Train, Visualize and Inspect Computer Vision Models on a no-code MLOps platform

https://miro.medium.com/max/1200/0*lYXiiJ2DVbJ34APC

Original Source Here

Training a Mask Detection model with Nexus

Upload images

First things first, I upload the images to the Images section; that part is pretty self-explanatory.

I took out 10 percent of the images and saved them later for use in Portal for inspection to validate my model.

Upload annotations

The dataset came with annotations in PASCAL VOC format, so uploading them to Nexus just requires selecting the right format.

Uploading annotations

Annotate Images

If you don’t have annotations or want to edit and add more annotations, Nexus has a sleek and rich interface you can use. One cool feature is you can label them collaboratively in real-time.

With images and annotations done, you get a nice little summary on the home page.

Distribution of labels, with nice statistics of images and annotations.

Creating the workflow

The visual way of building the workflow is super intuitive. The basic idea is basically:

dataset → augmentations → selecting the model architecture

simple workflow for my mask detection model

For the dataset, I used the default value 0.3 for the train-test split ratio, and you can also choose a specific random seed for reproducibility.

I chose a couple of basic random position augmentations — center crop, horizontal and vertical flip. There are also tons of other options within color space augmentations and switching to advanced mode gives you the option to choose the probability of those augmentations.

For the model, I chose a FasterRCNN ResNet50 640×640 and used the default values of 2 images per batch and 1000 training epochs, so training won’t take too much time since this is only a fun little project.

There’s also a preview augmentations feature, which is useful to test the selected augmentations.

Hyperparameter Tuning

Before the workflow is initiated, you’ll be able to see a preview of all the parameters you selected which I found to be useful as it serves as a last-minute check before the neural network is initialized.

tune hyperparameters before training

Real-time monitoring of training

The cool thing about training a model on Nexus is you see real-time progress of how your model is doing on a graph, and you can stop it with a click of a button if you decide to.

The model training on the NVIDIA k80 engine only took ~30 minutes but there are different GPU settings for you to choose from depending on the size and type of model you’re training.

And once it’s done, you get a nice summary of all the vital metrics regarding your model.

AI/ML

Trending AI/ML Article Identified & Digested via Granola by Ramsey Elbasheer; a Machine-Driven RSS Bot

%d bloggers like this: