Migrating to Trackio

It’s easy to migrate to Trackio from other experiment tracking libraries with minimal code changes. This guide shows you how to migrate from popular experiment tracking tools.

Weights & Biases ( wandb )

Migrating from Weights & Biases to Trackio is extremely easy because Trackio uses the exact same API syntax as wandb. In most cases, you only need to change the import statement!

Simple Migration

The most basic migration requires just changing your import:

- import wandb
+ import trackio as wandb

wandb.init(project="my-project")
wandb.log({"loss": 0.5, "accuracy": 0.8})
wandb.finish()

Complete Example

Here’s a more complete example showing how the rest of your code stays exactly the same!

- import wandb
+ import trackio as wandb
import numpy as np

wandb.init(
    project="image-classification",
    name="experiment-1",
    config={
        "learning_rate": 0.01,
        "batch_size": 32,
        "epochs": 10
    }
)

for epoch in range(10):
    loss = np.random.random()
    accuracy = np.random.random()
    
    wandb.log({
        "epoch": epoch,
        "loss": loss,
        "accuracy": accuracy,
        "learning_rate": wandb.config.learning_rate
    })

wandb.finish()

Advanced Features

Trackio supports logging Tables, Images, Audio, etc. - same API as wandb:

- import wandb
+ import trackio as wandb
import numpy as np

wandb.init(project="data-analysis")

image_array = np.random.randint(0, 255, (100, 100, 3), dtype=np.uint8)
wandb.log({
    "sample_image": wandb.Image(image_array, caption="Generated sample"),
    "model_diagram": wandb.Image("architecture.png")
})

columns = ["epoch", "train_loss", "val_loss", "accuracy"]
data = [
    [1, 0.8, 0.6, 0.75],
    [2, 0.6, 0.5, 0.82],
    [3, 0.4, 0.45, 0.89]
]
table = wandb.Table(data=data, columns=columns)
wandb.log({"training_results": table})

wandb.finish()

Neptune ( neptune )

Migrating from Neptune requires a few more changes since Neptune has a different API structure, but the migration is still straightforward.

Basic Logging Migration

- import neptune

+ import trackio

- run = neptune.init_run(
-     project="my-workspace/my-project",
-     api_token="your-token"
- )

+ trackio.init(project="my-project")

- run["parameters"] = {"learning_rate": 0.01, "batch_size": 32}
- run["metrics/loss"].log(0.5)
- run["metrics/accuracy"].log(0.8)

+ trackio.config.update({"learning_rate": 0.01, "batch_size": 32})
+ trackio.log({"loss": 0.5, "accuracy": 0.8})

- run.stop()

+ trackio.finish()

Complete Training Loop Migration

- import neptune
+ import trackio
import numpy as np

- run = neptune.init_run(
-     project="my-workspace/classification-project",
-     name="experiment-1",
-     tags=["pytorch", "cnn"]
- )

+ trackio.init(
+     project="classification-project",
+     name="experiment-1",
+     tags=["pytorch", "cnn"]
+ )

config = {"learning_rate": 0.01, "epochs": 10, "batch_size": 32}

- run["parameters"] = config
+ trackio.config.update(config)

for epoch in range(config["epochs"]):
    # Simulate training
    train_loss = np.random.random()
    val_accuracy = np.random.random()
    
    # Neptune logging
-     run["metrics/train/loss"].log(train_loss)
-     run["metrics/val/accuracy"].log(val_accuracy)
-     run["metrics/epoch"].log(epoch)

    # Trackio logging
+     trackio.log({
+         "train/loss": train_loss,
+         "val/accuracy": val_accuracy,
+         "epoch": epoch
+     })

- run["model/weights"].upload("model.pth")
+ trackio.save("model.pth")

- run.stop()
+ trackio.finish()

Key Migration Points

Initialization: Replace neptune.init_run() with trackio.init()
Logging: Use trackio.log() with dictionaries instead of individual metric assignments
Cleanup: Replace run.stop() with trackio.finish()

Benefits to Migrating

Simpler API: Flat dictionary logging vs nested attribute access
Local development: Work offline by default
Free hosting: Deploy dashboards on Hugging Face Spaces at no cost
Familiar interface: If you’ve used wandb before in particular, the API is unchanged

Update on GitHub