Thanks to visit codestin.com
Credit goes to github.com

Skip to content

Commit 8e2a509

Browse files
ShichenLiuShichenLiu
committed
clean jmmd src
1 parent b8e2d06 commit 8e2a509

File tree

3 files changed

+43
-209
lines changed

3 files changed

+43
-209
lines changed

‎caffe/include/caffe/layers/jmmd_layer.hpp‎

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -57,7 +57,6 @@ class JMMDLossLayer : public LossLayer<Dtype> {
5757
int label_kernel_num_;
5858
Dtype label_kernel_mul_;
5959
bool auto_sigma_;
60-
bool label_back_propagate_;
6160
};
6261

6362
} // namespace caffe

‎caffe/src/caffe/layers/jmmd_layer.cpp‎

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -17,7 +17,6 @@ void JMMDLossLayer<Dtype>::LayerSetUp(
1717
fix_gamma_ = this->layer_param_.jmmd_param().fix_gamma();
1818
sigma_ = this->layer_param_.jmmd_param().sigma();
1919
auto_sigma_ = this->layer_param_.jmmd_param().auto_sigma();
20-
label_back_propagate_ = true;
2120
gamma_ = Dtype(-1);
2221
kernel_mul_ = this->layer_param_.jmmd_param().kernel_mul();
2322
label_kernel_mul_ = this->layer_param_.jmmd_param().label_kernel_mul();

‎caffe/src/caffe/layers/jmmd_layer.cu‎

Lines changed: 43 additions & 207 deletions
Original file line numberDiff line numberDiff line change
@@ -116,107 +116,6 @@ __global__ void CalculateProbRBFLabelKernel(const int n, const Dtype* source_p,
116116
}
117117
}
118118

119-
template <typename Dtype>
120-
__global__ void CalculateOneOfKRBFLabelKernel(const int n, const Dtype* source_label, const Dtype* target_p,
121-
const int source_num, const int total_num, const int label_dim, const Dtype sigma, Dtype* out) {
122-
CUDA_KERNEL_LOOP(index, n) {
123-
int index1 = index / total_num;
124-
int index2 = index % total_num;
125-
if(index1 < source_num && index2 < source_num){
126-
//ss
127-
out[index] = (source_label[index1] == source_label[index2]) ? Dtype(1) : exp(Dtype(-1));
128-
}
129-
else if(index1 < source_num && index2 >= source_num){
130-
int offset = label_dim * (index2 - source_num);
131-
Dtype sum = Dtype(0);
132-
for(int j = 0;j < label_dim;++j){
133-
sum += (j == source_label[index1]) ?
134-
(Dtype(1) - target_p[offset + j]) * (Dtype(1) - target_p[offset + j]):
135-
target_p[offset + j] * target_p[offset + j];
136-
}
137-
out[index] = exp(-sum / sigma);
138-
}
139-
else if(index1 >= source_num && index2 < source_num){
140-
//ts
141-
int offset = label_dim * (index1 - source_num);
142-
Dtype sum = Dtype(0);
143-
for(int j = 0;j < label_dim;++j){
144-
sum += (j == source_label[index2]) ?
145-
(Dtype(1) - target_p[offset + j]) * (Dtype(1) - target_p[offset + j]):
146-
target_p[offset + j] * target_p[offset + j];
147-
}
148-
out[index] = exp(-sum / sigma);
149-
}
150-
else{
151-
//tt
152-
int offset1 = label_dim * (index1 - source_num);
153-
int offset2 = label_dim * (index2 - source_num);
154-
Dtype sum = Dtype(0);
155-
for(int j = 0;j < label_dim;++j){
156-
sum += (target_p[offset1 + j] - target_p[offset2 + j]) * (target_p[offset1 + j] - target_p[offset2 + j]);
157-
}
158-
out[index] = exp(-sum / sigma);
159-
}
160-
}
161-
}
162-
163-
template <typename Dtype>
164-
__global__ void CalculateIdentifyLabelKernel(const int n, const Dtype* source_label, const Dtype* target_p,
165-
const int source_num, const int total_num, const int label_dim, Dtype* out) {
166-
CUDA_KERNEL_LOOP(index, n) {
167-
int index1 = index / total_num;
168-
int index2 = index % total_num;
169-
if(index1 < source_num && index2 < source_num){
170-
//ss
171-
out[index] = (source_label[index1] == source_label[index2]) ? Dtype(1) : Dtype(0);
172-
}
173-
else if(index1 < source_num && index2 >= source_num){
174-
int offset = label_dim * (index2 - source_num);
175-
Dtype max = target_p[offset + (int)source_label[index1]];
176-
out[index] = Dtype(1);
177-
for(int j = 0;j < label_dim;++j){
178-
if(target_p[offset + j] > max){
179-
out[index] = Dtype(0);
180-
break;
181-
}
182-
}
183-
}
184-
else if(index1 >= source_num && index2 < source_num){
185-
//ts
186-
int offset = label_dim * (index1 - source_num);
187-
Dtype max = target_p[offset + (int)source_label[index2]];
188-
out[index] = Dtype(1);
189-
for(int j = 0;j < label_dim;++j){
190-
if(target_p[offset + j] > max){
191-
out[index] = Dtype(0);
192-
break;
193-
}
194-
}
195-
}
196-
else{
197-
//tt
198-
int offset1 = label_dim * (index1 - source_num);
199-
int offset2 = label_dim * (index2 - source_num);
200-
Dtype max = Dtype(-1);
201-
int max_index;
202-
for(int j = 0;j < label_dim;++j){
203-
if(target_p[offset1 + j] > max){
204-
max = target_p[offset1 + j];
205-
max_index = j;
206-
}
207-
}
208-
out[index] = Dtype(1);
209-
max = target_p[offset2 + max_index];
210-
for(int j = 0;j < label_dim;++j){
211-
if(target_p[offset2 + j] > max){
212-
out[index] = Dtype(0);
213-
break;
214-
}
215-
}
216-
}
217-
}
218-
}
219-
220119
template <typename Dtype>
221120
void JMMDLossLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
222121
const vector<Blob<Dtype>*>& top) {
@@ -289,9 +188,9 @@ void JMMDLossLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
289188
}
290189
sigma_ = bandwidth / total_num_;
291190
}
292-
CalculateProbRBFLabelKernel<Dtype><<<CAFFE_GET_BLOCKS(nthreads), CAFFE_CUDA_NUM_THREADS>>>(
293-
nthreads, bottom[2]->gpu_data(), bottom[3]->gpu_data(),
294-
source_num_, total_num_, label_dim, sigma_, label_kernel_num_, label_kernel_mul_, delta_.mutable_gpu_data());
191+
CalculateProbRBFLabelKernel<Dtype><<<CAFFE_GET_BLOCKS(nthreads), CAFFE_CUDA_NUM_THREADS>>>(
192+
nthreads, bottom[2]->gpu_data(), bottom[3]->gpu_data(),
193+
source_num_, total_num_, label_dim, sigma_, label_kernel_num_, label_kernel_mul_, delta_.mutable_gpu_data());
295194
}
296195
else{
297196
caffe_gpu_set(total_num_ * total_num_, Dtype(1), delta_.mutable_gpu_data());
@@ -301,84 +200,29 @@ void JMMDLossLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
301200
top[0]->mutable_cpu_data()[0] = Dtype(0);
302201
}
303202

304-
template <typename Dtype>
305-
__global__ void CalculateDiff(const int n, const Dtype* source, const Dtype* target,
306-
const int source_num, const int total_num, const int dim, const Dtype* kernel_val,
307-
const Dtype gamma, const Dtype* label_kernel, const Dtype* data, const int data_index,
308-
Dtype* source_diff, Dtype* target_diff) {
309-
CUDA_KERNEL_LOOP(index, n) {
310-
int oppo_index = index / dim;
311-
int dim_offset = index % dim;
312-
Dtype data1 = data[dim_offset];
313-
Dtype data2 = (oppo_index >= source_num) ?
314-
target[dim * (oppo_index - source_num) + dim_offset]:
315-
source[dim * oppo_index + dim_offset];
316-
int kernel_index = data_index * total_num + oppo_index;
317-
Dtype factor_of_diff = -2 * gamma * kernel_val[kernel_index] * label_kernel[kernel_index];
318-
if(oppo_index >= source_num){
319-
oppo_index -= source_num;
320-
if(data_index >= source_num)
321-
target_diff[oppo_index * dim + dim_offset] += factor_of_diff * (data2 - data1);
322-
else
323-
target_diff[oppo_index * dim + dim_offset] += -0.5 * factor_of_diff * (data2 - data1);
324-
}
325-
else{
326-
if(data_index < source_num)
327-
source_diff[oppo_index * dim + dim_offset] += 0.25 * factor_of_diff * (data2 - data1);
328-
else
329-
source_diff[oppo_index * dim + dim_offset] += -0.5 * factor_of_diff * (data2 - data1);
330-
}
331-
}
332-
}
333-
334203
template <typename Dtype>
335204
void JMMDLossLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
336205
const vector<bool>& propagate_down,
337206
const vector<Blob<Dtype>*>& bottom) {
338207
Dtype* source_diff = bottom[0]->mutable_gpu_diff();
339208
Dtype* target_diff = bottom[1]->mutable_gpu_diff();
340-
Dtype* label_source_diff = (label_back_propagate_) ? bottom[2]->mutable_gpu_diff() : NULL;
341-
Dtype* label_target_diff = (label_back_propagate_) ? bottom[3]->mutable_gpu_diff() : NULL;
342-
int label_dim = (label_back_propagate_) ? bottom[2]->count() / bottom[2]->count(0, 1) : 0;
343-
if(label_back_propagate_){
344-
caffe_gpu_set(source_num_ * label_dim, Dtype(0), label_source_diff);
345-
caffe_gpu_set(target_num_ * label_dim, Dtype(0), label_target_diff);
346-
}
209+
Dtype* label_source_diff = bottom[2]->mutable_gpu_diff();
210+
Dtype* label_target_diff = bottom[3]->mutable_gpu_diff();
211+
int label_dim = bottom[2]->count() / bottom[2]->count(0, 1);
212+
caffe_gpu_set(source_num_ * label_dim, Dtype(0), label_source_diff);
213+
caffe_gpu_set(target_num_ * label_dim, Dtype(0), label_target_diff);
347214

348215
caffe_gpu_set(source_num_ * dim_, Dtype(0), source_diff);
349216
caffe_gpu_set(target_num_ * dim_, Dtype(0), target_diff);
350-
/*
351-
for(int i = 0;i < total_num_;++i){
352-
int nthreads = total_num_ * dim_;
353-
const Dtype* data = (i >= source_num_) ?
354-
(bottom[1]->gpu_data() + dim_ * (i - source_num_)) :
355-
(bottom[0]->gpu_data() + dim_ * i);
356-
357-
Dtype gamma_times = pow(kernel_mul_, (Dtype)(kernel_num_ / 2));
358-
Dtype kernel_gamma = gamma_ / gamma_times;
359-
360-
for(int j = 0;j < kernel_num_;++j){
361-
CalculateDiff<Dtype><<<CAFFE_GET_BLOCKS(nthreads), CAFFE_CUDA_NUM_THREADS>>>(
362-
nthreads, bottom[0]->gpu_data(), bottom[1]->gpu_data(),
363-
source_num_, total_num_, dim_, kernel_val_[j]->gpu_data(),
364-
kernel_gamma, delta_.gpu_data(), data, i,
365-
bottom[0]->mutable_gpu_diff(), bottom[1]->mutable_gpu_diff());
366-
kernel_gamma *= kernel_mul_;
367-
}
368-
}
369-
caffe_gpu_scal(source_num_ * dim_, loss_weight_ / total_num_ / total_num_, bottom[0]->mutable_gpu_diff());
370-
caffe_gpu_scal(target_num_ * dim_, loss_weight_ / total_num_ / total_num_, bottom[1]->mutable_gpu_diff());
371-
*/
217+
372218
if(source_num_ <= 1 || target_num_ <= 1) return;
373219
int sample_num = (source_num_ > target_num_) ? source_num_ : target_num_;
374220
int s1, s2, t1, t2;
375221
Dtype* tempX1 = diff_.mutable_gpu_diff() + total_num_ * total_num_;
376222
Dtype* tempX2 = diff_.mutable_gpu_diff() + total_num_ * total_num_ + dim_;
377223
Dtype* tempY1 = NULL, *tempY2 = NULL;
378-
if(label_back_propagate_){
379-
tempY1 = diff_.mutable_gpu_diff() + total_num_ * total_num_ + dim_ + dim_;
380-
tempY2 = diff_.mutable_gpu_diff() + total_num_ * total_num_ + dim_ + dim_ + label_dim;
381-
}
224+
tempY1 = diff_.mutable_gpu_diff() + total_num_ * total_num_ + dim_ + dim_;
225+
tempY2 = diff_.mutable_gpu_diff() + total_num_ * total_num_ + dim_ + dim_ + label_dim;
382226
for(int i = 0;i < sample_num;++i){
383227
s1 = rand() % source_num_;
384228
s2 = rand() % source_num_;
@@ -394,10 +238,10 @@ void JMMDLossLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
394238
const Dtype* x_s2 = bottom[0]->gpu_data() + s2 * dim_;
395239
const Dtype* x_t1 = bottom[1]->gpu_data() + t1 * dim_;
396240
const Dtype* x_t2 = bottom[1]->gpu_data() + t2 * dim_;
397-
const Dtype* y_s1 = (label_back_propagate_) ? bottom[2]->gpu_data() + s1 * label_dim : NULL;
398-
const Dtype* y_s2 = (label_back_propagate_) ? bottom[2]->gpu_data() + s2 * label_dim : NULL;
399-
const Dtype* y_t1 = (label_back_propagate_) ? bottom[3]->gpu_data() + t1 * label_dim : NULL;
400-
const Dtype* y_t2 = (label_back_propagate_) ? bottom[3]->gpu_data() + t2 * label_dim : NULL;
241+
const Dtype* y_s1 = bottom[2]->gpu_data() + s1 * label_dim;
242+
const Dtype* y_s2 = bottom[2]->gpu_data() + s2 * label_dim;
243+
const Dtype* y_t1 = bottom[3]->gpu_data() + t1 * label_dim;
244+
const Dtype* y_t2 = bottom[3]->gpu_data() + t2 * label_dim;
401245

402246
caffe_gpu_sub<Dtype>(dim_, x_s1, x_s2, tempX1);
403247
caffe_gpu_sub<Dtype>(dim_, x_s2, x_s1, tempX2);
@@ -417,15 +261,13 @@ void JMMDLossLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
417261
caffe_gpu_scal(dim_, loss_weight_ * factor_for_diff / sample_num, tempX2);
418262
caffe_gpu_add(dim_, tempX1, source_diff + s1 * dim_, source_diff + s1 * dim_);
419263
caffe_gpu_add(dim_, tempX2, source_diff + s2 * dim_, source_diff + s2 * dim_);
420-
if(label_back_propagate_){
421-
caffe_gpu_sub<Dtype>(label_dim, y_s1, y_s2, tempY1);
422-
caffe_gpu_sub<Dtype>(label_dim, y_s2, y_s1, tempY2);
423-
factor_for_diff = (-2) / sigma_ * x_kernel * delta_.cpu_data()[s1 * total_num_ + s2];
424-
caffe_gpu_scal(label_dim, label_loss_weight_ * factor_for_diff / sample_num, tempY1);
425-
caffe_gpu_scal(label_dim, label_loss_weight_ * factor_for_diff / sample_num, tempY2);
426-
caffe_gpu_add(label_dim, tempY1, label_source_diff + s1 * label_dim, label_source_diff + s1 * label_dim);
427-
caffe_gpu_add(label_dim, tempY2, label_source_diff + s2 * label_dim, label_source_diff + s2 * label_dim);
428-
}
264+
caffe_gpu_sub<Dtype>(label_dim, y_s1, y_s2, tempY1);
265+
caffe_gpu_sub<Dtype>(label_dim, y_s2, y_s1, tempY2);
266+
factor_for_diff = (-2) / sigma_ * x_kernel * delta_.cpu_data()[s1 * total_num_ + s2];
267+
caffe_gpu_scal(label_dim, label_loss_weight_ * factor_for_diff / sample_num, tempY1);
268+
caffe_gpu_scal(label_dim, label_loss_weight_ * factor_for_diff / sample_num, tempY2);
269+
caffe_gpu_add(label_dim, tempY1, label_source_diff + s1 * label_dim, label_source_diff + s1 * label_dim);
270+
caffe_gpu_add(label_dim, tempY2, label_source_diff + s2 * label_dim, label_source_diff + s2 * label_dim);
429271

430272
factor_for_diff = 0;
431273
caffe_gpu_sub<Dtype>(dim_, x_s1, x_t2, tempX1);
@@ -445,15 +287,13 @@ void JMMDLossLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
445287
caffe_gpu_scal(dim_, loss_weight_ * factor_for_diff / sample_num, tempX2);
446288
caffe_gpu_add(dim_, tempX1, source_diff + s1 * dim_, source_diff + s1 * dim_);
447289
caffe_gpu_add(dim_, tempX2, target_diff + t2 * dim_, target_diff + t2 * dim_);
448-
if(label_back_propagate_){
449-
caffe_gpu_sub<Dtype>(label_dim, y_s1, y_t2, tempY1);
450-
caffe_gpu_sub<Dtype>(label_dim, y_t2, y_s1, tempY2);
451-
factor_for_diff = 2 / sigma_ * x_kernel * delta_.cpu_data()[s1 * total_num_ + source_num_ + t2];
452-
caffe_gpu_scal(label_dim, label_loss_weight_ * factor_for_diff / sample_num, tempY1);
453-
caffe_gpu_scal(label_dim, label_loss_weight_ * factor_for_diff / sample_num, tempY2);
454-
caffe_gpu_add(label_dim, tempY1, label_source_diff + s1 * label_dim, label_source_diff + s1 * label_dim);
455-
caffe_gpu_add(label_dim, tempY2, label_target_diff + t2 * label_dim, label_target_diff + t2 * label_dim);
456-
}
290+
caffe_gpu_sub<Dtype>(label_dim, y_s1, y_t2, tempY1);
291+
caffe_gpu_sub<Dtype>(label_dim, y_t2, y_s1, tempY2);
292+
factor_for_diff = 2 / sigma_ * x_kernel * delta_.cpu_data()[s1 * total_num_ + source_num_ + t2];
293+
caffe_gpu_scal(label_dim, label_loss_weight_ * factor_for_diff / sample_num, tempY1);
294+
caffe_gpu_scal(label_dim, label_loss_weight_ * factor_for_diff / sample_num, tempY2);
295+
caffe_gpu_add(label_dim, tempY1, label_source_diff + s1 * label_dim, label_source_diff + s1 * label_dim);
296+
caffe_gpu_add(label_dim, tempY2, label_target_diff + t2 * label_dim, label_target_diff + t2 * label_dim);
457297

458298
factor_for_diff = 0;
459299
caffe_gpu_sub<Dtype>(dim_, x_t1, x_s2, tempX1);
@@ -473,15 +313,13 @@ void JMMDLossLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
473313
caffe_gpu_scal(dim_, loss_weight_ * factor_for_diff / sample_num, tempX2);
474314
caffe_gpu_add(dim_, tempX1, target_diff + t1 * dim_, target_diff + t1 * dim_);
475315
caffe_gpu_add(dim_, tempX2, source_diff + s2 * dim_, source_diff + s2 * dim_);
476-
if(label_back_propagate_){
477-
caffe_gpu_sub<Dtype>(label_dim, y_s2, y_t1, tempY1);
478-
caffe_gpu_sub<Dtype>(label_dim, y_t1, y_s2, tempY2);
479-
factor_for_diff = 2 / sigma_ * x_kernel * delta_.cpu_data()[(t1 + source_num_) * total_num_ + s2];
480-
caffe_gpu_scal(label_dim, label_loss_weight_ * factor_for_diff / sample_num, tempY1);
481-
caffe_gpu_scal(label_dim, label_loss_weight_ * factor_for_diff / sample_num, tempY2);
482-
caffe_gpu_add(label_dim, tempY1, label_source_diff + s2 * label_dim, label_source_diff + s2 * label_dim);
483-
caffe_gpu_add(label_dim, tempY2, label_target_diff + t1 * label_dim, label_target_diff + t1 * label_dim);
484-
}
316+
caffe_gpu_sub<Dtype>(label_dim, y_s2, y_t1, tempY1);
317+
caffe_gpu_sub<Dtype>(label_dim, y_t1, y_s2, tempY2);
318+
factor_for_diff = 2 / sigma_ * x_kernel * delta_.cpu_data()[(t1 + source_num_) * total_num_ + s2];
319+
caffe_gpu_scal(label_dim, label_loss_weight_ * factor_for_diff / sample_num, tempY1);
320+
caffe_gpu_scal(label_dim, label_loss_weight_ * factor_for_diff / sample_num, tempY2);
321+
caffe_gpu_add(label_dim, tempY1, label_source_diff + s2 * label_dim, label_source_diff + s2 * label_dim);
322+
caffe_gpu_add(label_dim, tempY2, label_target_diff + t1 * label_dim, label_target_diff + t1 * label_dim);
485323

486324
factor_for_diff = 0;
487325
caffe_gpu_sub<Dtype>(dim_, x_t1, x_t2, tempX1);
@@ -501,15 +339,13 @@ void JMMDLossLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
501339
caffe_gpu_scal(dim_, loss_weight_ * factor_for_diff / sample_num, tempX2);
502340
caffe_gpu_add(dim_, tempX1, target_diff + t1 * dim_, target_diff + t1 * dim_);
503341
caffe_gpu_add(dim_, tempX2, target_diff + t2 * dim_, target_diff + t2 * dim_);
504-
if(label_back_propagate_){
505-
caffe_gpu_sub<Dtype>(label_dim, y_t1, y_t2, tempY1);
506-
caffe_gpu_sub<Dtype>(label_dim, y_t2, y_t1, tempY2);
507-
factor_for_diff = (-2) / sigma_ * x_kernel * delta_.cpu_data()[(t1 + source_num_) * total_num_ + t2 + source_num_];
508-
caffe_gpu_scal(label_dim, label_loss_weight_ * factor_for_diff / sample_num, tempY1);
509-
caffe_gpu_scal(label_dim, label_loss_weight_ * factor_for_diff / sample_num, tempY2);
510-
caffe_gpu_add(label_dim, tempY1, label_target_diff + t1 * label_dim, label_target_diff + t1 * label_dim);
511-
caffe_gpu_add(label_dim, tempY2, label_target_diff + t2 * label_dim, label_target_diff + t2 * label_dim);
512-
}
342+
caffe_gpu_sub<Dtype>(label_dim, y_t1, y_t2, tempY1);
343+
caffe_gpu_sub<Dtype>(label_dim, y_t2, y_t1, tempY2);
344+
factor_for_diff = (-2) / sigma_ * x_kernel * delta_.cpu_data()[(t1 + source_num_) * total_num_ + t2 + source_num_];
345+
caffe_gpu_scal(label_dim, label_loss_weight_ * factor_for_diff / sample_num, tempY1);
346+
caffe_gpu_scal(label_dim, label_loss_weight_ * factor_for_diff / sample_num, tempY2);
347+
caffe_gpu_add(label_dim, tempY1, label_target_diff + t1 * label_dim, label_target_diff + t1 * label_dim);
348+
caffe_gpu_add(label_dim, tempY2, label_target_diff + t2 * label_dim, label_target_diff + t2 * label_dim);
513349
}
514350
}
515351

0 commit comments

Comments
 (0)