Text Models
There are a number of built-in compiled text models in Analytics Zoo TFPark for Natural Language Processing (NLP) tasks based on KerasModel.
See this page for more details about how to construct built-in models for intent extraction, named entity extraction and pos tagging. etc.
In this page, we show the general steps how to train and evaluate an NER model in a distributed fashion and use this model for distributed inference. For other models, the steps are more or less quite similar.
Remarks:
- You need to install tensorflow==1.15.0 on your driver node.
- Your operating system (OS) is required to be one of the following 64-bit systems: Ubuntu 16.04 or later, macOS 10.12.6 or later and Windows 7 or later.
- To run on other systems, you need to manually compile the TensorFlow source code. Instructions can be found here.
Model Construction
You can easily construct a model for named entity recognition using the following API.
from zoo.tfpark.text.keras import NER
model = NER(num_entities, word_vocab_size, char_vocab_size, word_length)
Data Preparation
The NER model has two inputs: word indices and character indices.
Thus, each raw text record needs to go through word-wise tokenization, character-wise segmentation and alignment to the same target length for preprocessing.
If you are using numpy arrays, then the input x
should be a list of two numpy arrays:
x_words
of shape (batch, sequence_length) for word indicesx_chars
of shape (batch, sequence_length, word_length) for character indices.x = [x_words, x_char]
If there are labels (for training and evaluation), y
should be another numpy array of shape (batch, sequence_length, word_length) for entity tags.
Alternatively, you can construct a TFDataSet directly if you are dealing with RDD. Each record in TFDataSet should contain word indices, character indices and labels (if any) as well.
Model Training
You can easily call fit to train the NER model in a distributed fashion. You don't need to specify y
if x
is already a TFDataSet.
model.fit(x, y, batch_size, epochs, distributed=True)
Model Evaluation
You can easily call evaluate to evaluate the NER model in a distributed fashion. You don't need to specify y
if x
is already a TFDataSet.
result = model.evaluate(x, y, distributed=True)
Model Save and Load
After training, you can save the NER
model to a single HDF5 file.
model.save_model(path)
For inference, you can load a directly trained NER
model (with weights) from HDF5 file.
from zoo.tfpark.text.keras import NER
model = NER.load_model(path)
Model Inference
You can easily call predict to use the trained NER model for distributed inference. Note that you don't necessarily need labels for prediction.
predictions = model.predict(x, distributed=True)