search

GPUOpen-LibrariesAndSDKs / RadeonML

https://gpuopen.com/radeon-prorender-suite/
Other
83 stars 11 forks source link

rmlLoadModel ERROR: Unsupported attribute type: 4 #6

Open Armin234 opened 4 years ago

Armin234 commented 4 years ago

Hi, with the MaskRnn model from https://github.com/onnx/models/tree/master/vision/object_detection_segmentation/mask-rcnn rmlLoadModel returns the error "Unsupported attribute type: 4" With my own model I get the same error.

What must I change In in my model ?

JohnBravo08am commented 4 years ago

Hello, @Armin234 , unfortunately we don't support several operations for MaskRnn model right now (Gather, ConstantOfShape, NonMaxSuppression and logical operations). You can chek list of supported operations in RadeonML_graph.h file.

Armin234 commented 4 years ago

Hi John, it would be top, if the error reports the name of the unsupported opration. I now checked the operations of my own model in the RadeonML_graph.h and it seems that all operations are supported. Only with the operator sum I'm not sure .... armin_2_1 onnx

JohnBravo08am commented 4 years ago

@Armin234 , we log operation type if it's not supported, however, if any error occurs before parsing unsupported operation, we indicate this as a key error. "Unsupported attribute type: 4" means that some attribute of onnx node has type AttributeProto_AttributeType_TENSOR, which we don't support. Supported attribute types are:

AttributeProto_AttributeType_FLOAT
AttributeProto_AttributeType_FLOATS
AttributeProto_AttributeType_INT
AttributeProto_AttributeType_INTS
AttributeProto_AttributeType_STRING
AttributeProto_AttributeType_STRINGS
Armin234 commented 4 years ago

Hi John, I' now checked my model and I can't find no tensor attribute. Without information about the layer in the error message, the error message doesn't help much.

https://we.tl/t-3X6IAQCwgq

JohnBravo08am commented 4 years ago

Hi, Armin, thanks for the attached model, it seems like the Reshape node has such type. More specifically, it's shape constant 8InputDummy_shape has int64[4] type, which is actually a tensor, not an array. I agree on poor logging and we'll improve it in the next update.

Armin234 commented 4 years ago

Hi @JohnBravo08am, it would be fine if you can solve this problem ... this problem doesn't permitt me to test your RadoenML with our application which currently uses the onnxruntime

JohnBravo08am commented 4 years ago

@Armin234 , we'll look into this issue and will try to provide support for that.

achernigin1987 commented 4 years ago

Fixed in the 0.9.7