* fixed crash related with loading model with fq and sigmoid
* renamed multiple_input.* to multiple_input_fq.*; removed two unnecessary FQ layers from smoke_fq_fusion_with_sigmoid test; moved FQ params to test params
* [GNA] Add fake quantize support for convolution padding
Combine seven ngraph matcher passes into two.
Remove max pool size checking.
Add ngraph reference tests for subgraphs which were processed by POT.
Fix remaining issues with Max Pooling output calculations.
Add setting of default compile target based on execution target.
* [GNA] Remove redundant subgraph matcher
* [GNA] Remove redundant subgraph matcher
* [GNA] Use built in ngraph padding calculation
* [GNA] Use less channels for tests
* [GNA] Rename tests file
* [GNA] Add ngraph reference tests and some minor fixes in the transformation itself
* [GNA] Add invalid ngraph reference tests
* [GNA] Align brace style
* [GNA] Remove unnecessary condition
* additional checks are added to prevent errors from occurring on the GNA library side
* fixing after review
* move AreLayersSupported from Cnn2D namespace
* GNA_DEBUG is commented
* batch size <= 8 test was added
* [GNA] Convert Matmul with batch size > 8 to pointwise convolution.
Support Eltwise split to more than 2 parts.
Fake Quantize support fixes.
* Put convolution split into a separate transformation
* Add separate transformations for cases with bias and fake quantize
* Rollback restriction for diagonal layer reshape
* Fix FakeQuantize weights fusion when multiple operations were using the same const
* [GNA] Fix scale factor calculation for Relu if it has source and doesn't have destination statistics
Co-authored-by: Dmitrii Khurtin <dmitrii.khurtin@intel.com>
* fix bug with broadcasting constant layer with fq layer
* BroadcastConstWithFakeQuantizePass is removed; BroadcastConstPass is moved up in pass list
* constLayer->outData.front()->setDims is moved to conditions
* prevLayer->outData.front()->setLayout(nextLayer->outData.front()->getLayout()); is added
* adding conv2d decomposition
* save point
* build of conv_2d factorization succeeds
* working 2d conv decomposing transform
* added pseudo code for handling larger kernels
* fix conv splitting due to size
* active work on convolution 1xK without dilation
* validated NHWC ordered networks with convolution kernel sizes:
3x3
3x1
5x1
1x5
1x3
TODO: 2d max pooling
* removed debug printouts
* fusing max pooling/bias/af when -disable_nhwc_to_nchw option is used
* code cleanup
* [GNA] Fixes for CI run
* [GNA] Add tests, fix transform
* [GNA] Fix padded2valid and conv2d decomposition coexistence
* [GNA] Temporarily disable tests due to mock call count issues
* [GNA] Split tests for different hw versions
Co-authored-by: prozen <piotr.rozen@intel.com>
* Swap inputs pass
* [GNA] Handle case Gemm layer
* [GNA] Convert Matmul to FC
* VS tests
* Move to common optimization
* Gemm execution such as FC
* Test calc scale_factor
* Some changes
* Working version
* [GNA] Insert transpose between convolution/pooling and reshape.
Insert copy layers after concat inputs with multiple connections to the concat.
Accept networks with input connected to layers with different orientations if one of input dimensions is 1.
Fix scale factor calculation for Eltwise layer.
Fixes for Gemm quantization.
* Insert transpose after Reshape and before Matmul
* Fix concat input alignment when it's the network input
* Comments applying
Co-authored-by: Andrey Dmitriev <andrey.dmitriev@intel.com>
* add conversion of padded to valid convolution without other parameters change
* [GNA] Fix graph loop when multiple connections exist from single layer to concat
* [GNA] Add 1d and 2d conv test cases
Add models covering all transform scenarios.
Add test cases covering 1d and 2d convolutions.
Update transform with the newest code.
Add minor fixes in transform and elsewhere.
* [GNA] Remove debug code
* [GNA] Fixes after review
* [GNA] Fix failing tests
Co-authored-by: prozen <piotr.rozen@intel.com>
Szymon Irzabek