ReduceMin is created even though only Min is used

[nicklasb]

Problem description:
Paddle2ONNX creates ReduceMin ops (which currently aren't supported by the destination platform) .
This even though only Min is used by the Paddle model as verified in Netron.
(Avoided ReduceMin by running PaddleDetection PIR mode).

• Architecture: PicoDet

  • backbone: LCNet
  • neck: LCPAN
  • head: PicoHeadV2

• act: set to Relu6 everywhere*

* Note that I had to patch PaddleDetection for it to stick to Relu6 for ESP-DL compatibility reasons.

More information:

• Inference engine for deployment:
  ESP-DL

• Why it needs to be converted to ONNX format:
  See above.

• Paddle2ONNX version:
  paddle2onnx-1.3.1 with python>=3.8, paddlepaddle>=2.0.0