`timm` is a deep-learning library created by Ross Wightman and is a collection of SOTA computer vision models, layers, utilities, optimizers, schedulers, data-loaders, augmentations and also training/validating scripts with ability to reproduce ImageNet training results.

Install

pip install timm

Or for an editable install,

git clone https://github.com/rwightman/pytorch-image-models
cd pytorch-image-models && pip install -e .

How to use

Create a model

import timm 
import torch

model = timm.create_model('resnet34')
x     = torch.randn(1, 3, 224, 224)
model(x).shape
torch.Size([1, 1000])

It is that simple to create a model using timm. The create_model function is a factory method that can be used to create over 300 models that are part of the timm library.

To create a pretrained model, simply pass in pretrained=True.

pretrained_resnet_34 = timm.create_model('resnet34', pretrained=True)
Downloading: "https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/resnet34-43635321.pth" to /home/tmabraham/.cache/torch/hub/checkpoints/resnet34-43635321.pth

To create a model with a custom number of classes, simply pass in num_classes=<number_of_classes>.

import timm 
import torch

model = timm.create_model('resnet34', num_classes=10)
x     = torch.randn(1, 3, 224, 224)
model(x).shape
torch.Size([1, 10])

List Models with Pretrained Weights

timm.list_models() returns a complete list of available models in timm. To have a look at a complete list of pretrained models, pass in pretrained=True in list_models.

avail_pretrained_models = timm.list_models(pretrained=True)
len(avail_pretrained_models), avail_pretrained_models[:5]
(592,
 ['adv_inception_v3',
  'bat_resnext26ts',
  'beit_base_patch16_224',
  'beit_base_patch16_224_in22k',
  'beit_base_patch16_384'])

There are a total of 271 models with pretrained weights currently available in timm!

Search for model architectures by Wildcard

It is also possible to search for model architectures using Wildcard as below:

all_densenet_models = timm.list_models('*densenet*')
all_densenet_models
['densenet121',
 'densenet121d',
 'densenet161',
 'densenet169',
 'densenet201',
 'densenet264',
 'densenet264d_iabn',
 'densenetblur121d',
 'tv_densenet121']

Fine-tune timm model in fastai

The fastai library has support for fine-tuning models from timm:

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,
    label_func=lambda x: x[0].isupper(), item_tfms=Resize(224))
    
# if a string is passed into the model argument, it will now use timm (if it is installed)
learn = vision_learner(dls, 'vit_tiny_patch16_224', metrics=error_rate)

learn.fine_tune(1)
epoch train_loss valid_loss error_rate time
0 0.201583 0.024980 0.006766 00:08
epoch train_loss valid_loss error_rate time
0 0.040622 0.024036 0.005413 00:10