# Training Code Builder

> Generate complete, runnable machine-learning training code in the browser. Free; no sign-in; nothing is uploaded.

URL: https://train-code-builder.skillsafe.ai/

## What it does
- Tasks: image classification, object detection, tabular classification, tabular regression, text classification.
- Frameworks: PyTorch, PyTorch Lightning, Keras 3 (TensorFlow) and scikit-learn (detection: PyTorch and Lightning).
- Models: torchvision ResNet-18/50, EfficientNet-B0, MobileNetV3, ConvNeXt-Tiny, ViT-B/16 or a small CNN; Keras Applications; torchvision Faster R-CNN v2, RetinaNet v2, FCOS, SSDlite; MLPs for tables; Hugging Face Transformers encoders (DistilBERT, BERT, RoBERTa, XLM-R or any model id) for text; TF-IDF linear models, random forests, gradient boosting, SVMs and k-NN in scikit-learn.
- Data formats: a folder per class, CSV, Parquet, a Hugging Face Hub dataset, NumPy arrays, COCO JSON and Pascal VOC XML.
- Training options: epochs, batch size, learning rate, AdamW/Adam/SGD/RMSprop, weight decay, constant/cosine/step/one-cycle/linear-warmup/plateau schedules, augmentation, mixed precision, gradient clipping, label smoothing, early stopping, cross-validation and grid search (scikit-learn).
- Logging and checkpoints: TensorBoard, Weights & Biases, CSV or console; best, last or every-epoch checkpoints.

## Output
- train.py: one readable script; every Config field is also a command-line flag.
- train.ipynb: the same code in explained notebook cells.
- requirements.txt and README.md (data layout, outputs, how to run), or all four as a zip.
- A live, syntax-highlighted preview marks the lines each option change touches. A share link reproduces a setup exactly.

## Free example
"See an example" loads ready setups: cats-vs-dogs ResNet-18 in PyTorch, a churn CSV with gradient boosting, movie-review sentiment with DistilBERT in Lightning, house prices with a Keras MLP, and Faster R-CNN on COCO labels.

## Quality
Every combination of task, framework, data format, model, schedule and logger is generated and compiled with Python before release, and a sample of them is trained for two epochs on tiny synthetic data.

## Licence
The generated code is yours to use, change and share. Pretrained weights the scripts download keep their own licences.