void ScalarLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
const ScalarParameter& param = this->layer_param_.scalar_param();
Blob<Dtype>* scalar = (bottom.size() > 1) ? bottom[1] : this->blobs_[0].get();
// Always set axis_ == 0 in special case where scalar is an actual scalar
// (num_axes == 0). Mathematically equivalent for any choice of axis_, so the
// actual setting can be safely ignored; and computation is most efficient
// with axis_ == 0 and (therefore) outer_dim_ == 1. (Setting axis_ to
// bottom[0]->num_axes() - 1, giving inner_dim_ == 1, would be equally
// performant.)
axis_ = (scalar->num_axes() == 0) ?
0 : bottom[0]->CanonicalAxisIndex(param.axis());
CHECK_GE(bottom[0]->num_axes(), axis_ + scalar->num_axes())
<< "scalar blob's shape extends past bottom[0]'s shape when applied "
<< "starting with bottom[0] axis = " << axis_;
for (int i = 0; i < scalar->num_axes(); ++i) {
CHECK_EQ(bottom[0]->shape(axis_ + i), scalar->shape(i))
<< "dimension mismatch between bottom[0]->shape(" << axis_ + i
<< ") and scalar->shape(" << i << ")";
}
outer_dim_ = bottom[0]->count(0, axis_);
scalar_dim_ = scalar->count();
inner_dim_ = bottom[0]->count(axis_ + scalar->num_axes());
if (bottom[0] == top[0]) { // in-place computation
temp_.ReshapeLike(*bottom[0]);
} else {
top[0]->ReshapeLike(*bottom[0]);
}
sum_result_.Reshape(vector<int>(1, outer_dim_ * scalar_dim_));
const int sum_mult_size = std::max(outer_dim_, inner_dim_);
sum_multiplier_.Reshape(vector<int>(1, sum_mult_size));
if (sum_multiplier_.cpu_data()[sum_mult_size - 1] != Dtype(1)) {
caffe_set(sum_mult_size, Dtype(1), sum_multiplier_.mutable_cpu_data());
}
}
TYPED_TEST(DataTransformTest, TestCropSize) {
TransformationParameter transform_param;
const bool unique_pixels = false; // all pixels the same equal to label
const int label = 0;
const int channels = 3;
const int height = 4;
const int width = 5;
const int crop_size = 2;
transform_param.set_crop_size(crop_size);
Datum datum;
FillDatum(label, channels, height, width, unique_pixels, &datum);
DataTransformer<TypeParam>* transformer =
new DataTransformer<TypeParam>(transform_param, TEST,
Caffe::GetDefaultDeviceContext());
transformer->InitRand();
Blob<TypeParam>* blob =
new Blob<TypeParam>(1, channels, crop_size, crop_size);
for (int iter = 0; iter < this->num_iter_; ++iter) {
transformer->Transform(datum, blob);
EXPECT_EQ(blob->num(), 1);
EXPECT_EQ(blob->channels(), datum.channels());
EXPECT_EQ(blob->height(), crop_size);
EXPECT_EQ(blob->width(), crop_size);
for (int j = 0; j < blob->count(); ++j) {
EXPECT_EQ(blob->cpu_data()[j], label);
}
}
}
Blob::Blob(const Blob &source) {
if (data_)
data_ = nullptr;
shape_ = source.shape();
count_ = source.count();
data_ = source.data();
}
TYPED_TEST(Col2ImgMaskLayerTest, TestForward_2) {
typedef typename TypeParam::Dtype Dtype;
LayerParameter layer_param;
ConvolutionParameter* convolution_param =
layer_param.mutable_convolution_param();
//convolution_param->set_kernel_size(0,3);
//convolution_param->set_stride(0,2);
convolution_param->add_kernel_size(3);
convolution_param->add_stride(2);
caffe_set(blob_bottom_->count(), (Dtype)2, blob_bottom_->mutable_cpu_data());
Blob<Dtype> mask;
mask.ReshapeLike(*blob_bottom_);
caffe_set(mask.count(), (Dtype)1, mask.mutable_cpu_data());
vector<Blob<Dtype>*> blob_bottom_vec_2_;
blob_bottom_vec_2_.push_back(blob_bottom_);
blob_bottom_vec_2_.push_back(&mask);
Col2imgMaskLayer<Dtype> layer(layer_param);
layer.SetUp(blob_bottom_vec_2_, blob_top_vec_);
EXPECT_EQ(this->blob_top_->num(), 2);
EXPECT_EQ(this->blob_top_->channels(), 2);
EXPECT_EQ(this->blob_top_->height(), 5);
EXPECT_EQ(this->blob_top_->width(), 5);
layer.Forward(blob_bottom_vec_, blob_top_vec_);
const Dtype min_precision = 1e-5;
for (int i = 0; i < blob_top_->count(); i++)
EXPECT_NEAR(blob_top_->mutable_cpu_data()[i], 2, min_precision);
}
void ConcatLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
const Dtype* top_diff = top[0]->cpu_diff();
if (concat_dim_ == 0) {
int offset_num = 0;
for (int i = 0; i < bottom.size(); ++i) {
Blob<Dtype>* blob = bottom[i];
if (propagate_down[i]) {
Dtype* bottom_diff = blob->mutable_cpu_diff();
caffe_copy(blob->count(), top_diff + top[0]->offset(offset_num),
bottom_diff);
}
offset_num += blob->num();
}
} else if (concat_dim_ == 1) {
int offset_channel = 0;
for (int i = 0; i < bottom.size(); ++i) {
Blob<Dtype>* blob = bottom[i];
if (propagate_down[i]) {
Dtype* bottom_diff = blob->mutable_cpu_diff();
int num_elem = blob->channels()*blob->height()*blob->width();
for (int n = 0; n < num_; ++n) {
caffe_copy(num_elem, top_diff + top[0]->offset(n, offset_channel),
bottom_diff + blob->offset(n));
}
}
offset_channel += blob->channels();
}
} // concat_dim_ is guaranteed to be 0 or 1 by LayerSetUp.
}
Dtype SliceLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
vector<Blob<Dtype>*>* top) {
const Dtype* bottom_data = bottom[0]->mutable_cpu_data();
if (slice_dim_ == 0) {
int offset_num = 0;
for (int i = 0; i < top->size(); ++i) {
Blob<Dtype>* blob = (*top)[i];
Dtype* top_data = blob->mutable_cpu_data();
caffe_copy(blob->count(), bottom_data + bottom[0]->offset(offset_num),
top_data);
offset_num += blob->num();
}
} else if (slice_dim_ == 1) {
int offset_channel = 0;
for (int i = 0; i < top->size(); ++i) {
Blob<Dtype>* blob = (*top)[i];
Dtype* top_data = blob->mutable_cpu_data();
const int num_elem = blob->channels() * blob->height() * blob->width()*blob->depth();
for (int n = 0; n < num_; ++n) {
caffe_copy(num_elem, bottom_data + bottom[0]->offset(n, offset_channel),
top_data + blob->offset(n));
}
offset_channel += blob->channels();
}
} // slice_dim_ is guaranteed to be 0 or 1 by SetUp.
return Dtype(0.);
}
void InfogainLossLayer<Dtype, MItype, MOtype>::Reshape(
const vector<Blob<MItype>*>& bottom,
const vector<Blob<MOtype>*>& top) {
LossLayer<Dtype, MItype, MOtype>::Reshape(bottom, top);
softmax_layer_->Reshape(softmax_bottom_vec_, softmax_top_vec_);
infogain_axis_ =
bottom[0]->CanonicalAxisIndex(
this->layer_param_.infogain_loss_param().axis());
outer_num_ = bottom[0]->count(0, infogain_axis_);
inner_num_ = bottom[0]->count(infogain_axis_ + 1);
CHECK_EQ(outer_num_ * inner_num_, bottom[1]->count())
<< "Number of labels must match number of predictions; "
<< "e.g., if infogain axis == 1 and prediction shape is (n, c, H, W), "
<< "label count (number of labels) must be n*H*W, "
<< "with integer values in {0, 1, ..., c-1}.";
num_labels_ = bottom[0]->shape(infogain_axis_);
Blob<Dtype>* infogain = NULL;
if (bottom.size() < 3) {
infogain = &infogain_;
} else {
infogain = bottom[2];
}
CHECK_EQ(infogain->count(), num_labels_*num_labels_);
sum_rows_H_.Reshape(vector<int_tp>(1, num_labels_));
if (bottom.size() == 2) {
// H is provided as a parameter and will not change. sum rows once
sum_rows_of_H(infogain);
}
if (top.size() >= 2) {
// softmax output
top[1]->ReshapeLike(*bottom[0]);
}
}
void SliceLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, vector<Blob<Dtype>*>* bottom) {
if (!propagate_down[0]) { return; }
Dtype* bottom_diff = (*bottom)[0]->mutable_cpu_diff();
if (slice_dim_ == 0) {
int offset_num = 0;
for (int i = 0; i < top.size(); ++i) {
Blob<Dtype>* blob = top[i];
const Dtype* top_diff = blob->cpu_diff();
caffe_copy(blob->count(), top_diff,
bottom_diff + (*bottom)[0]->offset(offset_num));
offset_num += blob->num();
}
} else if (slice_dim_ == 1) {
int offset_channel = 0;
for (int i = 0; i < top.size(); ++i) {
Blob<Dtype>* blob = top[i];
const Dtype* top_diff = blob->cpu_diff();
const int num_elem = blob->channels() * blob->height() * blob->width() * blob->depth();
for (int n = 0; n < num_; ++n) {
caffe_copy(num_elem, top_diff + blob->offset(n),
bottom_diff + (*bottom)[0]->offset(n, offset_channel));
}
offset_channel += blob->channels();
}
} // slice_dim_ is guaranteed to be 0 or 1 by SetUp.
}
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.";
}
}
void BiasLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
const BiasParameter& param = this->layer_param_.bias_param();
Blob<Dtype>* bias = (bottom.size() > 1) ? bottom[1] : this->blobs_[0].get();
// Always set axis == 0 in special case where bias is a scalar
// (num_axes == 0). Mathematically equivalent for any choice of axis, so the
// actual setting can be safely ignored; and computation is most efficient
// with axis == 0 and (therefore) outer_dim_ == 1.
const int axis = (bias->num_axes() == 0) ?
0 : bottom[0]->CanonicalAxisIndex(param.axis());
CHECK_GE(bottom[0]->num_axes(), axis + bias->num_axes())
<< "bias blob's shape extends past bottom[0]'s shape when applied "
<< "starting with bottom[0] axis = " << axis;
for (int i = 0; i < bias->num_axes(); ++i) {
CHECK_EQ(bottom[0]->shape(axis + i), bias->shape(i))
<< "dimension mismatch between bottom[0]->shape(" << axis + i
<< ") and bias->shape(" << i << ")";
}
outer_dim_ = bottom[0]->count(0, axis);
bias_dim_ = bias->count();
inner_dim_ = bottom[0]->count(axis + bias->num_axes());
dim_ = bias_dim_ * inner_dim_;
if (bottom[0] != top[0]) {
top[0]->ReshapeLike(*bottom[0]);
}
bias_multiplier_.Reshape(vector<int>(1, inner_dim_));
if (bias_multiplier_.cpu_data()[inner_dim_ - 1] != Dtype(1)) {
caffe_set(inner_dim_, Dtype(1), bias_multiplier_.mutable_cpu_data());
}
}
void Blob<Dtype>::ShareDiff(const Blob& other) {
// SID MEMORY COMPACT LIGHT WEIGHT CAFFE <BEGIN>
if(_is_diff_initialized==0){
diff_.reset(new SyncedMemory(capacity_*sizeof(Dtype)));
_is_diff_initialized=1;
}
// SID MEMORY COMPACT LIGHT WEIGHT CAFFE <END>
CHECK_EQ(count_, other.count());
diff_ = other.diff();
}
int NumSequenceMatches(const TransformationParameter transform_param,
const Datum& datum, Phase phase) {
// Get crop sequence with Caffe seed 1701.
DataTransformer<Dtype>* transformer =
new DataTransformer<Dtype>(transform_param, phase);
const int crop_size = transform_param.crop_size();
int crop_h = transform_param.crop_h();
int crop_w = transform_param.crop_w();
if (crop_size > 0) {
crop_h = crop_w = crop_size;
}
Caffe::set_random_seed(seed_);
transformer->InitRand();
Blob<Dtype>* blob =
new Blob<Dtype>(1, datum.channels(), datum.height(), datum.width());
if (crop_h > 0 || crop_w > 0) {
blob->Reshape(1, datum.channels(), crop_h, crop_w);
}
vector<vector<Dtype> > crop_sequence;
for (int iter = 0; iter < this->num_iter_; ++iter) {
vector<Dtype> iter_crop_sequence;
transformer->Transform(datum, blob);
for (int j = 0; j < blob->count(); ++j) {
iter_crop_sequence.push_back(blob->cpu_data()[j]);
}
crop_sequence.push_back(iter_crop_sequence);
}
// Check if the sequence differs from the previous
int num_sequence_matches = 0;
for (int iter = 0; iter < this->num_iter_; ++iter) {
vector<Dtype> iter_crop_sequence = crop_sequence[iter];
transformer->Transform(datum, blob);
for (int j = 0; j < blob->count(); ++j) {
num_sequence_matches +=
(crop_sequence[iter][j] == blob->cpu_data()[j]);
}
}
return num_sequence_matches;
}
static void write_blob_to_file(const std::string& file_name,
const Blob<Dtype>& blob) {
std::ofstream file(file_name.c_str(), std::ios::out | std::ios::binary);
if (file.fail()) {
ASSERT_FALSE(true);
return;
}
file.write(reinterpret_cast<const char*>(&blob.shape()[0]), 4 * sizeof(int));
ASSERT_FALSE(file.fail());
file.write(reinterpret_cast<const char*>(blob.cpu_data()),
blob.count() * sizeof(Dtype));
ASSERT_FALSE(file.fail());
file.close();
}
void AsyncParamServer<Dtype>::ProcessUpdateTask() {
const vector<Blob<Dtype> *> &net_params = solver_->net()->learnable_params();
std::deque<TaskRequest> to_update;
update_queue_mutex_.lock();
to_update.swap(update_tasks_);
update_queue_mutex_.unlock();
while (!to_update.empty() ) {
TaskRequest task = to_update.front();
to_update.pop_front();
// copy to diff in solver
int root_rank = world_rank_to_root_rank(task.part_root_rank_);
Blob<Dtype>* blob = net_params[task.param_id_];
Dtype* solver_diff = blob->mutable_cpu_diff();
Dtype* mpi_buf =
recv_buf_[make_pair(root_rank, task.param_id_)].first;
int64_t count =
recv_buf_[make_pair(root_rank, task.param_id_)].second;
CHECK(count == blob->count() );
//copy MPI buffer to solver_diff
int64_t part_offset = task.part_id_ * count / task.num_parts_;
caffe_copy(count / task.num_parts_,
mpi_buf + part_offset, solver_diff + part_offset);
// apply update
int blob_wise_iter = async_iter_[make_pair(task.param_id_, task.part_id_) ];
solver_->set_iter(blob_wise_iter);
// TODO: supports partial param update per model parts
solver_->ApplyUpdate(task.param_id_);
DLOG(INFO) << "PS (iter " << blob_wise_iter << "): param id=" << task.param_id_ << " weight=" << net_params[task.param_id_]->sumsq_diff();
DLOG(INFO) << "PS (iter " << blob_wise_iter << "): param id=" << task.param_id_ << " data=" << net_params[task.param_id_]->sumsq_data();
//clean up
solver_->net()->ClearParamDiffs(task.param_id_);
async_iter_[ make_pair(task.param_id_, task.part_id_) ] += 1;
update_cnt_ += 1;
// copy model(data) in solver to mpi buffer
mpi_buf = send_buf_[make_pair(root_rank, task.param_id_)].first;
caffe_copy(count / task.num_parts_,
blob->cpu_data() + part_offset, mpi_buf + part_offset);
//ship off
send_queue_mutex_.lock();
send_tasks_.push_back(task);
send_queue_mutex_.unlock();
}
}
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 (device_->backend() == BACKEND_CUDA) {
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()));
}
} else {
#ifdef USE_GREENTEA
if (copy_diff) {
greentea_copy<Dtype>(
count_, (cl_mem) (source.gpu_diff()), 0,
(cl_mem) (diff_->mutable_gpu_data()), 0,
&viennacl::ocl::get_context(device_->id()));
} else {
greentea_copy<Dtype>(
count_, (cl_mem) (source.gpu_data()), 0,
(cl_mem) (data_->mutable_gpu_data()), 0,
&viennacl::ocl::get_context(device_->id()));
}
#endif
}
break;
}
case Caffe::CPU: {
if (copy_diff) {
caffe_cpu_copy(count_, source.cpu_diff(),
static_cast<Dtype*>(diff_->mutable_cpu_data()));
} else {
caffe_cpu_copy(count_, source.cpu_data(),
static_cast<Dtype*>(data_->mutable_cpu_data()));
}
break;
}
default:
LOG(FATAL)<< "Unknown caffe mode.";
}
}
void BatchTripletLossLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
if (propagate_down[1]) {
LOG(FATAL) << this->type()
<< " Layer cannot backpropagate to label inputs.";
}
if (propagate_down[0]) {
Blob<Dtype>* feat = bottom[0];
const Dtype* feat_data = feat->cpu_data();
Dtype* feat_diff = feat->mutable_cpu_diff();
int count = feat->count();
int num = feat->num();
int dim = count / num;
int agg_step = num * sizeof(Dtype);
Dtype * agg_data = (Dtype *)aggregator_->mutable_cpu_data();
caffe_memset(num * agg_step, 0, agg_data);
Dtype scale1 = Dtype(2) / triplets_.size() * mu_;
for (int i=0; i<triplets_.size(); ++i) {
int qry_id = triplets_[i].first_;
int pos_id = triplets_[i].second_;
int neg_id = triplets_[i].third_;
agg_data[qry_id * num + neg_id] += scale1;
agg_data[qry_id * num + pos_id] -= scale1;
agg_data[pos_id * num + pos_id] += scale1;
agg_data[pos_id * num + qry_id] -= scale1;
agg_data[neg_id * num + qry_id] += scale1;
agg_data[neg_id * num + neg_id] -= scale1;
}
Dtype scale2 = Dtype(2) / pos_pairs_.size() * (Dtype(1) - mu_);
for (int i=0; i<pos_pairs_.size(); ++i) {
int qry_id = pos_pairs_[i].first;
int pos_id = pos_pairs_[i].second;
agg_data[qry_id * num + qry_id] += scale2;
agg_data[qry_id * num + pos_id] -= scale2;
agg_data[pos_id * num + pos_id] += scale2;
agg_data[pos_id * num + qry_id] -= scale2;
}
caffe_cpu_gemm(CblasNoTrans, CblasNoTrans, num, dim, num,
Dtype(1), agg_data, feat_data, Dtype(0), feat_diff);
}
}
static void read_blob_from_file(const std::string& file_name,
Blob<Dtype>& blob) {
std::ifstream file(file_name.c_str(), std::ifstream::binary);
if (file.fail()) {
ASSERT_FALSE(true);
return;
}
vector<int> shape(4, 0);
file.read(reinterpret_cast<char*>(&shape[0]), 4 * sizeof(int));
ASSERT_FALSE(file.fail());
blob.Reshape(shape);
file.read(reinterpret_cast<char*>(blob.mutable_cpu_data()),
blob.count() * sizeof(Dtype));
ASSERT_FALSE(file.fail());
}
void CaffeMobile::putImage(AndroidBitmapInfo* info, void* pixels, const vector<Blob<float>*>& resImage) {
Blob<float> * srcBlob = *resImage.data();
LOG(DEBUG) << "srcBlob received";
vector<int> shape = {1, 3, (int) info->width, (int) info->height };
LOG(DEBUG) << "shape configured";
Blob<float>* imgBlob = new Blob<float>();
LOG(DEBUG) << "Blob created";
imgBlob->Reshape(shape);
LOG(DEBUG) << "imgBlob reshaped";
imgBlob->CopyFrom(*srcBlob, false, true);
LOG(DEBUG) << "imgBlob copied";
int size = imgBlob->count();
LOG(DEBUG) << "imgBlob size is: " << size;
/*Partially from https://github.com/ruckus/android-image-filter-ndk*/
uint32_t* pixelRow;
int ix, iy, red, green, blue;
for(iy = 0; iy < (int) info->height; iy++){
pixelRow = (uint32_t*) pixels;
for(ix =0; ix < (int) info->width; ix++){
red = (int) clip(imgBlob->data_at(0,0,iy,ix), 0, 255);
green = (int) clip(imgBlob->data_at(0,1,iy,ix), 0, 255);
blue = (int) clip(imgBlob->data_at(0,2,iy,ix), 0, 255);
pixelRow[ix] =
((red << 16) & 0x00FF0000) |
((green << 8) & 0x0000FF00) |
(blue & 0x000000FF);
}
pixels = (char*)pixels + info->stride;
}
LOG(DEBUG) << "before return putImage " << size;
return;
}
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.";
}
}
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()));
}
}
AsyncParamServer<Dtype>::AsyncParamServer(boost::shared_ptr<Solver<Dtype> > solver) :
recv_tasks_iter_(0),
solver_(solver),
send_cnt_(0), update_cnt_(0) {
// setup the mpi buffers and recv task vector
int mpi_rank = get_node_rank();
shared_ptr<Net<Dtype>> net = solver_->net();
const vector<Blob<Dtype> *> &net_params = net->learnable_params();
const std::vector<bool>& layer_need_backward{ net->layer_need_backward() };
for (int i = 0; i < get_num_groups(); i++) {
int root_rank = get_group_root_rank(i);
//iterate over layers and skip the ones without params
for (int j = 0; j < net->layers().size(); j++) {
shared_ptr<Layer<Dtype>> layer = net->layers()[j];
if (!layer_need_backward[j])
continue;
const MultinodeLayerParameter & mn_layer_param = layer->layer_param().multinode();
int model_parts = mn_layer_param.model_parts();
int mn_num_nodes = mn_layer_param.num_nodes();
GetCanonicalMnParam(mn_num_nodes, model_parts);
vector<int> layer_param_ids = net->get_layer_learnable_param_ids(j);
for (int k = 0; k < layer_param_ids.size(); k++) {
int param_id = layer_param_ids[k];
if (!layer->ParamNeedReduce(k)) continue;
if (param_to_server_rank(j, param_id) != mpi_rank) continue;
Blob<Dtype> *blob = net_params[param_id];
// Setup buf for recv
Dtype* buf = (Dtype*)std::malloc(sizeof(Dtype) * blob->count());
recv_buf_[make_pair(root_rank, param_id)] = make_pair(buf, blob->count());
for (int part_id = 0; part_id < model_parts; part_id++) {
int part_root_rank = get_group_root_rank(i, part_id, model_parts);
int64_t part_offset = part_id * blob->count() / model_parts;
TaskRequest recv_task(part_root_rank, j, param_id, part_id, model_parts);
recv_tasks_.push_back(recv_task);
rank_layer_blob_to_vec_pos[make_pair(part_root_rank, param_id)] =
recv_tasks_.size() - 1;
MPI_Irecv(buf + part_offset, blob->count() / model_parts,
DtypeToMPIDtype<Dtype>(), part_root_rank,
recv_task.GetTag(), MPI_COMM_WORLD,
&(recv_tasks_[recv_tasks_.size() - 1].mpi_request_));
async_iter_[make_pair(param_id, part_id)] = solver_->iter();
}
// Setup buf for send
buf = (Dtype*)std::malloc(sizeof(Dtype) * blob->count());
send_buf_[make_pair(root_rank, param_id)] = make_pair(buf, blob->count());
}
}
}
// assumed iter is started from 0
total_update_ = total_send_ = recv_tasks_.size() * solver_->param().max_iter();
}
void ConcatLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
const bool propagate_down, vector<Blob<Dtype>*>* bottom) {
const Dtype* top_diff = top[0]->cpu_diff();
if (concat_dim_ == 0) {
int offset_num = 0;
for (int i = 0; i < bottom->size(); ++i) {
Blob<Dtype>* blob = (*bottom)[i];
Dtype* bottom_diff = blob->mutable_cpu_diff();
caffe_copy(blob->count(),
top_diff+top[0]->offset(offset_num), bottom_diff);
offset_num += blob->num();
}
} else if (concat_dim_ == 1) {
int offset_channel = 0;
for (int i = 0; i < bottom->size(); ++i) {
Blob<Dtype>* blob = (*bottom)[i];
Dtype* bottom_diff = blob->mutable_cpu_diff();
int num_elem = blob->channels()*blob->height()*blob->width();
for (int n = 0; n < num_; ++n) {
caffe_copy(num_elem, top_diff+top[0]->offset(n, offset_channel),
bottom_diff+blob->offset(n));
}
offset_channel += blob->channels();
}
}else if (concat_dim_ == 4){// lipengyu add
int top_bias = 0;
for(int n = 0 ; n < num_ ; n++)
{
for(int i = 0 ; i < bottom->size() ; i++)
{
Blob<Dtype>* blob = (*bottom)[i];
Dtype* bottom_diff = blob->mutable_cpu_diff();
int num_elem = blob->channels()*blob->height()*blob->width();
caffe_copy(num_elem, top_diff+ top_bias,//top[0]->offset(n, offset_channel),
bottom_diff+blob->offset(n));
top_bias += num_elem;
}
}
}else {
LOG(FATAL) << "concat_dim along dim" << concat_dim_ <<
" not implemented yet";
}
}
TYPED_TEST(DataTransformTest, TestMeanValue) {
TransformationParameter transform_param;
const bool unique_pixels = false; // pixels are equal to label
const int label = 0;
const int channels = 3;
const int height = 4;
const int width = 5;
const int mean_value = 2;
transform_param.add_mean_value(mean_value);
Datum datum;
FillDatum(label, channels, height, width, unique_pixels, &datum);
Blob<TypeParam>* blob = new Blob<TypeParam>(1, channels, height, width);
DataTransformer<TypeParam>* transformer =
new DataTransformer<TypeParam>(transform_param, TEST);
transformer->InitRand();
transformer->Transform(datum, blob);
for (int j = 0; j < blob->count(); ++j) {
EXPECT_EQ(blob->cpu_data()[j], label - mean_value);
}
}
TYPED_TEST(DataTransformTest, TestMeanFile) {
TransformationParameter transform_param;
const bool unique_pixels = true; // pixels are consecutive ints [0,size]
const int_tp label = 0;
const int_tp channels = 3;
const int_tp height = 4;
const int_tp width = 5;
const int_tp size = channels * height * width;
// Create a mean file
string* mean_file = new string();
MakeTempFilename(mean_file);
BlobProto blob_mean;
blob_mean.set_num(1);
blob_mean.set_channels(channels);
blob_mean.set_height(height);
blob_mean.set_width(width);
for (int_tp j = 0; j < size; ++j) {
blob_mean.add_data(j);
}
LOG(INFO) << "Using temporary mean_file " << *mean_file;
WriteProtoToBinaryFile(blob_mean, *mean_file);
transform_param.set_mean_file(*mean_file);
Datum datum;
FillDatum(label, channels, height, width, unique_pixels, &datum);
Blob<TypeParam>* blob = new Blob<TypeParam>(1, channels, height, width);
DataTransformer<TypeParam>* transformer =
new DataTransformer<TypeParam>(transform_param, TEST,
Caffe::GetDefaultDevice());
transformer->InitRand();
transformer->Transform(datum, blob);
for (int_tp j = 0; j < blob->count(); ++j) {
EXPECT_EQ(blob->cpu_data()[j], 0);
}
}
TYPED_TEST(DataTransformTest, TestEmptyTransform) {
TransformationParameter transform_param;
const bool unique_pixels = false; // all pixels the same equal to label
const int label = 0;
const int channels = 3;
const int height = 4;
const int width = 5;
Datum datum;
FillDatum(label, channels, height, width, unique_pixels, &datum);
Blob<TypeParam>* blob = new Blob<TypeParam>(1, channels, height, width);
DataTransformer<TypeParam>* transformer =
new DataTransformer<TypeParam>(transform_param, TEST);
transformer->InitRand();
transformer->Transform(datum, blob);
EXPECT_EQ(blob->num(), 1);
EXPECT_EQ(blob->channels(), datum.channels());
EXPECT_EQ(blob->height(), datum.height());
EXPECT_EQ(blob->width(), datum.width());
for (int j = 0; j < blob->count(); ++j) {
EXPECT_EQ(blob->cpu_data()[j], label);
}
}
TYPED_TEST(DataTransformTest, TestEmptyTransformUniquePixels) {
TransformationParameter transform_param;
const bool unique_pixels = true; // pixels are consecutive ints [0,size]
const int label = 0;
const int channels = 3;
const int height = 4;
const int width = 5;
Datum datum;
FillDatum(label, channels, height, width, unique_pixels, &datum);
Blob<TypeParam>* blob = new Blob<TypeParam>(1, 3, 4, 5);
DataTransformer<TypeParam>* transformer =
new DataTransformer<TypeParam>(transform_param, TEST);
transformer->InitRand();
transformer->Transform(datum, blob);
EXPECT_EQ(blob->num(), 1);
EXPECT_EQ(blob->channels(), datum.channels());
EXPECT_EQ(blob->height(), datum.height());
EXPECT_EQ(blob->width(), datum.width());
for (int j = 0; j < blob->count(); ++j) {
EXPECT_EQ(blob->cpu_data()[j], j);
}
}
void project(const Template &src, Template &dst) const
{
CaffeNet *net = caffeResource.acquire();
if (net->layers()[0]->layer_param().type() != "MemoryData")
qFatal("OpenBR requires the first layer in the network to be a MemoryDataLayer");
MemoryDataLayer<float> *dataLayer = static_cast<MemoryDataLayer<float> *>(net->layers()[0].get());
if (src.size() != dataLayer->batch_size())
qFatal("src should have %d (batch size) mats. It has %d mats.", dataLayer->batch_size(), src.size());
dataLayer->AddMatVector(src.toVector().toStdVector(), std::vector<int>(src.size(), 0));
net->ForwardPrefilled();
Blob<float> *output = net->blobs().back().get();
int dimFeatures = output->count() / dataLayer->batch_size();
for (int n = 0; n < dataLayer->batch_size(); n++)
dst += Mat(1, dimFeatures, CV_32FC1, output->mutable_cpu_data() + output->offset(n)).clone();
caffeResource.release(net);
}
void ScaleLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
if (bias_layer_ &&
this->param_propagate_down_[this->param_propagate_down_.size() - 1]) {
bias_layer_->Backward(top, bias_propagate_down_, bias_bottom_vec_);
}
const bool scale_param = (bottom.size() == 1);
Blob<Dtype>* scale = scale_param ? this->blobs_[0].get() : bottom[1];
if ((!scale_param && propagate_down[1]) ||
(scale_param && this->param_propagate_down_[0])) {
const Dtype* top_diff = top[0]->cpu_diff();
const bool in_place = (bottom[0] == top[0]);
const Dtype* bottom_data = (in_place ? &temp_ : bottom[0])->cpu_data();
// Hack: store big eltwise product in bottom[0] diff, except in the special
// case where this layer itself does the eltwise product, in which case we
// can store it directly in the scale diff, and we're done.
// If we're computing in-place (and not doing eltwise computation), this
// hack doesn't work and we store the product in temp_.
const bool is_eltwise = (bottom[0]->count() == scale->count());
Dtype* product = (is_eltwise ? scale->mutable_cpu_diff() :
(in_place ? temp_.mutable_cpu_data() : bottom[0]->mutable_cpu_diff()));
caffe_mul(top[0]->count(), top_diff, bottom_data, product);
if (!is_eltwise) {
Dtype* sum_result = NULL;
if (inner_dim_ == 1) {
sum_result = product;
} else if (sum_result_.count() == 1) {
const Dtype* sum_mult = sum_multiplier_.cpu_data();
Dtype* scale_diff = scale->mutable_cpu_diff();
if (scale_param) {
Dtype result = caffe_cpu_dot(inner_dim_, product, sum_mult);
*scale_diff += result;
} else {
*scale_diff = caffe_cpu_dot(inner_dim_, product, sum_mult);
}
} else {
const Dtype* sum_mult = sum_multiplier_.cpu_data();
sum_result = (outer_dim_ == 1) ?
scale->mutable_cpu_diff() : sum_result_.mutable_cpu_data();
caffe_cpu_gemv(CblasNoTrans, sum_result_.count(), inner_dim_,
Dtype(1), product, sum_mult, Dtype(0), sum_result);
}
if (outer_dim_ != 1) {
const Dtype* sum_mult = sum_multiplier_.cpu_data();
Dtype* scale_diff = scale->mutable_cpu_diff();
if (scale_dim_ == 1) {
if (scale_param) {
Dtype result = caffe_cpu_dot(outer_dim_, sum_mult, sum_result);
*scale_diff += result;
} else {
*scale_diff = caffe_cpu_dot(outer_dim_, sum_mult, sum_result);
}
} else {
caffe_cpu_gemv(CblasTrans, outer_dim_, scale_dim_,
Dtype(1), sum_result, sum_mult, Dtype(scale_param),
scale_diff);
}
}
}
}
if (propagate_down[0]) {
const Dtype* top_diff = top[0]->cpu_diff();
const Dtype* scale_data = scale->cpu_data();
Dtype* bottom_diff = bottom[0]->mutable_cpu_diff();
for (int n = 0; n < outer_dim_; ++n) {
for (int d = 0; d < scale_dim_; ++d) {
const Dtype factor = scale_data[d];
caffe_cpu_scale(inner_dim_, factor, top_diff, bottom_diff);
bottom_diff += inner_dim_;
top_diff += inner_dim_;
}
}
}
}
void Blob<Dtype>::ShareData(const Blob& other) {
CHECK_EQ(count_, other.count());
data_ = other.data();
}
void Blob<Dtype>::ShareDiff(const Blob& other) {
CHECK_EQ(count_, other.count());
diff_ = other.diff();
}
void EvalDetectionLayer<Dtype>::Forward_cpu(
const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top)
{
//const Dtype* input_data = bottom[0]->cpu_data();// 网络输出 N*13*13*125
const Dtype* label_data = bottom[1]->cpu_data(); // 真实标签数据 N*30*5
//LOG(INFO) << bottom[0]->data_at(0,0,0,0) << " " << bottom[0]->data_at(0,0,0,1);
Blob<Dtype> swap;// 网络输出数据 N * 13* 13* 125
// 变形为 N * (13 * 13) * 5 * 25
// N*(5*(5+num_class_))*13*13 -> N * (13*13) * 5 * (5+num_class_)
swap.Reshape(bottom[0]->num(),
bottom[0]->height()*bottom[0]->width(),
num_object_,
bottom[0]->channels()/num_object_);
Dtype* swap_data = swap.mutable_cpu_data();// cpu上的数据
caffe_set(swap.count(), Dtype(0.0), swap_data);// 设置为0
int index = 0;
for (int b = 0; b < bottom[0]->num(); ++b)// 图片数量
for (int h = 0; h < bottom[0]->height(); ++h) // 格子 13
for (int w = 0; w < bottom[0]->width(); ++w)// 格子 13
for (int c = 0; c < bottom[0]->channels(); ++c)// 5*25=125
{
swap_data[index++] = bottom[0]->data_at(b,c,h,w);
}
//*******************************************************************************//
//caffe_set(swap.count(), Dtype(0.0), swap_data);// 设置为0
//int p_index = (7*13+4)*125;
//swap_data[p_index]=-0.1020;
//swap_data[p_index+1]=2.0867;
//swap_data[p_index+2]=1.612;
//swap_data[p_index+3]=1.0515;
//swap_data[p_index+4]=1.0;
//swap_data[p_index+5+11]=100;
//*******************************************************************************//
Dtype* top_data = top[0]->mutable_cpu_data();// 层输出 cpu数据
caffe_set(top[0]->count(), Dtype(0), top_data);// 设置为0
Dtype all_mAP = 0.0;// 总 batch 的 mAP
for (int i = 0; i < bottom[0]->num(); ++i)
{ // N 图片数量
int input_index = i * bottom[0]->count(1);// 网络输出标签 i * 13*13*125
int true_index = i * bottom[1]->count(1);// 真实标签 i * 30*5
int top_index = i * top[0]->count(1); // 输出数据 // i * ( 20 + 13*13*5*4 + 1) -> i * (13*13*5*4 + 1)
// 前面20个为 真实标签 物体类别出现的次数
// 获取真实边框 =========================================
map<int, vector<BoxData> > gt_boxes;
// 从 对应图片的标签数据中 获取 真实边框 label_ + score_ + box_
// 返回一张图像的 标签 + 多边框数据的 标签
GetGTBox(side_, label_data + true_index, >_boxes);
// 在输出数据中 记录 真实标签 物体类别出现的次数=======================
for (std::map<int, vector<BoxData > >::iterator it = gt_boxes.begin(); it != gt_boxes.end(); ++it)
{
// 遍历 每一个 真实的标签
//int label = it->first;// 标签 类别
vector<BoxData>& g_boxes = it->second;// BoxData: label_ + score_ + box_
for (int j = 0; j < g_boxes.size(); ++j)
{// 边框数量
// 输出数据中的前 20个================================================
// =======================================
// top_data[top_index + label] += 1; // 真实标签 物体类别出现的次数
}
}
// 获取预测边框 =============================================
map<int, vector<BoxData> > pred_boxes;
// 获取预测边框 13*13*5 -> NMS抑制+置信度抑制 -> pred_boxes(数量少很多)
//GetPredBox(side_, num_object_, num_class_, input_data + input_index, &pred_boxes, sqrt_, constriant_, score_type_, nms_);
GetPredBox(side_, num_object_, num_class_, swap_data + input_index, &pred_boxes, score_type_, nms_, biases_);
// 输出数据 后面 的 13*13*5*4 =============================
// int index = top_index + num_class_ + 1;// 20 + 1 之后为 上面的 (label + score + tp + fp) 参数
//=============================
int index = top_index + 1;// 1个 map 之后为 上面的 (label + score + tp + fp) 参数
int pred_count(0);//
Dtype mAP = 0.0;
int pre_clas_num=0;
//float AP = 0.0;
// int tp
// 遍历预测值的 每一类，与最合适的标签边框计算AP,在计算总类别的mAP============
for (std::map<int, vector<BoxData> >::iterator it = pred_boxes.begin(); it != pred_boxes.end(); ++it)
{
Dtype AP = 0.0;// 该类AP
int tp=0;// 预测正确
int fp=0;// 预测错误
++pre_clas_num;// 该图片预测的总类别数量
int label = it->first;// 预测 边框标签
vector<BoxData>& p_boxes = it->second;// 多个边框 vector<BoxData>
// 真实标签中 未找到该 类别=======================================
if (gt_boxes.find(label) == gt_boxes.end()) {
for (int b = 0; b < p_boxes.size(); ++b) {// 该类别下的每一个预测边框
top_data[index + pred_count * 4 + 0] = p_boxes[b].label_;// 标签
top_data[index + pred_count * 4 + 1] = p_boxes[b].score_;// 得分
top_data[index + pred_count * 4 + 2] = 0; //tp
top_data[index + pred_count * 4 + 3] = 1; //fp 错误的预测为正确的
++pred_count;
++fp;// 预测错误============================================
}
if(tp + fp)
AP = tp / (tp + fp);// 计算该类别的 平均准确度 这里等于0
mAP += AP;// 所有类别的总 AP 这里可以省略 因为 AP为0
continue;// 跳过预测错误的，只记录fp
}
// 真实标签找到了该预测的类别======================================
vector<BoxData>& g_boxes = gt_boxes[label];// 真实标签中该 类别的 多个真实边框=====
vector<bool> records(g_boxes.size(), false);// 记录 真实类别的每一个 物体边框 是否已经被预测过
for (int k = 0; k < p_boxes.size(); ++k)
{ // 遍历 每个 预测边框==============
top_data[index + pred_count * 4 + 0] = p_boxes[k].label_;// 标签
top_data[index + pred_count * 4 + 1] = p_boxes[k].score_;// 得分
Dtype max_iou(-1);// 预测边框最接近的 真实边框 iou
int idx(-1);// 对应的 真实边框id
// 遍历每个真实边框 找到 预测边框最接近的 真实边框===========
for (int g = 0; g < g_boxes.size(); ++g) {
Dtype iou = Calc_iou(p_boxes[k].box_, g_boxes[g].box_);// 计算交并比
if (iou > max_iou) {
max_iou = iou;// 记录 每个预测边框 最接近的 真实边框 的 IOU
idx = g;// 对应的 真实边框id
}
}
// 根据 交并比 确定判断 预测 正确/错误
if (max_iou >= threshold_) {
if ( !records[idx] ) {
records[idx] = true;// 对应的 真实边框id
top_data[index + pred_count * 4 + 2] = 1; // tp 正->正确
top_data[index + pred_count * 4 + 3] = 0; // fp
++tp;// 预测正确 =================================
}
else
{// 同一个位置的物体之前已经被预测过，又一个框来预测，则认为是错误的
top_data[index + pred_count * 4 + 2] = 0;
top_data[index + pred_count * 4 + 3] = 1;// 错误 -> 正确
++fp;// 预测错误============================================
}
}
++pred_count;
}
if(tp + fp) AP = tp / (tp + fp);// 计算该类别的 平均准确度
mAP += AP;// 所有类别的总 AP
}
if(pre_clas_num){
mAP /= pre_clas_num;
// mAP /= num_class_; // 计算 mAP 是除以总类别数，还是总预测类别数
}
else mAP = 0.0;
// 输出对应图片的mAP
top_data[ index - 1 ] = mAP;
all_mAP += mAP;
}
if(bottom[0]->num()) all_mAP /= (bottom[0]->num());
top_data[0] = all_mAP;
}