void PaddingLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
if (!propagate_down[0]) { return; }
caffe_gpu_set(bottom[0]->count(), Dtype(0), bottom[0]->mutable_gpu_diff());
if (pad_pos_) {
for (int n = 0; n < num_; ++n) {
for (int c = 0; c < channels_; ++c) {
for (int h = 0; h < height_in_; ++h) {
// copy the width part
caffe_gpu_axpy(width_in_, (Dtype)1.,
top[0]->gpu_diff(n, c, h + pad_beg_, pad_beg_),
bottom[0]->mutable_gpu_diff(n, c, h));
}
}
}
}
else {
for (int n = 0; n < num_; ++n) {
for (int c = 0; c < channels_; ++c) {
for (int h = 0; h < height_out_; ++h) {
// copy the width part
caffe_gpu_axpy(width_out_, (Dtype)1.,
top[0]->gpu_diff(n, c, h),
bottom[0]->mutable_gpu_diff(n, c, h - pad_beg_, - pad_beg_));
}
}
}
}
}
void SGDSolver<Dtype>::Regularize(int param_id) {
const vector<Blob<Dtype>*>& net_params = this->net_->learnable_params();
const vector<float>& net_params_weight_decay =
this->net_->params_weight_decay();
Dtype weight_decay = this->param_.weight_decay();
string regularization_type = this->param_.regularization_type();
Dtype local_decay = weight_decay * net_params_weight_decay[param_id];
switch (Caffe::mode()) {
case Caffe::CPU: {
if (local_decay) {
if (regularization_type == "L2") {
// add weight decay
caffe_axpy(net_params[param_id]->count(),
local_decay,
net_params[param_id]->cpu_data(),
net_params[param_id]->mutable_cpu_diff());
} else if (regularization_type == "L1") {
caffe_cpu_sign(net_params[param_id]->count(),
net_params[param_id]->cpu_data(),
temp_[param_id]->mutable_cpu_data());
caffe_axpy(net_params[param_id]->count(),
local_decay,
temp_[param_id]->cpu_data(),
net_params[param_id]->mutable_cpu_diff());
} else {
LOG(FATAL) << "Unknown regularization type: " << regularization_type;
}
}
break;
}
case Caffe::GPU: {
#ifndef CPU_ONLY
if (local_decay) {
if (regularization_type == "L2") {
// add weight decay
caffe_gpu_axpy(net_params[param_id]->count(),
local_decay,
net_params[param_id]->gpu_data(),
net_params[param_id]->mutable_gpu_diff());
} else if (regularization_type == "L1") {
caffe_gpu_sign(net_params[param_id]->count(),
net_params[param_id]->gpu_data(),
temp_[param_id]->mutable_gpu_data());
caffe_gpu_axpy(net_params[param_id]->count(),
local_decay,
temp_[param_id]->gpu_data(),
net_params[param_id]->mutable_gpu_diff());
} else {
LOG(FATAL) << "Unknown regularization type: " << regularization_type;
}
}
#else
NO_GPU;
#endif
break;
}
default:
LOG(FATAL) << "Unknown caffe mode: " << Caffe::mode();
}
}
void SGDFeedbackSolver<Dtype>::ComputeUpdateValue() {
vector<shared_ptr<Blob<Dtype> > >& net_params = this->net_->params();
vector<float>& net_params_lr = this->net_->params_lr();
vector<float>& net_params_weight_decay = this->net_->params_weight_decay();
// get the learning rate
Dtype rate = GetLearningRate();
if (this->param_.display() && this->iter_ % this->param_.display() == 0) {
LOG(INFO) << "Iteration " << this->iter_ << ", lr = " << rate;
}
Dtype momentum = this->param_.momentum();
Dtype weight_decay = this->param_.weight_decay();
switch (Caffe::mode()) {
case Caffe::CPU:
for (int param_id = 0; param_id < net_params.size(); ++param_id) {
// Compute the value to history, and then copy them to the blob's diff.
Dtype local_rate = rate * net_params_lr[param_id];
Dtype local_decay = weight_decay * net_params_weight_decay[param_id];
caffe_cpu_axpby(net_params[param_id]->count(), local_rate,
net_params[param_id]->cpu_diff(), momentum,
history_[param_id]->mutable_cpu_data());
if (local_decay) {
// add weight decay
caffe_axpy(net_params[param_id]->count(),
local_decay * local_rate,
net_params[param_id]->cpu_data(),
history_[param_id]->mutable_cpu_data());
}
// copy
caffe_copy(net_params[param_id]->count(),
history_[param_id]->cpu_data(),
net_params[param_id]->mutable_cpu_diff());
}
break;
case Caffe::GPU:
for (int param_id = 0; param_id < net_params.size(); ++param_id) {
// Compute the value to history, and then copy them to the blob's diff.
Dtype local_rate = rate * net_params_lr[param_id];
Dtype local_decay = weight_decay * net_params_weight_decay[param_id];
caffe_gpu_axpby(net_params[param_id]->count(), local_rate,
net_params[param_id]->gpu_diff(), momentum,
history_[param_id]->mutable_gpu_data());
if (local_decay) {
// add weight decay
caffe_gpu_axpy(net_params[param_id]->count(),
local_decay * local_rate,
net_params[param_id]->gpu_data(),
history_[param_id]->mutable_gpu_data());
}
// copy
caffe_gpu_copy(net_params[param_id]->count(),
history_[param_id]->gpu_data(),
net_params[param_id]->mutable_gpu_diff());
}
break;
default:
LOG(FATAL) << "Unknown caffe mode: " << Caffe::mode();
}
}
void SGDSolver<Dtype>::ComputeUpdateValue() {
vector<shared_ptr<Blob<Dtype> > >& net_params = this->net_->params();
// get the learning rate
Dtype rate = GetLearningRate();
Dtype momentum = this->param_.momentum();
Dtype weight_decay = this->param_.weight_decay();
// LOG(ERROR) << "rate:" << rate << " momentum:" << momentum
// << " weight_decay:" << weight_decay;
switch (Caffe::mode()) {
case Caffe::CPU:
for (size_t param_id = 0; param_id < net_params.size(); ++param_id) {
// Compute the value to history, and then copy them to the blob's diff.
caffe_axpby(net_params[param_id]->count(), rate,
net_params[param_id]->cpu_diff(), momentum,
history_[param_id]->mutable_cpu_data());
if (weight_decay) {
// add weight decay
caffe_axpy(net_params[param_id]->count(), weight_decay * rate,
net_params[param_id]->cpu_data(),
history_[param_id]->mutable_cpu_data());
}
// copy
caffe_copy(net_params[param_id]->count(),
history_[param_id]->cpu_data(),
net_params[param_id]->mutable_cpu_diff());
}
break;
case Caffe::GPU:
for (size_t param_id = 0; param_id < net_params.size(); ++param_id) {
// Compute the value to history, and then copy them to the blob's diff.
caffe_gpu_axpby(net_params[param_id]->count(), rate,
net_params[param_id]->gpu_diff(), momentum,
history_[param_id]->mutable_gpu_data());
if (weight_decay) {
// add weight decay
caffe_gpu_axpy(net_params[param_id]->count(), weight_decay * rate,
net_params[param_id]->gpu_data(),
history_[param_id]->mutable_gpu_data());
}
// copy
caffe_gpu_copy(net_params[param_id]->count(),
history_[param_id]->gpu_data(),
net_params[param_id]->mutable_gpu_diff());
}
break;
default:
LOG(FATAL) << "Unknown caffe mode: " << Caffe::mode();
}
}
void SplitLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
if (!propagate_down[0]) { return; }
if (top.size() == 1) {
caffe_copy(count_, top[0]->gpu_diff(), bottom[0]->mutable_gpu_diff());
return;
}
caffe_gpu_add(count_, top[0]->gpu_diff(), top[1]->gpu_diff(),
bottom[0]->mutable_gpu_diff());
// Add remaining top blob diffs.
for (int i = 2; i < top.size(); ++i) {
const Dtype* top_diff = top[i]->gpu_diff();
Dtype* bottom_diff = bottom[0]->mutable_gpu_diff();
caffe_gpu_axpy(count_, Dtype(1.), top_diff, bottom_diff);
}
}
void SigmoidCrossEntropyLossLayer<Dtype>::Backward_gpu(
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]) {
// First, compute the diff
const int count = bottom[0]->count();
const int num = bottom[0]->num();
const Dtype* sigmoid_output_data = sigmoid_output_->gpu_data();
const Dtype* target = bottom[1]->gpu_data();
Dtype* bottom_diff = bottom[0]->mutable_gpu_diff();
caffe_copy(count, sigmoid_output_data, bottom_diff);
caffe_gpu_axpy(count, Dtype(-1), target, bottom_diff);
// Scale down gradient
const Dtype loss_weight = top[0]->cpu_diff()[0];
caffe_gpu_scal(count, loss_weight / num, bottom_diff);
}
}
void Tensor<Dtype>::AddMulFrom(const Tensor& source, Dtype alpha) {
if (source.count() != count_ || source.shape() != shape_) {
ASSERT(false, "Trying to add blobs of different sizes: "
<< source.count() << " != " << count_);
}
switch (mode()) {
case Caffe::CPU:
caffe_axpy(count_, alpha,
source.cpu_mem(),
this->mutable_cpu_mem());
break;
case Caffe::GPU:
#ifndef CPU_ONLY
caffe_gpu_axpy(count_, alpha,
source.gpu_mem(),
this->mutable_gpu_mem());
#else
NO_GPU;
#endif
break;
default:
ASSERT(false, "Unknown caffe mode.");
}
}
void AdaGradSolver<Dtype>::ComputeUpdateValue() {
vector<shared_ptr<Blob<Dtype> > >& net_params = this->net_->params();
vector<float>& net_params_lr = this->net_->params_lr();
vector<float>& net_params_weight_decay = this->net_->params_weight_decay();
// get the learning rate
Dtype rate = this->GetLearningRate();
Dtype delta = this->param_.delta();
if (this->param_.display() && this->iter_ % this->param_.display() == 0) {
LOG(INFO) << "Iteration " << this->iter_ << ", lr = " << rate;
}
Dtype weight_decay = this->param_.weight_decay();
string regularization_type = this->param_.regularization_type();
switch (Caffe::mode()) {
case Caffe::CPU:
for (int param_id = 0; param_id < net_params.size(); ++param_id) {
Dtype local_rate = rate * net_params_lr[param_id];
Dtype local_decay = weight_decay * net_params_weight_decay[param_id];
if (local_decay) {
if (regularization_type == "L2") {
// add weight decay
caffe_axpy(net_params[param_id]->count(),
local_decay,
net_params[param_id]->cpu_data(),
net_params[param_id]->mutable_cpu_diff());
} else if (regularization_type == "L1") {
caffe_cpu_sign(net_params[param_id]->count(),
net_params[param_id]->cpu_data(),
this->temp_[param_id]->mutable_cpu_data());
caffe_axpy(net_params[param_id]->count(),
local_decay,
this->temp_[param_id]->cpu_data(),
net_params[param_id]->mutable_cpu_diff());
} else {
LOG(FATAL) << "Unknown regularization type: " << regularization_type;
}
}
// compute square of gradient in update
caffe_powx(net_params[param_id]->count(),
net_params[param_id]->cpu_diff(), Dtype(2),
this->update_[param_id]->mutable_cpu_data());
// update history
caffe_add(net_params[param_id]->count(),
this->update_[param_id]->cpu_data(),
this->history_[param_id]->cpu_data(),
this->history_[param_id]->mutable_cpu_data());
// prepare update
caffe_powx(net_params[param_id]->count(),
this->history_[param_id]->cpu_data(), Dtype(0.5),
this->update_[param_id]->mutable_cpu_data());
caffe_add_scalar(net_params[param_id]->count(),
delta, this->update_[param_id]->mutable_cpu_data());
caffe_div(net_params[param_id]->count(),
net_params[param_id]->cpu_diff(),
this->update_[param_id]->cpu_data(),
this->update_[param_id]->mutable_cpu_data());
// scale and copy
caffe_cpu_axpby(net_params[param_id]->count(), local_rate,
this->update_[param_id]->cpu_data(), Dtype(0),
net_params[param_id]->mutable_cpu_diff());
}
break;
case Caffe::GPU:
#ifndef CPU_ONLY
for (int param_id = 0; param_id < net_params.size(); ++param_id) {
Dtype local_rate = rate * net_params_lr[param_id];
Dtype local_decay = weight_decay * net_params_weight_decay[param_id];
if (local_decay) {
if (regularization_type == "L2") {
// add weight decay
caffe_gpu_axpy(net_params[param_id]->count(),
local_decay,
net_params[param_id]->gpu_data(),
net_params[param_id]->mutable_gpu_diff());
} else if (regularization_type == "L1") {
caffe_gpu_sign(net_params[param_id]->count(),
net_params[param_id]->gpu_data(),
this->temp_[param_id]->mutable_gpu_data());
caffe_gpu_axpy(net_params[param_id]->count(),
local_decay,
this->temp_[param_id]->gpu_data(),
net_params[param_id]->mutable_gpu_diff());
} else {
LOG(FATAL) << "Unknown regularization type: " << regularization_type;
}
}
// compute square of gradient in update
caffe_gpu_powx(net_params[param_id]->count(),
net_params[param_id]->gpu_diff(), Dtype(2),
this->update_[param_id]->mutable_gpu_data());
// update history
caffe_gpu_add(net_params[param_id]->count(),
this->update_[param_id]->gpu_data(),
this->history_[param_id]->gpu_data(),
this->history_[param_id]->mutable_gpu_data());
// prepare update
caffe_gpu_powx(net_params[param_id]->count(),
this->history_[param_id]->gpu_data(), Dtype(0.5),
this->update_[param_id]->mutable_gpu_data());
caffe_gpu_add_scalar(net_params[param_id]->count(),
delta, this->update_[param_id]->mutable_gpu_data());
caffe_gpu_div(net_params[param_id]->count(),
net_params[param_id]->gpu_diff(),
this->update_[param_id]->gpu_data(),
this->update_[param_id]->mutable_gpu_data());
// scale and copy
caffe_gpu_axpby(net_params[param_id]->count(), local_rate,
this->update_[param_id]->gpu_data(), Dtype(0),
net_params[param_id]->mutable_gpu_diff());
}
#else
NO_GPU;
#endif
break;
default:
LOG(FATAL) << "Unknown caffe mode: " << Caffe::mode();
}
}
void NesterovSolver<Dtype>::ComputeUpdateValue() {
vector<shared_ptr<Blob<Dtype> > >& net_params = this->net_->params();
vector<float>& net_params_lr = this->net_->params_lr();
vector<float>& net_params_weight_decay = this->net_->params_weight_decay();
// get the learning rate
Dtype rate = this->GetLearningRate();
if (this->param_.display() && this->iter_ % this->param_.display() == 0) {
LOG(INFO) << "Iteration " << this->iter_ << ", lr = " << rate;
}
Dtype momentum = this->param_.momentum();
Dtype weight_decay = this->param_.weight_decay();
string regularization_type = this->param_.regularization_type();
switch (Caffe::mode()) {
case Caffe::CPU:
for (int param_id = 0; param_id < net_params.size(); ++param_id) {
// save history momentum for stepping back
caffe_copy(net_params[param_id]->count(),
this->history_[param_id]->cpu_data(),
this->update_[param_id]->mutable_cpu_data());
Dtype local_rate = rate * net_params_lr[param_id];
Dtype local_decay = weight_decay * net_params_weight_decay[param_id];
if (local_decay) {
if (regularization_type == "L2") {
// add weight decay
caffe_axpy(net_params[param_id]->count(),
local_decay,
net_params[param_id]->cpu_data(),
net_params[param_id]->mutable_cpu_diff());
} else if (regularization_type == "L1") {
caffe_cpu_sign(net_params[param_id]->count(),
net_params[param_id]->cpu_data(),
this->temp_[param_id]->mutable_cpu_data());
caffe_axpy(net_params[param_id]->count(),
local_decay,
this->temp_[param_id]->cpu_data(),
net_params[param_id]->mutable_cpu_diff());
} else {
LOG(FATAL) << "Unknown regularization type: " << regularization_type;
}
}
// update history
caffe_cpu_axpby(net_params[param_id]->count(), local_rate,
net_params[param_id]->cpu_diff(), momentum,
this->history_[param_id]->mutable_cpu_data());
// compute udpate: step back then over step
caffe_cpu_axpby(net_params[param_id]->count(), Dtype(1) + momentum,
this->history_[param_id]->cpu_data(), -momentum,
this->update_[param_id]->mutable_cpu_data());
// copy
caffe_copy(net_params[param_id]->count(),
this->update_[param_id]->cpu_data(),
net_params[param_id]->mutable_cpu_diff());
}
break;
case Caffe::GPU:
#ifndef CPU_ONLY
for (int param_id = 0; param_id < net_params.size(); ++param_id) {
// save history momentum for stepping back
caffe_copy(net_params[param_id]->count(),
this->history_[param_id]->gpu_data(),
this->update_[param_id]->mutable_gpu_data());
Dtype local_rate = rate * net_params_lr[param_id];
Dtype local_decay = weight_decay * net_params_weight_decay[param_id];
if (local_decay) {
if (regularization_type == "L2") {
// add weight decay
caffe_gpu_axpy(net_params[param_id]->count(),
local_decay,
net_params[param_id]->gpu_data(),
net_params[param_id]->mutable_gpu_diff());
} else if (regularization_type == "L1") {
caffe_gpu_sign(net_params[param_id]->count(),
net_params[param_id]->gpu_data(),
this->temp_[param_id]->mutable_gpu_data());
caffe_gpu_axpy(net_params[param_id]->count(),
local_decay,
this->temp_[param_id]->gpu_data(),
net_params[param_id]->mutable_gpu_diff());
} else {
LOG(FATAL) << "Unknown regularization type: " << regularization_type;
}
}
// update history
caffe_gpu_axpby(net_params[param_id]->count(), local_rate,
net_params[param_id]->gpu_diff(), momentum,
this->history_[param_id]->mutable_gpu_data());
// compute udpate: step back then over step
caffe_gpu_axpby(net_params[param_id]->count(), Dtype(1) + momentum,
this->history_[param_id]->gpu_data(), -momentum,
this->update_[param_id]->mutable_gpu_data());
// copy
caffe_copy(net_params[param_id]->count(),
this->update_[param_id]->gpu_data(),
net_params[param_id]->mutable_gpu_diff());
}
#else
NO_GPU;
#endif
break;
default:
LOG(FATAL) << "Unknown caffe mode: " << Caffe::mode();
}
}
void SGDSolver<Dtype>::ComputeUpdateValue() {
vector<shared_ptr<Blob<Dtype> > >& net_params = this->net_->params();
vector<float>& net_params_lr = this->net_->params_lr();
vector<float>& net_params_weight_decay = this->net_->params_weight_decay();
// get the learning rate
Dtype rate = GetLearningRate();
Dtype momentum = this->param_.momentum();
if (this->param_.momentum_burnin() > this->iter_) {
momentum = momentum * this->iter_ / this->param_.momentum_burnin();
}
if (this->param_.display() && this->iter_ % this->param_.display() == 0) {
LOG(INFO) << "Iteration " << this->iter_ << ", lr = " << rate << ", mom = " << momentum;
}
Dtype weight_decay = this->param_.weight_decay();
string regularization_type = this->param_.regularization_type();
switch (Caffe::mode()) {
case Caffe::CPU:
for (int param_id = 0; param_id < net_params.size(); ++param_id) {
// Compute the value to history, and then copy them to the blob's diff.
Dtype local_rate = rate * net_params_lr[param_id];
Dtype local_decay = weight_decay * net_params_weight_decay[param_id];
if (local_decay) {
if (regularization_type == "L2") {
// add weight decay
caffe_axpy(net_params[param_id]->count(),
local_decay,
net_params[param_id]->cpu_data(),
net_params[param_id]->mutable_cpu_diff());
} else if (regularization_type == "L1") {
caffe_cpu_sign(net_params[param_id]->count(),
net_params[param_id]->cpu_data(),
temp_[param_id]->mutable_cpu_data());
caffe_axpy(net_params[param_id]->count(),
local_decay,
temp_[param_id]->cpu_data(),
net_params[param_id]->mutable_cpu_diff());
} else {
LOG(FATAL) << "Unknown regularization type: " << regularization_type;
}
}
caffe_cpu_axpby(net_params[param_id]->count(), local_rate,
net_params[param_id]->cpu_diff(), momentum,
history_[param_id]->mutable_cpu_data());
// copy
caffe_copy(net_params[param_id]->count(),
history_[param_id]->cpu_data(),
net_params[param_id]->mutable_cpu_diff());
}
break;
case Caffe::GPU:
#ifndef CPU_ONLY
for (int param_id = 0; param_id < net_params.size(); ++param_id) {
// Compute the value to history, and then copy them to the blob's diff.
Dtype local_rate = rate * net_params_lr[param_id];
Dtype local_decay = weight_decay * net_params_weight_decay[param_id];
if (local_decay) {
if (regularization_type == "L2") {
// add weight decay
caffe_gpu_axpy(net_params[param_id]->count(),
local_decay,
net_params[param_id]->gpu_data(),
net_params[param_id]->mutable_gpu_diff());
} else if (regularization_type == "L1") {
caffe_gpu_sign(net_params[param_id]->count(),
net_params[param_id]->gpu_data(),
temp_[param_id]->mutable_gpu_data());
caffe_gpu_axpy(net_params[param_id]->count(),
local_decay,
temp_[param_id]->gpu_data(),
net_params[param_id]->mutable_gpu_diff());
} else {
LOG(FATAL) << "Unknown regularization type: " << regularization_type;
}
}
caffe_gpu_axpby(net_params[param_id]->count(), local_rate,
net_params[param_id]->gpu_diff(), momentum,
history_[param_id]->mutable_gpu_data());
// copy
caffe_copy(net_params[param_id]->count(),
history_[param_id]->gpu_data(),
net_params[param_id]->mutable_gpu_diff());
}
#else
NO_GPU;
#endif
break;
default:
LOG(FATAL) << "Unknown caffe mode: " << Caffe::mode();
}
}
void SGDSolver<Dtype>::ComputeUpdateValue() {
vector<shared_ptr<Blob<Dtype> > >& net_params = this->net_->params();
vector<float>& net_params_lr = this->net_->params_lr();
vector<string>& net_params_lr_policy = this->net_->params_lr_policy();
vector<float>& net_params_weight_decay = this->net_->params_weight_decay();
// get the learning rate
Dtype rate = GetLearningRate();
if (this->param_.display() && this->iter_ % this->param_.display() == 0) {
LOG(INFO) << "Iteration " << this->iter_ << ", lr = " << rate;
}
Dtype momentum = this->param_.momentum();
Dtype weight_decay = this->param_.weight_decay();
switch (Caffe::mode()) {
case Caffe::CPU:
for (int param_id = 0; param_id < net_params.size(); ++param_id) {
// Compute the value to history, and then copy them to the blob's diff.
Dtype local_rate = rate * net_params_lr[param_id];
Dtype local_decay = weight_decay * net_params_weight_decay[param_id];
caffe_axpby(net_params[param_id]->count(), local_rate,
net_params[param_id]->cpu_diff(), momentum,
history_[param_id]->mutable_cpu_data());
if (local_decay) {
// add weight decay
caffe_axpy(net_params[param_id]->count(),
local_decay * local_rate,
net_params[param_id]->cpu_data(),
history_[param_id]->mutable_cpu_data());
}
// copy
caffe_copy(net_params[param_id]->count(),
history_[param_id]->cpu_data(),
net_params[param_id]->mutable_cpu_diff());
}
break;
case Caffe::GPU:
//LOG(INFO) << "Installing local lr policy";
for (int param_id = 0; param_id < net_params.size(); ++param_id) {
// Compute the value to history, and then copy them to the blob's diff.
Dtype local_rate;
if(net_params_lr_policy[param_id] == "naive_inv") {
local_rate = rate * net_params_lr[param_id] * Dtype(1.0)/(this->iter_/500 + 1);
//LOG(INFO) << "rate: " << rate << " local rate: " << net_params_lr[param_id] << " inv coeff: " << Dtype(1.0)/(this->iter_/500 + 1) << " hehe: " << (this->iter_/500 + 1);
}
else if (net_params_lr_policy[param_id] == "power_inv") {
local_rate = rate * net_params_lr[param_id] * pow(Dtype(1.0) + this->param_.localgamma() * this->iter_, - this->param_.localpower());
//LOG(INFO) << "local rate: " << local_rate;
}
else if (net_params_lr_policy[param_id] == "step") {
int current_step = this->iter_ / this->param_.localstepsize();
local_rate = rate * net_params_lr[param_id] *
pow(this->param_.localgamma(), current_step);
}
else if (net_params_lr_policy[param_id] == "nothing")
local_rate = rate * net_params_lr[param_id];
else LOG(FATAL) << "Unknown caffe local policy: " << net_params_lr_policy[param_id];
Dtype local_decay = weight_decay * net_params_weight_decay[param_id];
caffe_gpu_axpby(net_params[param_id]->count(), local_rate,
net_params[param_id]->gpu_diff(), momentum,
history_[param_id]->mutable_gpu_data());
if (local_decay) {
// add weight decay
caffe_gpu_axpy(net_params[param_id]->count(),
local_decay * local_rate,
net_params[param_id]->gpu_data(),
history_[param_id]->mutable_gpu_data());
}
// copy
caffe_gpu_copy(net_params[param_id]->count(),
history_[param_id]->gpu_data(),
net_params[param_id]->mutable_gpu_diff());
}
break;
default:
LOG(FATAL) << "Unknown caffe mode: " << Caffe::mode();
}
}
void SGDSolver<Dtype>::Regularize(int param_id) {
const vector<Blob<Dtype>*>& net_params = this->net_->learnable_params();
const vector<float>& net_params_weight_decay =
this->net_->params_weight_decay();
Dtype weight_decay = this->param_.weight_decay();
string regularization_type = this->param_.regularization_type();
Dtype local_decay = weight_decay * net_params_weight_decay[param_id];
switch (Caffe::mode()) {
case Caffe::CPU: {
if (local_decay) {
if (regularization_type == "L2") {
// add weight decay
caffe_axpy(net_params[param_id]->count(), local_decay,
net_params[param_id]->cpu_data(),
net_params[param_id]->mutable_cpu_diff());
} else if (regularization_type == "L1") {
caffe_cpu_sign(net_params[param_id]->count(),
net_params[param_id]->cpu_data(),
temp_[param_id]->mutable_cpu_data());
caffe_axpy(net_params[param_id]->count(), local_decay,
temp_[param_id]->cpu_data(),
net_params[param_id]->mutable_cpu_diff());
} else {
LOG(FATAL)<< "Unknown regularization type: " << regularization_type;
}
}
break;
}
case Caffe::GPU: {
#ifndef CPU_ONLY
if (this->device_->backend() == BACKEND_CUDA) {
#ifdef USE_CUDA
if (local_decay) {
if (regularization_type == "L2") {
// add weight decay
caffe_gpu_axpy(net_params[param_id]->count(),
local_decay,
net_params[param_id]->gpu_data(),
net_params[param_id]->mutable_gpu_diff());
} else if (regularization_type == "L1") {
caffe_gpu_sign(net_params[param_id]->count(),
net_params[param_id]->gpu_data(),
temp_[param_id]->mutable_gpu_data());
caffe_gpu_axpy(net_params[param_id]->count(),
local_decay,
temp_[param_id]->gpu_data(),
net_params[param_id]->mutable_gpu_diff());
} else {
LOG(FATAL)<< "Unknown regularization type: "
<< regularization_type;
}
}
#endif // USE_CUDA
} else {
#ifdef USE_GREENTEA
if (local_decay) {
if (regularization_type == "L2") {
// add weight decay
greentea_gpu_axpy<Dtype>(this->device_->id(),
net_params[param_id]->count(),
local_decay,
(cl_mem)(net_params[param_id]->gpu_data()), 0,
(cl_mem)(net_params[param_id]->mutable_gpu_diff()), 0);
} else if (regularization_type == "L1") {
greentea_gpu_sign<Dtype>(this->device_->id(),
net_params[param_id]->count(),
(cl_mem)(net_params[param_id]->gpu_data()), 0,
(cl_mem)(temp_[param_id]->mutable_gpu_data()), 0);
greentea_gpu_axpy<Dtype>(this->device_->id(),
net_params[param_id]->count(),
local_decay,
(cl_mem)(temp_[param_id]->gpu_data()), 0,
(cl_mem)(net_params[param_id]->mutable_gpu_diff()), 0);
} else {
LOG(FATAL)<< "Unknown regularization type: "
<< regularization_type;
}
}
#endif // USE_GREENTEA
}
#else
NO_GPU;
#endif
break;
}
default: {
LOG(FATAL) << "Unknown caffe mode: " << Caffe::mode();
}
}
}
