tesseract/lstm/series.h

98 lines
4.2 KiB
C++

///////////////////////////////////////////////////////////////////////
// File: series.h
// Description: Runs networks in series on the same input.
// Author: Ray Smith
// Created: Thu May 02 08:20:06 PST 2013
//
// (C) Copyright 2013, 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 TESSERACT_LSTM_SERIES_H_
#define TESSERACT_LSTM_SERIES_H_
#include "plumbing.h"
namespace tesseract {
// Runs two or more networks in series (layers) on the same input.
class Series : public Plumbing {
public:
// ni_ and no_ will be set by AddToStack.
explicit Series(const STRING& name);
virtual ~Series();
// Returns the shape output from the network given an input shape (which may
// be partially unknown ie zero).
virtual StaticShape OutputShape(const StaticShape& input_shape) const;
virtual STRING spec() const {
STRING spec("[");
for (int i = 0; i < stack_.size(); ++i)
spec += stack_[i]->spec();
spec += "]";
return spec;
}
// Sets up the network for training. Initializes weights using weights of
// scale `range` picked according to the random number generator `randomizer`.
// Returns the number of weights initialized.
virtual int InitWeights(float range, TRand* randomizer);
// Changes the number of outputs to the size of the given code_map, copying
// the old weight matrix entries for each output from code_map[output] where
// non-negative, and uses the mean (over all outputs) of the existing weights
// for all outputs with negative code_map entries. Returns the new number of
// weights. Only operates on Softmax layers with old_no outputs.
int RemapOutputs(int old_no, const std::vector<int>& code_map) override;
// Sets needs_to_backprop_ to needs_backprop and returns true if
// needs_backprop || any weights in this network so the next layer forward
// can be told to produce backprop for this layer if needed.
virtual bool SetupNeedsBackprop(bool needs_backprop);
// Returns an integer reduction factor that the network applies to the
// time sequence. Assumes that any 2-d is already eliminated. Used for
// scaling bounding boxes of truth data.
// WARNING: if GlobalMinimax is used to vary the scale, this will return
// the last used scale factor. Call it before any forward, and it will return
// the minimum scale factor of the paths through the GlobalMinimax.
virtual int XScaleFactor() const;
// Provides the (minimum) x scale factor to the network (of interest only to
// input units) so they can determine how to scale bounding boxes.
virtual void CacheXScaleFactor(int factor);
// Runs forward propagation of activations on the input line.
// See Network for a detailed discussion of the arguments.
virtual void Forward(bool debug, const NetworkIO& input,
const TransposedArray* input_transpose,
NetworkScratch* scratch, NetworkIO* output);
// Runs backward propagation of errors on the deltas line.
// See Network for a detailed discussion of the arguments.
virtual bool Backward(bool debug, const NetworkIO& fwd_deltas,
NetworkScratch* scratch,
NetworkIO* back_deltas);
// Splits the series after the given index, returning the two parts and
// deletes itself. The first part, up to network with index last_start, goes
// into start, and the rest goes into end.
void SplitAt(int last_start, Series** start, Series** end);
// Appends the elements of the src series to this, removing from src and
// deleting it.
void AppendSeries(Network* src);
};
} // namespace tesseract.
#endif // TESSERACT_LSTM_SERIES_H_