Visual One
Introducing VEDX: GPT-3 For Vision
Empower your products with computer vision models that can learn complex tasks using only a few samples.
Set up and use our Python API in 5 minutes. Or integrate VEDX into your products easily using our REST API.
A Novel Few Shot Learning Framework
Deep learning models require hundreds or thousands of samples to learn a new task. Collecting and labeling such amount of data can be prohibitively costly and time consuming.
Our brain can learn simple visual tasks, say recognize a specific animal, after seeing only a few examples. It can also learn very complex tasks such as recognizing a specific abnormality in an X-Ray image after seeing 10-20 examples.
We have built a novel Few Shot Learning framework, named VEDX, which is capable of doing just that. For simple tasks, you can train a model using only a few samples and for more complex tasks, you can train a model using 10-20 samples to achieve high accuracy.
VEDX is a versatile framework and can be used to train models for any computer vision task posed as a binary or multi-class classification problem.
Set up, train & apply in 5 min!
from visualone import vedx client = vedx.client(public_key, private_key) # Train a model using a few samples model = client.train(positive_samples, negative_samples) # Apply the trained model to a new image client.predict(model['task_id'], image_file)
Install our Python package: pip install visualone
You can access the docs on our PyPI page.
Submit your email here to receive your public and private keys.
Train a model with a few positive samples and a few negative samples in 10 sec. Then apply the model on a new image and get the result in a fraction of a second.
You can use this IPython notebook to run and evaluate the results quickly.
Integrate via Our REST API
You can use our REST API to train a model and apply it to new images.
To train a model for a new task, upload the training samples into a S3 bucket and send a POST request to our Task Creator endpoint.
To apply the trained model to a new image, upload the image into a S3 bucket and send a GET request to our Predictor endpoint.
For more details, visit our our API docs page.
Case Studies
Dataset from Kaggle.
Best result achieved on Kaggle as reported here:
Accuracy: 82%
Number of samples used for training: 5,216
VEDX accuracy for different number of samples used for training:
| n_positives | n_negatives | Accuracy |
|---|---|---|
| 5 | 5 | 71% |
| 25 | 25 | 81% |
| 50 | 50 | 87% |
| 100 | 100 | 85% |
Summary:
Best accuracy achieved with much fewer samples, 50 samples vs. 5000. Significantly higher accuracy achieved (85% vs. 82%) with 200 training samples.
Pneumonia
Normal
Dataset from Kaggle.
Best result achieved on Kaggle as reported here:
Accuracy: 96%
Number of samples used for training: 2,475
VEDX accuracy for different number of samples used for training:
| n_positives | n_negatives | Accuracy |
|---|---|---|
| 3 | 3 | 86% |
| 5 | 5 | 86% |
| 15 | 15 | 94% |
| 25 | 25 | 100% |
Summary:
Higher accuracy achieved (100% vs. 96%) with significantly lower number of training samples, 50 samples vs. ~2500 samples.
Bacterial Disease
Healthy
Dataset from Kaggle.
Best result achieved on Kaggle as reported here:
Accuracy: 98%
Number of samples used for training: 10,900
VEDX accuracy for different number of samples used for training:
| n_positives | n_negatives | Accuracy |
|---|---|---|
| 3 | 3 | 81% |
| 5 | 5 | 85% |
| 15 | 15 | 98% |
| 25 | 25 | 98% |
| 50 | 50 | 99% |
Summary:
Same accuracy achieved using much fewer samples (30 vs 10,900.)
Good accuracy achieved using only 10 samples.
Rotten
Fresh
You can see more case studies here.
Event Detection
You can see more demos or run a live experiment using your webcam here.