Completely transition of MO to Interpolate-4 (#970)
* Refactored infer function and function supported_attrs for the layer Interpolate. * Small change. * Deleted unneeded checks in transformations ResizeToInterpolate2D and ResizeToInterpolate3D. * Small fix in the extractor of ONNX Resize. * Now the extractor of TF ResizeBilinear generates Interpolate-1 again, because 'axis' in final version of Interpolate-4 specification is an input but is not attribute. * Now the extractor of TF ResizeNearest generates Interpolate-1 again, because 'axis' in final version of Interpolate-4 specification is an input but is not attribute. * Added static method get_axis into class Interpolate. * Refactored class CanBeFused in the transformation InterpolateSequenceToInterpolate. * Fixed transformation InterpolateSequenceToInterpolate according to the last version of the specification of Interpolate-4. * Started to write support of Interpolate-4 in the transformation InterpolateWithConcat. * Added support for Interpolate-4 into the transformation InterpolateWithConcat. * Added support for Interpolate-4 into the transformation InterpolateConcat. * Added support for Interpolate-4 into the transformation InterpolateReshapeWA. * Added support for Interpolate-4 into the transformation InterpolateTranspose. * Started to add test for opset4 case of the transformation InterpolateSequenceToInterpolate. * Added test for InterpolateSequenceToInterpolate (test_2d_interpolate_sequence_1_opset4_case). * Added test for InterpolateSequenceToInterpolate (test_2d_interpolate_sequence_4_opset4_case). * Added another test for InterpolateSequenceToInterpolate (test_2d_interpolate_sequence_5_opset4_case). * Added another test for InterpolateSequenceToInterpolate (test_3d_interpolate_sequence_1_opset4_case). * Finished addition of tests for opset4 case of InterpolateSequenceToInterpolate. * Small change. * Now opset is only opset1 or opset4 in the transformation InterpolateTranspose. * Small fixes in transformations ResizeToInterpolate2D and ResizeToInterpolate3D. * Deleted reading of unused ONNX attributes. * Fixed docstring of the transformation InterpolateV1ToInterpolateV4. * Added node name in assert about axes input. * Fixes in the definition of the operation ONNXResize11. * Now Interpolate-4 cannot have 'extension' as opset. * Now the transformation InterpolateV1ToInterpolateV4 uses find_and_replace_pattern but not replace_sub_graph. * Fixed tests for transformations InterpolateReshapeWA and InterpolateConcat. * Fixed some tests. * Rewritten operation Interpolate-4 class according to new variant of documentation. * Some fixes in ONNXResize11 operation class. * Now the transformation ONNXResize11ToInterpolate generates Interpolate-4 with 4 inputs. * Now the transformation UpsampleToResample generates Interpolate-4 with 4 inputs. * Now the transformation NearestNeighborUpsampling generates Interpolate-4 with 4 inputs. * Now transformations ResizeToInterpolate2D and ResizeToInterpolate3D generate Interpolate-4 with 4 inputs. * Now the transformation SplitConcatPairToInterpolate generates Interpolate-4 with 4 inputs. * Now the transformation UnsqueezeTileReshapeBlockToInterpolate generates Interpolate-4 with 4 inputs. * Now the transformation InterpolateV1ToInterpolateV4 generates Interpolate-4 with 4 inputs. * Some fixes. * Fixed the transformation InterpolateSequenceToInterpolate according to new variant of Interpolate-4 specification. * Fixed typos. * Added shape_calculation_mode to supported_attrs. * Small fixes. * Added operation ONNXResize10 and the transformation ONNXResize10ToInterpolate4. * Fixed function correct_scales_using_dst_shape. * Some fixes in InterpolateSequenceToInterpolate. * Fixed bug in the method __call__ of the class CanBeFused: now self.accumulated_axes is correctly cleared in all cases. * Small change. * Fixed tests for the transformation SplitConcatPairToInterpolate. * Now transformations InterpolateWithConcat, InterpolateReshapeWA, InterpolateConcat support Interpolate-4. * Fixed the transformation InterpolateTranspose for the case of Interpolate-4. * Written the back transformation InterpolateV4AxesCorrection to convert 'axes' input of Interpolate-4 from NHWC to NCHW layout. * Added PermuteInput in Interpolate-4 infer. * Fixed typos. * Deleted the transformation InterpolateAxesCorrection. * Now Interpolate-4 permutes axis, not shape in input port 3. * Small fix. * Some fix. * Fixed bug in the transformation UpsampleToResample. * Added some debug prints. * Added more debug prints. * Now ONNX Upsample-9 operation is read as ONNXResize10. * Small fix. * Small fixes. * Fixed tests for the transformation SplitConcatPairToInterpolate. * Deleted debug prints. * Deleted some debug prints. * Fixes in the transformation UnsqueezeTileReshapeBlockToInterpolate and its tests. * Small fix in the transformation InterpolateSequenceToInterpolate. * Started to write nGraph transformation to convert Interpolate-1 to Interpolate-4. * Deleted redundant files. * Small fixes. * Small fix. * Written draft of the transformation Interpolate-1 -> Interpolate-4. * Small fix. * Now ONNX Importer reads Resize-10 as Interpolate-4. * Fixes in the test onnx_model_resize10_import_only. * Small fix in the test for the conversion Interpolate-1 -> Interpolate-4. * Small fixes. * Fixed NGraphReaderTests for Interpolate. * Some fixes. * Deleted class for Resample operation. * Fix in the transformation NearestNeighborUpsampling: fixed precision of the input 'scales' of generated Interpolate-4. * Fixed typo. * Now the TF operations ResizeBilinear is readed as internal MO operation TFResizeBilinear. This internal operation is converted into Interpolate-4. * Small fix in BOM-file. * Added checks of existence of attributes of TF ResizeBilinear operation. * Small fixes in the conversion of the internal MO operation TFResizeBilinear to Interpolate-4. * Small fixes. * Small fixes. * Now the transformation ONNXResize10ToInterpolateV4 calculates sizes input as input_shape * (scales + epsilon). * Added the internal MO operation TFResizeNearestNeighbor. * Fixes in the transformation SplitConcatPairToInterpolate and its tests. * Fixes in the transformation UnsqueezeTileReshapeBlockToInterpolate and its tests. * Written the transformation that converts the internal operation TFResizeNearestNeighbor into Interpolate-4. * Now MO reads the TF operation ResizeNearestNeighbor as the internal MO operation TFResizeNearestNeighbor. * Small fix. * Now the specification of Interpolate-4 clarifies that the mode linear_onnx supports only 2D or 4D input tensors. * Small fix. * Some fixes. * Moved the transformation ONNXResize10ToInterpolateV4 to the front stage. * Deleted infer function and function supported_attrs for ONNXResize10 operation. * Deleted supported_attrs() for TFResizeBilinear and TFResizeNearestNeighbor. * Some fixes. * Fixes in the shape infer function of the nGraph operation Interpolate-4. Now 'axes' input can be non-constant. In the such case, all elements of the output shape are Dimension::dynamic(). * Deleted corner cases processing in transformations TFResizeBilinearToInterpolateV4 and TFResizeNearestNeighborToInterpolateV4. * Rewritten the function replace_resize_bilinear. * Written inner MO operation TFResize that covers TF operations ResizeBilinear and ResizeNearestNeighbor. * Now TF operations ResizeBilinear and ResizeNearestNeighbor are read as an internal operation TFResize in MO. Transformations TFResizeNearestNeighborToInterpolateV4 and TFResizeBilinearToInterpolateV4 are fused into one transformation TFResizeToInterpolateV4. * Some changes in the shape infer function of nGraph op Interpolate-4. * Small fix. * Some changes. * The transformation TFResizeToInterpolateV4 is moved to the front stage. * Deleted redundant assert. * Deleted transformations ResizeToInterpolate2D and ResizeToInterpolate3D. * Some renaming. * Small change. * Deleted .copy() in the shape infer function of the internal operation TFResize. * Small fix. * Small fixes. * Added comment about the case when the input 'axes' of Interpolate-4 is non-constant. * Written test for Interpolate-4 shape infer, for the case when the input 'axes' is non-constant and shape_calculation_mode = scales. * Some fixes. * Small fixes. * Small fix. * Added yet another test for the case of non-constant 'axes' input of Interpolate-4 (when shape_calculation_mode = sizes). * Added some comment. * Small fix. * Reverted changes for InterpolateWithConcat. * Added type checks for all inputs of nGraph operation Interpolate-4. * Added u32 and u64 to supported element types of sizes and axes inputs of nGraph operation Interpolate-4. * Fixed some functional tests. * Some changes. * Added helper function float32_array. * Now the MO transformation InterpolateV1ToInterpolate preserves names of layers. * Small fix. * Small fix. * Reverted some change. * Small fixes. * Small fix. * Small fix. * Small fix. * Small fix. * Reverted changes in the nGraph reader tests for Interpolate-1. * Some revert. * Fixed some copyright year.
This commit is contained in:
Vladimir Gavrilov
committed by
GitHub
GitHub
parent
bd3884b602
commit
c1136cd7b0
@@ -28,6 +28,10 @@ def float_array(l: list):
|
||||
return np.array(l, dtype=np.float64)
|
||||
|
||||
|
||||
def float32_array(l: list):
|
||||
return np.array(l, dtype=np.float32)
|
||||
|
||||
|
||||
def mark_input_bins(node, names=('weights', 'biases'), start_port: int = 1):
|
||||
"""
|
||||
Preparing necessary attributes for edges at input ports starting from start_port.
|
||||
|
||||
Reference in New Issue
Block a user