NGC, pretrained models. Converting pre-trained hdf5 models to caffemodels

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

**• GPU **
**• DeepStream 5.0 **
**• TensorRT 7 **
**• NVIDIA GPU Driver Version 450.36.06 **
Hi,

From the transfer learning toolkit documentations I found out that the pre-trained weights in hdf5 format can be used for training the models so that models need not be trained from scratch. I would like to convert the hdf5 weights to caffemodels so that I can create the caffemodel engine using TensorRT and use it for nvinfer plugins. But I cannot find the “prototxt” for the hdf5 files which is required to convert it to caffemodel. Where can I find the protoxt for those files? I am testing with resnet101.hdf5.

Thank you.

3

Move this topic into TLT forum because it is related to TLT.

1 Like
4

Hi v.snippy,
The hdf5 file cannot be converted to caffemodel. It can be used as a starting point while training via Transfer Learning Toolkit (TLT). Then user can train with their own data to get an etlt model, and then generate a TRT engine.

TLT also provides several Purpose-built models (Integrating TAO Models into DeepStream — TAO Toolkit 3.22.05 documentation)

You can find pruned etlt files for each kind of Purpose-built models. Then you can generate trt engine via the tool “tlt-converter”. See more in Integrating TAO Models into DeepStream — TAO Toolkit 3.22.05 documentation