Batch Ml Training

A machine learning platform needs to run training jobs in batches: prepare the dataset, configure hyperparameters, execute the training run, evaluate the model, and register the result. Each step depends on the previous one, and a failed training run should not require re-preparing the entire dataset.

Pipeline

[bml_prepare_data]
     |
     v
[bml_split_data]
     |
     v
     +───────────────────────────────────────────+
     | [bml_train_model_1] | [bml_train_model_2] |
     +───────────────────────────────────────────+
     [join]
     |
     v
[bml_evaluate]

Workflow inputs: datasetId, experimentName

Workers

BmlEvaluateWorker (task: bml_evaluate)

• Clamps with math.max()
• Reads model1Accuracy, model2Accuracy. Writes bestModel, bestAccuracy

BmlPrepareDataWorker (task: bml_prepare_data)

• Writes dataPath, samples, features

BmlSplitDataWorker (task: bml_split_data)

• Writes trainPath, testPath, trainSamples, testSamples

BmlTrainModel1Worker (task: bml_train_model_1)

• Writes modelPath, accuracy, f1Score

BmlTrainModel2Worker (task: bml_train_model_2)

• Writes modelPath, accuracy, f1Score

---

20 tests | Workflow: batch_ml_training_demo | Timeout: 60s

See RUNNING.md for setup and usage.