while i'm converting my onnx model
it reported like :
2022-08-10 23:44:44,705 - 214 - INFO - INFO_DLC_SAVE_LOCATION: Saving model at /model_repo/498937/snpe/params Encountered Error: Unsupported Asymmetric padding values [3, 3, 2, 2] for node pnasnet0_pnasunit5_dwsbranch0_nasdwsconv0_dwsconv0_conv0 Stack Trace: Traceback (most recent call last): File "/opt/source/easyedge-snpe/bin/x86_64-linux-clang/snpe-onnx-to-dlc", line 53, in <module> backend.save(optimized_graph) File "/opt/source/easyedge-snpe/lib/python/qti/aisw/converters/backend/ir_to_dlc.py", line 151, in save DlcTranslations.apply_method_to_all_ops(BackendTranslationBase.ADD_OP_TO_BACKEND, graph, self) File "/opt/source/easyedge-snpe/lib/python/qti/aisw/converters/common/converter_ir/translation.py", line 71, in apply_method_to_all_ops self.apply_method_to_op(node.op.type, method_name, node, graph, *args, **kwargs) File "/opt/source/easyedge-snpe/lib/python/qti/aisw/converters/common/converter_ir/translation.py", line 51, in apply_method_to_op return translation.apply_method(method_name, *args, **kwargs) File "/opt/source/easyedge-snpe/lib/python/qti/aisw/converters/common/converter_ir/translation.py", line 18, in apply_method return self.indexed_methods[method_name](*args, **kwargs) File "/opt/source/easyedge-snpe/lib/python/qti/aisw/converters/backend/ir_to_dlc.py", line 271, in add_op_to_backend node.op.padding_size_strategy = adjust_padding_strategy(self, node, input_w, input_h, kernel_w, kernel_h) File "/opt/source/easyedge-snpe/lib/python/qti/aisw/converters/backend/snpe_translation_utils.py", line 61, in adjust_padding_strategy raise ValueError("Unsupported Asymmetric padding values {} for node {}".format(pads_list, node.op.name)) ValueError: Unsupported Asymmetric padding values [3, 3, 2, 2] for node pnasnet0_pnasunit5_dwsbranch0_nasdwsconv0_dwsconv0_conv0
but i think asymmetric padding between height and width is very common, especially in OCR or SEG model
Will SNPE support onnx asymmetric padding(only in height and width, not in begin and end) in the future?
Dear custuomer,
What's your convertsion and quatization commands used? Started from SM8350 also known as 888 external, SNPE has supported to symmetric default. But asymmetric aslo support on SNPE internal already.
But we're still recommending to quantize with symmetric for your training model.
BR.
Wei