Is there any ways to get input/ouput name and dimensions info of ETLT models and SerializedEngine models

[ta.sugiyama]

Please provide the following information when requesting support.

• Hardware (T4/V100/Xavier/Nano/etc) : Any
• Network Type (Detectnet_v2/Faster_rcnn/Yolo_v4/LPRnet/Mask_rcnn/Classification/etc) : Any
• TLT Version (Please run “tlt info --verbose” and share “docker_tag” here) : “nvidia/tao/tao-toolkit-tf: v3.21.11-tf1.15.5-py3”
• Training spec file(If have, please share here) : None
• How to reproduce the issue ? (This is for errors. Please share the command line and the detailed log here.) : None

---

I have various etlt-models and serialized-engine-models (which was previously converted by tao-converter).

I have to know input-name and dimensions of etlt models because tao-converter requires -p option for dynamic batch models.
I also have to know input/output name & shapes of serialized-engine-models for triton configuration.

Is there any (regular) ways to get input/ouput name and dimensions info of etlt/serialized-engine models?

trtexec --loadEngine=a.engine --exportOutput=abc.json prints some info but

• this solution can be applied only to serialized engine file
• got dimensions sometimes include batchdim and sometimes not depending on models
• this solution does not print input name and shape

[Morganh]

There is a simple way as below.

$ python -m pip install colored
$ python -m pip install polygraphy --index-url https://pypi.ngc.nvidia.com
$ polygraphy inspect model your_trt_engine

[ta.sugiyama]

Thank you very much.
I confirmed that I can get name, shape explicitBatch/ImplicitBatch infomations for serialized engine models. I would have liked to have json output if possible, but it is ok because it can be parsed.

I still have two questions.

1. Is there any ways for etlt models?

2. This is an unimportant issue for me, but the output dtype seems to be float32 for a INT8 model. Is it a expected behavior? The model was converted by trtexec -t int8 command.

    ---- 1 Engine Input(s) ----
    {input_1 [dtype=float32, shape=(3, 544, 960)]}

    ---- 2 Engine Output(s) ----
    {output_bbox/BiasAdd [dtype=float32, shape=(12, 34, 60)],
     output_cov/Sigmoid [dtype=float32, shape=(3, 34, 60)]}

[Morganh]

1), For etlt model, you can refer to tao user guide or its config file. For example,
https://github.com/NVIDIA-AI-IOT/deepstream_tao_apps/blob/master/configs/yolov4_tao/pgie_yolov4_tao_config.txt

uff-input-blob-name=Input
output-blob-names=BatchedNMS

2), It is expected. The output layer has float32 format.

[ta.sugiyama]

1), For etlt model, you can refer to tao user guide or its config file. For example,

Some models do not provide input/ouput name infomatation.
for example: https://github.com/NVIDIA-AI-IOT/deepstream_tao_apps/blob/master/configs/unet_tao/pgie_unet_tao_config.txt does not provide input name

If there is a way to know it for any etlt models, please let me know.
Input name (and shape) is needed to convert with tao-converter.

2), It is expected. The output layer has float32 format.

I got it.