fastai

name: fastai description: Build, debug, explain, review, and productionize fastai deep learning workflows using current fastai v2 docs and source. Use when working with Practical Deep Learning for Coders notebooks, fastai.*.all imports, installation/version checks, DataBlock, Datasets, DataLoaders, Transform, Learner, vision_learner, cnn_learner, unet_learner, language_model_learner, text_classifier_learner, tabular_learner, collab_learner, medical imaging helpers, transfer learning, LR finder, callbacks, metrics, interpretation, mixed precision, distributed training, export/load_learner, or fastai/PyTorch/timm/Accelerate interop.

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 *

path = untar_data(URLs.PETS) / "images"
dls = ImageDataLoaders.from_name_func(
    path, get_image_files(path), valid_pct=0.2, seed=42,
    label_func=lambda x: x[0].isupper(),
    item_tfms=Resize(224),
)

dls.show_batch(max_n=9)

learn = vision_learner(dls, resnet34, metrics=error_rate)
learn.fine_tune(1)
learn.show_results()

Adapt the import, dataloader factory, model factory, metric, and target extraction to the domain. For custom data layouts, use DataBlock instead of forcing a high-level factory.

Read api-levels.md when choosing between high-level factories, DataBlock, mid-level transforms, Learner, callbacks, optimizers, metrics, distributed/mixed precision, or raw PyTorch interop.

Read domain-apis.md when the task depends on a domain surface: vision, segmentation, text/language modeling, tabular, collaborative filtering, medical imaging, or integrations.

Data Rules

• Always make the split explicit enough to be reproducible: RandomSplitter(seed=...), GrandparentSplitter, ColSplitter, or a project-specific splitter.
• Inspect data before training with dls.show_batch(), DataBlock.summary(...), dataframe checks, dls.one_batch(), and label/vocab sanity checks.
• Put item shape normalization in item_tfms and randomized batch augmentation in batch_tfms.
• Keep target extraction small and testable: parent_label, RegexLabeller, ColReader, using_attr, or a named function when logic is non-trivial.
• Match blocks, loss, output shape, and metrics to the target: CategoryBlock for single label, MultiCategoryBlock for multi-label, RegressionBlock for continuous targets, MaskBlock/segmentation loaders for masks.

Training Rules

• Use vision_learner, unet_learner, language_model_learner, text_classifier_learner, tabular_learner, or collab_learner when the domain fits.
• Treat cnn_learner as a deprecated alias for vision_learner: preserve it in existing code if churn is not useful, but do not introduce it in new examples.
• Use learn.lr_find() to choose a learning rate when training is unstable, slow, or user asks how to improve results. Treat suggestions as starting points, not guarantees.
• Use learn.fine_tune(...) for pretrained transfer learning. Use learn.fit_one_cycle(...) for training from scratch or after custom freezing/unfreezing decisions.
• Prefer metrics that match the real task: accuracy/error_rate for balanced classification, BalancedAccuracy/F1/ROC-AUC for imbalanced classification, threshold-aware *Multi metrics for multi-label work, RMSE/MAE/R2 for regression, Dice/Jaccard for segmentation.
• Use SaveModelCallback, EarlyStoppingCallback, logging/tracking callbacks, mixed precision, and distributed training deliberately; do not add them as decoration.
• For PyTorch interop, create fastai DataLoaders when possible. If using plain PyTorch dataloaders, call out which fastai display and interpretation features may be unavailable.

Inference And Deployment

• Use learn.predict(item) for single-item inference and learn.get_preds(...) for batch predictions.
• Use learn.export("export.pkl") only for trusted deployment artifacts. Custom functions/classes used by transforms, models, losses, or metrics must be importable in the deployment environment.
• Warn that load_learner(...) uses pickle semantics and must not load untrusted files. Use Learner.load(...) for weights/checkpoints when a full exported learner is not needed.
• Keep the inference path faithful to training transforms. Do not replace learn.predict(...) with direct learn.model(...) unless preprocessing, decoding, activation, and device handling are explicitly reproduced.
• For service deployment, load the learner/model once at startup, validate inputs, bound batch size and image/text lengths, return calibrated or clearly documented scores, and monitor drift and failures.

Read course-workflows.md for Practical Deep Learning for Coders-oriented recipes, study workflow, common domain templates, and deployment notes.

Read production-models.md when building, reviewing, deploying, or improving a model intended to run beyond a notebook.

Source Selection

• Prefer https://docs.fast.ai/ for current API behavior.
• Prefer PyPI and GitHub releases for latest released version, Python requirement, and release notes.
• Prefer https://github.com/fastai/fastai source when behavior depends on implementation details, signatures, version differences, or docs ambiguity.
• Prefer https://course.fast.ai/ and the fastai book/course notebooks when the user asks in a Practical Deep Learning for Coders context.
• Avoid copying old answers from forums or Stack Overflow without checking against current docs.

Open official-sources.md for the exact docs, repo, course, and source links to consult.