timm (PyTorch Image Models) Cheat Sheet

1Install & create_model1.x

pip install timm★
Pulls in PyTorch. Weights download from the HF Hub on first use.
import timm m = timm.create_model("resnet50", pretrained=True)★
The one factory for every architecture. Without pretrained=True you get random init.
m = timm.create_model("convnext_base", pretrained=True, num_classes=10)★
num_classes swaps the classifier head to your task; num_classes=0 removes it (feature extractor).
timm.create_model(name, in_chans=1, drop_rate=0.2, drop_path_rate=0.1)
Common knobs: input channels (grayscale/multispectral), classifier & stochastic-depth dropout.

2Find models & weightsdiscover

timm.list_models("*convnext*", pretrained=True)★
Glob-search available architectures; pretrained=True restricts to those with weights.
timm.list_pretrained("resnet50*")
A model can have several pretrained tags (e.g. resnet50.a1_in1k, .tv2_in1k) — different training recipes/datasets.
timm.create_model("vit_base_patch16_224.augreg2_in21k_ft_in1k", pretrained=True)
The model.tag syntax pins an exact weight set — reproducible and explicit.
m.default_cfg · m.pretrained_cfg # input size, mean/std, num_classes...
Every model carries its own config — the source of truth for preprocessing.

5Pooled embeddingsone vector per image

m = timm.create_model("resnet50", pretrained=True, num_classes=0) emb = m(x) # (B, feature_dim) — head removed★
num_classes=0 returns the pooled feature vector — perfect for retrieval, kNN, or a custom head.
m.forward_features(x) · m.forward_head(feat, pre_logits=True)
Split the pass: forward_features = pre-pool feature map; forward_head(..., pre_logits=True) = pooled embedding before the classifier.
m.num_features · m.reset_classifier(num_classes=5, global_pool="avg")
num_features is the embedding size; reset_classifier re-heads a loaded model in place.

6Feature maps (features_only)FPN / detection

m = timm.create_model("resnet50", features_only=True, pretrained=True) feats = m(x) # list of multi-scale feature maps★
Returns intermediate feature maps at several strides — the backbone for FPNs, detectors and segmenters.
timm.create_model(name, features_only=True, out_indices=(2,3,4))
Choose which stages to output. m.feature_info.channels() / .reduction() give channel counts & strides.
m.feature_info.channels() # e.g. [512, 1024, 2048]
Use these to size the neck of your downstream model — no hardcoding.

7Fine-tune / transferadapt to your data

m = timm.create_model("convnext_tiny", pretrained=True, num_classes=N)★
Load pretrained, new head — then train as normal PyTorch (timm gives the model, not the loop).
for p in m.parameters(): p.requires_grad = False for p in m.get_classifier().parameters(): p.requires_grad = True
Linear-probe by freezing the backbone and training only the head; m.get_classifier() returns it.
timm.create_model(name, pretrained=True, img_size=384)
Many ViT/Swin models can be created at a different resolution than they were trained on (weights are interpolated).

8Optimizers, schedulers, EMArecipe helpers

from timm.optim import create_optimizer_v2 opt = create_optimizer_v2(m, opt="adamw", lr=1e-3, weight_decay=0.05)★
timm ships tuned optimizers (Lamb, Lion, AdamW, ...) with sensible no-decay grouping for norm/bias params.
from timm.scheduler import create_scheduler_v2 # cosine w/ warmup
Schedulers step per-epoch and per-update — call scheduler.step(epoch) / step_update() as documented.
from timm.data import Mixup from timm.utils import ModelEmaV3
MixUp/CutMix augmentation and EMA weights — the ingredients behind timm's strong ImageNet recipes.
timm.create_model(name, scriptable=True) · torch.compile(m)
Export/speed: many models are TorchScript-able; torch.compile works broadly.

9Hugging Face Hubload & share

timm.create_model("hf_hub:timm/convnext_base.fb_in22k", pretrained=True)★
Load any timm-compatible checkpoint straight from the Hub by id with the hf_hub: prefix.
timm.models.push_to_hf_hub(m, "me/my-model")
Publish your fine-tuned model (weights + config + model card) to the Hub.
timm.models.load_model_config_from_hf(...) · save via torch.save(m.state_dict())
Configs travel with the weights so create_model reconstructs the exact architecture.

!Common gotchasread before shipping

use resolve_data_config, not fixed mean/std
Different weights use different normalization & input size. Preprocess with the model's own config or accuracy drops silently.
tag = exact weights
A bare name may map to a default tag that changes over releases. Pin model.tag for reproducibility.
features_only isn't supported by every model
Most CNNs support it; some ViT variants need forward_intermediates() instead. Check before relying on it.
timm gives the model, not the loop
There's no .fit() — write the training loop yourself (or use Lightning/HF Trainer with the timm model).

timm cheat sheet