1 /******************************************************************************/\r
2 /*!\r
3 * \file timlCNNConvUpdateParams.c\r
4 */\r
5 /* Copyright (C) 2015 Texas Instruments Incorporated - http://www.ti.com/\r
6 *\r
7 * Redistribution and use in source and binary forms, with or without\r
8 * modification, are permitted provided that the following conditions\r
9 * are met:\r
10 *\r
11 * Redistributions of source code must retain the above copyright\r
12 * notice, this list of conditions and the following disclaimer.\r
13 *\r
14 * Redistributions in binary form must reproduce the above copyright\r
15 * notice, this list of conditions and the following disclaimer in the\r
16 * documentation and/or other materials provided with the\r
17 * distribution.\r
18 *\r
19 * Neither the name of Texas Instruments Incorporated nor the names of\r
20 * its contributors may be used to endorse or promote products derived\r
21 * from this software without specific prior written permission.\r
22 *\r
23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS\r
24 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT\r
25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR\r
26 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT\r
27 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,\r
28 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT\r
29 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,\r
30 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY\r
31 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\r
32 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE\r
33 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\r
34 *\r
35 ******************************************************************************/\r
36 \r
37 \r
38 /*******************************************************************************\r
39 *\r
40 * INCLUDES\r
41 *\r
42 ******************************************************************************/\r
43 \r
44 #include "../api/timl.h"\r
45 \r
46 \r
47 /******************************************************************************/\r
48 /*!\r
49 * \ingroup cnn\r
50 * \brief Update the parameters of the conv layer\r
51 * \param[in] layer Layer\r
52 * \return Error code\r
53 */\r
54 /******************************************************************************/\r
55 \r
56 int timlCNNConvUpdateParams(timlCNNLayer *layer)\r
57 {\r
58 int err;\r
59 timlConvNeuralNetwork *cnn;\r
60 int prevFeatureMapChannel;\r
61 int currFeatureMapChannel;\r
62 int kernelRow;\r
63 int kernelCol;\r
64 int batchUpdate;\r
65 float kernelLearningRate;\r
66 float biasLearningRate;\r
67 float kernelDecay;\r
68 int deviceId;\r
69 int threadId;\r
70 \r
71 // init\r
72 err = 0;\r
73 cnn = layer->cnn;\r
74 prevFeatureMapChannel = layer->convParams.inputFeatureMapChannel;\r
75 currFeatureMapChannel = layer->convParams.outputFeatureMapChannel;\r
76 kernelRow = layer->convParams.kernelRow;\r
77 kernelCol = layer->convParams.kernelCol;\r
78 batchUpdate = cnn->params.batchUpdate;\r
79 kernelLearningRate = layer->convParams.kernelLearningFactor * cnn->params.learningRate;\r
80 biasLearningRate = layer->convParams.biasLearningFactor * cnn->params.learningRate;\r
81 kernelDecay = layer->convParams.kernelDecayFactor * cnn->params.weightDecay;\r
82 deviceId = cnn->deviceId;\r
83 threadId = cnn->threadId;\r
84 \r
85 // kernelGradAccum = kernelGradAccum/batchUpdate\r
86 timlUtilBLASsscal(kernelRow*kernelCol*prevFeatureMapChannel*currFeatureMapChannel, 1.0/batchUpdate, layer->convParams.kernelGradAccum, deviceId, threadId);\r
87 // kernelGradAccum = kernelGradAccum + kernelDecay * kernel\r
88 timlUtilBLASsaxpy(kernelRow*kernelCol*prevFeatureMapChannel*currFeatureMapChannel, kernelDecay, layer->convParams.kernel, layer->convParams.kernelGradAccum, deviceId, threadId);\r
89 // kernelGradAccum = kernelGradAccum * learningRate\r
90 timlUtilBLASsscal(kernelRow*kernelCol*prevFeatureMapChannel*currFeatureMapChannel, kernelLearningRate, layer->convParams.kernelGradAccum, deviceId, threadId);\r
91 // kernelGradAccum = kernelGradAccum + momentum * kernelInc\r
92 timlUtilBLASsaxpy(kernelRow*kernelCol*prevFeatureMapChannel*currFeatureMapChannel, cnn->params.momentum, layer->convParams.kernelInc, layer->convParams.kernelGradAccum, deviceId, threadId);\r
93 // kernelInc = kernelGradAccum\r
94 timlUtilBLASscopy(kernelRow*kernelCol*prevFeatureMapChannel*currFeatureMapChannel, layer->convParams.kernelGradAccum, layer->convParams.kernelInc, deviceId, threadId);\r
95 // kernel = kernel + (-1) * kernelInc\r
96 timlUtilBLASsaxpy(kernelRow*kernelCol*prevFeatureMapChannel*currFeatureMapChannel, -1.0, layer->convParams.kernelInc, layer->convParams.kernel, deviceId, threadId);\r
97 // reset kernelGradAccum\r
98 timlUtilVectorResetFloat(layer->convParams.kernelGradAccum, kernelRow*kernelCol*prevFeatureMapChannel*currFeatureMapChannel, 0.0, deviceId, threadId);\r
99 \r
100 // biasGradAccum = learningRate * biasGradAccum/batchUpdate\r
101 timlUtilBLASsscal(currFeatureMapChannel, biasLearningRate/batchUpdate, layer->convParams.biasGradAccum, deviceId, threadId);\r
102 // biasInc = biasInc * momentum\r
103 timlUtilBLASsscal(currFeatureMapChannel, cnn->params.momentum, layer->convParams.biasInc, deviceId, threadId);\r
104 // biasInc = biasInc + biasGradAccum\r
105 timlUtilBLASsaxpy(currFeatureMapChannel, 1.0, layer->convParams.biasGradAccum, layer->convParams.biasInc, deviceId, threadId);\r
106 // bias = bias - biasInc\r
107 timlUtilBLASsaxpy(currFeatureMapChannel, -1.0, layer->convParams.biasInc, layer->convParams.bias, deviceId, threadId);\r
108 // reset biasGradAccum\r
109 timlUtilVectorResetFloat(layer->convParams.biasGradAccum, currFeatureMapChannel, 0.0, deviceId, threadId);\r
110 \r
111 return err;\r
112 }\r