fastai

Use current fastai v2 patterns from docs.fast.ai, PyPI, and the fastai/fastai repo. As of 2026-06-03, PyPI's latest release is fastai 2.8.7, released 2026-02-14, requiring Python >=3.10. Verify PyPI/GitHub again before making version-sensitive recommendations.

Start Here

1. Inspect the local project first: imports, installed fastai version, notebook style, data layout, and whether the user is following a course notebook.
2. Prefer application-level fastai APIs for course, baseline, and standard domain work; drop to DataBlock, mid-level transforms, Learner, callbacks, optimizers, or raw PyTorch only when the task needs that control.
3. In notebooks, follow fastai's domain import style such as from fastai.vision.all import *, from fastai.text.all import *, from fastai.tabular.all import *, or from fastai.collab import *. In production/package code, prefer explicit imports if the repo already avoids wildcard imports.
4. Build the workflow around the fastai loop: create DataLoaders, inspect a batch, create a Learner, find or choose an LR, train/fine-tune, inspect results, then save/export.
5. For production models, require reproducible splits, version pins, dataset checks, appropriate metrics, callback-based checkpointing, a documented inference path, and a security stance for serialized artifacts.
6. When answering Practical Deep Learning for Coders questions, connect the code to the course concepts without replacing the user's experimentation loop.

Default Workflow

Use this shape for most training tasks:

from fastai.vision.all import *