Ludwig

Since 2020, Uber has been developing Ludwig to streamline the use of Deep Learning and make the application of Deep Learning easier for applied projects. These models require comparison among different architectures and fast iteration. Many projects at Uber have used Ludwig and witnessed its value. To name some of the projects, Customer Obsession Ticket Assistant (COTA), information extraction from driver licenses, identification of points of interest during conversation between driver and riders, food delivery time prediction, etc. uses Ludwig. Ludwig has the flexibility to be used on multiple Deep Learning architectures and also it is easy to use.

Ludwig was designed to be a tool for simplifying model development and comparing different processes when dealing with new applied Machine Learning problems. Ludwig draws inspiration from other Machine Learning software like:

from Weka and MLlib it draws the idea of working directly with raw data with some pre-built models,

the declarative programming style for Caffe, it adopts for definition file, from scikit-learn, simplistic API.

It is a powerful tool that offers tensor algebra primitives and other tools for code modelling, but it is also more generalised than other specialised libraries like StanfordNLP, AllenNLP, PyText, and OpenCV due to the combination of multiple inspirations.

Ludwig contains a number of architectures that, when used in concert, may produce an end-to-end model for a particular use case. If Deep Learning libraries are the building bricks, Ludwig offers the structures needed to construct a city. In this comparison, Ludwig may be compared to a city with a variety of structures.

The properties of Ludwig which makes it so robust are:

• Coding skills are not required:Model training and making predictions don’t require coding skills.
• Generality:It is usable across a wide variety of use cases as it has a new data type-based approach to Deep Learning models.
• Flexibility: It is easy for both newcomers and experienced users. The newcomers will find it easy to use and experienced users will have more control over the model building.
• Extensibility: If it is needed to build new model architecture or new feature data types, with Ludwig it is very easy to get it done.
• Understandability: Generally Deep Learning models are considered black boxes, but with Ludwig we can get standard visualization which makes it easier to understand their performance and compare their predictions.

By supplying a tabular data file (perhaps in CSV or excel) with the data and a YAML (Yet Another Markup Language) configuration file that specifies which columns are predictors and which are target variables, we can train a Deep Learning model in Ludwig. Ludwig can produce speedier prototypes thanks to the configuration file's simplicity, which cuts down on the amount of coding time from hours to only a few minutes. Ludwig may concurrently work on all output variables if there are several output variables provided.

The model definition can also contain additional information like preprocessing information for each feature in the dataset, encoder or decoder information, parameter for each encoder or decoder, training parameters, etc. Default values are also provided based on experience or they can be adapted from academic literature at the same time Ludwig also allows the user the ability to set each of the above values in the configuration file. So it is useful for both novices and experts alike. Each model trained in Ludwig gets saved and can be loaded at a later time.

Different input and output features can be combined to accomplish a wide variety of tasks

To create a system of data types with a particular preprocessing function, Ludwig created the idea of data type-specific encoders and decoders. For instance, preparation for text-type data differs from that for image-type data. In a nutshell, encoders translate input into tensors, while decoders translate tensors into output.

This design lets users access combiners that combine all the tensors from input encoders, process them, and provide the processed tensors to output decoders. Ludwig has a default concat combiner that concatenates the outputs of different encoders. Other combiners can also be added and some default combiners are also available for different use cases.

For E.g., we can combine an image encoder with a text decoder to get an image captioning model. By using these data type-specific components, Ludwig can be used on a wide variety of tasks.

Each data type might have more than one encoder and decoder, like a text can be encoded using CNN or RNN or other encoders. The user has the flexibility to mention the encoders to be used and hyper parameters to be used in the model definition file.

Ludwig came up with the concept of data type-specific encoders and decoders to build a system of data types with a specified preprocessing function. For instance, preparation for data of the text type differs from that for data of the picture kind. Encoders convert input into tensors, while decoders convert tensors into output, to put it simply.

At present time Ludwig supports only encoders and decoders for float numbers, categories, sets, binary values, images, text,, etc. along with the ability to load pre-trained models. But soon more data-types are to be added.

In addition to being accessible and having a flexible design, Ludwig offers advantages for non-programmers. A few CLI tools for training, testing, and making predictions are also included with Ludwig. Additionally, this toolkit offers more programmable API. With just a few lines of code, the user can now train and use a model.

Not only that, but additionally it contains model evaluation tools for comparing model performance and predictions by the use of visualization and also model weights and activations of models.

Last but not least, utilising Horovod, an open-source distributed training framework, it also allows us to train models on many GPU locally and in a distributed manner. As a consequence, it becomes simpler to loop over models and get results rapidly.

1. Training the model:
   • YAML model definition file
   • All the changes to be done can be done in this file only.
2. Visualizing the training result:
   • After training, Ludwig creates a result directory containing details about the model.
   • Visualization is done using available visualization tools.
3. Predicting results with trained model:
   • We can predict with new data on old models.

Beginning users, casual users, and experts (such as developers or academics) may all gain a lot from Ludwig. Beginners may quickly train and test the Deep Learning model without writing any code. Experts are allowed to experiment with various methods of model construction and test out novel concepts by modifying features, hyperparameters, encoder-decoders, etc.

A number of new encoders will soon be added to the list of encoders, which currently includes Transformer, BERT, ELMo, and Transformer for text and DenseNET and FractalNet for images. technologies for processing large data volumes, like Petastorm, are also available.

When designing Ludwig, extensibility was taken into account. To get feedback from the community, a developer's guide that shows how simple it is to add additional data-types has also been created. It also discusses how new encoders and decoders might build upon current data types.