Problem with deconvolution layer validation

[allnes]

Hello!

When I use deconvolution validate function I catch exception instead check status:

This code:

unsigned int pad_l = 0;
unsigned int pad_r = 5;
unsigned int pad_t = 0;
unsigned int pad_b = 7;
unsigned int stride_x = 1;
unsigned int stride_y = 1;

PadStrideInfo deconv_info(stride_x, stride_y, pad_l, pad_r, pad_t, pad_b,DimensionRoundingType::FLOOR);

Status status = NEDeconvolutionLayer::validate(&srcTensorInfo, &weiTensorInfo, nullptr, &dstTensorInfo, deconv_info);
if (!status) {
    std::cout << status.error_description() << std::endl;
    return false;
}

Layout - NCHW Precision - F32

And receive message:

in deconvolution_output_dimensions src/core/Utils.cpp:422: ((in_height - 1) * stride_y + kernel_height) < (pad_top + pad_bottom)

Why status of error description doesn't work?

[morgolock]

Hi @allnes

Could you please share the code initializing all the tensors including the shapes? I tried to reproduce but I'm missing srcTensorInfo, weiTensorInfo and dstTensorIn

[allnes]

Hi, this example create problem with exception:

TensorInfo srcTensorInfo = TensorInfo({7, 7, 12, 1}, 1, F32, NCHW);
TensorInfo weiTensorInfo = TensorInfo({3, 3, 6, 12}, 1, F32, NCHW);
TensorInfo dstTensorInfo = TensorInfo({15, 15, 6, 1}, 1, F32, NCHW);

And I have question, do you support negative pads?

[morgolock]

Hi @allnes

I tried the shapes with a new build using the latest main and I don't get the exception

  1 #include "arm_compute/core/Types.h"
  2 #include "utils/Utils.h"
  3 #include "tests/SimpleTensor.h"
  4 #include "arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h"
  5 
  6 
  7 using namespace std;
  8 using namespace arm_compute;
  9 using namespace arm_compute::test;
 10 
 11 
 12 
 13 int main()
 14 {   
 15     
 16     TensorInfo srcTensorInfo = TensorInfo({7, 7, 12, 1}, 1, DataType::F32, DataLayout::NCHW);
 17     TensorInfo weiTensorInfo = TensorInfo({3, 3, 6, 12}, 1, DataType::F32, DataLayout::NCHW);
 18     TensorInfo dstTensorInfo = TensorInfo({15, 15, 6, 1}, 1, DataType::F32, DataLayout::NCHW);
 19     unsigned int pad_l = 0;
 20     unsigned int pad_r = 5;
 21     unsigned int pad_t = 0;
 22     unsigned int pad_b = 7;
 23     unsigned int stride_x = 1;
 24     unsigned int stride_y = 1;
 25     
 26     PadStrideInfo padinfo(stride_x, stride_y, pad_l, pad_r, pad_t, pad_b,DimensionRoundingType::FLOOR);
 27      
 28     const auto status =  NEDeconvolutionLayer::validate(&srcTensorInfo, &weiTensorInfo,0, &dstTensorInfo,padinfo);
 29     std::cout << "validate " << bool(status) << " " << status.error_description() <<  std::endl;
 30    
 31     return 0;
 32 }

When I run the program I get

# LD_LIBRARY_PATH=.:$LD_LIBRARY_PATH ./test_deconv_validate
validate 0 in validate src/runtime/NEON/functions/NEDeconvolutionLayer.cpp:145: Output's width is invalid.

What version of ACL are you using?

Deconvolution does not support negative padding.

Hope this helps.

[allnes]

Hi @allnes

I tried the shapes with a new build using the latest main and I don't get the exception

  1 #include "arm_compute/core/Types.h"
  2 #include "utils/Utils.h"
  3 #include "tests/SimpleTensor.h"
  4 #include "arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h"
  5 
  6 
  7 using namespace std;
  8 using namespace arm_compute;
  9 using namespace arm_compute::test;
 10 
 11 
 12 
 13 int main()
 14 {   
 15     
 16     TensorInfo srcTensorInfo = TensorInfo({7, 7, 12, 1}, 1, DataType::F32, DataLayout::NCHW);
 17     TensorInfo weiTensorInfo = TensorInfo({3, 3, 6, 12}, 1, DataType::F32, DataLayout::NCHW);
 18     TensorInfo dstTensorInfo = TensorInfo({15, 15, 6, 1}, 1, DataType::F32, DataLayout::NCHW);
 19     unsigned int pad_l = 0;
 20     unsigned int pad_r = 5;
 21     unsigned int pad_t = 0;
 22     unsigned int pad_b = 7;
 23     unsigned int stride_x = 1;
 24     unsigned int stride_y = 1;
 25     
 26     PadStrideInfo padinfo(stride_x, stride_y, pad_l, pad_r, pad_t, pad_b,DimensionRoundingType::FLOOR);
 27      
 28     const auto status =  NEDeconvolutionLayer::validate(&srcTensorInfo, &weiTensorInfo,0, &dstTensorInfo,padinfo);
 29     std::cout << "validate " << bool(status) << " " << status.error_description() <<  std::endl;
 30    
 31     return 0;
 32 }

When I run the program I get

# LD_LIBRARY_PATH=.:$LD_LIBRARY_PATH ./test_deconv_validate
validate 0 in validate src/runtime/NEON/functions/NEDeconvolutionLayer.cpp:145: Output's width is invalid.

What version of ACL are you using?

Deconvolution does not support negative padding.

Hope this helps.

Hi, @morgolock Thanks for your reply, I will try to prepare a reproducer that shows the original error I found. When I do, I will post in this task. ACL version was 23.08.

[allnes]

I closed this issue because when I rechecked this case, I saw that we set negative values in unsigned int and get unpredictable behaviour.