void PowerLayer<Dtype>::Backward_gpu(
const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down,
const vector<Blob<Dtype>*>& bottom) {
if (propagate_down[0]) {
Dtype* bottom_diff = (bottom)[0]->mutable_gpu_diff();
const int count = (bottom)[0]->count();
const Dtype* top_diff = top[0]->gpu_diff();
if (diff_scale_ == Dtype(0) || power_ == Dtype(1)) {
caffe_gpu_set(count, diff_scale_, bottom_diff);
} else {
const Dtype* bottom_data = (bottom)[0]->gpu_data();
// Compute dy/dx = scale * power * (shift + scale * x)^(power - 1)
// = diff_scale * y / (shift + scale * x)
if (power_ == Dtype(2)) {
// Special case for y = (shift + scale * x)^2
// -> dy/dx = 2 * scale * (shift + scale * x)
// = diff_scale * shift + diff_scale * scale * x
caffe_gpu_axpby(
count,
diff_scale_ * scale_,
bottom_data,
Dtype(0),
bottom_diff);
if (shift_ != Dtype(0)) {
caffe_gpu_add_scalar(count, diff_scale_ * shift_, bottom_diff);
}
} else if (shift_ == Dtype(0)) {
// Special case for y = (scale * x)^power
// -> dy/dx = scale * power * (scale * x)^(power - 1)
// = scale * power * (scale * x)^power * (scale * x)^(-1)
// = power * y / x
const Dtype* top_data = top[0]->gpu_data();
caffe_gpu_div(count, top_data, bottom_data, bottom_diff);
caffe_gpu_scal(count, power_, bottom_diff);
} else {
caffe_copy(count, bottom_data, bottom_diff);
if (scale_ != Dtype(1)) {
caffe_gpu_scal(count, scale_, bottom_diff);
}
if (shift_ != Dtype(0)) {
caffe_gpu_add_scalar(count, shift_, bottom_diff);
}
const Dtype* top_data = top[0]->gpu_data();
caffe_gpu_div<Dtype>(count, top_data, bottom_diff, bottom_diff);
if (diff_scale_ != Dtype(1)) {
caffe_gpu_scal(count, diff_scale_, bottom_diff);
}
}
}
caffe_gpu_mul(count, top_diff, bottom_diff, bottom_diff);
}
}
void AbsValLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top, const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
const int count = top[0]->count();
const Dtype* top_data = top[0]->gpu_data();
const Dtype* top_diff = top[0]->gpu_diff();
if (propagate_down[0]) {
const Dtype* bottom_data = (bottom)[0]->gpu_data();
Dtype* bottom_diff = (bottom)[0]->mutable_gpu_diff();
caffe_gpu_div(count, top_data, bottom_data, bottom_diff);
caffe_gpu_mul(count, bottom_diff, top_diff, bottom_diff);
}
}
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 BatchNormLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down,
const vector<Blob<Dtype>*>& bottom) {
CHECK(!use_global_stats_);
const Dtype* top_diff;
if (bottom[0] != top[0]) {
top_diff = top[0]->gpu_diff();
}
else {
caffe_copy(x_norm_.count(), top[0]->gpu_diff(), x_norm_.mutable_gpu_diff());
top_diff = x_norm_.gpu_diff();
}
const Dtype* top_data = x_norm_.gpu_data();
Dtype* bottom_diff = bottom[0]->mutable_gpu_diff();
int num = bottom[0]->shape()[0];
int spatial_dim = bottom[0]->count() / (channels_*bottom[0]->shape(0));
// if Y = (X-mean(X))/(sqrt(var(X)+eps)), then
//
// dE(Y)/dX =
// (dE/dY - mean(dE/dY) - mean(dE/dY \cdot Y) \cdot Y)
// ./ sqrt(var(X) + eps)
//
// where \cdot and ./ are hadamard product and elementwise division,
// respectively, dE/dY is the top diff, and mean/var/sum are all computed
// along all dimensions except the channels dimension. In the above
// equation, the operations allow for expansion (i.e. broadcast) along all
// dimensions except the channels dimension where required.
// sum(dE/dY \cdot Y)
caffe_gpu_mul(temp_.count(), top_data, top_diff, bottom_diff);
caffe_gpu_gemv<Dtype>(CblasNoTrans, channels_ * num, spatial_dim, 1.,
bottom_diff, spatial_sum_multiplier_.gpu_data(), 0.,
num_by_chans_.mutable_gpu_data());
caffe_gpu_gemv<Dtype>(CblasTrans, num, channels_, 1.,
num_by_chans_.gpu_data(), batch_sum_multiplier_.gpu_data(), 0.,
mean_.mutable_gpu_data());
// reshape (broadcast) the above
caffe_gpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, num, channels_, 1, 1,
batch_sum_multiplier_.gpu_data(), mean_.gpu_data(), 0.,
num_by_chans_.mutable_gpu_data());
caffe_gpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, channels_ * num,
spatial_dim, 1, 1., num_by_chans_.gpu_data(),
spatial_sum_multiplier_.gpu_data(), 0., bottom_diff);
// sum(dE/dY \cdot Y) \cdot Y
caffe_gpu_mul(temp_.count(), top_data, bottom_diff, bottom_diff);
// sum(dE/dY)-sum(dE/dY \cdot Y) \cdot Y
caffe_gpu_gemv<Dtype>(CblasNoTrans, channels_ * num, spatial_dim, 1.,
top_diff, spatial_sum_multiplier_.gpu_data(), 0.,
num_by_chans_.mutable_gpu_data());
caffe_gpu_gemv<Dtype>(CblasTrans, num, channels_, 1.,
num_by_chans_.gpu_data(), batch_sum_multiplier_.gpu_data(), 0.,
mean_.mutable_gpu_data());
// reshape (broadcast) the above to make
// sum(dE/dY)-sum(dE/dY \cdot Y) \cdot Y
caffe_gpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, num, channels_, 1, 1,
batch_sum_multiplier_.gpu_data(), mean_.gpu_data(), 0.,
num_by_chans_.mutable_gpu_data());
caffe_gpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, num * channels_,
spatial_dim, 1, 1., num_by_chans_.gpu_data(),
spatial_sum_multiplier_.gpu_data(), 1., bottom_diff);
// dE/dY - mean(dE/dY)-mean(dE/dY \cdot Y) \cdot Y
caffe_gpu_axpby(temp_.count(), Dtype(1), top_diff,
Dtype(-1. / (num * spatial_dim)), bottom_diff);
// note: temp_ still contains sqrt(var(X)+eps), computed during the forward
// pass.
caffe_gpu_div(temp_.count(), bottom_diff, temp_.gpu_data(), bottom_diff);
}
void BatchNormLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
const Dtype* bottom_data = bottom[0]->gpu_data();
Dtype* top_data = top[0]->mutable_gpu_data();
int num = bottom[0]->shape(0);
int spatial_dim = bottom[0]->count() / (channels_*bottom[0]->shape(0));
if (bottom[0] != top[0]) {
caffe_copy(bottom[0]->count(), bottom_data, top_data);
}
if (use_global_stats_) {
// use the stored mean/variance estimates. TODO(cdoersch): allow an option
// to use an unbiased variance estimate, like the paper does.
caffe_copy(mean_.count(), this->blobs_[0]->gpu_data(),
mean_.mutable_gpu_data());
int m = bottom[0]->count() / channels_;
Dtype scale_factor = m > 1 ? Dtype(m) / (m - 1) : 1;
caffe_gpu_scale(variance_.count(), scale_factor,
this->blobs_[1]->gpu_data(), variance_.mutable_gpu_data());
}
else {
// compute mean
caffe_gpu_gemv<Dtype>(CblasNoTrans, channels_ * num, spatial_dim,
1. / (num * spatial_dim), bottom_data,
spatial_sum_multiplier_.gpu_data(), 0.,
num_by_chans_.mutable_gpu_data());
caffe_gpu_gemv<Dtype>(CblasTrans, num, channels_, 1.,
num_by_chans_.gpu_data(), batch_sum_multiplier_.gpu_data(), 0.,
mean_.mutable_gpu_data());
}
// subtract mean
caffe_gpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, num, channels_, 1, 1,
batch_sum_multiplier_.gpu_data(), mean_.gpu_data(), 0.,
num_by_chans_.mutable_gpu_data());
caffe_gpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, channels_ * num,
spatial_dim, 1, -1, num_by_chans_.gpu_data(),
spatial_sum_multiplier_.gpu_data(), 1., top_data);
if (!use_global_stats_) {
// compute variance using var(X) = E((X-EX)^2)
caffe_gpu_powx(top[0]->count(), top_data, Dtype(2),
temp_.mutable_gpu_data()); // (X-EX)^2
caffe_gpu_gemv<Dtype>(CblasNoTrans, channels_ * num, spatial_dim,
1. / (num * spatial_dim), temp_.gpu_data(),
spatial_sum_multiplier_.gpu_data(), 0.,
num_by_chans_.mutable_gpu_data());
caffe_gpu_gemv<Dtype>(CblasTrans, num, channels_, 1.,
num_by_chans_.gpu_data(), batch_sum_multiplier_.gpu_data(), 0.,
variance_.mutable_gpu_data()); // E((X_EX)^2)
// compute and save moving average
Dtype scale_factor = this->blobs_[2]->cpu_data()[0] == 0 ?
1 : 1 - moving_average_fraction_;
caffe_gpu_axpby(mean_.count(), scale_factor,
mean_.gpu_data(), moving_average_fraction_,
this->blobs_[0]->mutable_gpu_data());
caffe_gpu_axpby(variance_.count(), scale_factor,
variance_.gpu_data(), moving_average_fraction_,
this->blobs_[1]->mutable_gpu_data());
this->blobs_[2]->mutable_cpu_data()[0] += 1;
}
// normalize variance
caffe_gpu_add_scalar(variance_.count(), eps_, variance_.mutable_gpu_data());
caffe_gpu_powx(variance_.count(), variance_.gpu_data(), Dtype(0.5),
variance_.mutable_gpu_data());
// replicate variance to input size
caffe_gpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, num, channels_, 1, 1,
batch_sum_multiplier_.gpu_data(), variance_.gpu_data(), 0.,
num_by_chans_.mutable_gpu_data());
caffe_gpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, channels_ * num,
spatial_dim, 1, 1., num_by_chans_.gpu_data(),
spatial_sum_multiplier_.gpu_data(), 0., temp_.mutable_gpu_data());
caffe_gpu_div(temp_.count(), top_data, temp_.gpu_data(), top_data);
// TODO(cdoersch): The caching is only needed because later in-place layers
// might clobber the data. Can we skip this if they won't?
caffe_copy(x_norm_.count(), top_data,
x_norm_.mutable_gpu_data());
}
void AdaGradSolver<Dtype>::ComputeUpdateValue(uint_tp param_id, Dtype rate) {
CHECK(Caffe::root_solver());
const vector<Blob<Dtype>*>& net_params = this->net_->learnable_params();
const vector<float>& net_params_lr = this->net_->params_lr();
Dtype delta = this->param_.delta();
Dtype local_rate = rate * net_params_lr[param_id];
switch (Caffe::mode()) {
case Caffe::CPU: {
// 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
if (this->device_->backend() == BACKEND_CUDA) {
#ifdef USE_CUDA
// 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());
#endif // USE_CUDA
} else {
#ifdef USE_GREENTEA
// compute square of gradient in update
greentea_gpu_powx<Dtype>(
this->device_->id(), net_params[param_id]->count(),
(cl_mem) (net_params[param_id]->gpu_diff()), 0, Dtype(2),
(cl_mem) (this->update_[param_id]->mutable_gpu_data()), 0);
// update history
greentea_gpu_add<Dtype>(
this->device_->id(), net_params[param_id]->count(),
(cl_mem) (this->update_[param_id]->gpu_data()), 0,
(cl_mem) (this->history_[param_id]->gpu_data()), 0,
(cl_mem) (this->history_[param_id]->mutable_gpu_data()), 0);
// prepare update
greentea_gpu_powx<Dtype>(
this->device_->id(), net_params[param_id]->count(),
(cl_mem) (this->history_[param_id]->gpu_data()), 0, Dtype(0.5),
(cl_mem) (this->update_[param_id]->mutable_gpu_data()), 0);
greentea_gpu_add_scalar<Dtype>(
this->device_->id(), net_params[param_id]->count(), delta,
(cl_mem) (this->update_[param_id]->mutable_gpu_data()), 0);
greentea_gpu_div<Dtype>(
this->device_->id(), net_params[param_id]->count(),
(cl_mem) (net_params[param_id]->gpu_diff()), 0,
(cl_mem) (this->update_[param_id]->gpu_data()), 0,
(cl_mem) (this->update_[param_id]->mutable_gpu_data()), 0);
// scale and copy
greentea_gpu_axpby<Dtype>(
this->device_->id(), net_params[param_id]->count(),
local_rate, (cl_mem) (this->update_[param_id]->gpu_data()), 0,
Dtype(0), (cl_mem) (net_params[param_id]->mutable_gpu_diff()), 0);
#endif // USE_GREENTEA
}
#else
NO_GPU;
#endif
break;
}
default:
LOG(FATAL)<< "Unknown caffe mode: " << Caffe::mode();
}
}