Can not convert yolact model to TensorRT

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

Hi,I’m trying to convert yolact model from https://github.com/Ma-Dan/yolact/tree/onnx this repository and facing number of issues.

Here is my working environment :

  1. pytorch == 1.4.0
  2. TensorRT == 7.0.0(official docker release 20.01)

There are links to yolact model in onnx format
https://drive.google.com/open?id=16kEAPQGhmoI_C5KIV5jJ8mgfU5YVJE8n(opset version == 9),

The Error is :
UNSUPPORTED_NODE: Assertion failed: scales_input.is_weights()

https://drive.google.com/open?id=1wvZzKTG6Sjq3EvfCyGDig7tQU_2HqnJY (opset version == 11).
The Error is :
INVALID_GRAPH: Assertion failed: ctx->tensors().count(inputName)

Thank you in advance for your help,
Alexander

2

Hi,

If you build the TensorRT OSS components on top of the 20.01 docker container, you’ll get this error instead which is the real issue:

While parsing node number 410 [Resize]:
ERROR: /mnt/TensorRT/parsers/onnx/builtin_op_importers.cpp:2452 In function importResize:
[8] Assertion failed: (transformationMode == "asymmetric") && "This version of TensorRT only supports asymmetric resize!"
[03/03/2020-06:43:51] [E] Failed to parse onnx file
[03/03/2020-06:43:51] [E] Parsing model failed
[03/03/2020-06:43:51] [E] Engine creation failed
[03/03/2020-06:43:51] [E] Engine set up failed
&&&& FAILED TensorRT.trtexec # trtexec --explicitBatch --onnx=yolact_opset11.onnx

There is a similar post here: (Upsample) How can I use onnx parser with opset 11 ? · Issue #284 · NVIDIA/TensorRT · GitHub

Unfortunately, this type of padding/resize isn’t supported at the moment.

3

Thanks for prompt reply, what kind of padding/resize supporting in TensorRT ? I’ll adjust my model accordingly.

4

Hi,

You can probably refer to this code snippet for guidance: https://github.com/onnx/onnx-tensorrt/blob/c536d726f309376faa407433d0bf5cef75e89ce1/builtin_op_importers.cpp#L2445-L2458

I’m not sure how this maps exactly from framework → ONNX. But these statements are comparing against the ops created in the ONNX graph.

This might also help: Layers — NVIDIA TensorRT Standard Python API Documentation 8.4.3 documentation