mdanilow
mdanilow
"strings" attribute support added to CustomOp and tested
- Feature input shapes of external weights layers are saved to io_shape_dict driver dictionary under "external_weights_input_shapes" key. - These shapes are then used to recognize convolutional layers and send the...
A quick fix for the InsertDWC transform, the current implementation assumes all of the outputs of the producer and inputs of the consumer to have the same folded shapes, stream...
Different input datatypes were already supported by the HLS backend from the finn-hlslib. This case was added to the unit test. The unit test now also tests the InferAddStreamsLayer transform.
This PR depends on a PR from the finn-hlslib where the Concat operation was implemented (https://github.com/Xilinx/finn-hlslib/pull/143) - The Concat op code is not hardcoded in the compiler anymore and it...
This PR depends on a PR from the finn-hlslib where the Concat operation was implemented (https://github.com/Xilinx/finn-hlslib/pull/143) - The new Split custom op is added. - HLS implementation is defined for...
New transformations in finn/transformation/streamline/reorder.py needed for yolov8 streamlining, some refactoring and generalization of existing transformations included. - The MoveLinearPastEltwiseAdd() is now generalized by a refactored MoveIdenticalOpPastJoinOp and was removed -...
This version accepts rectangular inputs, however for some upsampling scales (3 for example) the output is not as expected. Works for scale of 2 used in the yolov8.
Faster implementation of finn.util.data_packing.packed_bytearray_to_finnpy. It utilizes bitshifts and is necessary for realtime processing on pynq.