tesseract/lstm/recodebeam.h

393 lines
19 KiB
C
Raw Normal View History

///////////////////////////////////////////////////////////////////////
// File: recodebeam.h
// Description: Beam search to decode from the re-encoded CJK as a sequence of
// smaller numbers in place of a single large code.
// Author: Ray Smith
// Created: Fri Mar 13 09:12:01 PDT 2015
//
// (C) Copyright 2015, Google Inc.
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
// http://www.apache.org/licenses/LICENSE-2.0
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
//
///////////////////////////////////////////////////////////////////////
#ifndef THIRD_PARTY_TESSERACT_LSTM_RECODEBEAM_H_
#define THIRD_PARTY_TESSERACT_LSTM_RECODEBEAM_H_
#include "dawg.h"
#include "dict.h"
#include "genericheap.h"
#include "kdpair.h"
#include "networkio.h"
#include "ratngs.h"
#include "unicharcompress.h"
namespace tesseract {
// Enum describing what can follow the current node.
// Consider the following softmax outputs:
// Timestep 0 1 2 3 4 5 6 7 8
// X-score 0.01 0.55 0.98 0.42 0.01 0.01 0.40 0.95 0.01
// Y-score 0.00 0.01 0.01 0.01 0.01 0.97 0.59 0.04 0.01
// Null-score 0.99 0.44 0.01 0.57 0.98 0.02 0.01 0.01 0.98
// Then the correct CTC decoding (in which adjacent equal classes are folded,
// and then all nulls are dropped) is clearly XYX, but simple decoding (taking
// the max at each timestep) leads to:
// Null@0.99 X@0.55 X@0.98 Null@0.57 Null@0.98 Y@0.97 Y@0.59 X@0.95 Null@0.98,
// which folds to the correct XYX. The conversion to Tesseract rating and
// certainty uses the sum of the log probs (log of the product of probabilities)
// for the Rating and the minimum log prob for the certainty, but that yields a
// minimum certainty of log(0.55), which is poor for such an obvious case.
// CTC says that the probability of the result is the SUM of the products of the
// probabilities over ALL PATHS that decode to the same result, which includes:
// NXXNNYYXN, NNXNNYYN, NXXXNYYXN, NNXXNYXXN, and others including XXXXXYYXX.
// That is intractable, so some compromise between simple and ideal is needed.
// Observing that evenly split timesteps rarely happen next to each other, we
// allow scores at a transition between classes to be added for decoding thus:
// N@0.99 (N+X)@0.99 X@0.98 (N+X)@0.99 N@0.98 Y@0.97 (X+Y+N)@1.00 X@0.95 N@0.98.
// This works because NNX and NXX both decode to X, so in the middle we can use
// N+X. Note that the classes either side of a sum must stand alone, i.e. use a
// single score, to force all paths to pass through them and decode to the same
// result. Also in the special case of a transition from X to Y, with only one
// timestep between, it is possible to add X+Y+N, since XXY, XYY, and XNY all
// decode to XY.
// An important condition is that we cannot combine X and Null between two
// stand-alone Xs, since that can decode as XNX->XX or XXX->X, so the scores for
// X and Null have to go in separate paths. Combining scores in this way
// provides a much better minimum certainty of log(0.95).
// In the implementation of the beam search, we have to place the possibilities
// X, X+N and X+Y+N in the beam under appropriate conditions of the previous
// node, and constrain what can follow, to enforce the rules explained above.
// We therefore have 3 different types of node determined by what can follow:
enum NodeContinuation {
NC_ANYTHING, // This node used just its own score, so anything can follow.
NC_ONLY_DUP, // The current node combined another score with the score for
// itself, without a stand-alone duplicate before, so must be
// followed by a stand-alone duplicate.
NC_NO_DUP, // The current node combined another score with the score for
// itself, after a stand-alone, so can only be followed by
// something other than a duplicate of the current node.
NC_COUNT
};
// Enum describing the top-n status of a code.
enum TopNState {
TN_TOP2, // Winner or 2nd.
TN_TOPN, // Runner up in top-n, but not 1st or 2nd.
TN_ALSO_RAN, // Not in the top-n.
TN_COUNT
};
// Lattice element for Re-encode beam search.
struct RecodeNode {
RecodeNode()
: code(-1),
unichar_id(INVALID_UNICHAR_ID),
permuter(TOP_CHOICE_PERM),
start_of_dawg(false),
start_of_word(false),
end_of_word(false),
duplicate(false),
certainty(0.0f),
score(0.0f),
prev(nullptr),
dawgs(nullptr),
code_hash(0) {}
RecodeNode(int c, int uni_id, PermuterType perm, bool dawg_start,
bool word_start, bool end, bool dup, float cert, float s,
const RecodeNode* p, DawgPositionVector* d, uint64_t hash)
: code(c),
unichar_id(uni_id),
permuter(perm),
start_of_dawg(dawg_start),
start_of_word(word_start),
end_of_word(end),
duplicate(dup),
certainty(cert),
score(s),
prev(p),
dawgs(d),
code_hash(hash) {}
// NOTE: If we could use C++11, then this would be a move constructor.
// Instead we have copy constructor that does a move!! This is because we
// don't want to copy the whole DawgPositionVector each time, and true
// copying isn't necessary for this struct. It does get moved around a lot
// though inside the heap and during heap push, hence the move semantics.
RecodeNode(RecodeNode& src) : dawgs(nullptr) {
*this = src;
ASSERT_HOST(src.dawgs == nullptr);
}
RecodeNode& operator=(RecodeNode& src) {
delete dawgs;
memcpy(this, &src, sizeof(src));
src.dawgs = nullptr;
return *this;
}
~RecodeNode() { delete dawgs; }
// Prints details of the node.
void Print(int null_char, const UNICHARSET& unicharset, int depth) const;
// The re-encoded code here = index to network output.
int code;
// The decoded unichar_id is only valid for the final code of a sequence.
int unichar_id;
// The type of permuter active at this point. Intervals between start_of_word
// and end_of_word make valid words of type given by permuter where
// end_of_word is true. These aren't necessarily delimited by spaces.
PermuterType permuter;
// True if this is the initial dawg state. May be attached to a space or,
// in a non-space-delimited lang, the end of the previous word.
bool start_of_dawg;
// True if this is the first node in a dictionary word.
bool start_of_word;
// True if this represents a valid candidate end of word position. Does not
// necessarily mark the end of a word, since a word can be extended beyond a
// candidate end by a continuation, eg 'the' continues to 'these'.
bool end_of_word;
// True if this->code is a duplicate of prev->code. Some training modes
// allow the network to output duplicate characters and crush them with CTC,
// but that would mess up the dictionary search, so we just smash them
// together on the fly using the duplicate flag.
bool duplicate;
// Certainty (log prob) of (just) this position.
float certainty;
// Total certainty of the path to this position.
float score;
// The previous node in this chain. Borrowed pointer.
const RecodeNode* prev;
// The currently active dawgs at this position. Owned pointer.
DawgPositionVector* dawgs;
// A hash of all codes in the prefix and this->code as well. Used for
// duplicate path removal.
uint64_t code_hash;
};
typedef KDPairInc<double, RecodeNode> RecodePair;
typedef GenericHeap<RecodePair> RecodeHeap;
// Class that holds the entire beam search for recognition of a text line.
class RecodeBeamSearch {
public:
// Borrows the pointer, which is expected to survive until *this is deleted.
RecodeBeamSearch(const UnicharCompress& recoder, int null_char,
bool simple_text, Dict* dict);
// Decodes the set of network outputs, storing the lattice internally.
// If charset is not null, it enables detailed debugging of the beam search.
void Decode(const NetworkIO& output, double dict_ratio, double cert_offset,
double worst_dict_cert, const UNICHARSET* charset);
void Decode(const GENERIC_2D_ARRAY<float>& output, double dict_ratio,
double cert_offset, double worst_dict_cert,
const UNICHARSET* charset);
// Returns the best path as labels/scores/xcoords similar to simple CTC.
void ExtractBestPathAsLabels(GenericVector<int>* labels,
GenericVector<int>* xcoords) const;
// Returns the best path as unichar-ids/certs/ratings/xcoords skipping
// duplicates, nulls and intermediate parts.
void ExtractBestPathAsUnicharIds(bool debug, const UNICHARSET* unicharset,
GenericVector<int>* unichar_ids,
GenericVector<float>* certs,
GenericVector<float>* ratings,
GenericVector<int>* xcoords) const;
// Returns the best path as a set of WERD_RES.
void ExtractBestPathAsWords(const TBOX& line_box, float scale_factor,
bool debug, const UNICHARSET* unicharset,
PointerVector<WERD_RES>* words);
// Generates debug output of the content of the beams after a Decode.
void DebugBeams(const UNICHARSET& unicharset) const;
// Clipping value for certainty inside Tesseract. Reflects the minimum value
// of certainty that will be returned by ExtractBestPathAsUnicharIds.
// Supposedly on a uniform scale that can be compared across languages and
// engines.
static const float kMinCertainty;
// Number of different code lengths for which we have a separate beam.
static const int kNumLengths = RecodedCharID::kMaxCodeLen + 1;
// Total number of beams: dawg/nodawg * number of NodeContinuation * number
// of different lengths.
static const int kNumBeams = 2 * NC_COUNT * kNumLengths;
// Returns the relevant factor in the beams_ index.
static int LengthFromBeamsIndex(int index) { return index % kNumLengths; }
static NodeContinuation ContinuationFromBeamsIndex(int index) {
return static_cast<NodeContinuation>((index / kNumLengths) % NC_COUNT);
}
static bool IsDawgFromBeamsIndex(int index) {
return index / (kNumLengths * NC_COUNT) > 0;
}
// Computes a beams_ index from the given factors.
static int BeamIndex(bool is_dawg, NodeContinuation cont, int length) {
return (is_dawg * NC_COUNT + cont) * kNumLengths + length;
}
private:
// Struct for the Re-encode beam search. This struct holds the data for
// a single time-step position of the output. Use a PointerVector<RecodeBeam>
// to hold all the timesteps and prevent reallocation of the individual heaps.
struct RecodeBeam {
// Resets to the initial state without deleting all the memory.
void Clear() {
for (int i = 0; i < kNumBeams; ++i) {
beams_[i].clear();
}
RecodeNode empty;
for (int i = 0; i < NC_COUNT; ++i) {
best_initial_dawgs_[i] = empty;
}
}
// A separate beam for each combination of code length,
// NodeContinuation, and dictionary flag. Separating out all these types
// allows the beam to be quite narrow, and yet still have a low chance of
// losing the best path.
// We have to keep all these beams separate, since the highest scoring paths
// come from the paths that are most likely to dead-end at any time, like
// dawg paths, NC_ONLY_DUP etc.
// Each heap is stored with the WORST result at the top, so we can quickly
// get the top-n values.
RecodeHeap beams_[kNumBeams];
// While the language model is only a single word dictionary, we can use
// word starts as a choke point in the beam, and keep only a single dict
// start node at each step (for each NodeContinuation type), so we find the
// best one here and push it on the heap, if it qualifies, after processing
// all of the step.
RecodeNode best_initial_dawgs_[NC_COUNT];
};
typedef KDPairInc<float, int> TopPair;
// Generates debug output of the content of a single beam position.
void DebugBeamPos(const UNICHARSET& unicharset, const RecodeHeap& heap) const;
// Returns the given best_nodes as unichar-ids/certs/ratings/xcoords skipping
// duplicates, nulls and intermediate parts.
static void ExtractPathAsUnicharIds(
const GenericVector<const RecodeNode*>& best_nodes,
GenericVector<int>* unichar_ids, GenericVector<float>* certs,
GenericVector<float>* ratings, GenericVector<int>* xcoords);
// Sets up a word with the ratings matrix and fake blobs with boxes in the
// right places.
WERD_RES* InitializeWord(bool leading_space, const TBOX& line_box,
int word_start, int word_end, float space_certainty,
const UNICHARSET* unicharset,
const GenericVector<int>& xcoords,
float scale_factor);
// Fills top_n_flags_ with bools that are true iff the corresponding output
// is one of the top_n.
void ComputeTopN(const float* outputs, int num_outputs, int top_n);
// Adds the computation for the current time-step to the beam. Call at each
// time-step in sequence from left to right. outputs is the activation vector
// for the current timestep.
void DecodeStep(const float* outputs, int t, double dict_ratio,
double cert_offset, double worst_dict_cert,
const UNICHARSET* charset);
// Adds to the appropriate beams the legal (according to recoder)
// continuations of context prev, which is from the given index to beams_,
// using the given network outputs to provide scores to the choices. Uses only
// those choices for which top_n_flags[code] == top_n_flag.
void ContinueContext(const RecodeNode* prev, int index, const float* outputs,
TopNState top_n_flag, double dict_ratio,
double cert_offset, double worst_dict_cert,
RecodeBeam* step);
// Continues for a new unichar, using dawg or non-dawg as per flag.
void ContinueUnichar(int code, int unichar_id, float cert,
float worst_dict_cert, float dict_ratio, bool use_dawgs,
NodeContinuation cont, const RecodeNode* prev,
RecodeBeam* step);
// Adds a RecodeNode composed of the args to the correct heap in step if
// unichar_id is a valid dictionary continuation of whatever is in prev.
void ContinueDawg(int code, int unichar_id, float cert, NodeContinuation cont,
const RecodeNode* prev, RecodeBeam* step);
// Sets the correct best_initial_dawgs_ with a RecodeNode composed of the args
// if better than what is already there.
void PushInitialDawgIfBetter(int code, int unichar_id, PermuterType permuter,
bool start, bool end, float cert,
NodeContinuation cont, const RecodeNode* prev,
RecodeBeam* step);
// Adds a RecodeNode composed of the args to the correct heap in step for
// partial unichar or duplicate if there is room or if better than the
// current worst element if already full.
void PushDupOrNoDawgIfBetter(int length, bool dup, int code, int unichar_id,
float cert, float worst_dict_cert,
float dict_ratio, bool use_dawgs,
NodeContinuation cont, const RecodeNode* prev,
RecodeBeam* step);
// Adds a RecodeNode composed of the args to the correct heap in step if there
// is room or if better than the current worst element if already full.
void PushHeapIfBetter(int max_size, int code, int unichar_id,
PermuterType permuter, bool dawg_start, bool word_start,
bool end, bool dup, float cert, const RecodeNode* prev,
DawgPositionVector* d, RecodeHeap* heap);
// Adds a RecodeNode to heap if there is room
// or if better than the current worst element if already full.
void PushHeapIfBetter(int max_size, RecodeNode* node, RecodeHeap* heap);
// Searches the heap for an entry matching new_node, and updates the entry
// with reshuffle if needed. Returns true if there was a match.
bool UpdateHeapIfMatched(RecodeNode* new_node, RecodeHeap* heap);
// Computes and returns the code-hash for the given code and prev.
uint64_t ComputeCodeHash(int code, bool dup, const RecodeNode* prev) const;
// Backtracks to extract the best path through the lattice that was built
// during Decode. On return the best_nodes vector essentially contains the set
// of code, score pairs that make the optimal path with the constraint that
// the recoder can decode the code sequence back to a sequence of unichar-ids.
void ExtractBestPaths(GenericVector<const RecodeNode*>* best_nodes,
GenericVector<const RecodeNode*>* second_nodes) const;
// Helper backtracks through the lattice from the given node, storing the
// path and reversing it.
void ExtractPath(const RecodeNode* node,
GenericVector<const RecodeNode*>* path) const;
// Helper prints debug information on the given lattice path.
void DebugPath(const UNICHARSET* unicharset,
const GenericVector<const RecodeNode*>& path) const;
// Helper prints debug information on the given unichar path.
void DebugUnicharPath(const UNICHARSET* unicharset,
const GenericVector<const RecodeNode*>& path,
const GenericVector<int>& unichar_ids,
const GenericVector<float>& certs,
const GenericVector<float>& ratings,
const GenericVector<int>& xcoords) const;
static const int kBeamWidths[RecodedCharID::kMaxCodeLen + 1];
// The encoder/decoder that we will be using.
const UnicharCompress& recoder_;
// The beam for each timestep in the output.
PointerVector<RecodeBeam> beam_;
// The number of timesteps valid in beam_;
int beam_size_;
// A flag to indicate which outputs are the top-n choices. Current timestep
// only.
GenericVector<TopNState> top_n_flags_;
// A record of the highest and second scoring codes.
int top_code_;
int second_code_;
// Heap used to compute the top_n_flags_.
GenericHeap<TopPair> top_heap_;
// Borrowed pointer to the dictionary to use in the search.
Dict* dict_;
// True if the language is space-delimited, which is true for most languages
// except chi*, jpn, tha.
bool space_delimited_;
// True if the input is simple text, ie adjacent equal chars are not to be
// eliminated.
bool is_simple_text_;
// The encoded (class label) of the null/reject character.
int null_char_;
};
} // namespace tesseract.
#endif // THIRD_PARTY_TESSERACT_LSTM_RECODEBEAM_H_