Announcing Masterful 0.4.1

We’re excited to announce the v0.4.1 release of Masterful, the training platform for computer vision models!

This release introduces a major new capability: support for Object Detection. The platform now also accepts all the data types that `tf.keras.Model.fit()` accepts. We’ve also updated our comprehensive optimization module to deliver even faster training times. Finally, we’ve expanded our documentation to with many more guides and examples. Read on for more!

Object Detection

You can now train Object Detection models with Masterful!

We married the state-of-the-art research models from the Tensorflow Object Detection API (TFOD) with the ease of training production-grade models with Masterful.

Experienced users of the TFOD know that the config files of the TFOD are unwieldy and scattered, to say the least. It’s confusing enough that the usual answer is to just give up and use whatever defaults are provided by TFOD. But that means you aren’t getting the most out of your model and data because the default regularization and optimization configs will lead to a less accurate model and longer training times.

This is where Masterful helps out - Masterful’s built-in metalearners will automatically figure out the right hyperparameters to get your model as accurate as possible and trained fast. 

Follow the Masterful guide to training object detection models to modernize your object detector training.

Train Your Object Detection Model with Raw, Unlabeled Images

As part of our launch of Object Detection, we also include Semi-Supervised Learning support for Object Detection (as we do for all our supported computer vision tasks: binary classification, single label classification, multi-label classification, object detection, and semantic segmentation).

This is another feature unique to Masterful, based on our proprietary and third-party research. Now, you can improve your object detection model's mAP score using raw, unlabeled images. 

Additional Data Input Types

Although Tensorflow strongly recommends using the `tf.data.Dataset` data type,  for many applications a simpler approach like a Numpy array or Keras Sequence works just as well. Masterful now supports these data input types, as well as Python generators. This means Masterful’s `masterful.training.train()` function supports all of the input types as `tf.keras.Model.fit()`.

Learn more with our tutorial on data types. 

Faster Training with Optimization Metalearner

After extensive testing, we're rolling out our latest version of the Optimization metalearner. We’ve been able to more finely calculate the right learning rate, given the relatively large batch sizes that the batch size metalearner calculates. Some of our own internal testing benchmarks now train in half the time, with no loss of accuracy. We’ll have more to say about this on our blog soon. You'll get the benefit of this on your next call to masterful.optimization.learn_optimization_params().

More Documentation

We have rounded out our documentation with more guides and examples to make it easier than ever to get started with Masterful. Visit www.masterfulai.com/docs to find:

• An installation guide including troubleshooting for many common TF bugs you might run into.
• A Quickstart tutorial and video to get you started with the platform quickly.
• A review of the visualization toolkit. You can run this alongside Tensorboard so you can peek under the hood of Masterful.
• A tutorial on classification, modeled on the Tensorflow Image Classification tutorial. See how to build a state-of-the-art model to classify species of flowers.
• A tutorial on object detection, using the Pascal VOC dataset.
• A tutorial on semantic segmentation, using the Oxford Pets dataset.
• An introduction to the field of SSL, a hot field of research that represents the state-of-the-art method to train models today. The secret is learning from unlabeled data. 
• A two part walk through of combining two complementary SSL techniques available in the Masterful platform to train a production-ready model with minimal data.
• A guide to our Simple SSL recipe, which is a stand-alone product to apply semi-supervised learning within your current training pipeline.

Conclusion

This release is a response to the user feedback we've heard. Our backlog is long but please keep sending your thoughts to us so we can prioritize what you need. And run `pip install —upgrade masterful` to get going with v0.4.1 now.