1 ///////////////////////////////////////////////////////////////////////
2 // File: convolve.cpp
3 // Description: Convolutional layer that stacks the inputs over its rectangle
4 // and pulls in random data to fill out-of-input inputs.
5 // Output is therefore same size as its input, but deeper.
6 // Author: Ray Smith
7 //
8 // (C) Copyright 2014, Google Inc.
9 // Licensed under the Apache License, Version 2.0 (the "License");
10 // you may not use this file except in compliance with the License.
11 // You may obtain a copy of the License at
12 // http://www.apache.org/licenses/LICENSE-2.0
13 // Unless required by applicable law or agreed to in writing, software
14 // distributed under the License is distributed on an "AS IS" BASIS,
15 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
16 // See the License for the specific language governing permissions and
17 // limitations under the License.
18 ///////////////////////////////////////////////////////////////////////
19
20 #ifdef HAVE_CONFIG_H
21 # include "config_auto.h"
22 #endif
23
24 #include "convolve.h"
25
26 #include "networkscratch.h"
27 #include "serialis.h"
28
29 namespace tesseract {
30
Convolve(const std::string & name,int ni,int half_x,int half_y)31 Convolve::Convolve(const std::string &name, int ni, int half_x, int half_y)
32 : Network(NT_CONVOLVE, name, ni, ni * (2 * half_x + 1) * (2 * half_y + 1))
33 , half_x_(half_x)
34 , half_y_(half_y) {}
35
36 // Writes to the given file. Returns false in case of error.
Serialize(TFile * fp) const37 bool Convolve::Serialize(TFile *fp) const {
38 return Network::Serialize(fp) && fp->Serialize(&half_x_) && fp->Serialize(&half_y_);
39 }
40
41 // Reads from the given file. Returns false in case of error.
DeSerialize(TFile * fp)42 bool Convolve::DeSerialize(TFile *fp) {
43 if (!fp->DeSerialize(&half_x_)) {
44 return false;
45 }
46 if (!fp->DeSerialize(&half_y_)) {
47 return false;
48 }
49 no_ = ni_ * (2 * half_x_ + 1) * (2 * half_y_ + 1);
50 return true;
51 }
52
53 // Runs forward propagation of activations on the input line.
54 // See NetworkCpp for a detailed discussion of the arguments.
Forward(bool debug,const NetworkIO & input,const TransposedArray * input_transpose,NetworkScratch * scratch,NetworkIO * output)55 void Convolve::Forward(bool debug, const NetworkIO &input, const TransposedArray *input_transpose,
56 NetworkScratch *scratch, NetworkIO *output) {
57 output->Resize(input, no_);
58 int y_scale = 2 * half_y_ + 1;
59 StrideMap::Index dest_index(output->stride_map());
60 do {
61 // Stack x_scale groups of y_scale * ni_ inputs together.
62 int t = dest_index.t();
63 int out_ix = 0;
64 for (int x = -half_x_; x <= half_x_; ++x, out_ix += y_scale * ni_) {
65 StrideMap::Index x_index(dest_index);
66 if (!x_index.AddOffset(x, FD_WIDTH)) {
67 // This x is outside the image.
68 output->Randomize(t, out_ix, y_scale * ni_, randomizer_);
69 } else {
70 int out_iy = out_ix;
71 for (int y = -half_y_; y <= half_y_; ++y, out_iy += ni_) {
72 StrideMap::Index y_index(x_index);
73 if (!y_index.AddOffset(y, FD_HEIGHT)) {
74 // This y is outside the image.
75 output->Randomize(t, out_iy, ni_, randomizer_);
76 } else {
77 output->CopyTimeStepGeneral(t, out_iy, ni_, input, y_index.t(), 0);
78 }
79 }
80 }
81 }
82 } while (dest_index.Increment());
83 #ifndef GRAPHICS_DISABLED
84 if (debug) {
85 DisplayForward(*output);
86 }
87 #endif
88 }
89
90 // Runs backward propagation of errors on the deltas line.
91 // See NetworkCpp for a detailed discussion of the arguments.
Backward(bool debug,const NetworkIO & fwd_deltas,NetworkScratch * scratch,NetworkIO * back_deltas)92 bool Convolve::Backward(bool debug, const NetworkIO &fwd_deltas, NetworkScratch *scratch,
93 NetworkIO *back_deltas) {
94 back_deltas->Resize(fwd_deltas, ni_);
95 NetworkScratch::IO delta_sum;
96 delta_sum.ResizeFloat(fwd_deltas, ni_, scratch);
97 delta_sum->Zero();
98 int y_scale = 2 * half_y_ + 1;
99 StrideMap::Index src_index(fwd_deltas.stride_map());
100 do {
101 // Stack x_scale groups of y_scale * ni_ inputs together.
102 int t = src_index.t();
103 int out_ix = 0;
104 for (int x = -half_x_; x <= half_x_; ++x, out_ix += y_scale * ni_) {
105 StrideMap::Index x_index(src_index);
106 if (x_index.AddOffset(x, FD_WIDTH)) {
107 int out_iy = out_ix;
108 for (int y = -half_y_; y <= half_y_; ++y, out_iy += ni_) {
109 StrideMap::Index y_index(x_index);
110 if (y_index.AddOffset(y, FD_HEIGHT)) {
111 fwd_deltas.AddTimeStepPart(t, out_iy, ni_, delta_sum->f(y_index.t()));
112 }
113 }
114 }
115 }
116 } while (src_index.Increment());
117 back_deltas->CopyAll(*delta_sum);
118 return true;
119 }
120
121 } // namespace tesseract.
122