TensorRT 3.0 RC uff-converter-tf tool doesn't support ExpandDims, Square, SplitV, RealDiv, etc.

We have some TensorFlow model pb file, and want to accelerate it inference speed, so we found the uff-converter-tf tool in TensorRT 3.0 RC SDK.

uff-converter-tf generate too many customer node, and all of them is come from TensorFlow frozen pb file. We event update the converter.py file to make it go further, but we still stuck.

def convert_tf2uff_field(cls, code, val):
        if isinstance(val, AttrValue):
            val = getattr(val, code)
        if code == 'i':
            if hasattr(val, 'i'):
                return int(val.i)
            else:
                return int (val)
        elif code == 'f':
            return float(val)
        elif code == 's':
            return str(val)
        elif code == 'b':
            return bool(val)
        elif code == 'type':
            return TensorFlowToUFFConverter.convert_tf2numpy_dtype(val.type)
        elif code == 'list':
            fields = val.list.ListFields()
            if len(fields) == 0:
                return None
            elif len(fields) > 1:
                raise ValueError("Invalid list field")
            else:
                field_desc, field_value = fields[0]
                sub_code = field_desc.name
                uff_code = {'i': 'i', 'f': 'f', 's': 's', 'b': 'b',
                            'type': 'dtype', 'list': 'dim_orders'}[code]
                return uff.List(uff_code, [{1:cls.convert_tf2uff_field(sub_code, v) for v in field_value}])
        elif code == 'shape':
            shp = val.dim
            if hasattr(shp, 'unknown_rank') and shp.unknown_rank:
                raise ValueError(
                    "Unsupported: shape attribute with unknown rank")
            return uff.List('i', [dim.size for dim in shp])
        else:
            print(val)
            raise TypeError("Unsupported field type:" + code)

we stuck at:split_3 output to Mul[0-63], but the tool execute

raise UFFException('.../split_3:63' input doesn't exist)

the Mul[63] input should connect split_3 the 63 indexed output, what we should do to make pb2uff happen?

Can replace ExpandDims, Square node with reshape layer?

Hi,

Guess that there is some non-supported layer in your model.
Could you check our document if all the layers are valid for TensorRT 3?
/usr/share/doc/tensorrt/TensorRT-3-User-Guide.pdf.gz

2.3.2.2.4. Supported TensorFlow Operations

It will be very tricky to parse a non-supported layer from TensorFlow to UFF format.

Thanks.

Hmm, we checked the supported native layer with our team leader, it seems possible to replace ExpandDims, Square node with Reshape, but we haven’t done it. Once we checked, we will make a note here.

Hi, were you able to create an alternate for RealDiv?

+1