Forums - snpe-onnx-to-dlc: Wrong allowed_data_types

Hello,

1. Problem

I'm using snpe-onnx-to-dlc from snpe-1.61.0.3358 to convert my onnx model to dlc using some UDOs.

I find out that if I specify the data-type for of input tensors for each core for my UDOs, the data type verification is not correct.

This is the Traceback that I receive

Traceback (most recent call last):

  File "/workspace/qualcomm-base/snpe-1.61.0.3358/lib/python/qti/aisw/converters/onnx/onnx_to_ir.py", line 223, in convert

    self.graph)

  File "/workspace/qualcomm-base/snpe-1.61.0.3358/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 "/workspace/qualcomm-base/snpe-1.61.0.3358/lib/python/qti/aisw/converters/common/converter_ir/translation.py", line 17, in apply_method

    return self.indexed_methods[method_name](*args, **kwargs)

  File "/workspace/qualcomm-base/snpe-1.61.0.3358/lib/python/qti/aisw/converters/common/converter_ir/translation.py", line 120, in add_op

    op = self.extract_parameters(src_op, graph)

  File "/workspace/qualcomm-base/snpe-1.61.0.3358/lib/python/qti/aisw/converters/onnx/custom_op_translations.py", line 44, in extract_parameters

    param.data_type = get_internal_dtype(param.data, param)

  File "/workspace/qualcomm-base/snpe-1.61.0.3358/lib/python/qti/aisw/converters/backend/custom_ops/helpers/udo_module_helpers.py", line 185, in get_internal_dtype

    .format(src_type, op_attr.name, op_attr.allowed_data_types))

TypeError: The provided datatype: uint16 is not a valid datatype defined for: 672. Expected one of ['SNPE_UDO_CORETYPE_CPU', 'SNPE_UDO_CORETYPE_DSP']

2. Finding the cause

Below is parameters of the operation and the UDO in my UDO configuration. 

Parameters: (i can get these values by printing variable param in method extract_parameters in snpe-1.61.0.3358/lib/python/qti/aisw/converters/onnx/custom_op_translations.py

{'name': '672', 'allowed_data_types': ['SNPE_UDO_CORETYPE_CPU', 'SNPE_UDO_CORETYPE_DSP'], 'shape': '', 'default_value': None, 'layout': 'SNPE_UDO_LAYOUT_LAST', 'dimensions': [], 'repeated': False, 'static': True, 'rank': 1, 'data_type': 'SNPE_UDO_DATATYPE_UINT_16', 'data': None, 'param_type': 'SNPE_UDO_PARAMTYPE_TENSOR', 'per_core_data_types': None}

UDO:

If we look at the parameters at allowed_data_types': ['SNPE_UDO_CORETYPE_CPU', 'SNPE_UDO_CORETYPE_DSP'], the value of allowed_data_types does not seem correct. 

I traced back to the staticmethod create_per_core_tensor_infos in snpe-1.61.0.3358/lib/python/qti/aisw/converters/backend/custom_ops/core.py I found that the way snpe gets the allowed dtype of tensor is not correct:

- At line 77 of that file, the value of per_core_dict is {'SNPE_UDO_CORETYPE_CPU': 'SNPE_UDO_DATATYPE_FLOAT_32', 'SNPE_UDO_CORETYPE_DSP': 'SNPE_UDO_DATATYPE_UINT_8'}. And just after that, snpe gets tensor_info.allowed_data_types by getting it keys instead of its values. That's why we get ['SNPE_UDO_CORETYPE_CPU', 'SNPE_UDO_CORETYPE_DSP'] as the dtype. That causes the whole problem of dtype verification after.

3. Fix Suggestion

Change line 77 of snpe-1.61.0.3358/lib/python/qti/aisw/converters/backend/custom_ops/core.py from: 

tensor_info.allowed_data_types = list(per_core_dict.keys())

to 

tensor_info.allowed_data_types = [per_core_dict[k] for k in per_core_dict]