Torchvision¶

This page describes how to use Torchvision models with LightlyTrain.

Pretrain and Fine-tune a Torchvision Model¶

Pretrain¶

Pretraining Torchvision models with LightlyTrain is straightforward. Below we provide the minimum scripts for pretraining using torchvision/resnet18 as an example:

import lightly_train

if __name__ == "__main__":
    lightly_train.train(
        out="out/my_experiment",                # Output directory.
        data="my_data_dir",                     # Directory with images.
        model="torchvision/resnet18",           # Pass the Torchvision model.
    )

Or alternatively, pass directly a Torchvision model instance:

from torchvision.models import resnet18

import lightly_train

if __name__ == "__main__":
    model = resnet18()                        # Load the Torchvision model.
    lightly_train.train(
        out="out/my_experiment",              # Output directory.
        data="my_data_dir",                   # Directory with images.
        model=model,                          # Pass the Torchvision model.
    )

lightly-train train out="out/my_experiment" data="my_data_dir" model="torchvision/resnet18"

Fine-tune¶

After pretraining, you can load the exported model for fine-tuning with Torchvision:

import torch
from torchvision.models import resnet18

model = resnet18()
state_dict = torch.load("out/my_experiment/exported_models/exported_last.pt")
model.load_state_dict(state_dict, weights_only=True)

Supported Models¶

The following Torchvision models are supported:

  • ResNet

    • torchvision/resnet18

    • torchvision/resnet34

    • torchvision/resnet50

    • torchvision/resnet101

    • torchvision/resnet152

  • ConvNext

    • torchvision/convnext_base

    • torchvision/convnext_large

    • torchvision/convnext_small

    • torchvision/convnext_tiny

  • ShuffleNetV2

    • torchvision/shufflenet_v2_x0_5

    • torchvision/shufflenet_v2_x1_0

    • torchvision/shufflenet_v2_x1_5

    • torchvision/shufflenet_v2_x2_0