Evaluation/Prediction Performance enhancements and ONNX support

Mixed Precision Support for evaluation and prediction

Mixed precision (fp16) inference is now supported for evaluation and prediction in the following models:

• ClassificationModel
• ConvAI
• MultiModalClassificationModel
• NERModel
• QuestionAnsweringModel
• Seq2Seq
• T5Model

You can disable fp16 by setting fp16 = False in the model_args.

Multi-GPU support for evaluation and prediction

Set the number of GPUs with n_gpu. in model_args Currently supported in the following models:

• ClassificationModel
• ConvAI
• MultiModalClassificationModel
• NERModel
• QuestionAnsweringModel
• Seq2Seq
• T5Model

Native ONNX support for Classification and NER tasks (Beta)

Please note that ONNX support is still experimental.

See docs for details.