Logging Data in Job Runs

Prerequisites

To log data from your job runs, you need access to an ML Repository (MLRepo). ML Repositories store models, data, artifacts, and prompts and are backed by blob storage like S3, GCS, or Azure Blob Storage.

Setting up ML Repository Access

Step 1: Create or Access an ML Repository

If you don’t have an ML Repository yet, you’ll need to create one:

Prerequisite - Blob Storage Integration: Before creating a Repository, connect one or more Blob Storages to TrueFoundry:
- AWS S3
- Google Cloud Storage
- Azure Blob Storage
- Any S3 API Compatible Storage
Create Repository: Go to Platform → Repositories tab and create a new ML Repository

Grant ML Repository Access to Workspace

To enable your job to log data, you need to grant access to the ML Repository for your workspace:

Go to Platform → Workspaces tab
Edit your workspace
In the “ML Repositories” section, grant access to your ML Repository
Choose appropriate permissions (Viewer or Editor)

Creating Run and Logging Data

A run is used to represent a single ML experiment. You can create a run at the beginning of your script or notebook, log parameters, metrics, artifacts, models, tags and finally end the run. This provides an easy to keep track of all data related to ML experiments. A quick code snippet to create a run and end it:

from truefoundry.ml import get_client

client = get_client()
run = client.create_run(ml_repo="iris-demo", run_name="svm-model")
# Your code here.
run.end()

You can organize multiple runs under a single ml_repo. For example, the run svm-model will be created under the ml_repo iris-demo. Once you’ve created runs and logged data, you can view them in the TrueFoundry dashboard. Navigate to your job in the Platform → Applications tab, click on the job name, and go to the “Job Runs” tab to see all executions with their status, metrics, and parameters.

Job runs dashboard showing multiple runs with different statuses including finished, terminated, and failed runs — Job Runs Dashboard - Example showing multiple runs with different statuses

Logging Different Types of Data

Creating a run and ending it

Python

from truefoundry.ml import get_client

client = get_client()
run = client.create_run(ml_repo="iris-demo", run_name="svm-model")
# Your code here.
run.end()

Adding tags to a run

from truefoundry.ml import get_client

client = get_client()
run = client.create_run(ml_repo="iris-demo", run_name="svm-model")
run.set_tags({"env": "development", "task": "classification"})
# Your code here.
run.end()

You can view the tags from the dashboard and also create new tags.

Logging parameters

Parameters are used to store the configuration of a run. This can be either the inputs to your script or the hyperparameters of your model during training like learning_rate, cache_size. The parameter values are stringified before storing.You can log parameters using the log_params as shown below:

from truefoundry.ml import get_client

client = get_client()
run = client.create_run(ml_repo="iris-demo", run_name="svm-model")

run.log_params({"cache_size": 200.0, "kernel": "linear"})

run.end()

Parameters are immutable and you cannot change the value of param once logged. If you need to change the value of param, it basically means that you are changing your input configuration and it’s best to create a new run for that.

Viewing logged parameter in the dashboard

Filtering runs based on parameter value

To filters runs, click on top right corner of the screen to apply the required filter.

Capturing command-line arguments in the run

We can capture command-line arguments directly from the argparse.Namespace object.

import argparse
from truefoundry.ml import get_client

parser = argparse.ArgumentParser()
parser.add_argument("--batch_size", type=int, required=True)
args = parser.parse_args()

client = get_client()
run = client.create_run(ml_repo="iris-demo")

run.log_params(args)

run.end()

Logging metrics

Metrics are values that help you to evaluate and compare different runs - for e.g. accuracy, f1 score. You can log any output of your script as a metric.You can capture metrics using the log_metrics method.

from truefoundry.ml import get_client

client = get_client()
run = client.create_run(ml_repo="iris-demo", run_name="svm-model")
run.log_params({"cache_size": 200.0, "kernel": "linear"})
run.log_metrics(metric_dict={"accuracy": 0.7, "loss": 0.6})

run.end()

These metrics can be seen in Truefoundry dashboard. Filters can be used on metrics values to filter out runs as shown in the figure.

Step-wise metric logging in the run

You can capture step-wise metrics too using the step argument.

for global_step in range(1000):
    run.log_metrics(metric_dict={"accuracy": 0.7, "loss": 0.6}, step=global_step)

The stepwise-metrics can be visualized as graphs in the dashboard.

Should I use epoch or global step as a value for the `step` argument in the run?

If available you should use the global step as a value for the step argument. To capture epoch-level metric aggregates, you can use the following pattern.

run.log_metrics(
  (metric_dict = { "epoch/train_accuracy": 0.7, epoch: epoch }),
  (step = global_step)
);

Log Artifacts

import os
from truefoundry.ml import get_client
from truefoundry.ml import ArtifactPath

client = get_client()
run = client.create_run(ml_repo="iris-demo", run_name="svm-model")
run.log_params({"cache_size": 200.0, "kernel": "linear"})
run.log_metrics(metric_dict={"accuracy": 0.7, "loss": 0.6})

# Just creating sample files to log as artifacts
# os.makedirs("my-folder", exist_ok=True)
# with open("my-folder/file-inside-folder.txt", "w") as f:
#     f.write("Hello!")

# with open("just-a-file.txt", "w") as f:
#     f.write("Hello from file!")

artifact_version = run.log_artifact(
    name="my-artifact",
    artifact_paths=[
        # Add files and folders here, `ArtifactPath` takes source and destination
        # source can be single file path or folder path
        # destination can be file path or folder path
        # Note: When source is a folder path, destination is always interpreted as folder path
        ArtifactPath(src="just-a-file.txt"),
        ArtifactPath(src="my-folder/", dest="cool-dir"),
        ArtifactPath(src="just-a-file.txt", dest="cool-dir/copied-file.txt")
    ],
    description="This is a sample artifact",
    metadata={"created_by": "my-username"}
)
print(artifact_version.fqn)
run.end()

Log Models

from truefoundry.ml import get_client

client = get_client()
run = client.create_run(ml_repo="iris-demo", run_name="svm-model")
run.log_params({"cache_size": 200.0, "kernel": "linear"})
run.log_metrics(metric_dict={"accuracy": 0.7, "loss": 0.6})

model_version = run.log_model(
    name="name-for-the-model",
    model_file_or_folder="path/to/model/file/or/folder/on/disk",
    framework=<None or Framework> # Check
)
run.end()

Complete Examples

Here are comprehensive examples that demonstrate how to deploy a job and log data during machine learning training:

MNIST Training Script with Logging

This example shows the complete training script that logs parameters, metrics, and models:

import argparse
import tensorflow as tf
from tensorflow.keras.datasets import mnist
from truefoundry.ml import get_client

# Parse command line arguments
parser = argparse.ArgumentParser()
parser.add_argument("--num_epochs", type=int, default=4)
parser.add_argument("--learning_rate", type=float, default=0.01)
parser.add_argument(
    "--ml_repo",
    type=str,
    required=True,
    help="The name of the ML Repo to track metrics and models"
)
args = parser.parse_args()

# Load and preprocess the MNIST dataset
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train = x_train / 255.0
x_test = x_test / 255.0

print(f"The number of train images: {len(x_train)}")
print(f"The number of test images: {len(x_test)}")

# Initialize TrueFoundry client and create a run
client = get_client()
run = client.create_run(ml_repo=args.ml_repo, run_name="mnist-training")

try:
    # Define and compile the model
    model = tf.keras.Sequential([
        tf.keras.layers.Flatten(input_shape=(28, 28)),
        tf.keras.layers.Dense(128, activation="relu"),
        tf.keras.layers.Dense(10, activation="softmax"),
    ])

    optimizer = tf.keras.optimizers.Adam(learning_rate=args.learning_rate)
    model.compile(
        optimizer=optimizer,
        loss="sparse_categorical_crossentropy",
        metrics=["accuracy"]
    )

    # Log hyperparameters
    run.log_params({
        "optimizer": "adam",
        "learning_rate": args.learning_rate,
        "num_epochs": args.num_epochs,
        "loss": "sparse_categorical_crossentropy",
        "metrics": ["accuracy"],
        "model_architecture": "Sequential with Dense layers"
    })

    # Train the model
    history = model.fit(
        x_train, y_train,
        epochs=args.num_epochs,
        validation_data=(x_test, y_test)
    )

    # Evaluate the model
    loss, accuracy = model.evaluate(x_test, y_test)
    print(f"Test loss: {loss}")
    print(f"Test accuracy: {accuracy}")

    # Log metrics for each epoch
    history_dict = history.history
    for epoch in range(args.num_epochs):
        run.log_metrics({
            "train_accuracy": history_dict['accuracy'][epoch],
            "val_accuracy": history_dict['val_accuracy'][epoch],
            "train_loss": history_dict['loss'][epoch],
            "val_loss": history_dict['val_loss'][epoch]
        }, step=epoch + 1)

    # Log final metrics
    run.log_metrics({
        "final_test_accuracy": accuracy,
        "final_test_loss": loss
    })

    # Save and log the trained model
    model.save("mnist_model.h5")
    run.log_model(
        name="mnist-handwritten-digits",
        model_file_or_folder="mnist_model.h5",
        framework="tensorflow",
        description="Neural network model to recognize handwritten digits from MNIST dataset",
        metadata={
            "test_accuracy": accuracy,
            "test_loss": loss,
            "num_epochs": args.num_epochs,
            "learning_rate": args.learning_rate
        }
    )

    # Add tags for organization
    run.set_tags({
        "dataset": "mnist",
        "model_type": "neural_network",
        "framework": "tensorflow",
        "task": "classification"
    })

    print("Training completed successfully!")

except Exception as e:
    print(f"Training failed: {e}")
    raise
finally:
    # End the run
    run.end()

Key Features Demonstrated:

Parameter Logging: Hyperparameters like learning rate, epochs, and model configuration
Metrics Logging: Training and validation accuracy/loss for each epoch
Model Logging: Saved model with metadata and framework information
Tag Organization: Categorizing the run for easy filtering
Error Handling: Proper exception handling to ensure runs are marked correctly

Deploy MNIST Training Job

This example shows how to deploy a parameterized job that can be run multiple times with different configurations:

import logging
import argparse

from truefoundry.deploy import Build, Job, LocalSource, Param, PythonBuild, Resources

logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(name)s] %(levelname)-8s %(message)s")

# Parse command line arguments
parser = argparse.ArgumentParser()
parser.add_argument("--workspace_fqn", type=str, required=True, help="FQN of the workspace to deploy to")
args = parser.parse_args()

# Define the job specifications
job = Job(
    name="mnist-train-job",
    image=Build(
        build_source=LocalSource(local_build=False),
        build_spec=PythonBuild(
            python_version="3.11",
            command="python train.py --num_epochs {{num_epochs}} --ml_repo {{ml_repo}}",
            requirements_path="requirements.txt",
        ),
    ),
    params=[
        Param(name="num_epochs", default="4"),
        Param(name="ml_repo", param_type="ml_repo"),
    ],
    resources=Resources(
        cpu_request=0.5,
        cpu_limit=0.5,
        memory_request=1500,
        memory_limit=2000
    ),
)

# Deploy the job
deployment = job.deploy(workspace_fqn=args.workspace_fqn, wait=False)
print(f"Job deployed successfully! FQN: {deployment.fqn}")

Key Features:

Parameterized Job: Uses Param objects to make the job configurable
Python Build: Automatically builds the container from source code
Resource Configuration: Specifies CPU and memory requirements
ML Repository Integration: Links to an ML repository for logging

Accessing Detailed Run Information

In the Job Runs table, you’ll notice that the “RUN DETAILS” column contains clickable links. When you click on any run details link, you’ll be taken to a comprehensive view of that specific run, which includes:

Overview Tab: Key metrics and hyperparameters used in the run
Results Tab: Detailed metrics and performance data
Models Tab: All logged models with their metadata
Artifacts Tab: Files and artifacts associated with the run

Pro Tip: Use the run details view to analyze your experiments, compare different hyperparameter configurations, and track the progress of your machine learning projects over time.

Getting Started

Train and Deploy Models

Service Deployment

Job Deployment

LLM Deployment

LLM Finetuning

MCP Server Deployment

Workflow Deployment

Async Service Deployment

Volumes

Experiment Tracking

Advanced Features

Prerequisites

Step 1: Create or Access an ML Repository

Grant ML Repository Access to Workspace

Creating Run and Logging Data

Logging Different Types of Data

Viewing logged parameter in the dashboard

Filtering runs based on parameter value

Capturing command-line arguments in the run

Step-wise metric logging in the run

Should I use epoch or global step as a value for the `step` argument in the run?

Complete Examples

Key Features Demonstrated:

Key Features:

Accessing Detailed Run Information

​Prerequisites

​Step 1: Create or Access an ML Repository

​Grant ML Repository Access to Workspace

​Creating Run and Logging Data

​Logging Different Types of Data

​Viewing logged parameter in the dashboard

​Filtering runs based on parameter value

​Capturing command-line arguments in the run

​Step-wise metric logging in the run

​Should I use epoch or global step as a value for the step argument in the run?

​Complete Examples

​Key Features Demonstrated:

​Key Features:

​Accessing Detailed Run Information

Prerequisites

Step 1: Create or Access an ML Repository

Grant ML Repository Access to Workspace

Creating Run and Logging Data

Logging Different Types of Data

Viewing logged parameter in the dashboard

Filtering runs based on parameter value

Capturing command-line arguments in the run

Step-wise metric logging in the run

Should I use epoch or global step as a value for the `step` argument in the run?

Complete Examples

Key Features Demonstrated:

Key Features:

Accessing Detailed Run Information