Skip to content

Align output shape formulas across conv implementations #237

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
aobolensk merged 1 commit into from
Nov 23, 2025
Merged

Align output shape formulas across conv implementations #237

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
58 changes: 22 additions & 36 deletions include/layers/ConvLayer.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,16 @@

namespace it_lab_ai {

inline size_t ComputeConvOutputDim(size_t input_size, size_t kernel_size,
size_t stride, size_t padding,
size_t dilation) {
const size_t effective_kernel = dilation * (kernel_size - 1) + 1;
if (stride == 0 || input_size + 2 * padding < effective_kernel) {
return 0;
}
return (input_size + 2 * padding - effective_kernel) / stride + 1;
}

class ConvolutionalLayer : public Layer {
private:
size_t stride_;
Expand Down Expand Up @@ -165,12 +175,10 @@ void Conv4D(const Tensor& input, const Tensor& kernel_, const Tensor& bias_,
}
}

size_t out_height =
(in_height + 2 * pads_ - dilations_ * (kernel_height - 1) - 1) / stride_ +
1;
size_t out_height = ComputeConvOutputDim(in_height, kernel_height, stride_,
pads_, dilations_);
size_t out_width =
(in_width + 2 * pads_ - dilations_ * (kernel_width - 1) - 1) / stride_ +
1;
ComputeConvOutputDim(in_width, kernel_width, stride_, pads_, dilations_);

std::vector<std::vector<std::vector<std::vector<ValueType>>>> padded_input(
batch_size,
Expand Down Expand Up @@ -352,20 +360,10 @@ void Conv4DSTL(const Tensor& input, const Tensor& kernel_, const Tensor& bias_,
for (auto& t : threads) t.join();
threads.clear();

size_t crat = 0;
if ((in_height + 2 * pads_ - dilations_ * (kernel_height - 1)) % stride_ != 0)
crat = 1;

size_t out_height =
(in_height + 2 * pads_ - dilations_ * (kernel_height - 1)) / stride_ +
crat;

crat = 0;
if ((in_width + 2 * pads_ - dilations_ * (kernel_width - 1)) % stride_ != 0)
crat = 1;

size_t out_height = ComputeConvOutputDim(in_height, kernel_height, stride_,
pads_, dilations_);
size_t out_width =
(in_width + 2 * pads_ - dilations_ * (kernel_width - 1)) / stride_ + crat;
ComputeConvOutputDim(in_width, kernel_width, stride_, pads_, dilations_);

std::vector<std::vector<std::vector<std::vector<ValueType>>>> output_tensor(
batch_size, std::vector<std::vector<std::vector<ValueType>>>(
Expand Down Expand Up @@ -474,12 +472,10 @@ void DepthwiseConv4D(const Tensor& input, const Tensor& kernel_,
throw std::runtime_error("Invalid kernel shape for depthwise convolution");
}

size_t out_height =
(in_height + 2 * pads_ - dilations_ * (kernel_height - 1) - 1) / stride_ +
1;
size_t out_height = ComputeConvOutputDim(in_height, kernel_height, stride_,
pads_, dilations_);
size_t out_width =
(in_width + 2 * pads_ - dilations_ * (kernel_width - 1) - 1) / stride_ +
1;
ComputeConvOutputDim(in_width, kernel_width, stride_, pads_, dilations_);

Tensor output_tensor(Shape({batch_size, channels, out_height, out_width}),
input.get_type());
Expand Down Expand Up @@ -568,20 +564,10 @@ void Conv4D_Legacy(const Tensor& input, const Tensor& kernel_,
}
}

size_t crat = 0;
if ((in_height + 2 * pads_ - dilations_ * (kernel_height - 1)) % stride_ != 0)
crat = 1;

size_t out_height =
(in_height + 2 * pads_ - dilations_ * (kernel_height - 1)) / stride_ +
crat;

crat = 0;
if ((in_width + 2 * pads_ - dilations_ * (kernel_width - 1)) % stride_ != 0)
crat = 1;

size_t out_height = ComputeConvOutputDim(in_height, kernel_height, stride_,
pads_, dilations_);
size_t out_width =
(in_width + 2 * pads_ - dilations_ * (kernel_width - 1)) / stride_ + crat;
ComputeConvOutputDim(in_width, kernel_width, stride_, pads_, dilations_);

std::vector<std::vector<std::vector<std::vector<ValueType>>>> output_tensor(
batch_size, std::vector<std::vector<std::vector<ValueType>>>(
Expand Down
Loading