diff options
author | Jan Eilers <jan.eilers@arm.com> | 2021-06-02 12:01:25 +0100 |
---|---|---|
committer | Jan Eilers <jan.eilers@arm.com> | 2021-06-16 11:31:42 +0000 |
commit | 53ef79504b4c881c572735393c2eede5fa556c46 (patch) | |
tree | f6e0cd27c4d03075fa154074c5b12d7c8c3149f7 /src/backends/backendsCommon/WorkloadData.hpp | |
parent | 77fe76bfa8cb798943821d1f3e432c228e1cdee3 (diff) | |
download | armnn-53ef79504b4c881c572735393c2eede5fa556c46.tar.gz |
IVGCVSW-5826 Change weights layout for depthwise to [1,H,W,I*M]
* This change is necessary because tflite uses a [1,H,W,I*M] format
and uses the I*M dimension for per axis quantization. Our previous
layout [M,I,H,W] can't handle the correlating quantization scales.
* Updates Onnx-, TfLiteParser and TfliteDelegate
* Updates the CpuRef, CpuAcc and GpuAcc backends
* Adjusts unit tests
* Adds test to ensure models with old layout can still be read and
executed
* Adds conversion function to previous layout [1,H,W,I*M] --> [M,I,H,W]
which can be used by backend developers
!android-nn-driver:5553
Signed-off-by: Jan Eilers <jan.eilers@arm.com>
Change-Id: Ifef23368b8c3702cf315a5838d214f7dc13c0152
Diffstat (limited to 'src/backends/backendsCommon/WorkloadData.hpp')
-rw-r--r-- | src/backends/backendsCommon/WorkloadData.hpp | 14 |
1 files changed, 13 insertions, 1 deletions
diff --git a/src/backends/backendsCommon/WorkloadData.hpp b/src/backends/backendsCommon/WorkloadData.hpp index 77d4209657..11ce2cb44f 100644 --- a/src/backends/backendsCommon/WorkloadData.hpp +++ b/src/backends/backendsCommon/WorkloadData.hpp @@ -208,7 +208,19 @@ struct Convolution2dQueueDescriptor : QueueDescriptorWithParameters<Convolution2 void Validate(const WorkloadInfo& workloadInfo) const; }; -// Depthwise Convolution 2D layer workload data. +/// Depthwise Convolution 2D layer workload data. +/// +/// @note +/// The weights are in the format [1, H, W, I*M]. Where I is the input channel size, M the depthwise mutliplier and +/// H, W is the height and width of the filter kernel. If per channel quantization is applied +/// the weights will be quantized along the last dimension/axis (I*M) which corresponds to the output channel size. +/// If per channel quantization is applied the weights tensor will have I*M scales, one for each dimension +/// of the quantization axis. You have to be aware of this when reshaping the weights tensor. +/// Splitting the I*M axis, e.g. [1, H, W, I*M] --> [H, W, I, M], won't work without taking care of the +/// corresponding quantization scales. +/// If there is no per channel quantization applied reshaping the weights tensor won't cause any issues. There are +/// preconfigured permutation functions available @link WorkloadUtils.hpp here. +/// struct DepthwiseConvolution2dQueueDescriptor : QueueDescriptorWithParameters<DepthwiseConvolution2dDescriptor> { DepthwiseConvolution2dQueueDescriptor() |