timm (PyTorch Image Models)

The timm repository provides implementations for a wide variety of computer vision architectures. Each model implementation is typically located in the timm/models/ directory. Below is a summary of key model families and their implementation files:

Common Architectures

• ResNet / ResNeXt: timm/models/resnet.py (includes variants like Squeeze-and-Excitation and ECAResNet)
• EfficientNet: timm/models/efficientnet.py (includes NoisyStudent, AdvProp, and EdgeTPU variants)
• Vision Transformer (ViT): timm/models/vision_transformer.py
• MobileNet-V3: timm/models/mobilenetv3.py
• Inception family: inception_v3.py, inception_v4.py, inception_resnet_v2.py
• VGG: timm/models/vgg.py
• Xception: timm/models/xception.py

For specific model details, you can refer to the corresponding implementation file in the source code to see the exact architecture and the paper it is based on.

Most timm pretrained models include an identifier in their pretrained tag that corresponds to a specific family or version of hyperparameters (HParams) used during training. You can use these tags to understand the training recipe (optimizer, learning rate schedule, and augmentation strategies) applied to a model.

Common tags include:

• a1h, ah: Based on ResNet Strikes Back A1 recipe (uses LAMB, Cosine with warmup).
• a1, a2, a3: ResNet Strikes Back A{1,2,3} recipes (uses LAMB with BCE loss).
• b1, b2, b1k, b2k: Based on ResNet Strikes Back B recipe (uses RMSProp, Step decay).
• c, c1, c2, c3, ch: Based on ResNet Strikes Back C recipes (uses SGD Nesterov with AGC).
• d, d1, d2: Based on ResNet Strikes Back D recipe (uses AdamW with BCE loss).
• sw: Swin Transformer recipe (basis for DeiT and ConvNeXt; uses AdamW with gradient clipping and EMA).
• ra, ra2, ra3, racm, raa: RandAugment recipes.
• am: AugMix recipe.
• ram: AugMix with RandAugment recipe.
• bt: Bag-of-Tricks recipe.

The results/ directory contains performance metrics for models in the timm collection. Validation scores are provided for models with pretrained weights and ImageNet-1k heads, while benchmark numbers are available for all models.

Available Datasets

• ImageNet Validation (results-imagenet.csv): The standard 50,000 image ImageNet-1k validation set.
• ImageNet-'Real Labels' (results-imagenet-real.csv): The standard validation set but with reassessed labels to correct original annotation errors.
• ImageNetV2 Matched Frequency (results-imagenetv2-matched-frequency.csv): A 10,000 image test set sampled to replicate the original ImageNet curation process.
• ImageNet-Sketch (results-sketch.csv): 50,000 non-photographic images (sketches/doodles) covering all 1000 ImageNet classes.
• ImageNet-Adversarial (results-imagenet-a.csv): 7,500 naturally occurring adversarial images covering 200 classes. For a baseline comparison using the same 200 classes, refer to results-imagenet-a-clean.csv.
• ImageNet-Rendition (results-imagenet-r.csv): 30,000 images representing renditions of 200 ImageNet classes. For a baseline comparison using the same 200 classes, refer to results-imagenet-r-clean.csv.

To use a Res2Net model for inference, load a pretrained variant using timm.create_model, resolve the appropriate data configuration for preprocessing, and apply the transformation to your input image. Ensure you add a batch dimension before passing the tensor to the model.

Common Res2Net variants include:

• res2net101_26w_4s
• res2net50_14w_8s
• res2net50_26w_4s
• res2net50_26w_6s
• res2net50_26w_8s
• res2net50_48w_2s

import timm
import torch
import urllib
from PIL import Image
from timm.data import resolve_data_config
from timm.data.transforms_factory import create_transform

# 1. Load pretrained model
model = timm.create_model('res2net101_26w_4s', pretrained=True)
model.eval()

# 2. Preprocess image
config = resolve_data_config({}, model=model)
transform = create_transform(**config)

url, filename = ("https://github.com/pytorch/hub/raw/master/images/dog.jpg", "dog.jpg")
urllib.request.urlretrieve(url, filename)
img = Image.open(filename).convert('RGB')
tensor = transform(img).unsqueeze(0) # transform and add batch dimension

# 3. Get predictions
with torch.inference_mode():
    out = model(tensor)
probabilities = torch.nn.functional.softmax(out[0], dim=0)