NER

Available models

• Model Description

• Model:

  • How to train the JulioCP model

Main assumptions

- OLD model

• Framework: [Java-Scala]
  • Pyspark 2.4.4 and spark-nlp 2.5.3
  • Pyspark 3.1.1 and spark-nlp 3.0.2 (This version has been tested by JulioCP)
• Neural Network architecture:
  • LSTMs (Bert model from google)
• Requirements:
  • Training: 500Gb-700Gb RAM to train ~300.000 samples
  • Prediction: Depends on the memory RAM , and therefore, the amount of samples
• Inference:
  • The inference will be performed by a JAR artifact, which call to the model trained (Only tested with: Pyspark 2.4.4 and spark-nlp 2.5.3).
  • This model can not be serialized by ONNX (directly).

- JulioCP model

• Framework: [Python]
  • torch 1.8.1
• Neural Network architecture:
  • Transformers (Bert model from google)
• Requirements:
  • Training: Data will be taken according to the batch which is going to be trained
  • Prediction: Data will be taken according to the batch which is going to be predicted
• Inference:
  • This model can be serialized by ONNX directly.