[ad_1]
A step-by-step tutorial in PyTorch Lightning
Picture created by the creator utilizing Midjourney.
The Imaginative and prescient Transformer is a strong AI mannequin for picture classification. Launched in 2020, it introduced the environment friendly transformer structure to laptop imaginative and prescient.
In pretraining, an AI mannequin ingests massive quantities of knowledge and learns frequent patterns. The Imaginative and prescient Transformer was pretrained on ImageNet-21K, a dataset of 14 million photos and 21,000 courses.
Satellite tv for pc photos will not be coated in ImageNet-21K, and the Imaginative and prescient Transformer would carry out poorly if utilized out-of-the-box.
Right here, I’ll present you how one can fine-tune a pretrained Imaginative and prescient Transformer on 27,000 satellite tv for pc photos from the EuroSat dataset. We are going to predict land cowl, similar to forests, crops, and industrial areas.
Instance photos from the EuroSAT RGB dataset. Sentinel knowledge is free and open to the general public below EU regulation.
We are going to work in PyTorch Lightning, a deep studying library that builds on PyTorch. Lightning reduces the quantity of code one has to write down, and lets us deal with modeling.
All code is offered on GitHub.
Organising the challenge
The pretrained Imaginative and prescient Transformer is offered on Huggingface. The mannequin structure and weights could be put in from GitHub. We may even want to put in PyTorch Lightning. I used model 2.2.1 for this tutorial, however any model > 2.0 ought to work.
pip set up -q git+https://github.com/huggingface/transformers
pip set up lightning=2.2.1
We will cut up our challenge in 4 steps, which we’ll cowl intimately:
Pretrained Imaginative and prescient Transformer: Lightning ModuleEuroSAT datasetTrain the Imaginative and prescient Transformer on the EuroSAT datasetCalculate the accuracy on the take a look at set
Adapting the Imaginative and prescient Transformer to our dataset
The Imaginative and prescient Transformer from Huggingface is optimized for a subset of ImageNet with 1,000 courses.
[ad_2]
Supply hyperlink
