blob: d21def43a847baed3d624d35f6d3671da7242d09 (
plain)
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
|
/*
* 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 <math.h>
#include "aom_dsp/aom_dsp_common.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;
}
}
void av1_nn_softmax(const float *input, float *output, int n) {
// Softmax function is invariant to adding the same constant
// to all input values, so we subtract the maximum input to avoid
// possible overflow.
float max_inp = input[0];
for (int i = 1; i < n; i++) max_inp = AOMMAX(max_inp, input[i]);
float sum_out = 0.0f;
for (int i = 0; i < n; i++) {
output[i] = (float)exp(input[i] - max_inp);
sum_out += output[i];
}
for (int i = 0; i < n; i++) output[i] /= sum_out;
}
|
