tesseract  4.0.0-1-g2a2b
networkio.h
Go to the documentation of this file.
1 // File: networkio.h
3 // Description: Network input/output data, allowing float/int implementations.
4 // Author: Ray Smith
5 // Created: Tue Jun 17 08:43:11 PST 2014
6 //
7 // (C) Copyright 2014, Google Inc.
8 // Licensed under the Apache License, Version 2.0 (the "License");
9 // you may not use this file except in compliance with the License.
10 // You may obtain a copy of the License at
11 // http://www.apache.org/licenses/LICENSE-2.0
12 // Unless required by applicable law or agreed to in writing, software
13 // distributed under the License is distributed on an "AS IS" BASIS,
14 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15 // See the License for the specific language governing permissions and
16 // limitations under the License.
18 
19 #ifndef TESSERACT_LSTM_NETWORKIO_H_
20 #define TESSERACT_LSTM_NETWORKIO_H_
21 
22 #include <cmath>
23 #include <cstdio>
24 #include <vector>
25 
26 #include "genericvector.h"
27 #include "helpers.h"
28 #include "static_shape.h"
29 #include "stridemap.h"
30 #include "weightmatrix.h"
31 
32 struct Pix;
33 
34 namespace tesseract {
35 
36 // Class to contain all the input/output of a network, allowing for fixed or
37 // variable-strided 2d to 1d mapping, and float or int8_t values. Provides
38 // enough calculating functions to hide the detail of the implementation.
39 class NetworkIO {
40  public:
41  NetworkIO() : int_mode_(false) {}
42  // Resizes the array (and stride), avoiding realloc if possible, to the given
43  // size from various size specs:
44  // Same stride size, but given number of features.
45  void Resize(const NetworkIO& src, int num_features) {
46  ResizeToMap(src.int_mode(), src.stride_map(), num_features);
47  }
48  // Resizes to a specific size as a 2-d temp buffer. No batches, no y-dim.
49  void Resize2d(bool int_mode, int width, int num_features);
50  // Resizes forcing a float representation with the stridemap of src and the
51  // given number of features.
52  void ResizeFloat(const NetworkIO& src, int num_features) {
53  ResizeToMap(false, src.stride_map(), num_features);
54  }
55  // Resizes to a specific stride_map.
56  void ResizeToMap(bool int_mode, const StrideMap& stride_map,
57  int num_features);
58  // Shrinks image size by x_scale,y_scale, and use given number of features.
59  void ResizeScaled(const NetworkIO& src, int x_scale, int y_scale,
60  int num_features);
61  // Resizes to just 1 x-coord, whatever the input.
62  void ResizeXTo1(const NetworkIO& src, int num_features);
63  // Initialize all the array to zero.
64  void Zero();
65  // Initializes to zero all elements of the array that do not correspond to
66  // valid image positions. (If a batch of different-sized images are packed
67  // together, then there will be padding pixels.)
68  void ZeroInvalidElements();
69  // Sets up the array from the given image, using the currently set int_mode_.
70  // If the image width doesn't match the shape, the image is truncated or
71  // padded with noise to match.
72  void FromPix(const StaticShape& shape, const Pix* pix, TRand* randomizer);
73  // Sets up the array from the given set of images, using the currently set
74  // int_mode_. If the image width doesn't match the shape, the images are
75  // truncated or padded with noise to match.
76  void FromPixes(const StaticShape& shape, const std::vector<const Pix*>& pixes,
77  TRand* randomizer);
78  // Copies the given pix to *this at the given batch index, stretching and
79  // clipping the pixel values so that [black, black + 2*contrast] maps to the
80  // dynamic range of *this, ie [-1,1] for a float and (-127,127) for int.
81  // This is a 2-d operation in the sense that the output depth is the number
82  // of input channels, the height is the height of the image, and the width
83  // is the width of the image, or truncated/padded with noise if the width
84  // is a fixed size.
85  void Copy2DImage(int batch, Pix* pix, float black, float contrast,
86  TRand* randomizer);
87  // Copies the given pix to *this at the given batch index, as Copy2DImage
88  // above, except that the output depth is the height of the input image, the
89  // output height is 1, and the output width as for Copy2DImage.
90  // The image is thus treated as a 1-d set of vertical pixel strips.
91  void Copy1DGreyImage(int batch, Pix* pix, float black, float contrast,
92  TRand* randomizer);
93  // Helper stores the pixel value in i_ or f_ according to int_mode_.
94  // t: is the index from the StrideMap corresponding to the current
95  // [batch,y,x] position
96  // f: is the index into the depth/channel
97  // pixel: the value of the pixel from the image (in one channel)
98  // black: the pixel value to map to the lowest of the range of *this
99  // contrast: the range of pixel values to stretch to half the range of *this.
100  void SetPixel(int t, int f, int pixel, float black, float contrast);
101  // Converts the array to a Pix. Must be pixDestroyed after use.
102  Pix* ToPix() const;
103  // Prints the first and last num timesteps of the array for each feature.
104  void Print(int num) const;
105 
106  // Returns the timestep width.
107  int Width() const {
108  return int_mode_ ? i_.dim1() : f_.dim1();
109  }
110  // Returns the number of features.
111  int NumFeatures() const {
112  return int_mode_ ? i_.dim2() : f_.dim2();
113  }
114  // Accessor to a timestep of the float matrix.
115  float* f(int t) {
116  ASSERT_HOST(!int_mode_);
117  return f_[t];
118  }
119  const float* f(int t) const {
120  ASSERT_HOST(!int_mode_);
121  return f_[t];
122  }
123  const int8_t* i(int t) const {
124  ASSERT_HOST(int_mode_);
125  return i_[t];
126  }
127  bool int_mode() const {
128  return int_mode_;
129  }
130  void set_int_mode(bool is_quantized) {
131  int_mode_ = is_quantized;
132  }
133  const StrideMap& stride_map() const {
134  return stride_map_;
135  }
136  void set_stride_map(const StrideMap& map) {
137  stride_map_ = map;
138  }
139  const GENERIC_2D_ARRAY<float>& float_array() const { return f_; }
141 
142  // Copies a single time step from src.
143  void CopyTimeStepFrom(int dest_t, const NetworkIO& src, int src_t);
144  // Copies a part of single time step from src.
145  void CopyTimeStepGeneral(int dest_t, int dest_offset, int num_features,
146  const NetworkIO& src, int src_t, int src_offset);
147  // Zeroes a single time step.
148  void ZeroTimeStep(int t) { ZeroTimeStepGeneral(t, 0, NumFeatures()); }
149  void ZeroTimeStepGeneral(int t, int offset, int num_features);
150  // Sets the given range to random values.
151  void Randomize(int t, int offset, int num_features, TRand* randomizer);
152 
153  // Helper returns the label and score of the best choice over a range.
154  int BestChoiceOverRange(int t_start, int t_end, int not_this, int null_ch,
155  float* rating, float* certainty) const;
156  // Helper returns the rating and certainty of the choice over a range in t.
157  void ScoresOverRange(int t_start, int t_end, int choice, int null_ch,
158  float* rating, float* certainty) const;
159  // Returns the index (label) of the best value at the given timestep,
160  // and if not null, sets the score to the log of the corresponding value.
161  int BestLabel(int t, float* score) const {
162  return BestLabel(t, -1, -1, score);
163  }
164  // Returns the index (label) of the best value at the given timestep,
165  // excluding not_this and not_that, and if not null, sets the score to the
166  // log of the corresponding value.
167  int BestLabel(int t, int not_this, int not_that, float* score) const;
168  // Returns the best start position out of range (into which both start and end
169  // must fit) to obtain the highest cumulative score for the given labels.
170  int PositionOfBestMatch(const GenericVector<int>& labels, int start,
171  int end) const;
172  // Returns the cumulative score of the given labels starting at start, and
173  // using one label per time-step.
174  double ScoreOfLabels(const GenericVector<int>& labels, int start) const;
175  // Helper function sets all the outputs for a single timestep, such that
176  // label has value ok_score, and the other labels share 1 - ok_score.
177  // Assumes float mode.
178  void SetActivations(int t, int label, float ok_score);
179  // Modifies the values, only if needed, so that the given label is
180  // the winner at the given time step t.
181  // Assumes float mode.
182  void EnsureBestLabel(int t, int label);
183  // Helper function converts prob to certainty taking the minimum into account.
184  static float ProbToCertainty(float prob);
185  // Returns true if there is any bad value that is suspiciously like a GT
186  // error. Assuming that *this is the difference(gradient) between target
187  // and forward output, returns true if there is a large negative value
188  // (correcting a very confident output) for which there is no corresponding
189  // positive value in an adjacent timestep for the same feature index. This
190  // allows the box-truthed samples to make fine adjustments to position while
191  // stopping other disagreements of confident output with ground truth.
192  bool AnySuspiciousTruth(float confidence_thr) const;
193 
194  // Reads a single timestep to floats in the range [-1, 1].
195  void ReadTimeStep(int t, double* output) const;
196  // Adds a single timestep to floats.
197  void AddTimeStep(int t, double* inout) const;
198  // Adds part of a single timestep to floats.
199  void AddTimeStepPart(int t, int offset, int num_features, float* inout) const;
200  // Writes a single timestep from floats in the range [-1, 1].
201  void WriteTimeStep(int t, const double* input);
202  // Writes a single timestep from floats in the range [-1, 1] writing only
203  // num_features elements of input to (*this)[t], starting at offset.
204  void WriteTimeStepPart(int t, int offset, int num_features,
205  const double* input);
206  // Maxpools a single time step from src.
207  void MaxpoolTimeStep(int dest_t, const NetworkIO& src, int src_t,
208  int* max_line);
209  // Runs maxpool backward, using maxes to index timesteps in *this.
210  void MaxpoolBackward(const NetworkIO& fwd,
211  const GENERIC_2D_ARRAY<int>& maxes);
212  // Returns the min over time of the maxes over features of the outputs.
213  float MinOfMaxes() const;
214  // Returns the min over time.
215  float Max() const { return int_mode_ ? i_.Max() : f_.Max(); }
216  // Computes combined results for a combiner that chooses between an existing
217  // input and itself, with an additional output to indicate the choice.
218  void CombineOutputs(const NetworkIO& base_output,
219  const NetworkIO& combiner_output);
220  // Computes deltas for a combiner that chooses between 2 sets of inputs.
221  void ComputeCombinerDeltas(const NetworkIO& fwd_deltas,
222  const NetworkIO& base_output);
223 
224  // Copies the array checking that the types match.
225  void CopyAll(const NetworkIO& src);
226  // Adds the array to a float array, with scaling to [-1, 1] if the src is int.
227  void AddAllToFloat(const NetworkIO& src);
228  // Subtracts the array from a float array. src must also be float.
229  void SubtractAllFromFloat(const NetworkIO& src);
230 
231  // Copies src to *this, with maxabs normalization to match scale.
232  void CopyWithNormalization(const NetworkIO& src, const NetworkIO& scale);
233  // Multiplies the float data by the given factor.
234  void ScaleFloatBy(float factor) { f_ *= factor; }
235  // Copies src to *this with independent reversal of the y dimension.
236  void CopyWithYReversal(const NetworkIO& src);
237  // Copies src to *this with independent reversal of the x dimension.
238  void CopyWithXReversal(const NetworkIO& src);
239  // Copies src to *this with independent transpose of the x and y dimensions.
240  void CopyWithXYTranspose(const NetworkIO& src);
241  // Copies src to *this, at the given feature_offset, returning the total
242  // feature offset after the copy. Multiple calls will stack outputs from
243  // multiple sources in feature space.
244  int CopyPacking(const NetworkIO& src, int feature_offset);
245  // Opposite of CopyPacking, fills *this with a part of src, starting at
246  // feature_offset, and picking num_features. Resizes *this to match.
247  void CopyUnpacking(const NetworkIO& src, int feature_offset,
248  int num_features);
249  // Transposes the float part of *this into dest.
250  void Transpose(TransposedArray* dest) const;
251 
252  // Clips the content of a single time-step to +/-range.
253  void ClipVector(int t, float range);
254 
255  // Applies Func to timestep t of *this (u) and multiplies the result by v
256  // component-wise, putting the product in *product.
257  // *this and v may be int or float, but must match. The outputs are double.
258  template <class Func>
259  void FuncMultiply(const NetworkIO& v_io, int t, double* product) {
260  Func f;
261  ASSERT_HOST(!int_mode_);
262  ASSERT_HOST(!v_io.int_mode_);
263  int dim = f_.dim2();
264  if (int_mode_) {
265  const int8_t* u = i_[t];
266  const int8_t* v = v_io.i_[t];
267  for (int i = 0; i < dim; ++i) {
268  product[i] = f(u[i] / static_cast<double>(INT8_MAX)) * v[i] /
269  static_cast<double>(INT8_MAX);
270  }
271  } else {
272  const float* u = f_[t];
273  const float* v = v_io.f_[t];
274  for (int i = 0; i < dim; ++i) {
275  product[i] = f(u[i]) * v[i];
276  }
277  }
278  }
279  // Applies Func to *this (u) at u_t, and multiplies the result by v[v_t] * w,
280  // component-wise, putting the product in *product.
281  // All NetworkIOs are assumed to be float.
282  template <class Func>
283  void FuncMultiply3(int u_t, const NetworkIO& v_io, int v_t, const double* w,
284  double* product) const {
285  ASSERT_HOST(!int_mode_);
286  ASSERT_HOST(!v_io.int_mode_);
287  Func f;
288  const float* u = f_[u_t];
289  const float* v = v_io.f_[v_t];
290  int dim = f_.dim2();
291  for (int i = 0; i < dim; ++i) {
292  product[i] = f(u[i]) * v[i] * w[i];
293  }
294  }
295  // Applies Func to *this (u) at u_t, and multiplies the result by v[v_t] * w,
296  // component-wise, adding the product to *product.
297  // All NetworkIOs are assumed to be float.
298  template <class Func>
299  void FuncMultiply3Add(const NetworkIO& v_io, int t, const double* w,
300  double* product) const {
301  ASSERT_HOST(!int_mode_);
302  ASSERT_HOST(!v_io.int_mode_);
303  Func f;
304  const float* u = f_[t];
305  const float* v = v_io.f_[t];
306  int dim = f_.dim2();
307  for (int i = 0; i < dim; ++i) {
308  product[i] += f(u[i]) * v[i] * w[i];
309  }
310  }
311  // Applies Func1 to *this (u), Func2 to v, and multiplies the result by w,
312  // component-wise, putting the product in product, all at timestep t, except
313  // w, which is a simple array. All NetworkIOs are assumed to be float.
314  template <class Func1, class Func2>
315  void Func2Multiply3(const NetworkIO& v_io, int t, const double* w,
316  double* product) const {
317  ASSERT_HOST(!int_mode_);
318  ASSERT_HOST(!v_io.int_mode_);
319  Func1 f;
320  Func2 g;
321  const float* u = f_[t];
322  const float* v = v_io.f_[t];
323  int dim = f_.dim2();
324  for (int i = 0; i < dim; ++i) {
325  product[i] = f(u[i]) * g(v[i]) * w[i];
326  }
327  }
328 
329  private:
330  // Returns the padding required for the given number of features in order
331  // for the SIMD operations to be safe.
332  static int GetPadding(int num_features);
333 
334  // Choice of float vs 8 bit int for data.
337  // Which of f_ and i_ are we actually using.
338  bool int_mode_;
339  // Stride for 2d input data.
340  StrideMap stride_map_;
341  // Holds the optimal integer multiplier for this machine.
342  // This is a leaked, lazily initialized singleton, and is used for computing
343  // padding to apply to i_ for SIMD use.
344  static IntSimdMatrix* multiplier_;
345 };
346 
347 } // namespace tesseract.
348 
349 #endif // TESSERACT_LSTM_NETWORKIO_H_
void set_int_mode(bool is_quantized)
Definition: networkio.h:130
int BestLabel(int t, float *score) const
Definition: networkio.h:161
void ZeroTimeStepGeneral(int t, int offset, int num_features)
Definition: networkio.cpp:412
GENERIC_2D_ARRAY< float > * mutable_float_array()
Definition: networkio.h:140
void CopyWithXReversal(const NetworkIO &src)
Definition: networkio.cpp:877
void CombineOutputs(const NetworkIO &base_output, const NetworkIO &combiner_output)
Definition: networkio.cpp:741
void Print(int num) const
Definition: networkio.cpp:371
void AddTimeStepPart(int t, int offset, int num_features, float *inout) const
Definition: networkio.cpp:634
const float * f(int t) const
Definition: networkio.h:119
void SetActivations(int t, int label, float ok_score)
Definition: networkio.cpp:542
void set_stride_map(const StrideMap &map)
Definition: networkio.h:136
void MaxpoolBackward(const NetworkIO &fwd, const GENERIC_2D_ARRAY< int > &maxes)
Definition: networkio.cpp:700
void ClipVector(int t, float range)
Definition: networkio.cpp:976
void CopyTimeStepFrom(int dest_t, const NetworkIO &src, int src_t)
Definition: networkio.cpp:388
int CopyPacking(const NetworkIO &src, int feature_offset)
Definition: networkio.cpp:922
int BestChoiceOverRange(int t_start, int t_end, int not_this, int null_ch, float *rating, float *certainty) const
Definition: networkio.cpp:436
void MaxpoolTimeStep(int dest_t, const NetworkIO &src, int src_t, int *max_line)
Definition: networkio.cpp:673
void AddAllToFloat(const NetworkIO &src)
Definition: networkio.cpp:822
void WriteTimeStep(int t, const double *input)
Definition: networkio.cpp:650
void SubtractAllFromFloat(const NetworkIO &src)
Definition: networkio.cpp:829
void Copy2DImage(int batch, Pix *pix, float black, float contrast, TRand *randomizer)
Definition: networkio.cpp:213
const int8_t * i(int t) const
Definition: networkio.h:123
void Copy1DGreyImage(int batch, Pix *pix, float black, float contrast, TRand *randomizer)
Definition: networkio.cpp:251
void ResizeScaled(const NetworkIO &src, int x_scale, int y_scale, int num_features)
Definition: networkio.cpp:67
void Resize(const NetworkIO &src, int num_features)
Definition: networkio.h:45
void ZeroInvalidElements()
Definition: networkio.cpp:93
void Transpose(TransposedArray *dest) const
Definition: networkio.cpp:969
void CopyWithXYTranspose(const NetworkIO &src)
Definition: networkio.cpp:896
void CopyWithNormalization(const NetworkIO &src, const NetworkIO &scale)
Definition: networkio.cpp:836
T Max() const
Definition: matrix.h:342
void ZeroTimeStep(int t)
Definition: networkio.h:148
void FromPixes(const StaticShape &shape, const std::vector< const Pix *> &pixes, TRand *randomizer)
Definition: networkio.cpp:175
Pix * ToPix() const
Definition: networkio.cpp:291
void Func2Multiply3(const NetworkIO &v_io, int t, const double *w, double *product) const
Definition: networkio.h:315
void ComputeCombinerDeltas(const NetworkIO &fwd_deltas, const NetworkIO &base_output)
Definition: networkio.cpp:775
const StrideMap & stride_map() const
Definition: networkio.h:133
static float ProbToCertainty(float prob)
Definition: networkio.cpp:573
void ScoresOverRange(int t_start, int t_end, int choice, int null_ch, float *rating, float *certainty) const
Definition: networkio.cpp:455
void AddTimeStep(int t, double *inout) const
Definition: networkio.cpp:618
void WriteTimeStepPart(int t, int offset, int num_features, const double *input)
Definition: networkio.cpp:656
int dim1() const
Definition: matrix.h:206
float * f(int t)
Definition: networkio.h:115
void CopyAll(const NetworkIO &src)
Definition: networkio.cpp:816
void FuncMultiply(const NetworkIO &v_io, int t, double *product)
Definition: networkio.h:259
void FuncMultiply3(int u_t, const NetworkIO &v_io, int v_t, const double *w, double *product) const
Definition: networkio.h:283
void FromPix(const StaticShape &shape, const Pix *pix, TRand *randomizer)
Definition: networkio.cpp:166
void CopyTimeStepGeneral(int dest_t, int dest_offset, int num_features, const NetworkIO &src, int src_t, int src_offset)
Definition: networkio.cpp:398
bool int_mode() const
Definition: networkio.h:127
int PositionOfBestMatch(const GenericVector< int > &labels, int start, int end) const
Definition: networkio.cpp:512
void SetPixel(int t, int f, int pixel, float black, float contrast)
Definition: networkio.cpp:280
void CopyUnpacking(const NetworkIO &src, int feature_offset, int num_features)
Definition: networkio.cpp:950
void FuncMultiply3Add(const NetworkIO &v_io, int t, const double *w, double *product) const
Definition: networkio.h:299
double ScoreOfLabels(const GenericVector< int > &labels, int start) const
Definition: networkio.cpp:530
int dim2() const
Definition: matrix.h:207
void ResizeToMap(bool int_mode, const StrideMap &stride_map, int num_features)
Definition: networkio.cpp:51
void ReadTimeStep(int t, double *output) const
Definition: networkio.cpp:603
void CopyWithYReversal(const NetworkIO &src)
Definition: networkio.cpp:858
void EnsureBestLabel(int t, int label)
Definition: networkio.cpp:554
float Max() const
Definition: networkio.h:215
void ScaleFloatBy(float factor)
Definition: networkio.h:234
void ResizeXTo1(const NetworkIO &src, int num_features)
Definition: networkio.cpp:75
float MinOfMaxes() const
Definition: networkio.cpp:717
int Width() const
Definition: networkio.h:107
void Resize2d(bool int_mode, int width, int num_features)
Definition: networkio.cpp:40
void ResizeFloat(const NetworkIO &src, int num_features)
Definition: networkio.h:52
bool AnySuspiciousTruth(float confidence_thr) const
Definition: networkio.cpp:584
int NumFeatures() const
Definition: networkio.h:111
#define ASSERT_HOST(x)
Definition: errcode.h:84
const GENERIC_2D_ARRAY< float > & float_array() const
Definition: networkio.h:139
void Randomize(int t, int offset, int num_features, TRand *randomizer)
Definition: networkio.cpp:421