TFLite Models for MobileBERT for MLPerf Inference

• Application: Question & Answering
• ML Task: MobileBERT
• Framework: TensorFlow (Lite) 2.2
• Training Information: See source for float model @ https://github.com/google-research/google-research/tree/master/mobilebert. The quant model source will be made available soon.
• Quality: Float: 90 F1, Quant: 88 F1
• Precision: Float32, Int8
• Is Quantized: Yes
• Is ONNX: No
• Dataset: Squad v1.1

Additional Model Details:

• Model: Vocab Size: 30k, Sequence Length: 384
• Inputs: input_ids (int32), input_mask (int32), segment_ids (int32)
• Outputs: start_logits, end_logits
• NNAPI Compat: No. Additional work required for the quantized model.