Deconvolution: ERROR: Internal error: could not find any implementation for node conv6_2_upsample, try increasing the workspace size with IBuilder::setMaxWorkspaceSize()

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

I try to implement FeaturePyramidNetwork(FPN) with caffe Deconvolution.
When I use TensorRT3.04 to parse the caffe prototxt, there are some errors:

[GIE] building using kFLOAT
ERROR: Internal error: could not find any implementation for node conv6_2_upsample, try increasing the workspace size with IBuilder::setMaxWorkspaceSize()

layer {
  name: "conv6_2_upsample"
  type: "Deconvolution"
  bottom: "conv6_2"
  top: "conv6_2_upsample"
  convolution_param {
    num_output: 64
    bias_term: false
    pad: 1
    kernel_size: 4
    group: 64
    stride: 2
  }
}

I have tried to increase the workspace (16 << 24), but no any help.

sorry, my mistake.

just follow the error instruction: increase the workspace

builder->setMaxBatchSize(maxBatchSize);
	builder->setMaxWorkspaceSize(16<<24);
	builder->setDebugSync(mEnableDebug);
	builder->setMinFindIterations(3);// allow time for TX1 GPU to spin up
	builder->setAverageFindIterations(2);

	std::size_t max_work_space_size = builder->getMaxWorkspaceSize();
	printf("MaxWorkspaceSize: %ld\n", max_work_space_size);

MaxWorkspaceSize: 268435456
[GIE] building using kFLOAT

1 Like
2

Thanks!