Tutorials/MLflow/How-to-get-started-with-MLflow.md

+# How to get started with MLflow
+
+*work in progress*
+
+The purpose of this page is to give you an idea what MLflow can be used for, which aspects to consider and best practices. 
+
+---
+
+#### 0. Set Up Your ML Project with [Tracking](https://mlflow.org/docs/latest/tracking/) in Mind
+- Use a clean project structure with separate folders for data, code, outputs, and models.
+- Use a virtual environment to manage dependencies.
+- [Install MLflow and test that the tracking UI works locally or on a server.](Tutorials/MLflow/How-to-set-up-MLflow)
+
+#### 1. Organise Experiments 
+- Define an experiment name to group all related model runs. Think of an [experiment as a folder](https://mlflow.org/docs/latest/traditional-ml/hyperparameter-tuning-with-child-runs/part1-child-runs/) containing different versions of your model with varying settings.
+- Use descriptive run names and tags.
+- Organise runs using the concept of [Parent and Child Runs](https://mlflow.org/docs/latest/traditional-ml/hyperparameter-tuning-with-child-runs/part1-child-runs/) in MLflow, a feature that provides a hierarchical structure when doing hyperparameter searches. 
+
+#### 3. Track Runs During Training
+- Log everything relevant to reproduce or understand the run later. <br>
+Info: Most ML libraries (e.g. [Keras](https://mlflow.org/docs/latest/deep-learning/keras/)/[TensorFlow](https://mlflow.org/docs/latest/deep-learning/tensorflow/), [PyTorch](https://mlflow.org/docs/latest/deep-learning/pytorch/guide/), Scikit-learn, ...) support the [autologging feature of MLflow](https://mlflow.org/docs/latest/tracking/autolog/): 
+- For each model training run: <br>
+  * Log the model's parameters (e.g., learning rate)<br>
+  * Log performance metrics (e.g., accuracy or loss)<br>
+  * Save and log any important artifacts (e.g., plots, config files, or trained models)<br>
+Note: Avoid displaying figures inside notebooks, instead [log the files and view them in the UI](https://mlflow.org/docs/latest/traditional-ml/hyperparameter-tuning-with-child-runs/notebooks/logging-plots-in-mlflow). Use scripts or pipeline functions that can be reused and versioned.
+
+#### 4. Use the MLflow UI to Compare Results
+- Browse and filter model runs
+- Compare different configurations side by side
+- View plots and metrics to [choose the best model](https://mlflow.org/docs/latest/getting-started/quickstart-2/)
+
+#### 6. Collaborate and Share Results
+- Use a shared tracking server so others on your team can log to and view the same experiment database.
+- Store artifacts and models in shared or centralized locations.
+
+#### 7. Reuse and Deploy Models
+- Save your trained models via MLflow for later use.
+- Optionally, register the best models and serve them with MLflow's model serving features.
\ No newline at end of file