How to generate an ONNX file for Diffusion Planner

[JokerWhy233]

in autoware universe

diffusion_planner

• name: Create diffusion_planner directory inside {{ data_dir }}
  ansible.builtin.file:
  path: "{{ data_dir }}/diffusion_planner"
  mode: "755"
  state: directory

• name: Download diffusion_planner/diffusion_planner.onnx
  become: true
  ansible.builtin.get_url:
  url: https://awf.ml.dev.web.auto/planning/models/diffusion_planner/v0.2/diffusion_planner.onnx
  dest: "{{ data_dir }}/diffusion_planner/diffusion_planner.onnx"
  mode: "644"
  checksum: sha256:f8c5734499db27e745a28695b4d85d1425c4668ba5fb125d436d3bcf1cc1f241

How was this model generated ? we generated onnx by [email protected]:tier4/Diffusion-Planner.git /tier4-main , but faild

[SakodaShintaro]

Sorry, I forgot which commit corresponds to the v0.X.

tier/Diffusion-Planner	autoware_universe	comment
?	v0.X
tier4/main	v1.X	I'm not sure if these versions are properly matched.
feat/turn_indicators	v2.X	WIP https://github.com/tier4/autoware_universe/tree/feat/diffusion_planner/turn_indicators

[SakodaShintaro]

This is my environment.

pip_list.txt

$ nvidia-smi -L | head -n 1
GPU 0: NVIDIA H100 80GB HBM3 (UUID: GPU-fc569cc5-ae49-e998-d4c4-d227aa04e800)

[JokerWhy233]

Thanks!!

in https://github.com/autowarefoundation/autoware/blob/main/ansible/roles/artifacts/tasks/main.yaml,

we found

• name: Download diffusion_planner/v1.0/diffusion_planner.onnx
  become: true
  ansible.builtin.get_url:
  url: https://awf.ml.dev.web.auto/planning/models/diffusion_planner/v1.0/diffusion_planner.onnx
  dest: "{{ data_dir }}/diffusion_planner/v1.0/diffusion_planner.onnx"
  mode: "644"
  checksum: sha256:35094c876426655ccd384f1046f84a2bf0ac32a935335c6b659c1364870dc928

• name: Download diffusion_planner/v0.1/diffusion_planner.onnx
  become: true
  ansible.builtin.get_url:
  url: https://awf.ml.dev.web.auto/planning/models/diffusion_planner/v0.1/diffusion_planner.onnx
  dest: "{{ data_dir }}/diffusion_planner/v0.1/diffusion_planner.onnx"
  mode: "644"
  checksum: sha256:7028f68639abc9a3fb96f2ef73cb3f7795e0471a9a1e88201527331168abe8c1

May i ask , in your table,v2.X is for which branch , or commit ?
feat/turn_indicators v2.X WIP https://github.com/tier4/autoware_universe/tree/feat/diffusion_planner/turn_indicators

[SakodaShintaro]

v2 is currently under trial and error. v1 is the latest published weight.

The branches are exactly as linked.

[mcudsm]

THANKS！！
When i set the parameter "dynamo=True" of "torch.onnx.export" function，I can't generate onnx.
But if i set the parameter "dynamo=True" of "torch.onnx.export" function , the generted onnx can't be converted .engine file.
Can you tell me how to solve this problem？

[SakodaShintaro]

Sorry, I never changed the parameter "dynamo=True", so I don't know about that.

[JokerWhy233]

when we use tier4-main branch to generate onnx and diffusion planner .param.json , we found some different with

• name: Download diffusion_planner/v1.0/diffusion_planner.onnx
  become: true
  ansible.builtin.get_url:
  url: https://awf.ml.dev.web.auto/planning/models/diffusion_planner/v1.0/diffusion_planner.onnx
  dest: "{{ data_dir }}/diffusion_planner/v1.0/diffusion_planner.onnx"
  mode: "644"
  checksum: sha256:35094c876426655ccd384f1046f84a2bf0ac32a935335c6b659c1364870dc928

• name: Download diffusion_planner/v1.0/diffusion_planner.param.json
  become: true
  ansible.builtin.get_url:
  url: https://awf.ml.dev.web.auto/planning/models/diffusion_planner/v1.0/diffusion_planner.param.json
  dest: "{{ data_dir }}/diffusion_planner/v1.0/diffusion_planner.param.json"
  mode: "644"
  checksum: sha256:215493daac110055990318b7916fbf1a7d6b08feea9c27cdd6403df31557d992

---

in tier4-main branch , have
"polygon_num": 10,
"polygon_len": 40,
"line_string_num": 10,
"line_string_len": 20,

but these param isnt in autoware universe /v1.0/diffusion_planner.param.json ,

---

i have no idea that which commit or branch can generate url: https://awf.ml.dev.web.auto/planning/models/diffusion_planner/v1.0/ diffusion_planner.param.json

[SakodaShintaro]

It might work if you revert the following pull request:

tier4/Diffusion-Planner#8

[JokerWhy233]

hi , in [email protected]:tier4/Diffusion-Planner.git , In the tier4-main branch, the training model uses a dataset converted from rosbags. What is the scale of this dataset?

Is it necessary to use the original diffusion_planner model for knowledge distillation?

[SakodaShintaro]

Although we are not yet satisfied with the performance, our current efforts:

(1) We are training the model from scratch.

(2) In addition to real data, we are also partially using the results of running standard Autoware on a planning simulator as training data.

[SakodaShintaro]

We have updated diffusion_planner to v2

• feat(diffusion_planner): diffusion_planner v2 autoware_universe#11690
• https://github.com/tier4/Diffusion-Planner