Common Experiment

Last updated:2026-01-15 17:31:35
Common Experiment
Last updated: 2026-01-15 17:31:35
Feature Overview
Experiment management is one of the key modules in WeData's implementation model, ensuring manageable processes in production and reproducible experiments. It supports the following core features:
Support submitting, storing, managing, and viewing key information during experiment/operation, such as experiment/operation name, model file, code package, hyperparameter, environment, dataset/feature, training metrics, and creation time.
Support comparing key information between experiments/operations.
Support version management and association records for easy information tracking and problem localization.
The experiment management module of WeData is based on the mainstream MLFlow toolkit in the industry. Its implementation ensures manageable processes and reproducible experiments. The general workflow and module functionality are as follows:
Operation Steps
Experimental List
1.  Click to enter the "Model Experiment" feature menu.
﻿
2. The homepage displays the create experiment and experiment list. (Currently, AutoML type experiments are only available in the Chongqing region and will gradually cover all regions in the future.)
﻿
Creating an Experiment
1. You can create a machine learning experiment on the current page by clicking the create button.
﻿
2. Click Create Machine Learning Experiment, fill in the Experiment Name in the popup to create the instance successfully.
﻿
Note：
After creating an experiment, go to Studio to associate it before registering the training task under the linked experiment.
3. You can use the mlflow.create_experiment() function in the training script to create an experiment, for example, define the experiment name as "experiment_202512261509". After running the code, the experiment will appear in the list.
﻿
Experimental Details
1. Click "Experiment Name", and the running list of the experiment will be displayed on the right, showing relevant information including: run name, running state, experiment name, creation time, running time, dataset, code file, registered model, etc.
2. In the current list, you can click the code source file and registered model to navigate to Studio and the model details page for viewing.
﻿
3. You can also delete tasks in the operation column.
﻿
4. Running tasks chart view.
﻿
Display model metrics of running tasks. Support click to show or hide the histogram of the running task. Global search for chart name and parameters is supported at the top.
Support downloading (default PNG file) and full screen display in the top-right corner.
Note：
When the number of running tasks exceeds 10, the first 10 running tasks are displayed sequentially by default to ensure a smooth reading experience. Tasks exceeding the limit are automatically hidden.
5. Compare tasks.
﻿
6. Select tasks via checkboxes, click the Compare tasks button to view parameter and object comparison for running tasks. This allows detailed comparison of experimental running tasks to finally pick over for application.
﻿
Numeric value comparison display view of selected running task details, parameters, and metrics.
Parallel coordinates
﻿
Scatter chart
﻿
Boxplot
﻿
Chart comparison allows you to manually select three types of visual charts: parallel coordinates, scatter chart, and boxplot. The X-axis and Y-axis can be customized as needed in your scenario.
Running Details
1.  Overview Information.
Click "Running Name" to enter the running details page. By default, it enters the overview page and displays the following content:
﻿
Involved queries:
Key information of operation: Creation Time, creator, experiment ID, running state, running ID, Training Duration, dataset, tag, code file, register model.
Hyperparameters Configuration: model hyperparameters in the training process, such as weight in this example.
Metrics: metrics submitted during the model training process (based on the training test set or training validation set).
2. View model metrics.
Click Model Metrics to view model metrics. This feature supports recording and assessing the model's performance metrics, such as accuracy, precision, and recall rate. These metrics help you understand the model's performance and make comparisons.
﻿
You can use the mlflow.log_metric() function in the training script to record model performance metrics, for example:
import mlflow
from sklearn.metrics import accuracy_score, precision_score, recall_score
 
# Assume you have a trained model and test data
y_true = [...]  # actual tag
y_pred = [...]  # prediction tag
 
# Start MLflow run
with mlflow.start_run():
    # Computation model performance metric
    accuracy = accuracy_score(y_true, y_pred)
    precision = precision_score(y_true, y_pred)
    recall = recall_score(y_true, y_pred)
 
    # Log model performance metrics
    mlflow.log_metric("accuracy", accuracy)
    mlflow.log_metric("precision", precision)
    mlflow.log_metric("recall", recall)
...
﻿
3. View system metrics.
System Metrics are used to monitor and log system performance metrics during the model training process, such as CPU utilization and memory usage. These metrics help you understand the resource consumption status of model training. The system metrics recorded by MLFlow by default are as follows:
cpu_utilization_percentage
system_memory_usage_megabytes
system_memory_usage_percentage
network_receive_megabytes
network_transmit_megabytes
disk_usage_megabytes
disk_available_megabytes
You can use three methods in the training script to record system performance metrics, for example:
import mlflow
import psutil  # to obtain system performance metrics
 
#Enable method 1: use environment variables
import os
os.environ["MLFLOW_ENABLE_SYSTEM_METRICS_LOGGING"] = "true"
 
#Enable method 2: use mlflow.enable_system_metrics_logging() to log
mlflow.enable_system_metrics_logging()
 
#Enable method 3: target a designated run to start
with mlflow.start_run(log_system_metrics=True):
    ...
4. View model file.
Artifacts refer to files related to the model, such as model files, charts, datasets, and log files.
As shown in the example diagram below, this is the model file generated this time, including the model path, input and output metadata, and example code for debugging and running. It also allows you to click to navigate to model management and view the published model.
﻿