Command-Line Interface (CLI)¶

AutoTimm provides a YAML-config-driven command-line interface built on LightningCLI. Train, validate, test, and run predictions without writing Python scripts.

CLI Workflow¶

graph LR
    A[<b>YAML Config</b><br/>model, data,<br/>trainer] --> B[<b>autotimm</b><br/>CLI]
    B --> C{Subcommand}
    C --> D[fit]
    C --> E[validate]
    C --> F[test]
    C --> G[predict]

    style A fill:#1565C0,stroke:#0D47A1
    style B fill:#1976D2,stroke:#1565C0
    style C fill:#FF9800,stroke:#F57C00
    style D fill:#4CAF50,stroke:#388E3C
    style E fill:#4CAF50,stroke:#388E3C
    style F fill:#4CAF50,stroke:#388E3C
    style G fill:#4CAF50,stroke:#388E3C

Installation¶

The CLI is available after installing AutoTimm. The only additional dependency is jsonargparse, which is included automatically:

pip install autotimm

Quick Start¶

1. Create a Config File¶

model:
  class_path: autotimm.ImageClassifier
  init_args:
    backbone: resnet18
    num_classes: 10

data:
  class_path: autotimm.ImageDataModule
  init_args:
    dataset_name: CIFAR10
    data_dir: ./data
    batch_size: 32
    image_size: 224
    num_workers: 4

trainer:
  max_epochs: 10
  accelerator: auto
  devices: auto
  precision: 32

2. Train¶

autotimm fit --config config.yaml

Or using the Python module:

python -m autotimm fit --config config.yaml

3. Override Parameters¶

Any config value can be overridden from the command line:

autotimm fit --config config.yaml \
    --model.init_args.lr 0.001 \
    --trainer.max_epochs 20 \
    --trainer.precision "bf16-mixed"

Subcommands¶

Training¶

autotimm fit --config config.yaml

Validation & Testing¶

autotimm validate --config config.yaml --ckpt_path path/to/checkpoint.ckpt
autotimm test --config config.yaml --ckpt_path path/to/checkpoint.ckpt

Prediction¶

autotimm predict --config config.yaml --ckpt_path path/to/checkpoint.ckpt

Getting Help¶

# Show available subcommands
autotimm --help

# Show all options for training
autotimm fit --help

# Print the resolved config (useful for debugging)
autotimm fit --config config.yaml --print_config

Config File Structure¶

A config file has three main sections: model, data, and trainer.

Model Section¶

Specify the task class and its arguments:

model:
  class_path: autotimm.ImageClassifier  # or ObjectDetector, SemanticSegmentor, etc.
  init_args:
    backbone: resnet50
    num_classes: 10
    lr: 0.001

Available task classes:

Data Section¶

Specify the data module and its arguments:

data:
  class_path: autotimm.ImageDataModule  # or DetectionDataModule, etc.
  init_args:
    data_dir: ./data
    batch_size: 32
    image_size: 224

Available data modules:

Trainer Section¶

All AutoTrainer parameters are available:

trainer:
  max_epochs: 10
  accelerator: auto
  devices: auto
  precision: 32
  seed: 42
  deterministic: true
  tuner_config: false
  enable_checkpointing: true
  logger: false
  fast_dev_run: false

Task-Specific Examples¶

Image Classification¶

model:
  class_path: autotimm.ImageClassifier
  init_args:
    backbone: efficientnet_b0
    num_classes: 10
    lr: 0.001

data:
  class_path: autotimm.ImageDataModule
  init_args:
    dataset_name: CIFAR10
    data_dir: ./data
    batch_size: 64
    image_size: 224
    num_workers: 4

trainer:
  max_epochs: 20
  accelerator: auto
  tuner_config: false

Object Detection¶

model:
  class_path: autotimm.ObjectDetector
  init_args:
    backbone: resnet50
    num_classes: 80
    detection_arch: fcos
    fpn_channels: 256
    lr: 0.01

data:
  class_path: autotimm.DetectionDataModule
  init_args:
    data_dir: ./coco
    image_size: 640
    batch_size: 8

trainer:
  max_epochs: 50
  accelerator: auto
  precision: "16-mixed"

Semantic Segmentation¶

model:
  class_path: autotimm.SemanticSegmentor
  init_args:
    backbone: resnet50
    num_classes: 19
    head_type: deeplabv3plus
    lr: 0.01

data:
  class_path: autotimm.SegmentationDataModule
  init_args:
    data_dir: ./cityscapes
    format: cityscapes
    image_size: 512
    batch_size: 4

trainer:
  max_epochs: 100
  accelerator: auto

Tips¶

Quick Smoke Test¶

Use fast_dev_run to verify your config works before a full training run:

autotimm fit --config config.yaml --trainer.fast_dev_run true

Disable Auto-Tuning¶

AutoTrainer enables auto-tuning by default. Disable it in the config or from the CLI:

autotimm fit --config config.yaml --trainer.tuner_config false

Use HuggingFace Hub Models¶

Specify HF Hub models in the backbone field:

model:
  class_path: autotimm.ImageClassifier
  init_args:
    backbone: "hf-hub:timm/convnext_base.fb_in22k_ft_in1k"
    num_classes: 100

Combine Multiple Configs¶

LightningCLI supports multiple --config flags for composing configs:

autotimm fit --config base.yaml --config overrides.yaml

Standalone CLI Modules¶

Beyond the autotimm training CLI, AutoTimm provides standalone CLI modules for common workflows:

Export to TorchScript¶

Export a trained checkpoint to TorchScript (JIT) format for production deployment or model visualization:

python -m autotimm.export.export_jit \
    --checkpoint logs/run_1/checkpoints/best.ckpt \
    --output model.pt \
    --task-class ImageClassifier

The exported .pt file can be loaded with torch.jit.load() without any AutoTimm dependency. Input size is auto-detected from model hparams when available.

Run Model Interpretation¶

Run interpretation methods on a trained checkpoint and save heatmap overlays:

python -m autotimm.cli.interpret_cli \
    --checkpoint logs/run_1/checkpoints/best.ckpt \
    --image test_image.jpg \
    --methods gradcam,gradcampp,integrated_gradients \
    --output-dir ./interpretations \
    --task-class ImageClassifier

Outputs JSON to stdout with heatmap file paths and predicted class. Attention methods are automatically skipped for non-ViT models.

See CLI API Reference for full argument details.

See also: AutoTrainer API | CLI API Reference | Export API | Example Configs