kubeflow-custom-tinyYOLOV3

Kubeflow training a tiny YoloV3 model with custom dataset. In this example code, it is for PCB data. LakeFS, MLFlow are utilized in this repo.

Requirements

• Python 3.8
• kfp library. Install by: pip install kfp

Custom parameters

There are 12 parameters in total that need to be configured before starting a run. Some of them have no default values, some have. However, we still can change it if needed.

1. with_gpu | int
   • meaning: train with GPUs or not
   • values: 0 or 1
   • default: 0 (no GPU)
2. print_training | int
   • meaning: print the training output or not (iter, loss, avg loss, etc.,)
   • values: 0 or 1
   • default: 1 (print)
3. n_iters | int
   • meaning: number of training iteration. After this iter, the training process will stop/done
   • default: 500
4. n_clasess | int
   • meaning: number of classes in the dataset
   • default: 13
5. batch_size | int
   • meaning: number of images used in one iteration
   • default: 8
6. subdiv | int
   • meaning: number of mini-batches. Batch will be subdivided into this many smaller batches. If the GPU/RAM strong and has enough memory, we can set this value small (eg. 1, 2, 4)
   • default: 4
7. lake_host | str <Need to fill this field>
   • meaning: host (address) of the LakeFS
   • default: ''
8. lake_user | str <Need to fill this field>
   • meaning: registered account's user on LakeFS system
   • default: ''
9. lake_pwd | str <Need to fill this field>
   • meaning: password of registered account on LakeFS system
   • default: ''
10. lake_repo | str <Need to fill this field>
    • meaning: the repository that store the data
    • default: ''
11. mlflow_server | str <Need to fill this field>
    • meaning: the tracking server of MLFlow. Should be something like this: http://x.x.x.x:8001
    • default: ''
12. threshold | float
    • meaning: model will not return predictions that have confidence scores lower than this threshold
    • default: 0.6 (60%)

Usage

1. Generate the pipeline.yaml

   python3 pipeline.py
   

2. Access kubeflow dashboard and login with registered account (email, password)

   

3. After login, click on Pipeline on the left panel to show up all the examples or created pipelines

   

4. Click + Upload pipeline button near the top right corner

   

5. Fill in the mandatory fields (name, description), upload the pipeline.yaml file then click Create

   

6. After creating the pipeline, on the left panel, click on Experiments (KFP) to create an experiment. This will used for running the pipeline later on.

7. Click + Create experiment on the top of the window, fill the Experiment name and click next.

   

8. Go back to the Pipelines list by clicking on the Pipelines tab on the left panel

   

   Click on the pipeline that we've just created and it will jump to the summary view. In here Click the + Create run button from top right corner.

   

9. In here, just choose the experiment we've just created and change the parameters at the end of the form (if needed) before clicking start.

   

---

DONE!! NOW JUST WAIT FOR IT TO RUN ....OR TO DUMP AN ERROR :)

---