tesseract/lstm/fullyconnected.cpp

///////////////////////////////////////////////////////////////////////
// File:        fullyconnected.cpp
// Description: Simple feed-forward layer with various non-linearities.
// Author:      Ray Smith
// Created:     Wed Feb 26 14:49:15 PST 2014
//
// (C) Copyright 2014, 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.
///////////////////////////////////////////////////////////////////////

#include "fullyconnected.h"

#ifdef _OPENMP
#include <omp.h>
#endif
#include <stdio.h>
#include <stdlib.h>

#include "functions.h"
#include "networkscratch.h"

// Number of threads to use for parallel calculation of Forward and Backward.
const int kNumThreads = 4;

namespace tesseract {

FullyConnected::FullyConnected(const STRING& name, int ni, int no,
                               NetworkType type)
  : Network(type, name, ni, no), external_source_(NULL), int_mode_(false) {
}

FullyConnected::~FullyConnected() {
}

// Returns the shape output from the network given an input shape (which may
// be partially unknown ie zero).
StaticShape FullyConnected::OutputShape(const StaticShape& input_shape) const {
  LossType loss_type = LT_NONE;
  if (type_ == NT_SOFTMAX)
    loss_type = LT_CTC;
  else if (type_ == NT_SOFTMAX_NO_CTC)
    loss_type = LT_SOFTMAX;
  else if (type_ == NT_LOGISTIC)
    loss_type = LT_LOGISTIC;
  StaticShape result(input_shape);
  result.set_depth(no_);
  result.set_loss_type(loss_type);
  return result;
}

// Suspends/Enables training by setting the training_ flag. Serialize and
// DeSerialize only operate on the run-time data if state is false.
void FullyConnected::SetEnableTraining(TrainingState state) {
  if (state == TS_RE_ENABLE) {
    if (training_ == TS_DISABLED) weights_.InitBackward(false);
    training_ = TS_ENABLED;
  } else {
    training_ = state;
  }
}

// Sets up the network for training. Initializes weights using weights of
// scale `range` picked according to the random number generator `randomizer`.
int FullyConnected::InitWeights(float range, TRand* randomizer) {
  Network::SetRandomizer(randomizer);
  num_weights_ = weights_.InitWeightsFloat(no_, ni_ + 1, TestFlag(NF_ADA_GRAD),
                                           range, randomizer);
  return num_weights_;
}

// Converts a float network to an int network.
void FullyConnected::ConvertToInt() {
  weights_.ConvertToInt();
}

// Provides debug output on the weights.
void FullyConnected::DebugWeights() {
  weights_.Debug2D(name_.string());
}

// Writes to the given file. Returns false in case of error.
bool FullyConnected::Serialize(TFile* fp) const {
  if (!Network::Serialize(fp)) return false;
  if (!weights_.Serialize(IsTraining(), fp)) return false;
  return true;
}

// Reads from the given file. Returns false in case of error.
// If swap is true, assumes a big/little-endian swap is needed.
bool FullyConnected::DeSerialize(bool swap, TFile* fp) {
  if (!weights_.DeSerialize(IsTraining(), swap, fp)) return false;
  return true;
}

// Runs forward propagation of activations on the input line.
// See NetworkCpp for a detailed discussion of the arguments.
void FullyConnected::Forward(bool debug, const NetworkIO& input,
                             const TransposedArray* input_transpose,
                             NetworkScratch* scratch, NetworkIO* output) {
  int width = input.Width();
  if (type_ == NT_SOFTMAX)
    output->ResizeFloat(input, no_);
  else
    output->Resize(input, no_);
  SetupForward(input, input_transpose);
  GenericVector<NetworkScratch::FloatVec> temp_lines;
  temp_lines.init_to_size(kNumThreads, NetworkScratch::FloatVec());
  GenericVector<NetworkScratch::FloatVec> curr_input;
  curr_input.init_to_size(kNumThreads, NetworkScratch::FloatVec());
  for (int i = 0; i < temp_lines.size(); ++i) {
    temp_lines[i].Init(no_, scratch);
    curr_input[i].Init(ni_, scratch);
  }
#ifdef _OPENMP
#pragma omp parallel for num_threads(kNumThreads)
  for (int t = 0; t < width; ++t) {
    // Thread-local pointer to temporary storage.
    int thread_id = omp_get_thread_num();
#else
  for (int t = 0; t < width; ++t) {
    // Thread-local pointer to temporary storage.
    int thread_id = 0;
#endif
    double* temp_line = temp_lines[thread_id];
    const double* d_input = NULL;
    const inT8* i_input = NULL;
    if (input.int_mode()) {
      i_input = input.i(t);
    } else {
      input.ReadTimeStep(t, curr_input[thread_id]);
      d_input = curr_input[thread_id];
    }
    ForwardTimeStep(d_input, i_input, t, temp_line);
    output->WriteTimeStep(t, temp_line);
    if (IsTraining() && type_ != NT_SOFTMAX) {
      acts_.CopyTimeStepFrom(t, *output, t);
    }
  }
  // Zero all the elements that are in the padding around images that allows
  // multiple different-sized images to exist in a single array.
  // acts_ is only used if this is not a softmax op.
  if (IsTraining() && type_ != NT_SOFTMAX) {
    acts_.ZeroInvalidElements();
  }
  output->ZeroInvalidElements();
#if DEBUG_DETAIL > 0
  tprintf("F Output:%s\n", name_.string());
  output->Print(10);
#endif
  if (debug) DisplayForward(*output);
}

// Components of Forward so FullyConnected can be reused inside LSTM.
void FullyConnected::SetupForward(const NetworkIO& input,
                                  const TransposedArray* input_transpose) {
  // Softmax output is always float, so save the input type.
  int_mode_ = input.int_mode();
  if (IsTraining()) {
    acts_.Resize(input, no_);
    // Source_ is a transposed copy of input. It isn't needed if provided.
    external_source_ = input_transpose;
    if (external_source_ == NULL) source_t_.ResizeNoInit(ni_, input.Width());
  }
}

void FullyConnected::ForwardTimeStep(const double* d_input, const inT8* i_input,
                                     int t, double* output_line) {
  // input is copied to source_ line-by-line for cache coherency.
  if (IsTraining() && external_source_ == NULL && d_input != NULL)
    source_t_.WriteStrided(t, d_input);
  if (d_input != NULL)
    weights_.MatrixDotVector(d_input, output_line);
  else
    weights_.MatrixDotVector(i_input, output_line);
  if (type_ == NT_TANH) {
    FuncInplace<GFunc>(no_, output_line);
  } else if (type_ == NT_LOGISTIC) {
    FuncInplace<FFunc>(no_, output_line);
  } else if (type_ == NT_POSCLIP) {
    FuncInplace<ClipFFunc>(no_, output_line);
  } else if (type_ == NT_SYMCLIP) {
    FuncInplace<ClipGFunc>(no_, output_line);
  } else if (type_ == NT_RELU) {
    FuncInplace<Relu>(no_, output_line);
  } else if (type_ == NT_SOFTMAX || type_ == NT_SOFTMAX_NO_CTC) {
    SoftmaxInPlace(no_, output_line);
  } else if (type_ != NT_LINEAR) {
    ASSERT_HOST("Invalid fully-connected type!" == NULL);
  }
}

// Runs backward propagation of errors on the deltas line.
// See NetworkCpp for a detailed discussion of the arguments.
bool FullyConnected::Backward(bool debug, const NetworkIO& fwd_deltas,
                              NetworkScratch* scratch,
                              NetworkIO* back_deltas) {
  if (debug) DisplayBackward(fwd_deltas);
  back_deltas->Resize(fwd_deltas, ni_);
  GenericVector<NetworkScratch::FloatVec> errors;
  errors.init_to_size(kNumThreads, NetworkScratch::FloatVec());
  for (int i = 0; i < errors.size(); ++i) errors[i].Init(no_, scratch);
  GenericVector<NetworkScratch::FloatVec> temp_backprops;
  if (needs_to_backprop_) {
    temp_backprops.init_to_size(kNumThreads, NetworkScratch::FloatVec());
    for (int i = 0; i < kNumThreads; ++i) temp_backprops[i].Init(ni_, scratch);
  }
  int width = fwd_deltas.Width();
  NetworkScratch::GradientStore errors_t;
  errors_t.Init(no_, width, scratch);
#ifdef _OPENMP
#pragma omp parallel for num_threads(kNumThreads)
  for (int t = 0; t < width; ++t) {
    int thread_id = omp_get_thread_num();
#else
  for (int t = 0; t < width; ++t) {
    int thread_id = 0;
#endif
    double* backprop = NULL;
    if (needs_to_backprop_) backprop = temp_backprops[thread_id];
    double* curr_errors = errors[thread_id];
    BackwardTimeStep(fwd_deltas, t, curr_errors, errors_t.get(), backprop);
    if (backprop != NULL) {
      back_deltas->WriteTimeStep(t, backprop);
    }
  }
  FinishBackward(*errors_t.get());
  if (needs_to_backprop_) {
    back_deltas->ZeroInvalidElements();
    back_deltas->CopyWithNormalization(*back_deltas, fwd_deltas);
#if DEBUG_DETAIL > 0
    tprintf("F Backprop:%s\n", name_.string());
    back_deltas->Print(10);
#endif
    return true;
  }
  return false;  // No point going further back.
}

void FullyConnected::BackwardTimeStep(const NetworkIO& fwd_deltas, int t,
                                      double* curr_errors,
                                      TransposedArray* errors_t,
                                      double* backprop) {
  if (type_ == NT_TANH)
    acts_.FuncMultiply<GPrime>(fwd_deltas, t, curr_errors);
  else if (type_ == NT_LOGISTIC)
    acts_.FuncMultiply<FPrime>(fwd_deltas, t, curr_errors);
  else if (type_ == NT_POSCLIP)
    acts_.FuncMultiply<ClipFPrime>(fwd_deltas, t, curr_errors);
  else if (type_ == NT_SYMCLIP)
    acts_.FuncMultiply<ClipGPrime>(fwd_deltas, t, curr_errors);
  else if (type_ == NT_RELU)
    acts_.FuncMultiply<ReluPrime>(fwd_deltas, t, curr_errors);
  else if (type_ == NT_SOFTMAX || type_ == NT_SOFTMAX_NO_CTC ||
           type_ == NT_LINEAR)
    fwd_deltas.ReadTimeStep(t, curr_errors);  // fwd_deltas are the errors.
  else
    ASSERT_HOST("Invalid fully-connected type!" == NULL);
  // Generate backprop only if needed by the lower layer.
  if (backprop != NULL) weights_.VectorDotMatrix(curr_errors, backprop);
  errors_t->WriteStrided(t, curr_errors);
}

void FullyConnected::FinishBackward(const TransposedArray& errors_t) {
  if (external_source_ == NULL)
    weights_.SumOuterTransposed(errors_t, source_t_, true);
  else
    weights_.SumOuterTransposed(errors_t, *external_source_, true);
}

// Updates the weights using the given learning rate and momentum.
// num_samples is the quotient to be used in the adagrad computation iff
// use_ada_grad_ is true.
void FullyConnected::Update(float learning_rate, float momentum,
                            int num_samples) {
  weights_.Update(learning_rate, momentum, num_samples);
}

// Sums the products of weight updates in *this and other, splitting into
// positive (same direction) in *same and negative (different direction) in
// *changed.
void FullyConnected::CountAlternators(const Network& other, double* same,
                                      double* changed) const {
  ASSERT_HOST(other.type() == type_);
  const FullyConnected* fc = reinterpret_cast<const FullyConnected*>(&other);
  weights_.CountAlternators(fc->weights_, same, changed);
}

}  // namespace tesseract.