS];
// Weight parameters, indexed by layer.
const float *weights[NN_MAX_HIDDEN_LAYERS + 1];
// Bias parameters, indexed by layer.
const float *bias[NN_MAX_HIDDEN_LAYERS + 1];
} NN_CONFIG;
// Calculate prediction based on the given input features and neural net config.
// Assume there are no more than NN_MAX_NODES_PER_LAYER nodes in each hidden
// layer.
void av1_nn_predict(const float *features, const NN_CONFIG *nn_config,
float *output);
#ifdef __cplusplus
} // extern "C"
#endif
#endif // AV1_ENCODER_RD_H_
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment