Recently,I need to change the code of softmax_loss_layer in caffe. The caffe version is: https://github.com/BVLC/caffe
The code between ... is my code,and the rest is the same.Here is how I modified the code First,I added rs_ in loss_layers.hpp
class SoftmaxWithLossLayer : public LossLayer<Dtype> {
...
Blob<Dtype> rs_;
...
}
Then,I reshape rs_ in softmax_loss_layer.cpp
void SoftmaxWithLossLayer<Dtype>::LayerSetUp(
const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
...
rs_.Reshape(bottom[2]->num(),bottom[2]->channels(),bottom[2]->height(),
bottom[2]->width());
}
Bottom[2] is from dense_image_data_layer.cpp
template <typename Dtype>
void DenseImageDataLayer<Dtype>::DataLayerSetUp(constvector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
top[2]->Reshape(batch_size, 1, height, width);
this->prefetch_weight_.Reshape(batch_size, 1, height, width);
this->transformed_weight_.Reshape(1, 1, height, width);
}
Dtype* top_data = top[2]->mutable_cpu_data();
const int count = top[2]->count();
if(crop_size > 0) {
for(int index = 0;index < count;++index)
{
const int ww = index % crop_size;
const int wh = (index / crop_size) % crop_size;
top_data[index]=static_cast<Dtype>(cv_weight.at<uchar>(wh,ww));
}
}
else{
for(int index = 0;index < count;++index)
{
const int ww = index % width;
const int wh = (index / width) % height;
top_data[index]=static_cast<Dtype>(cv_weight.at<uchar>(wh,ww));
}
}
At last,I added this code in softmax_loss_layer.cu
template <typename Dtype>
void SoftmaxWithLossLayer<Dtype>::Forward_gpu(
const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
...
Dtype loss;
Dtype em;
bool add_weight=this->layer_param_.loss_param().add_weight();
Dtype* weight=bottom[2]->mutable_gpu_data();
Dtype z=0;
Dtype* rs=rs_.mutable_gpu_data();
if(add_weight)
{
SoftmaxLossWeightForwardGPU<Dtype><<<CAFFE_GET_BLOCKS(nthreads),
CAFFE_CUDA_NUM_THREADS>>>(nthreads, prob_data, label,
weight_by_label_freqs_, label_count_data , loss_data,
outer_num_, dim, inner_num_,
has_ignore_label_, ignore_label_, counts,weight,rs);
const Dtype*rs1=rs_.gpu_data();
caffe_gpu_asum(nthreads, loss_data, &loss);
em=loss/nthreads;
count=nthreads;
Dtype am=1/2*log((1-em)/em);
CalculateZ<Dtype><<<CAFFE_GET_BLOCKS(nthreads),
CAFFE_CUDA_NUM_THREADS>>>(nthreads,weight,rs1,am,z,inner_num_);
WeightUpdate<Dtype><<<CAFFE_GET_BLOCKS(nthreads),
CAFFE_CUDA_NUM_THREADS>>>(nthreads,weight,rs1,am,z,inner_num_);
...
}
template <typename Dtype>
void SoftmaxWithLossLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
...
const Dtype* weight=bottom[2]->gpu_data();
bool add_weight=this->layer_param_.loss_param().add_weight();
if(add_weight)
SoftmaxLossWeightBackwardGPU<Dtype><<<CAFFE_GET_BLOCKS(nthreads),
CAFFE_CUDA_NUM_THREADS>>>(nthreads, top_data, label,
weight_by_label_freqs_, label_count_data, bottom_diff,
outer_num_, dim, inner_num_, has_ignore_label_,
ignore_label_, counts,weight);
...
}
The error is:
F0108 10:45:48.291290 2859 math_functions.cu:81] Check failed: error == cudaSuccess (77 vs. 0) an illegal memory access was encountered
*** Check failure stack trace: ***
@ 0x7f0b3bcfbdaa (unknown)
@ 0x7f0b3bcfbce4 (unknown)
@ 0x7f0b3bcfb6e6 (unknown)
@ 0x7f0b3bcfe687 (unknown)
@ 0x7f0b3c16dd48 caffe::caffe_gpu_memcpy()
@ 0x7f0b3c09b67e caffe::SyncedMemory::gpu_data()
@ 0x7f0b3c054472 caffe::Blob<>::gpu_data()
@ 0x7f0b3c0b0568 caffe::Net<>::ForwardFromTo()
@ 0x7f0b3c0b0947 caffe::Net<>::ForwardPrefilled()
@ 0x7f0b3c08e555 caffe::Solver<>::Step()
@ 0x7f0b3c08ee8f caffe::Solver<>::Solve()
@ 0x407806 train()
@ 0x405d41 main
@ 0x7f0b3b20dec5 (unknown)
@ 0x4062ed (unknown)
@ (nil) (unknown)
I think that this error is because I didn't use rs_ correctly or bottom[2] is not right. All the code that I changed is posted above,so can anybody tell me what to do? If you need extra information,please tell me?
It is difficult to follow your code and changes.
Some comments;
reshape
ing rs_
should occur in reshape()
method, rather than in setup()
.
It is best using rs_.ReshapeLike(*bottom[2])
than explicitly enumerating num
, channels
etc. What if you are going to have a Blob with different number of dimensions?
Have you tested your modified layer? From caffe wiki :
Write tests in
test/test_your_layer.cpp
. Usetest/test_gradient_check_util.hpp
to check that your Forward and Backward implementations are in numerical agreement.
The technical post webpages of this site follow the CC BY-SA 4.0 protocol. If you need to reprint, please indicate the site URL or the original address.Any question please contact:yoyou2525@163.com.