How one can Use ResNet-50

Launched within the paper “Deep Residual Studying for Picture Recognition” in 2015, ResNet-50 is an picture classification structure developed by Microsoft Analysis. The default ResNet50 checkpoint was educated on the ImageNet-1k dataset, which comprises information on 1,000 lessons of photos.

On this information, we’re going to stroll by means of methods to set up ResNet-50 classify photos utilizing ResNet-50.

By the tip of this information, we could have code that assigns the category “forklift” to the next picture:

With out additional ado, let’s get began!

What’s ResNet-50?

ResNet-50 is a picture classification mannequin structure. Launched in 2015, ResNet-50 received first place on the ILVRC 2015 picture classification job. Whereas many new mannequin architectures that obtain robust efficiency have since been launched, ResNet-50 continues to be a notable structure within the historical past of pc imaginative and prescient.

The default ResNet checkpoint can determine any of 1,000 lessons within the ImageNet-1k dataset.

How one can Set up ResNet-50

You’ll be able to set up ResNet-50 utilizing the HuggingFace Transformers Python bundle.

To get began, first set up Transformers:

pip set up transformers

After you have put in Transformers, you’ll be able to load the microsoft/resnet-50 mannequin in your code with the ResNetForImageClassification information loader.

How one can Use ResNet-50

To get began, create a brand new Python file and add the next code:

from transformers import AutoImageProcessor, ResNetForImageClassification
import torch
from datasets import load_dataset
from PIL import Picture picture = Picture.open(“picture.jpg”) processor = AutoImageProcessor.from_pretrained("microsoft/resnet-50")
mannequin = ResNetForImageClassification.from_pretrained("microsoft/resnet-50") inputs = processor(picture, return_tensors="pt") with torch.no_grad():
    logits = mannequin(**inputs).logits predicted_label = logits.argmax(-1).merchandise()
print(mannequin.config.id2label[predicted_label])

On this code, we first open a picture referred to as picture.jpg. Then, we load our mannequin. We run inference on our mannequin with the mannequin(**inputs) perform name. Lastly, we retrieve the category with the best confidence returned by our mannequin.

Within the code above, change picture.jpg with the identify of the picture on which you need to run inference.

Contemplate the next picture of a forklift:

After we run the picture by means of ResNet, the mannequin returns “forklift”.

Conclusion and The Present Classification Panorama

ResNet-50 is a picture classification structure launched in 2015 and was educated on the ImageNet-1k dataset. You’ll be able to practice fashions on a customized dataset utilizing the ResNet structure if you wish to determine your individual lessons.

Whereas ResNet is a number of years previous, the mannequin is established as a picture classification mannequin. Since then, many new architectures have been launched that can help you fine-tune a mannequin on a customized dataset, together with:

1. The Imaginative and prescient Transformer
2. FastViT
3. Ultralytics YOLOv8 
4. ResNext

There are additionally zero-shot classification fashions the place you should utilize the mannequin on arbitrary lessons with out fine-tuning fashions.

For instance, you should utilize OpenAI CLIP to assign labels to pictures with out fine-tuning the mannequin. It’s because CLIP has been educated on a big dataset with a variety of descriptions.

Zero-shot fashions like CLIP can be utilized on their very own (i.e. for classification, content material moderation), or used to auto-label framework like Autodistill to be used in coaching a sooner, fine-tuned imaginative and prescient mannequin.