Machine Learning models are slowly becoming democratised, as multiple marketplaces are emerging with the aim to decentralise AI development and lower the entry barrier for startups. These marketplaces provide some of the best models for ML development, whether for personal, educational use or for deployment in an enterprise setting.
With a mission to further the use of AI technology in multiple spaces, most of these services also provide opportunities to integrate into existing applications through APIs. Moreover, the models provided by these services are pre-trained, thus removing the need to tweak and train the model to fit individual specifications.
In the startup age, solutions like these provide a welcome break for companies that lack an ML team. An easy deployment, along with low, yet scalable, cost of operation, provide valuable incentives to startups looking to infuse AI-powered processes to their products.
In this article, we will explore four of the top ML marketplaces for accessing models that can be adopted by startups for use in their offerings such as apps.
Algorithmia is one of the first open marketplaces for algorithms. With the rise of ML models, Algorithmia began to provide services that catered to ML enthusiasts. This includes deep learning algorithms that handle tasks like facial recognition and character recognition, that are designed to slot into any application.
Algorithmia provides serverless microservices that utilise CPUs and GPUs hosted by the company. The company charges customers per API call, with calls being based on compute time charged at 1 credit per usage second. This does not include the cost incurred by royalty per call if the developer implements it.
The exchange rate is set at 10,000 credits for $1, but new users are given 50,000 credits upon creating a free account, along with 5000 credits every month. Additionally, there is also an enterprise plan with hosting on Algorithmia’s private cloud. On-premises and air-gapped options are also available.
ModelDepot is an ML marketplace that provides pre-trained weights, so as to avoid training ML models from scratch. It also allows for the deployment of a pre-trained image classification using the REST API. The services are provided in a self-contained Docker image for easy deployment, with transfer learning available for use with custom datasets.
The website offers a free option, wherein up to 10,000 images can be predicted and 20,000 images can be trained. This also has to be hosted by the user, with no option to save a custom model trained state to restore it in another container. However, they do offer a hosted plan, with $1 for every 1,000 predicted images and $2 for every 1,000 trained images. This plan comes with model persistence.
ModelDepot also offers a free trial for some of its selected models, so users can see whether to integrate similar features into their applications. With over 53,000 models available for use, ModelDepot’s offerings can be used in a situation where image recognition, text recognition, and generative models are required.
BigML offers a selection of robustly-engineered Machine Learning algorithms for budding ML researcher. With a single, standardised framework covering the entirety of the service, BigML comes with an easy-to-use web interface and REST API. The service claims to have an API-first approach, as all features are delivered first to the REST API, with libraries available for all popular languages.
The interface includes interactive visualisation for models, with explainability features to ensure clear communication of data. This also allows the user to serve local, offline predictions or deploy immediately. BigML allows for the automatic optimisation for model selection through a service called OptiML. They also provide a domain specific language for automating complex workflows called WhizzML.
The service also comes with team and project management capabilities, with transparency and collaboration at the heart of the experience.
Users are offered unlimited datasets and models for free, no credit card required. Total tasks and storage are also unlimited, with the maximum dataset size restricted to 16MB. On the free plan, users can only run 2 parallel tasks at maximum, with a bevy of options being available to paying customers.
BigML offers ML models for classification, regression and time series forecasting and unsupervised learning tasks such as cluster analysis, anomaly detection, and topic modeling. This makes its products easy for use in fields such as energy, electronics, financial services, food, IoT, media, automotive etc.
Amazon’s AWS Cloud-based services provide multiple ML models for purchase and deployment through the Amazon SageMaker software. Their selection includes algorithms for computer vision, NLP, speech recognition, images, text, structured data, audio, and video. The highlight of choosing AWS Marketplace for machine learning solution is the amazon SageMaker software. The service labels and prepares the data, choose an algorithm, trains it, tunes it and takes action for deployment.
This is also coupled with the Ground Truth feature, which allows users to build and manage highly accurate training datasets quickly. The feature offers access to labelers and provides them with pre-built workflows.
SageMaker automatically configures and optimises models in TensorFlow, Apache MXNet, PyTorch, Chainer, Scikit-learn, SparkML, Horovod, Keras, and Gluon. With over a hundred additional pre-trained models, AWS Marketplace also allows users to bring any other algorithm or framework by building it into a Docker container.
With a cloud-based deployment, users can use the built-in containers for any popular framework or bring their preferred framework for use with the software. The prices charged for the models vary from seller to seller, with Amazon providing cloud-based computing power.