Plugin InstanceNormalization failure of TensorRT 8.6 when running InstanceNorm on GPU V100

[DataXujing]

A bug? I use the Plugin of InstanceNorm plugin, when use trtexec to get the engine, Error show like:

[07/27/2023-06:18:54] [I] [TRT] No importer registered for op: InstanceNormalization_TRT. Attempting to import as plugin.
[07/27/2023-06:18:54] [I] [TRT] Searching for plugin: InstanceNormalization_TRT, plugin_version: 1, plugin_namespace:
[07/27/2023-06:18:54] [V] [TRT] Local registry did not find InstanceNormalization_TRT creator. Will try parent registry if enabled.
[07/27/2023-06:18:54] [V] [TRT] Global registry found InstanceNormalization_TRT creator.
[07/27/2023-06:18:54] [W] [TRT] builtin_op_importers.cpp:5221: Attribute scales not found in plugin node! Ensure that the plugin creator has a default value defined or the engine may fail to build.
[07/27/2023-06:18:54] [F] [TRT] Validation failed: scale.count == bias.count
plugin/instanceNormalizationPlugin/instanceNormalizationPlugin.cu:96

[07/27/2023-06:18:54] [E] [TRT] std::exception
[07/27/2023-06:18:54] [E] [TRT] ModelImporter.cpp:771: While parsing node number 3415 [InstanceNormalization_TRT -> "InstanceNormV-27"]:
[07/27/2023-06:18:54] [E] [TRT] ModelImporter.cpp:772: --- Begin node ---
[07/27/2023-06:18:54] [E] [TRT] ModelImporter.cpp:773: input: "/unet/input_blocks.1/input_blocks.1.0/in_layers/in_layers.0/Reshape_output_0"
output: "InstanceNormV-27"
name: "InstanceNormN-27"
op_type: "InstanceNormalization_TRT"
attribute {
  name: "epsilon"
  f: 1e-05
  type: FLOAT
}
attribute {
  name: "scale"
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  floats: 1
  type: FLOATS
}
attribute {
  name: "bias"
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  floats: 0
  type: FLOATS
}
attribute {
  name: "relu"
  i: 0
  type: INT
}
attribute {
  name: "alpha"
  f: 0
  type: FLOAT
}
attribute {
  name: "plugin_version"
  s: "1"
  type: STRING
}

[07/27/2023-06:18:54] [E] [TRT] ModelImporter.cpp:774: --- End node ---
[07/27/2023-06:18:54] [E] [TRT] ModelImporter.cpp:777: ERROR: builtin_op_importers.cpp:5412 In function importFallbackPluginImporter:
[8] Assertion failed: plugin && "Could not create plugin"
[07/27/2023-06:18:54] [E] Failed to parse onnx file
[07/27/2023-06:18:54] [I] Finished parsing network model. Parse time: 7.78809
[07/27/2023-06:18:54] [E] Parsing model failed
[07/27/2023-06:18:54] [E] Failed to create engine from model or file.
[07/27/2023-06:18:54] [E] Engine set up failed
&&&& FAILED TensorRT.trtexec [TensorRT v8601] # trtexec --onnx=./combine_0.onnx --saveEngine=combine_1.plan --verbose --workspace=3000 --fp16

I give five attributes to this plugin: epsilon,scale, bias, relu and alpha like the readme: https://github.com/NVIDIA/TensorRT/blob/release/8.6/plugin/instanceNormalizationPlugin/README.md#parameters ,

Therefore, I think there is a problem with our readme parameter. Should the scale parameter be replaced with scales?

[zerollzeng]

@samurdhikaru ^ ^

[ttyio]

@DataXujing I see this from the log attached , could you check your model, thanks!

Validation failed: scale.count == bias.count
plugin/instanceNormalizationPlugin/instanceNormalizationPlugin.cu:96

[DataXujing]

@DataXujing I see this from the log attached , could you check your model, thanks!

Validation failed: scale.count == bias.count
plugin/instanceNormalizationPlugin/instanceNormalizationPlugin.cu:96

I have reconfirmed that my model is correct, and I believe there is a bug in this line of plugin code:
https://github.com/NVIDIA/TensorRT/blob/35477bdb94eab72862ffbdf66d4419e408bef45f/plugin/instanceNormalizationPlugin/instanceNormalizationPlugin.cu#L621C1-L621C98

in line 621:

 mPluginAttributes.emplace_back(PluginField("scales", nullptr, PluginFieldType::kFLOAT32, 1));

should the attribute "scales" be written as "scale" ?