Episode 5 for my VIPrior challenge journey link!
Issue 1. change scheduler
Using this
mmclassification doc. link
Error
Error Message
../aten/src/ATen/native/cuda/Loss.cu:115: operator(): block: [176,0,0], thread: [108,0,0] Assertion `input_val >= zero && input_val <= one` failed.
Debug
Seems like a loss scale issue due to lr, so I re-enabled auto_lr.
Reference: https://discuss.pytorch.org/t/assertion-input-val-zero-input-val-one-failed/107554/3
FYI: There is a feature to receive notifications via email or Slack in case of crashes or completion of training.
Issue 2. Ensemble
I investigated and summarized it on my blog: Link, kor
The tool provided there takes image inputs, but there is no need to perform inference several times. Therefore, I will receive JSON files and code for ensemble.
Requirements
- Reference Code: https://github.com/ZFTurbo/Weighted-Boxes-Fusion
- output1: /raid/sghong/Detection/mmdetection/work_dirs/yoloxx_0826/inferences/bbox.json
- output2: /nas4/viprior/sbhong/yoloxx/mmcls/submission.json
Situation
- need to provide only bbox, category id, and score to Weighted-Boxes-Fusion.
- Perform ensemble for each image and save the results (the order should not change due to the submission format).
- Consensus will be added, similar to preprocessing as described in my blog.
Keep using my Docker environment used for running the detection. The only thing I need for these ensemble is:
pip install ensemble-boxes.
Implementation
- Using the ensemble-boxes module
- Take multiple JSON inference results as input -> perform ensemble -> then write the results as JSON.
- Follow the 2022 VIPriors Object Detection Challenge submission format.
- Consensus or other preprocessing(?) will be implemented tomorrow....
My code link: https://github.com/sghong977/Daily_AIML/blob/fc4eede51557f6870e70939c54f9db855a3f6741/Object_Detection/ensemble.py