autotimm 0.7.34

Automated deep learning for computer vision — train image classification, object detection, and segmentation models with 1000+ backbones in minimal code

Train state-of-the-art vision models with minimal code.

Production-ready computer vision framework powered by timm and PyTorch Lightning.

Overview · Features · Quick Start · Tasks · Architecture · Docs · Contributing · License

Overview

AutoTimm is a production-ready computer vision framework that combines timm (1000+ pretrained models) with PyTorch Lightning. Train image classifiers, object detectors, and segmentation models with any timm backbone using a simple, intuitive API.

Whether you're training an image classifier, an object detector, or a segmentation model, AutoTimm handles:

• Automated model training with auto-tuning of learning rate and batch size
• 4 vision tasks — classification (single & multi-label), object detection, semantic & instance segmentation
• 1000+ backbones — ResNet, EfficientNet, ViT, ConvNeXt, Swin, and more via timm + HuggingFace Hub
• Model interpretation — GradCAM, Integrated Gradients, Attention Rollout, and more
• Production export — TorchScript and ONNX for deployment anywhere

From prototype to production in minutes, not hours.

---

How It Compares

	Manual PyTorch	AutoTimm
Boilerplate	Hundreds of lines per task	5–10 lines for any task
Backbone variety	Wire up manually	1000+ backbones, swap with one arg
Auto-tuning	Implement yourself	LR + batch size finding built-in
Experiment tracking	External tooling needed	Multi-logger (TensorBoard, MLflow, W&B, CSV)
Reproducibility	Manual seed management	Opt-in deterministic mode (seed=42)
Model export	Custom export scripts	One-line TorchScript & ONNX
Interpretation	Separate libraries	6 methods + metrics built-in

---

Features

Vision Tasks Image Classification — single-label and multi-label with any timm backbone Object Detection — FCOS and YOLOX architectures (official + custom) Semantic Segmentation — DeepLabV3+ and FCN with combined losses Instance Segmentation — FCOS + Mask R-CNN style mask head	Smart Training Auto-tuning — automatic LR and batch size discovery torch.compile — PyTorch 2.0+ optimization enabled by default Reproducibility — deterministic seeding across Python, NumPy, PyTorch Mixed precision — 16-bit and bf16 training out of the box Multi-GPU — distributed training with zero config
Interpretation & Export 6 explanation methods — GradCAM, GradCAM++, Integrated Gradients, SmoothGrad, Attention Rollout, Attention Flow 6 quality metrics — faithfulness, sensitivity, localization, sanity checks Interactive visualizations — Plotly-powered HTML reports Model export — TorchScript (.pt) and ONNX (.onnx) for production	Data & Integration Flexible data loading — folder structure, COCO JSON, CSV, HuggingFace datasets Smart transforms — AI-powered backend selection, unified TransformConfig HuggingFace Hub — load models with hf-hub: prefix Multi-logger — TensorBoard, MLflow, W&B, CSV simultaneously CLI — YAML-driven training from the command line

Who Is It For? Researchers needing reproducible experiments · Engineers building production ML systems · Students learning computer vision · Startups rapidly prototyping vision applications

---

Quick Start

Installation

pip install autotimm

Everything included: PyTorch, timm, PyTorch Lightning, torchmetrics, albumentations, and more.

Optional extras

pip install autotimm[tensorboard]  # TensorBoard
pip install autotimm[wandb]        # Weights & Biases
pip install autotimm[mlflow]       # MLflow
pip install autotimm[onnx]         # ONNX export (onnx + onnxruntime + onnxscript)
pip install autotimm[all]          # Everything

Your First Model

import autotimm as at  # recommended alias
from autotimm import AutoTrainer, ImageClassifier, ImageDataModule, MetricConfig

# Data
data = ImageDataModule(
    data_dir="./data",
    dataset_name="CIFAR10",
    image_size=224,
    batch_size=64,
)

# Metrics
metrics = [
    MetricConfig(
        name="accuracy",
        backend="torchmetrics",
        metric_class="Accuracy",
        params={"task": "multiclass"},
        stages=["train", "val", "test"],
        prog_bar=True,
    )
]

# Model — try efficientnet_b0, vit_base_patch16_224, hf-hub:timm/resnet50.a1_in1k, etc.
model = ImageClassifier(backbone="resnet18", num_classes=10, metrics=metrics, lr=1e-3)

# Train with auto-tuning (finds optimal LR and batch size automatically)
trainer = AutoTrainer(max_epochs=10)
trainer.fit(model, datamodule=data)

Auto-tuning is enabled by default. Disable with tuner_config=False for manual control.

Command-Line Interface

Train from YAML configs — no Python scripts needed:

autotimm fit --config config.yaml
autotimm test --config config.yaml --ckpt_path best.ckpt

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

trainer:
  max_epochs: 10
  accelerator: auto

---

Task Examples

Image Classification

from autotimm import ImageClassifier

model = ImageClassifier(
    backbone="efficientnet_b0",  # or "hf-hub:timm/resnet50.a1_in1k"
    num_classes=10,
    metrics=metrics,
)

trainer = AutoTrainer(max_epochs=10)
trainer.fit(model, datamodule=data)

Object Detection with YOLOX

from autotimm import YOLOXDetector, DetectionDataModule

model = YOLOXDetector(
    model_name="yolox-s",  # nano, tiny, s, m, l, x
    num_classes=80,
    lr=0.01,
    optimizer="sgd",
    scheduler="yolox",
    total_epochs=300,
)

trainer = AutoTrainer(max_epochs=300, precision="16-mixed")
trainer.fit(model, datamodule=DetectionDataModule(data_dir="./coco", image_size=640))

Semantic Segmentation

from autotimm import SemanticSegmentor, SegmentationDataModule

model = SemanticSegmentor(
    backbone="resnet50",
    num_classes=19,
    head_type="deeplabv3plus",
    loss_type="combined",  # CE + Dice for better boundaries
)

data = SegmentationDataModule(data_dir="./cityscapes", format="cityscapes", image_size=512)

trainer = AutoTrainer(max_epochs=100)
trainer.fit(model, datamodule=data)

Instance Segmentation

from autotimm import InstanceSegmentor, InstanceSegmentationDataModule

model = InstanceSegmentor(backbone="resnet50", num_classes=80, mask_loss_weight=1.0)

trainer = AutoTrainer(max_epochs=100)
trainer.fit(model, datamodule=InstanceSegmentationDataModule(data_dir="./coco"))

Multi-Label Classification

from autotimm import ImageClassifier, MultiLabelImageDataModule, MetricConfig

data = MultiLabelImageDataModule(
    train_csv="train.csv", image_dir="./images", val_csv="val.csv",
    image_size=224, batch_size=32,
)
data.setup("fit")

model = ImageClassifier(
    backbone="resnet50",
    num_classes=data.num_labels,
    multi_label=True,
    threshold=0.5,
    metrics=[MetricConfig(
        name="accuracy", backend="torchmetrics", metric_class="MultilabelAccuracy",
        params={"num_labels": data.num_labels}, stages=["train", "val"], prog_bar=True,
    )],
)

trainer = AutoTrainer(max_epochs=10)
trainer.fit(model, datamodule=data)

---

Model Interpretation

from autotimm.interpretation import (
    GradCAM, GradCAMPlusPlus, IntegratedGradients,
    SmoothGrad, AttentionRollout, AttentionFlow,
    ExplanationMetrics, InteractiveVisualizer,
)

# Explain a prediction
explainer = GradCAM(model)
heatmap = explainer.explain(image, target_class=5)
explainer.visualize(image, heatmap, save_path="gradcam.png")

# Quantitative evaluation
metrics = ExplanationMetrics(model, explainer)
deletion = metrics.deletion(image, target_class=5, steps=50)
insertion = metrics.insertion(image, target_class=5, steps=50)

# Interactive comparison
viz = InteractiveVisualizer(model)
viz.compare_methods(image, {
    'GradCAM': GradCAM(model),
    'IntegratedGradients': IntegratedGradients(model),
}, save_path="comparison.html")

---

Model Export

from autotimm import ImageClassifier, export_to_torchscript, export_to_onnx
import torch

model = ImageClassifier.load_from_checkpoint("model.ckpt")
example_input = torch.randn(1, 3, 224, 224)

# TorchScript
export_to_torchscript(model, "model.pt", example_input=example_input)

# ONNX
export_to_onnx(model, "model.onnx", example_input=example_input)

Format	Use Case	Runtimes
TorchScript	PyTorch ecosystem, C++, mobile	LibTorch, PyTorch Mobile
ONNX	Cross-platform, hardware-optimized	ONNX Runtime, TensorRT, OpenVINO, CoreML

---

Loading from Checkpoints

All task classes inherit PyTorch Lightning's load_from_checkpoint. Most parameters are auto-restored from saved hyperparameters — only ignored parameters (non-serializable objects) need to be re-supplied.

import autotimm as at

# Basic — works for all tasks (ignored params default to None)
model = at.ImageClassifier.load_from_checkpoint("checkpoint.ckpt")

# For inference on CPU
model = at.ImageClassifier.load_from_checkpoint("checkpoint.ckpt", map_location="cpu")
model.eval()

# Override any saved hyperparameter
model = at.ImageClassifier.load_from_checkpoint("checkpoint.ckpt", num_classes=20)

# Disable torch.compile for faster loading (recommended for inference/export)
model = at.ImageClassifier.load_from_checkpoint("checkpoint.ckpt", compile_model=False)

Ignored parameters by task (not saved in checkpoint, re-supply if needed)

Task	Ignored Parameters
ImageClassifier	metrics, logging_config, transform_config, loss_fn
ObjectDetector	metrics, logging_config, transform_config, cls_loss_fn, reg_loss_fn
SemanticSegmentor	metrics, logging_config, transform_config, class_weights, loss_fn
InstanceSegmentor	metrics, logging_config, transform_config, cls_loss_fn, reg_loss_fn, mask_loss_fn
YOLOXDetector	metrics, logging_config, transform_config

All ignored parameters default to None, so load_from_checkpoint works without passing them. Re-supply only if you need custom losses or metrics for continued training.

# Resume training with custom metrics and loss
model = at.ObjectDetector.load_from_checkpoint(
    "checkpoint.ckpt",
    metrics=[at.MetricConfig(name="map", backend="torchmetrics", metric_class="MeanAveragePrecision")],
    cls_loss_fn="focal",
    reg_loss_fn="giou",
)

Note: backbone, num_classes, and all other constructor arguments are saved as hyperparameters and restored automatically — no need to re-specify them.

---

Architecture

graph TD
    subgraph API["Public API · autotimm"]
        Init["__init__.py"] --> Tasks
        Init --> Data
        Init --> Interpret
        Init --> Export
    end

    subgraph Core["Core · autotimm.core"]
        Backbone["backbone.py · 1000+ timm models"]
        Metrics["metrics.py · torchmetrics"]
        Loggers["loggers.py · TB / MLflow / W&B / CSV"]
        Utils["utils.py · seeding, optimizers"]
        Logging["logging.py · loguru"]
    end

    subgraph Tasks["Tasks · autotimm.tasks"]
        Cls["ImageClassifier"]
        Det["ObjectDetector"]
        Seg["SemanticSegmentor"]
        Inst["InstanceSegmentor"]
        YOLOX["YOLOXDetector"]
    end

    subgraph Data["Data · autotimm.data"]
        DM["DataModules"] --> DS["Datasets"]
        DM --> TX["Transforms"]
    end

    subgraph Interpret["Interpretation · autotimm.interpretation"]
        Methods["6 Methods"] --> Viz["Visualization"]
        Methods --> QM["Quality Metrics"]
    end

    subgraph Export["Export · autotimm.export"]
        JIT["TorchScript"]
        ONNX["ONNX"]
    end

    Tasks --> Core
    Data --> Core

Package Structure

Module	Purpose
autotimm	Public API — all exports via __init__.py
autotimm.core	Backbone factory, metrics, loggers, logging, utilities
autotimm.tasks	Task models (LightningModule subclasses)
autotimm.data	DataModules, datasets, and transform pipelines
autotimm.heads	Task-specific prediction heads (Classification, Detection, FPN, DeepLabV3+, Mask, YOLOX)
autotimm.losses	Loss functions and registry (Focal, GIoU, Dice, Tversky, etc.)
autotimm.models	YOLOX-specific components (CSPDarknet, PAFPN)
autotimm.interpretation	Explanation methods, metrics, visualization, callbacks
autotimm.export	TorchScript and ONNX export utilities + CLIs
autotimm.training	AutoTrainer wrapper with auto-tuning
autotimm.cli	Command-line interface (fit/test/validate + interpretation CLI)
autotimm.callbacks	JSON progress callback for frontend integration

---

Documentation & Examples

Section	Description
Quick Start	Get up and running in 5 minutes
User Guide	In-depth guides for all features
Interpretation Guide	Model explainability and visualization
YOLOX Guide	Complete YOLOX implementation guide
CLI Guide	Train from YAML configs on the command line
API Reference	Complete API documentation
Examples	50+ runnable code examples

Ready-to-Run Examples

Getting Started — classify_cifar10.py · classify_custom_folder.py · vit_finetuning.py

Computer Vision — yolox_official.py · object_detection_yolox.py · semantic_segmentation.py · instance_segmentation.py

HuggingFace Hub — huggingface_hub_models.py · hf_interpretation.py · hf_transfer_learning.py · hf_ensemble.py · hf_deployment.py

Data & Training — csv_classification.py · csv_detection.py · multilabel_classification.py · multi_gpu_training.py

Interpretation — interpretation_demo.py · interpretation_metrics_demo.py · interactive_visualization_demo.py

CLI Configs — classification.yaml · detection.yaml · segmentation.yaml

Browse all examples

---

Testing

# Run all tests
pytest tests/ -v

# Specific modules
pytest tests/test_classification.py
pytest tests/test_yolox.py
pytest tests/test_interpretation.py

# With coverage
pytest tests/ --cov=autotimm --cov-report=html

---

Contributing

Contributions, issues, and feature requests are welcome! See the issues page to get started.

1. Fork the repository
2. Create a feature branch — git checkout -b feat/your-feature
3. Commit your changes — git commit -m 'feat: add your feature'
4. Push to the branch — git push origin feat/your-feature
5. Open a Pull Request

git clone https://github.com/theja-vanka/AutoTimm.git && cd AutoTimm
pip install -e ".[dev,all]"
pytest tests/ -v

---

Citation

@software{autotimm,
  author = {Krishnatheja Vanka},
  title = {AutoTimm: Automatic PyTorch Image Models},
  url = {https://github.com/theja-vanka/AutoTimm},
  year = {2026},
}

---

License

Distributed under the Apache License 2.0. See LICENSE for details.

---

Built with care by Krishnatheja Vanka

If AutoTimm saves you time, consider giving it a ⭐