import torch

# === SEMI-WEAKLY SUPERVISED MODELSP RETRAINED WITH 940 HASHTAGGED PUBLIC CONTENT ===
model = torch.hub.load('facebookresearch/semi-supervised-ImageNet1K-models', 'resnet18_swsl')
# model = torch.hub.load('facebookresearch/semi-supervised-ImageNet1K-models', 'resnet50_swsl')
# model = torch.hub.load('facebookresearch/semi-supervised-ImageNet1K-models', 'resnext50_32x4d_swsl')
# model = torch.hub.load('facebookresearch/semi-supervised-ImageNet1K-models', 'resnext101_32x4d_swsl')
# model = torch.hub.load('facebookresearch/semi-supervised-ImageNet1K-models', 'resnext101_32x8d_swsl')
# model = torch.hub.load('facebookresearch/semi-supervised-ImageNet1K-models', 'resnext101_32x16d_swsl')
# ================= SEMI-SUPERVISED MODELS PRETRAINED WITH YFCC100M ==================
# model = torch.hub.load('facebookresearch/semi-supervised-ImageNet1K-models', 'resnet18_ssl')
# model = torch.hub.load('facebookresearch/semi-supervised-ImageNet1K-models', 'resnet50_ssl')
# model = torch.hub.load('facebookresearch/semi-supervised-ImageNet1K-models', 'resnext50_32x4d_ssl')
# model = torch.hub.load('facebookresearch/semi-supervised-ImageNet1K-models', 'resnext101_32x4d_ssl')
# model = torch.hub.load('facebookresearch/semi-supervised-ImageNet1K-models', 'resnext101_32x8d_ssl')
# model = torch.hub.load('facebookresearch/semi-supervised-ImageNet1K-models', 'resnext101_32x16d_ssl')
model.eval()

All pre-trained models expect input images normalized in the same way, i.e. mini-batches of 3-channel RGB images of shape (3 x H x W), where H and W are expected to be at least 224. The images have to be loaded in to a range of [0, 1] and then normalized using mean = [0.485, 0.456, 0.406] and std = [0.229, 0.224, 0.225].

Here’s a sample execution.

# Download an example image from the pytorch website
import urllib
url, filename = ("https://github.com/pytorch/hub/raw/master/dog.jpg", "dog.jpg")
try: urllib.URLopener().retrieve(url, filename)
except: urllib.request.urlretrieve(url, filename)

# sample execution (requires torchvision)
from PIL import Image
from torchvision import transforms
input_image = Image.open(filename)
preprocess = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
input_tensor = preprocess(input_image)
input_batch = input_tensor.unsqueeze(0) # create a mini-batch as expected by the model

# move the input and model to GPU for speed if available
if torch.cuda.is_available():
    input_batch = input_batch.to('cuda')
    model.to('cuda')

with torch.no_grad():
    output = model(input_batch)
# Tensor of shape 1000, with confidence scores over Imagenet's 1000 classes
print(output[0])
# The output has unnormalized scores. To get probabilities, you can run a softmax on it.
print(torch.nn.functional.softmax(output[0], dim=0))

Model Description

This project includes the semi-supervised and semi-weakly supervised ImageNet models introduced in “Billion-scale Semi-Supervised Learning for Image Classification” https://arxiv.org/abs/1905.00546.

“Semi-supervised” (SSL) ImageNet models are pre-trained on a subset of unlabeled YFCC100M public image dataset and fine-tuned with the ImageNet1K training dataset, as described by the semi-supervised training framework in the paper mentioned above. In this case, the high capacity teacher model was trained only with labeled examples.

“Semi-weakly” supervised (SWSL) ImageNet models are pre-trained on 940 million public images with 1.5K hashtags matching with 1000 ImageNet1K synsets, followed by fine-tuning on ImageNet1K dataset. In this case, the associated hashtags are only used for building a better teacher model. During training the student model, those hashtags are ingored and the student model is pretrained with a subset of 64M images selected by the teacher model from the same 940 million public image dataset.

Semi-weakly supervised ResNet and ResNext models provided in the table below significantly improve the top-1 accuracy on the ImageNet validation set compared to training from scratch or other training mechanisms introduced in the literature as of September 2019. For example, We achieve state-of-the-art accuracy of 81.2% on ImageNet for the widely used/adopted ResNet-50 model architecture.

Architecture Supervision #Parameters FLOPS Top-1 Acc. Top-5 Acc.
ResNet-18 semi-supervised 14M 2B 72.8 91.5
ResNet-50 semi-supervised 25M 4B 79.3 94.9
ResNeXt-50 32x4d semi-supervised 25M 4B 80.3 95.4
ResNeXt-101 32x4d semi-supervised 42M 8B 81.0 95.7
ResNeXt-101 32x8d semi-supervised 88M 16B 81.7 96.1
ResNeXt-101 32x16d semi-supervised 193M 36B 81.9 96.2
ResNet-18 semi-weakly supervised 14M 2B 73.4 91.9
ResNet-50 semi-weakly supervised 25M 4B 81.2 96.0
ResNeXt-50 32x4d semi-weakly supervised 25M 4B 82.2 96.3
ResNeXt-101 32x4d semi-weakly supervised 42M 8B 83.4 96.8
ResNeXt-101 32x8d semi-weakly supervised 88M 16B 84.3 97.2
ResNeXt-101 32x16d semi-weakly supervised 193M 36B 84.8 97.4

Citation

If you use the models released in this repository, please cite the following publication (https://arxiv.org/abs/1905.00546).

@misc{yalniz2019billionscale,
    title={Billion-scale semi-supervised learning for image classification},
    author={I. Zeki Yalniz and Hervé Jégou and Kan Chen and Manohar Paluri and Dhruv Mahajan},
    year={2019},
    eprint={1905.00546},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}