BigARTM v0.7.1 Release notes¶
We are happy to introduce BigARTM v0.7.1, which brings you the following changes:
- BigARTM noteboks — new source of information about BigARTM
- ArtmModel — a brand new Python API
- Much faster retrieval of Phi and Theta matrices from Python
- Much faster dictionary imports from Python
- Auto-detect and use all CPU cores by default
- Fixed Import/Export of topic models (was broken in v0.7.0)
- New capability to implement Phi-regularizers in Python code
- Improvements in Coherence score
Before you upgrade to BigARTM v0.7.1 please review the changes that break backward compatibility.
BigARTM notebooks is your go-to links to read more ideas, examples and other information around BigARTM:
Best thing about ArtmModel is that this API had been designed by BigARTM users. Not by BigARTM programmers. This means that BigARTM finally has a nice, clean and easy-to-use programming interface for Python. Don’t believe it? Just take a look and some examples:
That is cool, right? This new API allows you to load input data from several file formats, infer topic model, find topic distribution for new documents, visualize scores, apply regularizers, and perform many other actions. Each action typically takes one line to write, which allows you to work with BigARTM interactively from Python command line.
ArtmModel exposes most of BigARTM functionality, and it should be sufficiently powerful to cover 95% of all BigARTM use-cases. However, for the most advanced scenarios you might still need to go through the previous API (artm.library). When in doubt which API to use, ask email@example.com — we are there to help!
Coding Phi-regularizers in Python code¶
This is of course one of those very advanced scenarios where you need to go down to the old API :) Take a look at this example:
First one tells how to use Phi regularizers, built into BigARTM. Second one provides a new capability to manipulate Phi matrix from Python. We call this Attach numpy matrix to the model, because this is similar to attaching debugger (like gdb or Visual Studio) to a running application.
To implement your own Phi regularizer in Python you need to to attach to
rwt model from the first example, and update its values.
Fast retrieval of Phi and Theta matrices. In BigARTM v0.7.1 dense Phi and Theta matrices will be retrieved to Python as numpy matrices. All copying work will be done in native C++ code. This is much faster comparing to current solution, where all data is transferred in a large Protobuf message which needs to be deserialized in Python. ArtmModel already takes advantage of this performance improvements.
Fast dictionary import. BigARTM core now supports importing dictionary files from disk, so you no longer have to load them to Python. ArtmModel already take advantage of this performance improvement.
Auto-detect number of CPU cores.
You no longer need to specify
By default BigARTM will detect the number of cores on your machine and load all of them.
num_processors still can be used to limit CPU resources used by BigARTM.
Fixed Import/Export of topic models. Export and Import of topic models will now work. As simple as this:
master.ExportModel("pwt", "file_on_disk.model") master.ImportModel("pwt", "file_on_disk.model")
This will also take care of very large models above 1 GB that does not fit into single protobuf message.
Coherence scores. Ask firstname.lastname@example.org if you are interested :)
Changes in Python methods
From BigARTM v0.7.1 and onwards method
MasterComponent.GetTopicModelof the low-level Python API will return a tuple, where first argument is of type TopicModel (protobuf message), and second argument is a numpy matrix. TopicModel message will keep all fields as usual, except token_weights field which will became empty. Information from token_weights field had been moved to numpy matrix (rows = tokens, columns = topics).
MasterComponent.GetThetaMatrixwill also return a tuple, where first argument is of type ThetaMatrix (protobuf message), and second argument is a numpy matrix. ThetaMatrix message will keep all fields as usual, except item_weights field which will became empty. Information from item_weights field had been moved to numpy matrix (rows = items, columns = topics).
Use the followign syntax to restore the old behaviour:
- MasterComponent.GetTopicModel(use_matrix = False)
- MasterComponent.GetThetaMatrix(use_matrix = False)
This will return a complete protobuf message, without numpy matrix.
Python method ParseCollectionOrLoadDictionary is now obsolete
- Use ParseCollection method to convert collection into a set of batches
- Use MasterComponent.ImportDictionary to load dictionary into BigARTM
- Updated example: example06_use_dictionaries.py