21天实战caffe（4）数据结构 blob

namespace caffe {

 //变维函数，将（num，channels，height，width）参数转换维vector<int>，然后调用重载的变维函数void Blob<Dtype>::Reshape(const vector<int>& shape)

template <typename Dtype>

void Blob<Dtype>::Reshape(const int num, const int channels, const int height,

    const int width) {

  vector<int> shape(4);

  shape[0] = num;

  shape[1] = channels;

  shape[2] = height;

  shape[3] = width;

  Reshape(shape);

}

 

//真正变维函数

template <typename Dtype>

void Blob<Dtype>::Reshape(const vector<int>& shape) {

  CHECK_LE(shape.size(), kMaxBlobAxes);//保证vector维度<=kMaxBlobAxes

  count_ = 1;//用于计算元素总数 = num*channels*height*width

  shape_.resize(shape.size()); // 成员变量维度也被重置

  if (!shape_data_ || shape_data_->size() < shape.size() * sizeof(int)) {

    shape_data_.reset(new SyncedMemory(shape.size() * sizeof(int)));

  }

  int* shape_data = static_cast<int*>(shape_data_->mutable_cpu_data());

  for (int i = 0; i < shape.size(); ++i) {

    CHECK_GE(shape[i], 0);

    if (count_ != 0) {

      CHECK_LE(shape[i], INT_MAX / count_) << "blob size exceeds INT_MAX";

    }

    count_ *= shape[i];

    shape_[i] = shape[i];

    shape_data[i] = shape[i];

  }

  if (count_ > capacity_) {

    capacity_ = count_;

    data_.reset(new SyncedMemory(capacity_ * sizeof(Dtype)));

    diff_.reset(new SyncedMemory(capacity_ * sizeof(Dtype)));

  }

}

 

template <typename Dtype>

void Blob<Dtype>::Reshape(const BlobShape& shape) {

  CHECK_LE(shape.dim_size(), kMaxBlobAxes);

  vector<int> shape_vec(shape.dim_size());

  for (int i = 0; i < shape.dim_size(); ++i) {

    shape_vec[i] = shape.dim(i);

  }

  Reshape(shape_vec);

}

 

template <typename Dtype>

void Blob<Dtype>::ReshapeLike(const Blob<Dtype>& other) {

  Reshape(other.shape());

}

 

template <typename Dtype>

Blob<Dtype>::Blob(const int num, const int channels, const int height,

    const int width)

  // capacity_ must be initialized before calling Reshape

  : capacity_(0) {

  Reshape(num, channels, height, width);

}

 

template <typename Dtype>

Blob<Dtype>::Blob(const vector<int>& shape)

  // capacity_ must be initialized before calling Reshape

  : capacity_(0) {

  Reshape(shape);

}

 

template <typename Dtype>

const int* Blob<Dtype>::gpu_shape() const {

  CHECK(shape_data_);

  return (const int*)shape_data_->gpu_data();

}

 

template <typename Dtype>

const Dtype* Blob<Dtype>::cpu_data() const {

  CHECK(data_);

  return (const Dtype*)data_->cpu_data();

}

 

template <typename Dtype>

void Blob<Dtype>::set_cpu_data(Dtype* data) {

  CHECK(data);

  // Make sure CPU and GPU sizes remain equal

  size_t size = count_ * sizeof(Dtype);

  if (data_->size() != size) {

    data_.reset(new SyncedMemory(size));

    diff_.reset(new SyncedMemory(size));

  }

  data_->set_cpu_data(data);

}

 

template <typename Dtype>

const Dtype* Blob<Dtype>::gpu_data() const {

  CHECK(data_);

  return (const Dtype*)data_->gpu_data();

}

 

template <typename Dtype>

void Blob<Dtype>::set_gpu_data(Dtype* data) {

  CHECK(data);

  // Make sure CPU and GPU sizes remain equal

  size_t size = count_ * sizeof(Dtype);

  if (data_->size() != size) {

    data_.reset(new SyncedMemory(size));

    diff_.reset(new SyncedMemory(size));

  }

  data_->set_gpu_data(data);

}

 

template <typename Dtype>

const Dtype* Blob<Dtype>::cpu_diff() const {

  CHECK(diff_);

  return (const Dtype*)diff_->cpu_data();

}

 

template <typename Dtype>

const Dtype* Blob<Dtype>::gpu_diff() const {

  CHECK(diff_);

  return (const Dtype*)diff_->gpu_data();

}

 

template <typename Dtype>

Dtype* Blob<Dtype>::mutable_cpu_data() {

  CHECK(data_);

  return static_cast<Dtype*>(data_->mutable_cpu_data());

}

 

template <typename Dtype>

Dtype* Blob<Dtype>::mutable_gpu_data() {

  CHECK(data_);

  return static_cast<Dtype*>(data_->mutable_gpu_data());

}

 

template <typename Dtype>

Dtype* Blob<Dtype>::mutable_cpu_diff() {

  CHECK(diff_);

  return static_cast<Dtype*>(diff_->mutable_cpu_data());

}

 

template <typename Dtype>

Dtype* Blob<Dtype>::mutable_gpu_diff() {

  CHECK(diff_);

  return static_cast<Dtype*>(diff_->mutable_gpu_data());

}

 

template <typename Dtype>

void Blob<Dtype>::ShareData(const Blob& other) {

  CHECK_EQ(count_, other.count());

  data_ = other.data();

}

 

template <typename Dtype>

void Blob<Dtype>::ShareDiff(const Blob& other) {

  CHECK_EQ(count_, other.count());

  diff_ = other.diff();

}

 

// The "update" method is used for parameter blobs in a Net, which are stored

// as Blob<float> or Blob<double> -- hence we do not define it for

// Blob<int> or Blob<unsigned int>.

template <> void Blob<unsigned int>::Update() { NOT_IMPLEMENTED; }

template <> void Blob<int>::Update() { NOT_IMPLEMENTED; }

 

template <typename Dtype>

void Blob<Dtype>::Update() {

  // We will perform update based on where the data is located.

  switch (data_->head()) {

  case SyncedMemory::HEAD_AT_CPU:

    // perform computation on CPU

    caffe_axpy<Dtype>(count_, Dtype(-1),

        static_cast<const Dtype*>(diff_->cpu_data()),

        static_cast<Dtype*>(data_->mutable_cpu_data()));

    break;

  case SyncedMemory::HEAD_AT_GPU:

  case SyncedMemory::SYNCED:

#ifndef CPU_ONLY

    // perform computation on GPU

    caffe_gpu_axpy<Dtype>(count_, Dtype(-1),

        static_cast<const Dtype*>(diff_->gpu_data()),

        static_cast<Dtype*>(data_->mutable_gpu_data()));

#else

    NO_GPU;

#endif

    break;

  default:

    LOG(FATAL) << "Syncedmem not initialized.";

  }

}

 

template <> unsigned int Blob<unsigned int>::asum_data() const {

  NOT_IMPLEMENTED;

  return 0;

}

 

template <> int Blob<int>::asum_data() const {

  NOT_IMPLEMENTED;

  return 0;

}

 

template <typename Dtype>

Dtype Blob<Dtype>::asum_data() const {

  if (!data_) { return 0; }

  switch (data_->head()) {

  case SyncedMemory::HEAD_AT_CPU:

    return caffe_cpu_asum(count_, cpu_data());

  case SyncedMemory::HEAD_AT_GPU:

  case SyncedMemory::SYNCED:

#ifndef CPU_ONLY

  {

    Dtype asum;

    caffe_gpu_asum(count_, gpu_data(), &asum);

    return asum;

  }

#else

    NO_GPU;

#endif

  case SyncedMemory::UNINITIALIZED:

    return 0;

  default:

    LOG(FATAL) << "Unknown SyncedMemory head state: " << data_->head();

  }

  return 0;

}

 

template <> unsigned int Blob<unsigned int>::asum_diff() const {

  NOT_IMPLEMENTED;

  return 0;

}

 

template <> int Blob<int>::asum_diff() const {

  NOT_IMPLEMENTED;

  return 0;

}

 

template <typename Dtype>

Dtype Blob<Dtype>::asum_diff() const {

  if (!diff_) { return 0; }

  switch (diff_->head()) {

  case SyncedMemory::HEAD_AT_CPU:

    return caffe_cpu_asum(count_, cpu_diff());

  case SyncedMemory::HEAD_AT_GPU:

  case SyncedMemory::SYNCED:

#ifndef CPU_ONLY

  {

    Dtype asum;

    caffe_gpu_asum(count_, gpu_diff(), &asum);

    return asum;

  }

#else

    NO_GPU;

#endif

  case SyncedMemory::UNINITIALIZED:

    return 0;

  default:

    LOG(FATAL) << "Unknown SyncedMemory head state: " << diff_->head();

  }

  return 0;

}

 

template <> unsigned int Blob<unsigned int>::sumsq_data() const {

  NOT_IMPLEMENTED;

  return 0;

}

 

template <> int Blob<int>::sumsq_data() const {

  NOT_IMPLEMENTED;

  return 0;

}

 

template <typename Dtype>

Dtype Blob<Dtype>::sumsq_data() const {

  Dtype sumsq;

  const Dtype* data;

  if (!data_) { return 0; }

  switch (data_->head()) {

  case SyncedMemory::HEAD_AT_CPU:

    data = cpu_data();

    sumsq = caffe_cpu_dot(count_, data, data);

    break;

  case SyncedMemory::HEAD_AT_GPU:

  case SyncedMemory::SYNCED:

#ifndef CPU_ONLY

    data = gpu_data();

    caffe_gpu_dot(count_, data, data, &sumsq);

#else

    NO_GPU;

#endif

    break;

  case SyncedMemory::UNINITIALIZED:

    return 0;

  default:

    LOG(FATAL) << "Unknown SyncedMemory head state: " << data_->head();

  }

  return sumsq;

}

 

template <> unsigned int Blob<unsigned int>::sumsq_diff() const {

  NOT_IMPLEMENTED;

  return 0;

}

 

template <> int Blob<int>::sumsq_diff() const {

  NOT_IMPLEMENTED;

  return 0;

}

 

template <typename Dtype>

Dtype Blob<Dtype>::sumsq_diff() const {

  Dtype sumsq;

  const Dtype* diff;

  if (!diff_) { return 0; }

  switch (diff_->head()) {

  case SyncedMemory::HEAD_AT_CPU:

    diff = cpu_diff();

    sumsq = caffe_cpu_dot(count_, diff, diff);

    break;

  case SyncedMemory::HEAD_AT_GPU:

  case SyncedMemory::SYNCED:

#ifndef CPU_ONLY

    diff = gpu_diff();

    caffe_gpu_dot(count_, diff, diff, &sumsq);

    break;

#else

    NO_GPU;

#endif

  case SyncedMemory::UNINITIALIZED:

    return 0;

  default:

    LOG(FATAL) << "Unknown SyncedMemory head state: " << data_->head();

  }

  return sumsq;

}

 

template <> void Blob<unsigned int>::scale_data(unsigned int scale_factor) {

  NOT_IMPLEMENTED;

}

 

template <> void Blob<int>::scale_data(int scale_factor) {

  NOT_IMPLEMENTED;

}

 

template <typename Dtype>

void Blob<Dtype>::scale_data(Dtype scale_factor) {

  Dtype* data;

  if (!data_) { return; }

  switch (data_->head()) {

  case SyncedMemory::HEAD_AT_CPU:

    data = mutable_cpu_data();

    caffe_scal(count_, scale_factor, data);

    return;

  case SyncedMemory::HEAD_AT_GPU:

  case SyncedMemory::SYNCED:

#ifndef CPU_ONLY

    data = mutable_gpu_data();

    caffe_gpu_scal(count_, scale_factor, data);

    return;

#else

    NO_GPU;

#endif

  case SyncedMemory::UNINITIALIZED:

    return;

  default:

    LOG(FATAL) << "Unknown SyncedMemory head state: " << data_->head();

  }

}

 

template <> void Blob<unsigned int>::scale_diff(unsigned int scale_factor) {

  NOT_IMPLEMENTED;

}

 

template <> void Blob<int>::scale_diff(int scale_factor) {

  NOT_IMPLEMENTED;

}

 

template <typename Dtype>

void Blob<Dtype>::scale_diff(Dtype scale_factor) {

  Dtype* diff;

  if (!diff_) { return; }

  switch (diff_->head()) {

  case SyncedMemory::HEAD_AT_CPU:

    diff = mutable_cpu_diff();

    caffe_scal(count_, scale_factor, diff);

    return;

  case SyncedMemory::HEAD_AT_GPU:

  case SyncedMemory::SYNCED:

#ifndef CPU_ONLY

    diff = mutable_gpu_diff();

    caffe_gpu_scal(count_, scale_factor, diff);

    return;

#else

    NO_GPU;

#endif

  case SyncedMemory::UNINITIALIZED:

    return;

  default:

    LOG(FATAL) << "Unknown SyncedMemory head state: " << diff_->head();

  }

}

 

template <typename Dtype>

bool Blob<Dtype>::ShapeEquals(const BlobProto& other) {

  if (other.has_num() || other.has_channels() ||

      other.has_height() || other.has_width()) {

    // Using deprecated 4D Blob dimensions --

    // shape is (num, channels, height, width).

    // Note: we do not use the normal Blob::num(), Blob::channels(), etc.

    // methods as these index from the beginning of the blob shape, where legacy

    // parameter blobs were indexed from the end of the blob shape (e.g., bias

    // Blob shape (1 x 1 x 1 x N), IP layer weight Blob shape (1 x 1 x M x N)).

    return shape_.size() <= 4 &&

           LegacyShape(-4) == other.num() &&

           LegacyShape(-3) == other.channels() &&

           LegacyShape(-2) == other.height() &&

           LegacyShape(-1) == other.width();

  }

  vector<int> other_shape(other.shape().dim_size());

  for (int i = 0; i < other.shape().dim_size(); ++i) {

    other_shape[i] = other.shape().dim(i);

  }

  return shape_ == other_shape;

}

 

template <typename Dtype>

void Blob<Dtype>::CopyFrom(const Blob& source, bool copy_diff, bool reshape) {

  if (source.count() != count_ || source.shape() != shape_) {

    if (reshape) {

      ReshapeLike(source);

    } else {

      LOG(FATAL) << "Trying to copy blobs of different sizes.";

    }

  }

  switch (Caffe::mode()) {

  case Caffe::GPU:

    if (copy_diff) {

      caffe_copy(count_, source.gpu_diff(),

          static_cast<Dtype*>(diff_->mutable_gpu_data()));

    } else {

      caffe_copy(count_, source.gpu_data(),

          static_cast<Dtype*>(data_->mutable_gpu_data()));

    }

    break;

  case Caffe::CPU:

    if (copy_diff) {

      caffe_copy(count_, source.cpu_diff(),

          static_cast<Dtype*>(diff_->mutable_cpu_data()));

    } else {

      caffe_copy(count_, source.cpu_data(),

          static_cast<Dtype*>(data_->mutable_cpu_data()));

    }

    break;

  default:

    LOG(FATAL) << "Unknown caffe mode.";

  }

}

 

template <typename Dtype>

void Blob<Dtype>::FromProto(const BlobProto& proto, bool reshape) {

  if (reshape) {

    vector<int> shape;

    if (proto.has_num() || proto.has_channels() ||

        proto.has_height() || proto.has_width()) {

      // Using deprecated 4D Blob dimensions --

      // shape is (num, channels, height, width).

      shape.resize(4);

      shape[0] = proto.num();

      shape[1] = proto.channels();

      shape[2] = proto.height();

      shape[3] = proto.width();

    } else {

      shape.resize(proto.shape().dim_size());

      for (int i = 0; i < proto.shape().dim_size(); ++i) {

        shape[i] = proto.shape().dim(i);

      }

    }

    Reshape(shape);

  } else {

    CHECK(ShapeEquals(proto)) << "shape mismatch (reshape not set)";

  }

  // copy data

  Dtype* data_vec = mutable_cpu_data();

  if (proto.double_data_size() > 0) {

    CHECK_EQ(count_, proto.double_data_size());

    for (int i = 0; i < count_; ++i) {

      data_vec[i] = proto.double_data(i);

    }

  } else {

    CHECK_EQ(count_, proto.data_size());

    for (int i = 0; i < count_; ++i) {

      data_vec[i] = proto.data(i);

    }

  }

  if (proto.double_diff_size() > 0) {

    CHECK_EQ(count_, proto.double_diff_size());

    Dtype* diff_vec = mutable_cpu_diff();

    for (int i = 0; i < count_; ++i) {

      diff_vec[i] = proto.double_diff(i);

    }

  } else if (proto.diff_size() > 0) {

    CHECK_EQ(count_, proto.diff_size());

    Dtype* diff_vec = mutable_cpu_diff();

    for (int i = 0; i < count_; ++i) {

      diff_vec[i] = proto.diff(i);

    }

  }

}

 

template <>

void Blob<double>::ToProto(BlobProto* proto, bool write_diff) const {

  proto->clear_shape();

  for (int i = 0; i < shape_.size(); ++i) {

    proto->mutable_shape()->add_dim(shape_[i]);

  }

  proto->clear_double_data();

  proto->clear_double_diff();

  const double* data_vec = cpu_data();

  for (int i = 0; i < count_; ++i) {

    proto->add_double_data(data_vec[i]);

  }

  if (write_diff) {

    const double* diff_vec = cpu_diff();

    for (int i = 0; i < count_; ++i) {

      proto->add_double_diff(diff_vec[i]);

    }

  }

}

 

template <>

void Blob<float>::ToProto(BlobProto* proto, bool write_diff) const {

  proto->clear_shape();

  for (int i = 0; i < shape_.size(); ++i) {

    proto->mutable_shape()->add_dim(shape_[i]);

  }

  proto->clear_data();

  proto->clear_diff();

  const float* data_vec = cpu_data();

  for (int i = 0; i < count_; ++i) {

    proto->add_data(data_vec[i]);

  }

  if (write_diff) {

    const float* diff_vec = cpu_diff();

    for (int i = 0; i < count_; ++i) {

      proto->add_diff(diff_vec[i]);

    }

  }

}

 

INSTANTIATE_CLASS(Blob);

template class Blob<int>;

template class Blob<unsigned int>;

 

}  // namespace caffe

Blob是砖

转载于:https://www.cnblogs.com/mengmengmiaomiao/p/7516300.html