mmr/tkDNN
Archived
1
0
Fork 0
Code Issues 46 Pull Requests 12 Releases 4 Activity

Fix the inference operation of the deformable convolutional layer.

Browse Source 
This commit removes the malloc operation in the inference
method and adds the sigmoid kernel.

Signed-off-by: Davide Sapienza <sapienza.dav@gmail.com>
This commit is contained in:
Davide Sapienza
2019-12-19 14:47:57 +01:00
parent df888a3457
commit e99b353d8b
5 changed files with 97 additions and 163 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/tkDNN/Layer.h
+5
View File
@@ -219,7 +219,12 @@ public:
int kernelH, kernelW, strideH, strideW, paddingH, paddingW;
protected:
dnnType *ones_d1;
dnnType *ones_d2;
cudnnTensorDescriptor_t biasTensorDesc;
int chunk_dim;
dnnType *offset, *mask;
dnnType *output_conv;
void initCUDNN();
include/tkDNN/kernels.h
+1
View File
@@ -6,6 +6,7 @@
void activationELUForward(dnnType* srcData, dnnType* dstData, int size, cudaStream_t stream = cudaStream_t(0));
void activationLEAKYForward(dnnType* srcData, dnnType* dstData, int size, cudaStream_t stream = cudaStream_t(0));
void activationLOGISTICForward(dnnType* srcData, dnnType* dstData, int size, cudaStream_t stream = cudaStream_t(0));
void activationSIGMOIDForward(dnnType* srcData, dnnType* dstData, int size, cudaStream_t stream = cudaStream_t(0));
void fill(dnnType* data, int size, dnnType val, cudaStream_t stream = cudaStream_t(0));