Keras/tensorflow/onnx weight deploy to deepstream

I started work with deepstream and trying to convert weight file from keras to uff format. converted successfully but not working with deepstream docker 4.0
Is there any document that will indicate the correct version to use all the tools.
Tried the below path

  1. keras . >> frozen graph >> uff
  2. keras >> fronzen graph >> onnx
    Tried this path for resnet18 classification network
    uff conversion getting : DEBUG: convert reshape to flatten node
    and deepstream giving error for the same file: log(): UffParser: Parser error: bn_conv1/Reshape_7: You cannot reshape with a different volume

Just wanted the correct way to deploy models to deepstream.

Hi,

For Deepstream 4.0.2, it’s recommended to use TensorRT 6.0.1, TensorFlow 1.1x (TensorFlow backend) and ONNX opset-7 (PyTorch backend).

Then you also need to check if all the layers you used are supported by the TensorRT before the conversion:
https://docs.nvidia.com/deeplearning/sdk/tensorrt-archived/tensorrt-601/tensorrt-support-matrix/index.html

Thanks.

Thanks for reply . But I never found version details like tensorflow 1.1 with deepstream 4.0.2 anywhere. Good to know these things.

Hi,

The recommended library for Deepstream can be found here:

https://docs.nvidia.com/metropolis/deepstream/dev-guide/index.html

Thanks.

Hi AastaLLL,

I read this document, carefully but not found details that which version of tensorflow/keras/onnx model is compatible with deepstream 4.0.2.

Hi,

Sorry that the information is layered

You can find that it’s recommended to use TensorRT 6.0.1 for Deepstram 4.0.2.
https://docs.nvidia.com/metropolis/deepstream/dev-guide/index.html

Then in TensorRT 6.0.1 doc, you can find the supported version of each frameworks:
https://docs.nvidia.com/deeplearning/sdk/tensorrt-archived/tensorrt-601/tensorrt-release-notes/tensorrt-6.html#rel_6-0-1

Compatibility

Thanks.