timm

Warnings

• breaking timm.models.layers module moved to timm.layers in 0.9.x. Direct module imports (import timm.models.layers.module) fail. Only top-level from timm.models.layers import name still works via deprecation shim — which will be removed.
  Fix: Replace from timm.models.layers import X with from timm.layers import X throughout your codebase.
• breaking Model naming changed to architecture.pretrained_tag format in 0.9+. Old names like resnet50_21k still work via deprecation remapping but new weight variants are only accessible via the new format (e.g. resnet50.a1_in1k).
  Fix: Use timm.list_models() to discover available model names. For specific weight variants use the architecture.tag format.
• breaking Pretrained weights now loaded from Hugging Face Hub (https://huggingface.co/timm) not GitHub releases. Old GitHub release URLs hardcoded in custom code will 404.
  Fix: Use timm.create_model(name, pretrained=True) — weight URLs are managed automatically. Do not hardcode weight URLs.
• gotcha timm.create_model with num_classes=0 removes the classifier entirely and returns features. num_classes=None is NOT the same — it keeps the default head. Setting wrong num_classes silently produces wrong output shapes.
  Fix: For feature extraction: num_classes=0 (removes head). For fine-tuning with N classes: num_classes=N (replaces head with random init). Use model.reset_classifier(num_classes=N) to change after creation.
• gotcha Each model has its own expected preprocessing (mean, std, input size). Using generic ImageNet normalization values directly instead of model-specific config produces degraded accuracy.
  Fix: Always use timm's preprocessing utilities: config = timm.data.resolve_data_config({}, model=model); transform = timm.data.create_transform(**config).
• gotcha Not all model variants have pretrained weights — timm lists models without weights too. timm.create_model('some_model', pretrained=True) raises RuntimeError if no weights exist for that variant.
  Fix: Use timm.list_models(pretrained=True) to list only models with available weights. Or check: timm.list_pretrained().

Install

• pip install timm Standard

Imports

• timm.layers

  from timm.layers import PatchEmbed, Mlp, DropPath
  # or
  import timm.layers

  timm.models.layers was moved to timm.layers in 0.9.x. Direct module imports like import timm.models.layers.module no longer work — only the top-level from timm.models.layers import name still works via a deprecation shim.
• create_model

  import timm
  
  # Load with pretrained weights
  model = timm.create_model('resnet50', pretrained=True)
  
  # Load specific weight variant using architecture.tag format
  model = timm.create_model('resnet50.a1_in1k', pretrained=True)
  
  # Custom num_classes for fine-tuning
  model = timm.create_model('efficientnet_b0', pretrained=True, num_classes=10)
  
  # Feature extraction (removes classifier)
  model = timm.create_model('resnet50', features_only=True, pretrained=True)

  Pretrained weights now load from Hugging Face Hub (huggingface.co/timm) not GitHub releases. Model names with architecture.pretrained_tag (e.g. resnet50.a1_in1k) select specific weight variants.

Quickstart

Load pretrained model, apply model-specific preprocessing, run inference.

import timm
import torch
from PIL import Image
from timm.data import resolve_data_config, create_transform

# List available models
print(timm.list_models('resnet*')[:5])

# Load pretrained model
model = timm.create_model('efficientnet_b0.ra_in1k', pretrained=True)
model.eval()

# Get model-specific preprocessing
config = resolve_data_config({}, model=model)
transform = create_transform(**config)

# Inference
img = Image.open('image.jpg').convert('RGB')
tensor = transform(img).unsqueeze(0)

with torch.no_grad():
    output = model(tensor)          # [1, 1000] logits
    probs = torch.softmax(output, dim=1)
    top5 = torch.topk(probs, 5)

# Fine-tune with custom head
model = timm.create_model('resnet50', pretrained=True, num_classes=10)