Optimized serving endpoint for classical ML models #3377
Conversation
mzegla
added
the
GSoC
Co-authored-by: Miłosz Żeglarski <[email protected]>
Co-authored-by: Miłosz Żeglarski <[email protected]>
| @@ -0,0 +1,2 @@ | |||
| SepalLength,SepalWidth,PetalLength,PetalWidth | |||
There was a problem hiding this comment.
I guess we don't need that in the repo since we have an instruction on preparing the data.
| @@ -0,0 +1,508 @@ | |||
| SepalLength,SepalWidth,PetalLength,PetalWidth,Species | |||
There was a problem hiding this comment.
I guess we don't need that in the repo since we have an instruction on preparing the data.
|
|
||
| output_dir = sys.argv[1] | ||
|
|
||
| #os.makedirs(output_dir, exist_ok=True) |
There was a problem hiding this comment.
Pull Request Overview
This PR introduces a complete machine learning pipeline project for the Iris dataset using OpenVINO Model Server (OVMS) with MediaPipe. The project supports both logistic regression and k-means clustering models with training and inference capabilities through client scripts.
Key changes:
- Complete pipeline implementation with PyTorch logistic regression and scikit-learn k-means models
- Client scripts for training and inference with flexible parameter handling
- Docker-based deployment with Intel optimization libraries
- Data preprocessing utilities and example datasets
Reviewed Changes
Copilot reviewed 15 out of 15 changed files in this pull request and generated 5 comments.
Show a summary per file
| File | Description |
|---|---|
| extras/iris_pipeline_project/pipeline/ovmsmodel.py | Core OVMS Python model implementing training/inference logic for both model types |
| extras/iris_pipeline_project/pipeline/model.py | Abstract model classes with PyTorch logistic regression and scikit-learn k-means implementations |
| extras/iris_pipeline_project/pipeline/graph.pbtxt | MediaPipe graph configuration for the pipeline |
| extras/iris_pipeline_project/model_config.json | OVMS model configuration |
| extras/iris_pipeline_project/labelmap.json | Label mapping for Iris species |
| extras/iris_pipeline_project/kmeans_param.json | K-means hyperparameters |
| extras/iris_pipeline_project/hyperparams.json | Logistic regression hyperparameters |
| extras/iris_pipeline_project/data_preprocess.py | Data preprocessing script for Iris dataset |
| extras/iris_pipeline_project/data_folder/iris_train.csv | Training dataset |
| extras/iris_pipeline_project/data_folder/iris_test.csv | Test dataset |
| extras/iris_pipeline_project/client/client_train.py | Client script for training models |
| extras/iris_pipeline_project/client/client_inference.py | Client script for inference |
| extras/iris_pipeline_project/README.md | Project documentation |
| extras/iris_pipeline_project/Dockerfile | Docker configuration with Intel optimizations |
| extras/iris_pipeline_project/.gitignore | Git ignore file |
Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.
| import sys | ||
|
|
||
| if len(sys.argv) < 2: | ||
| print("Usage: python datapreprocess.py <output_directory>") |
There was a problem hiding this comment.
The filename in the usage message should match the actual filename: 'data_preprocess.py' instead of 'datapreprocess.py'.
| print("Usage: python datapreprocess.py <output_directory>") | |
| print("Usage: python data_preprocess.py <output_directory>") |
Copilot uses AI. Check for mistakes.
Co-authored-by: Copilot <[email protected]>
Co-authored-by: Copilot <[email protected]>
Co-authored-by: Copilot <[email protected]>
Co-authored-by: Copilot <[email protected]>
darksapien23151
changed the title
| For Enabling accelerator support: | ||
| Manually set the ```bool - (use_ipex/use_oneDAL)``` in model.py file under "pipeline" directory to either True/False depending on the necessity. |
There was a problem hiding this comment.
We needed that for benchmarks.In the guide we should not require user to modify the code.
Either remove it completely and use acceleration by default or pass it in hyperparameters JSON.
|
|
||
| model_obj = AVAILABLE_MODEL_CLASSES[model_class_name]() | ||
|
|
||
| if model_class_name == "KMeansSkLearn": |
There was a problem hiding this comment.
I don't like this conditional logic. execute method should be generic and wrap all possible implementations. Having any model specific code goes against this idea. Please move those implementation specific code blocks to its classes.
You define the interface and convention. You can make fit or infer method return not only results, but also metrics that can be used directly to create pyovms.Tensor object like:
trained_model, training_metrics = model_obj.fit(X, y, params))
...
return [Tensor("pipeline_output", training_metrics)]
🛠 Summary
GSOC Contribution for the project - Optimized serving endpoint for classical Machine Learning models.
Current status: Finished working on Logistic Regression and KMeans.
Mentors:
Zeglarski, Milosz
Trawinski, Dariusz