fastai Cheat Sheet

1Install & the 4-line model2.8

pip install fastai★
Pulls in PyTorch. Python 3.10+.
from fastai.vision.all import *★
The .all star-import is idiomatic fastai — brings in the whole application API.
dls = ImageDataLoaders.from_folder(path, valid_pct=0.2, item_tfms=Resize(224))★
Factory methods build DataLoaders fast: from_folder, from_name_func, from_csv, from_df.
learn = vision_learner(dls, resnet34, metrics=error_rate) learn.fine_tune(3)★
Transfer-learn a pretrained backbone in a few epochs. That's a full training run.

2Application learnersone API, many domains

vision_learner(dls, resnet34, metrics=accuracy)★
Image classification/regression on a pretrained CNN or timm backbone (vision_learner(dls, "convnext_tiny")).
tabular_learner(dls, metrics=accuracy) · unet_learner(dls, resnet34)
Tabular deep nets; unet_learner for segmentation.
text_classifier_learner(dls, AWD_LSTM) · language_model_learner(...)
ULMFiT-style text classification & LM fine-tuning.
collab_learner(dls, y_range=(0,5))
Collaborative filtering / recommendation.

3DataBlockflexible pipeline

dblock = DataBlock( blocks=(ImageBlock, CategoryBlock), get_items=get_image_files, splitter=RandomSplitter(0.2, seed=42), get_y=parent_label, item_tfms=Resize(224))★
Declares: input/output types, how to list items, train/valid split, how to label, and transforms.
dls = dblock.dataloaders(path, bs=64)★
Turn the blueprint into DataLoaders. dblock.summary(path) debugs the pipeline step by step.
blocks=(ImageBlock, MaskBlock) · (ImageBlock, BBoxBlock, BBoxLblBlock)
Blocks: CategoryBlock, MultiCategoryBlock, RegressionBlock, MaskBlock, TextBlock, BBox*.
get_y=ColReader("label") · splitter=ColSplitter() · GrandparentSplitter()
Common getters/splitters for DataFrame- or folder-structured data.

4Item vs batch transformsCPU vs GPU

item_tfms=Resize(224) # per-item, on CPU, before batching★
Runs on each item so they're all the same size — required before items can be collated into a batch.
batch_tfms=aug_transforms(mult=1.0) # per-batch, on GPU★
Augmentation applied to whole batches on the GPU (fast). aug_transforms() bundles flips/rotate/zoom/lighting.
batch_tfms=[*aug_transforms(), Normalize.from_stats(*imagenet_stats)]★
Normalize with the pretrained model's stats. RandomResizedCrop as an item_tfms for presizing.
dls.show_batch(max_n=9) · dls.vocab · dls.c★
Inspect a batch; vocab = class names, c = number of classes.

5fine_tune / fit_one_cyclethe schedule

learn.fine_tune(epochs, base_lr=1e-3)★
Transfer learning: trains the head 1 epoch frozen, then unfreezes and trains the whole net with discriminative LRs.
learn.fit_one_cycle(epochs, lr_max=1e-3)★
Train from scratch (or after manual unfreeze) with the one-cycle LR schedule — fastai's default optimizer path.
learn.fit(epochs, lr) # plain fixed-LR fit
The vanilla loop; you'll rarely want it over one-cycle.
learn.recorder.plot_loss() · learn.show_results()
Inspect the training curve and predictions after fitting.

6LR finder & freezingtuning

learn.lr_find()★
Sweep LRs and plot loss — pick a value on the steepest descent. Returns suggestions (valley, slide).
learn.freeze() · learn.unfreeze() · learn.freeze_to(-2)★
fine_tune does this for you; use manually for custom schedules. freeze_to unfreezes the last N groups.
learn.fit_one_cycle(4, lr_max=slice(1e-5, 1e-3))★
Discriminative LRs: a slice applies a low LR to early layers, high to late — the transfer-learning trick.

7Metrics & callbackscustomize the loop

vision_learner(dls, resnet34, metrics=[accuracy, error_rate, F1Score()])★
Metrics print each epoch (validation only). Many built-ins; wrap sklearn with skm_to_fastai.
learn.fit_one_cycle(10, cbs=[EarlyStoppingCallback(patience=3), SaveModelCallback(monitor="valid_loss")])★
Callbacks hook every training event. Common: EarlyStopping, SaveModel, ReduceLROnPlateau, MixedPrecision.
learn = vision_learner(...).to_fp16()★
One call enables mixed-precision training for speed/memory.

8Predict & exportdeploy

pred_class, pred_idx, probs = learn.predict("img.jpg")★
Single-item inference — runs the full pipeline (transforms + model + decode).
preds, targs = learn.get_preds(dl=test_dl)
Batched predictions over a DataLoader; learn.tta() for test-time augmentation.
learn.export("model.pkl")★
Pickles the model plus its inference pipeline (transforms, vocab). The production artifact.
learn = load_learner("model.pkl"); learn.predict(x)★
Reload anywhere for inference — no need to rebuild the DataBlock.

9Interpretationunderstand errors

interp = ClassificationInterpretation.from_learner(learn) interp.plot_confusion_matrix()★
Where the model confuses classes. interp.most_confused(min_val=5) ranks the worst pairs.
interp.plot_top_losses(9)★
The examples the model got most wrong (or right but unsure) — often reveals label errors.
learn.model · learn.summary() · learn.save("stage1") / learn.load("stage1")
The raw nn.Module; save/load persist weights only (vs export = full pipeline).

!Common gotchasread before shipping

item_tfms before batch_tfms
Items must be a uniform size (item_tfms=Resize) before they can be batched for GPU batch_tfms — mixing them up is the #1 DataBlock error.
export vs save
learn.export = model + inference transforms (deploy this). learn.save = weights only (needs the same DataLoaders to reload).
from fastai.vision.all import *
The star-import is expected — many symbols (metrics, transforms) are only convenient that way. Fighting it makes examples not work.
metrics are validation-only
The printed metric is on the validation set; the loss column shows train/valid. A metric needs a validation set to appear.

fastai cheat sheet