Pretrained Models

A repository of pretrained models required for reproducing experiments

Torch Hub Instructions

1. Save pretrained models directly using torch.save(<your torch model>, <save path>).
2. Push <save path> to this github repository
3. Add a function <model name> to hubconf.py that loads and returns your model (see the file for a simple example)
4. Load and run models using torch.hub.load('rgklab/pretrained_models', '<model name>', **kwargs)

import torch
model, tf = torch.hub.load('rgklab/pretrained_models', 'resnet18_cifar10', return_transform=True)

Torch Model Zoo Instructions

1. Save pretrained models directly using torch.save(<your torch model>, <save path>).
2. Push <save path> to this github repository
3. Load and run models using torch.utils.model_zoo.load_url(<url_for_raw_file>)

See example below:

import torch
from torchvision.datasets import CIFAR10

# download model
model = torch.utils.model_zoo.load_url(
        'https://github.com/rgklab/pretrained_models/blob/main/detectron/cifar10.pt?raw=true'
      ).cuda()

# download transform
tf = torch.utils.model_zoo.load_url(
      'https://github.com/rgklab/pretrained_models/blob/main/detectron/cifar10_input_transform.pt?raw=true'
    )
data = CIFAR10('/voyager/datasets', transform=tf)

model.eval()
logits = model(data[0][0].unsqueeze(0).cuda())[0]
print(f'Prediction: {data.classes[logits.argmax()]}, Probability: {100*logits.softmax(0).max():.2f}%, True: {data.classes[data[0][1]]}')
## Prediction: frog, Probability: 97.58%, True: frog