TFLite Models for MobileBERT for MLPerf Inference
Zhiqing Sun · Zenodo (CERN European Organization for Nuclear Research) · 2020
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.