There is a problem with Topk output

[lcx1874000]

Description

Environment

TensorRT Version:8.6.1

NVIDIA GPU:3080Ti

NVIDIA Driver Version:536.40

CUDA Version:12.1

CUDNN Version:6.5.0

Relevant Files

    nvinfer1::IBuilder*`` builder = nvinfer1::createInferBuilder(gLogger);
nvinfer1::IBuilderConfig* config = builder->createBuilderConfig();
nvinfer1::INetworkDefinition* network = builder->createNetworkV2(1);
nvinfer1::ITensor* input_data = network->addInput("input", nvinfer1::DataType::kFLOAT, nvinfer1::Dims4{ 1,2,1024,1024 });
int reduceAxis = 0x1;
nvinfer1::ITopKLayer* topk = network->addTopK(*input_data, nvinfer1::TopKOperation::kMIN, 1, reduceAxis);
int k = topk->getK();
assert(topk);
topk->getOutput(1)->setName("index");	
network->markOutput(*topk->getOutput(1));
topk->getOutput(1)->setType(nvinfer1::DataType::kINT32);
nvinfer1::IHostMemory* modelStream = builder->buildSerializedNetwork(*network, *config);
nvinfer1::IRuntime* runtime1 = nvinfer1::createInferRuntime(gLogger);
auto mEngine = std::shared_ptr<nvinfer1::ICudaEngine>(
	runtime1->deserializeCudaEngine(modelStream->data(), modelStream->size()), samplesCommon::InferDeleter());
//nvinfer1::ICudaEngine* engine1 = runtime1->deserializeCudaEngine(modelStream->data(),modelStream->size());
IExecutionContext* context1 = mEngine->createExecutionContext();
assert(context != nullptr);
int32_t nbBindings1 = mEngine->getNbBindings(); 
nvinfer1::Dims input_topk = mEngine->getBindingDimensions(0);
nvinfer1::Dims output_topk = mEngine->getBindingDimensions(1);

Model link:

Steps To Reproduce

[lcx1874000]

Description:
I created a Topk network layer separately to replace argmax in segmentation, but there was a problem with the output results, I entered the dimensions 1210241024, I only took the indices of the output, but the dimensions were still 1210241024, and the results were also wrong, all zeros

[zerollzeng]

Your input is float but the output is int32?

[zerollzeng]

More like a code issue to me, perhaps you can create an onnx and use our onnx parser.