From b9070a42a44ec1a0102e2f0b04523d2e96392903 Mon Sep 17 00:00:00 2001 From: Matthew Jackson Date: Thu, 22 Aug 2019 16:13:27 +0100 Subject: COMPMID-2605: Add asymmetric padding support for Deconvolution layer Change-Id: I63b773bdce25f1342ccd3a08ded623a1508f70fe Signed-off-by: Matthew Jackson Reviewed-on: https://review.mlplatform.org/c/1797 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Michele Di Giorgio Reviewed-by: Giuseppe Rossini --- .../CL/functions/CLDirectDeconvolutionLayer.cpp | 44 ++++++++++++++-------- 1 file changed, 29 insertions(+), 15 deletions(-) (limited to 'src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp') diff --git a/src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp b/src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp index c1a39ef26a..b8089d8229 100644 --- a/src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp +++ b/src/runtime/CL/functions/CLDirectDeconvolutionLayer.cpp @@ -63,13 +63,8 @@ Status CLDirectDeconvolutionLayer::validate(const ITensorInfo *input, const ITen ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_w) != weights->dimension(idx_h)); ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_w) < 1); - ARM_COMPUTE_RETURN_ERROR_ON(!info.padding_is_symmetric()); - const unsigned int stride_x = info.stride().first; - const unsigned int stride_y = info.stride().second; - - auto out_dims = deconvolution_output_dimensions(input->dimension(idx_w), input->dimension(idx_h), weights->dimension(idx_w), weights->dimension(idx_h), - info.pad().first, info.pad().second, stride_x, stride_y); + auto out_dims = deconvolution_output_dimensions(input->dimension(idx_w), input->dimension(idx_h), weights->dimension(idx_w), weights->dimension(idx_h), info); const TensorShape output_shape = compute_deconvolution_output_shape(out_dims, *input, *weights); @@ -92,9 +87,11 @@ Status CLDirectDeconvolutionLayer::validate(const ITensorInfo *input, const ITen ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(idx_h) != output_shape[idx_h], "Output's height is invalid."); ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(idx_c) != output_shape[idx_c], "Output's depth is invalid."); - unsigned int padx = 0; - unsigned int pady = 0; - const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*input, *weights, stride_x, stride_y, out_dims, padx, pady); + unsigned int deconv_pad_x = 0; + unsigned int deconv_pad_y = 0; + const unsigned int stride_x = info.stride().first; + const unsigned int stride_y = info.stride().second; + const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*input, *weights, stride_x, stride_y, out_dims, deconv_pad_x, deconv_pad_y); TensorInfo scale_out_info(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(scale_out_shape).set_data_layout(data_layout)); const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL); @@ -109,6 +106,10 @@ void CLDirectDeconvolutionLayer::configure(ICLTensor *input, ICLTensor *weights, { ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); + const unsigned int pad_left = info.pad_left(); + const unsigned int pad_right = info.pad_right(); + const unsigned int pad_top = info.pad_top(); + const unsigned int pad_bottom = info.pad_bottom(); const unsigned int stride_x = info.stride().first; const unsigned int stride_y = info.stride().second; @@ -122,8 +123,7 @@ void CLDirectDeconvolutionLayer::configure(ICLTensor *input, ICLTensor *weights, _weights_flipped.allocator()->init(weights->info()->clone()->set_data_layout(data_layout)); _flip_weights.configure(weights, &_weights_flipped, &_flip_axis); - auto out_dims = deconvolution_output_dimensions(input->info()->dimension(idx_w), input->info()->dimension(idx_h), weights->info()->dimension(idx_w), weights->info()->dimension(idx_h), - info.pad().first, info.pad().second, stride_x, stride_y); + auto out_dims = deconvolution_output_dimensions(input->info()->dimension(idx_w), input->info()->dimension(idx_h), weights->info()->dimension(idx_w), weights->info()->dimension(idx_h), info); const TensorShape output_shape = compute_deconvolution_output_shape(out_dims, *input->info(), *weights->info()); @@ -138,16 +138,30 @@ void CLDirectDeconvolutionLayer::configure(ICLTensor *input, ICLTensor *weights, _memory_group.manage(&_scaled_output); // Find the upsampled dimensions and the padding needed for the convolution with stride 1 in order to match output shape - unsigned int padx = 0; - unsigned int pady = 0; - const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*input->info(), *weights->info(), stride_x, stride_y, out_dims, padx, pady); + unsigned int deconv_pad_x = 0; + unsigned int deconv_pad_y = 0; + const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*input->info(), *weights->info(), stride_x, stride_y, out_dims, deconv_pad_x, deconv_pad_y); + + unsigned int deconv_pad_left = pad_right > pad_left ? pad_right - pad_left : 0; + unsigned int deconv_pad_right = pad_left > pad_right ? pad_left - pad_right : 0; + deconv_pad_x -= deconv_pad_left + deconv_pad_right; + ARM_COMPUTE_ERROR_ON((deconv_pad_x % 2) != 0); + deconv_pad_left += deconv_pad_x / 2; + deconv_pad_right += deconv_pad_x / 2; + + unsigned int deconv_pad_top = pad_bottom > pad_top ? pad_bottom - pad_top : 0; + unsigned int deconv_pad_bottom = pad_top > pad_bottom ? pad_top - pad_bottom : 0; + deconv_pad_y -= deconv_pad_top + deconv_pad_bottom; + ARM_COMPUTE_ERROR_ON((deconv_pad_y % 2) != 0); + deconv_pad_top += deconv_pad_y / 2; + deconv_pad_bottom += deconv_pad_y / 2; TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type(), input->info()->quantization_info()); scale_out_info.set_data_layout(data_layout); _scaled_output.allocator()->init(scale_out_info); // configure scale function - const PadStrideInfo upsample_info(stride_x, stride_y, padx / 2, pady / 2); + const PadStrideInfo upsample_info(stride_x, stride_y, deconv_pad_left, deconv_pad_right, deconv_pad_top, deconv_pad_bottom, DimensionRoundingType::FLOOR); _scale_f.configure(input, &_scaled_output, upsample_info); // Setup the function to convolve the upscaled output -- cgit v1.2.1