# MLOps

## Suivez vos expériences  Livrez de meilleurs modèles

Reproductibilité complète avec métriques, paramètres et artefacts. Comparez les runs, collaborez avec votre équipe et versionnez les modèles automatiquement.

Works with:

+more

[See It In Action](/content/demo/index.html) [Documentation](https://documentation.picsellia.com/docs/experiment)

## Best Model

mAP: 0.91 • exp-003

# Training on GPU

## Logging

### Log everything, miss nothing

Capture metrics, parameters, artifacts, and evaluations with a simple API. Full lineage tracking with minimal integration code.

### Scalar

Single values (loss, accuracy)

### Line

Time-series metrics

### Image

Visualizations & samples

### Table

Structured data

### Histogram

Distributions

### Confusion Matrix

Classification results

## train.py

```python
from picsellia import Client

client = Client()
project = client.get_project("my-project")
experiment = project.get_experiment(
  "my-experiment"
)

# Log hyperparameters
experiment.log_parameters({
  "learning_rate": 1e-4,
  "batch_size": 32
})

# Log metrics during training
for epoch in range(epochs):
  experiment.log(
    "train_loss",
    loss.item()
  )
```

#### Artifact Storage

Store checkpoints, configs, and outputs

`experiment.store("model.pt")`

#### Dataset Attachment

Link training data for reproducibility

`experiment.attach_dataset(dataset_version)`

#### Model Export

Push to model registry with one call

`experiment.export_as_model("my-model")`

## Compare

### Compare trainings instantly

See the exact training distribution, hyperparameters, and augmentations behind every performance change. Find your best model faster.

- Side-by-side metric comparison
- Parameter diff highlighting
- Dataset version tracking
- Collaborative comments

### EXPERIMENT COMPARISON

| Name   | mAP | Loss | LR   | Epochs | Status    |
| ------ | --- | ---  | ---  | ------ | --------- |
| exp-001| 0.87| 0.09 | 1e-4 | 100    | completed |
| exp-002| 0.82| 0.12 | 1e-3 | 80     | completed |
| exp-003| 0.91| 0.07 | 5e-5 | 150    | running   |

## CV Engine

### Build training pipelines with ease

Picsellia CV Engine is a modular toolkit for constructing computer vision workflows. Composable steps, framework extensions, and CLI automation.

### Training Pipelines

Data → Model → Results. Streamlined training processes with built-in logic.

### Processing Pipelines

Dataset transformation, pre-annotation, and data cleaning operations.

### Framework Extensions

Pluggable architecture supporting multiple training libraries.

## terminal

```bash
$pip install picsellia-cv-engine

# Initialize a new training pipeline
$pxl-pipeline init --type training

# Run locally for testing
$pxl-pipeline test

# Deploy to Picsellia cloud
$pxl-pipeline deploy --gpus 1
```

## MODEL REGISTRY
### Version your models automatically

Export experiments to the model registry with a single call. Track versions, compare performance, and deploy with confidence.

- Automatic version incrementation
- Framework metadata (TensorFlow, PyTorch, etc.)
- Docker configuration for deployment
- Lineage to training experiment

## Evaluation

### COCO metrics built-in

Add predictions, compare against ground truth, and compute standard evaluation metrics automatically.

- 0.91 mAP@50 Mean Average Precision
- 0.68 mAP@50:95 Strict mAP
- 0.89 Precision True Positives / Predicted
- 0.87 Recall True Positives / Actual

Supports rectangles, polygons, classifications, and keypoints

## Ready to track your experiments?

Start logging metrics, comparing runs, and shipping better models with full reproducibility.