Skip to content

Command Line Interface¤

Artifex provides a powerful CLI for training, generating, evaluating, and serving generative models without writing code.

Overview¤

  • Training

    Train models with artifex train using YAML configs

  • Generation

    Generate samples with artifex generate

  • Evaluation

    Benchmark models with artifex evaluate

  • Serving

    Deploy models with artifex serve

Installation¤

The CLI is installed automatically with Artifex:

# Verify installation
artifex --help

# Or run via uv
uv run artifex --help

Quick Start¤

Train a Model¤

# Train from config file
artifex train --config configs/vae_mnist.yaml

# Train with overrides
artifex train --config configs/vae_mnist.yaml \
    --batch-size 128 \
    --epochs 100 \
    --lr 1e-3

# Resume training from checkpoint
artifex train --config configs/vae_mnist.yaml \
    --resume checkpoints/vae_mnist/latest

Generate Samples¤

# Generate samples from trained model
artifex generate --checkpoint checkpoints/vae_mnist/best \
    --num-samples 100 \
    --output generated/

# Generate with specific seed
artifex generate --checkpoint model.ckpt \
    --num-samples 50 \
    --seed 42 \
    --output samples/

Evaluate Model¤

# Run evaluation
artifex evaluate --checkpoint checkpoints/vae_mnist/best \
    --dataset mnist \
    --metrics fid inception_score

# Evaluate with custom test set
artifex evaluate --checkpoint model.ckpt \
    --data-path /path/to/test/data \
    --output results/evaluation.json

Serve Model¤

# Start inference server
artifex serve --checkpoint checkpoints/vae_mnist/best \
    --port 8000

# Serve with specific host
artifex serve --checkpoint model.ckpt \
    --host 0.0.0.0 \
    --port 8080

Commands¤

artifex train¤

Train generative models from configuration.

artifex train [OPTIONS]

Options:
  --config PATH          Path to training config file [required]
  --batch-size INT       Override batch size
  --epochs INT           Override number of epochs
  --lr FLOAT             Override learning rate
  --resume PATH          Resume from checkpoint
  --output-dir PATH      Output directory for checkpoints
  --seed INT             Random seed
  --device TEXT          Device to use (auto/gpu/cpu)
  --wandb                Enable Weights & Biases logging
  --mlflow               Enable MLflow logging
  --help                 Show help message

Train Command Reference

artifex generate¤

Generate samples from trained models.

artifex generate [OPTIONS]

Options:
  --checkpoint PATH      Path to model checkpoint [required]
  --num-samples INT      Number of samples to generate [default: 10]
  --output PATH          Output directory for samples
  --seed INT             Random seed for reproducibility
  --temperature FLOAT    Sampling temperature [default: 1.0]
  --batch-size INT       Batch size for generation
  --format TEXT          Output format (png/npy/pt)
  --help                 Show help message

Generate Command Reference

artifex evaluate¤

Evaluate model performance with metrics.

artifex evaluate [OPTIONS]

Options:
  --checkpoint PATH      Path to model checkpoint [required]
  --dataset TEXT         Dataset name or path
  --data-path PATH       Custom data path
  --metrics TEXT         Metrics to compute (comma-separated)
  --output PATH          Output file for results
  --batch-size INT       Batch size for evaluation
  --help                 Show help message

Evaluate Command Reference

artifex serve¤

Deploy model as REST API.

artifex serve [OPTIONS]

Options:
  --checkpoint PATH      Path to model checkpoint [required]
  --host TEXT            Host address [default: localhost]
  --port INT             Port number [default: 8000]
  --workers INT          Number of workers [default: 1]
  --help                 Show help message

Serve Command Reference

artifex benchmark¤

Run comprehensive benchmarks.

artifex benchmark [OPTIONS]

Options:
  --config PATH          Benchmark configuration file
  --model PATH           Model checkpoint or config
  --suite TEXT           Benchmark suite to run
  --output PATH          Output directory for results
  --help                 Show help message

Benchmark Command Reference

artifex convert¤

Convert between model formats.

artifex convert [OPTIONS]

Options:
  --input PATH           Input model path [required]
  --output PATH          Output model path [required]
  --format TEXT          Target format (onnx/tflite/safetensors)
  --help                 Show help message

Convert Command Reference

Configuration Files¤

Training Config Example¤

# configs/vae_mnist.yaml
model:
  type: vae
  latent_dim: 32
  encoder:
    hidden_dims: [256, 128]
    activation: relu
  decoder:
    hidden_dims: [128, 256]
    activation: relu

training:
  batch_size: 128
  epochs: 100
  optimizer:
    type: adam
    learning_rate: 1e-3
  scheduler:
    type: cosine
    warmup_steps: 1000

data:
  dataset: mnist
  train_split: 0.9

logging:
  wandb: true
  project: artifex-experiments
  log_interval: 100

Using Environment Variables¤

# Set defaults via environment
export ARTIFEX_DEVICE=gpu
export ARTIFEX_SEED=42
export ARTIFEX_OUTPUT_DIR=./outputs

# These become defaults for all commands
artifex train --config config.yaml

Utility Modules¤

Configuration Utilities¤

  • config - Configuration loading and validation
  • formatting - Output formatting helpers

Logging Utilities¤

Examples¤

Complete Training Workflow¤

# 1. Train the model
artifex train --config configs/dcgan_cifar.yaml \
    --output-dir experiments/dcgan_001 \
    --wandb

# 2. Generate samples
artifex generate \
    --checkpoint experiments/dcgan_001/best.ckpt \
    --num-samples 1000 \
    --output experiments/dcgan_001/samples/

# 3. Evaluate quality
artifex evaluate \
    --checkpoint experiments/dcgan_001/best.ckpt \
    --dataset cifar10 \
    --metrics fid inception_score lpips \
    --output experiments/dcgan_001/metrics.json

# Run multiple training jobs
for lr in 1e-4 1e-3 1e-2; do
    artifex train --config configs/vae.yaml \
        --lr $lr \
        --output-dir experiments/vae_lr_$lr
done