Skip to content

N
ngraph
Switch branch/tag
  1. 18 Feb, 2020 4 commits
    • Adam Osewski's avatar
      Unify (static) auto-broadcasting helpers. (#4242) · fe054d67
      Adam Osewski authored 
      * Helper function get_axes_mapping.

* Enhance Broadcast:v1 NUMPY broadcasting.

- Enable NUMPY broadcasting mechanism to work in bothdirections:
    target_shape <-> arg_shape

* Add opset1:squeeze and fix bug in reading squeezed axis idx.

* Fix and enhance downgrade pass for Broadcast:v1

* Use Broadcast:v1 in ONNX Expand operator.

* Replace Broadcast:v0 with v1 in some helper functions.

* Remove call to deprecated legacy_broadcasting helper function.

* Add helper get_axes_mapping_output function.

* Use directly Broadcast:v1 instead of helper function.

* Get back operators from v0 in helper function.

* Use helper function and some refactoring.

* Add legacy style broadcast helper function for opset1.

* User helper broadcasting function for arithmetic operators.

* Add empty axis only if its size is equal to one.

* Aplly review remarks:

- Rename broadcasting function deleting _values_ infix
- Remove variables used only once.
- Use STL library where possible.
- Remove unnecessary conditions.

* Add helper for Broadcast:v1.

* Fix merge artifact and force unsigned type for argument.

* Review. Add additional check for static output.

* Apply clang-format.

* Fix: call v0 ops in ngraph::builder namespace.

* Move opset1 boradcasting helpers from util/broadcasting.hpp

* Use autobroadcast built-in mechanism for arithmetic operators in RNN.

* Move helper functions to autobroadcast.hpp file.

- Update calls with new namespace ngraph::builder
- Remove calls using shared_ptr<ngraph::Node> and replace them with
  one accepting Output<ngraph::Node>
- Some small formatting (remove unnecesary namespace prefix)

* Remove unused function.

* Rename error class to reflect it's NumPy related.

* Fix thrown error name in autobroadcast UT.

* Code refactoring.

- Use one one set of helpers to broadcast node according to NumPy scheme

* Documentation formatting.

* Remove include to deleted header.

* Apply style format.

* Remove std:: prefix.

* Do reshape and/or broadcast only when necessary.

* Remove std:: and ngraph:: prefixes.

* UT for numpy_broadcast_for_matmul and legacy boradcast.

* Rename helper function.

* UT for opset1 legacy broadcast helper function.

* Add more UT for get_axes_mapping and style-format.

* Review comments.

* Restrict if with NGRAPH_WARN to NGRAPH_CHECK.
Co-authored-by: 's avatarMichał Karzyński <postrational@users.noreply.github.com>
Co-authored-by: 's avatarSang Ik Lee <sang.ik.lee@intel.com>
      fe054d67
    • Diego Caballero's avatar
      [MLIR] Fix warning: variable used uninitialized. (#4337) · 639ff3f1
      Diego Caballero authored 
      Co-authored-by: 's avatarRobert Kimball <robert.kimball@intel.com>
Co-authored-by: 's avatarSang Ik Lee <sang.ik.lee@intel.com>
      639ff3f1
    • Michał Karzyński's avatar
      Update OneHot to use v1 ops only (#4333) · 5803d20f
      Michał Karzyński authored 
      Co-authored-by: 's avatarSang Ik Lee <sang.ik.lee@intel.com>
      5803d20f
    • Michał Karzyński's avatar
      Use default_opset in ONNX importer code (#4332) · 7f98942d
      Michał Karzyński authored 
      * Use default_opset in Gemm

* Use default_opset in Cos and Cosh

* Use default_opset in Slice

* Use default_opset in Log

* Use explicit opset0 in ScatterND

* Use default_opset in Hardmax

* Use default_opset in utils

* Remove redundant headers

* Style check
      7f98942d
  3. 16 Feb, 2020 1 commit
    • Tomasz Dołbniak's avatar
      POC enabling Resnet50 with dynamic batch dimension (#4298) · b2da4cee
      Tomasz Dołbniak authored 
      * Constify the onnx importer conv

* Extract and fix the groups attribute validation for Conv

* Check if the convolution's data input rank is static

* Validate the groups attribute against channels and filters

* Validate the conv operation in a separate function

* Dynamically broadcast the conv bias if needed

* Import a test model with dynamic batch conv op

* Run a conv test with dynamic batch

* Cleanup of conv bias handling code

* Use a proper Broadcast constructor for bias in onnx conv

* Handle dynamic ReduceMean with statically defined rank

* Use the target shape rank to construct the default output shape for Broadcast

* Handle ONNX Squeeze with dynamic input and static rank

* Handle ONNX Shape with dynamic input and static rank

* Handle the dynamic target shape in ONNX Reshape

* Fix for the ONNX Shape input validation

* Handle ONNX Softmax with dynamic input and static rank

* Fix the failing Broadcast type prop test

* Code formatting

* Dont broadcast bias before adding it to the conv node

* Drop the conv node validation and rely on the core op implementation checks

* Code review feedback

* Revert the Broadcast op changes

* More code review feedback

* Dynamic conv test using ng test case

* Obsolete headers removal

* Code formatting

* Variable names refactor

* Disable model_conv_with_dynamic_batch test on GPU

* Code formatting
Co-authored-by: 's avatarSang Ik Lee <sang.ik.lee@intel.com>
      b2da4cee
  5. 14 Feb, 2020 4 commits
  7. 13 Feb, 2020 7 commits
  9. 12 Feb, 2020 11 commits
  11. 11 Feb, 2020 7 commits
  13. 10 Feb, 2020 4 commits
  15. 07 Feb, 2020 2 commits