This example covers the Self supervised learning deep learning task.
Tutorial time: 5 minutes
Start New Run with 1 GPU
Update to run with 8 GPUs
In Self supervised learning (SSL), input data is not provided with labels. Rather, input data is divided into parts where some parts are suppressed with a mask and then model is trained to predict the data that is missing. Self supervised methods for image classification are becoming popular, this tutorial shows an example
The dataset used is CIFAR-10, collection of images in 10 different classes. However there is no need to upload data for this example, the repository includes functions to download
Use the Web Interface, select New Run and choose this script: https://github.com/gridai/lightning-simclr/blob/master/src/train.py
Add script arguments to set batch size, number of workers etc
--batch_size 256 \--num_workers 16 \--exclude_bn_bias \--max_epochs 800
Model starts to train.
Visualize training on the Web interface, see metrics and download checkpoints from the run
Navigate to experiment details and download the artifacts
Run same example with 8 V100, notice how fast model trains
grid run \--instance_type 1_v100_16gb \--gpus 8 \src/train.py \--gpus 8 \--batch_size 256 \--num_workers 16 \--exclude_bn_bias \--max_epochs 800