Skip to content

Latest commit

 

History

History
77 lines (56 loc) · 4.4 KB

bert-small-uncased-whole-word-masking-squad-int8-0002.md

File metadata and controls

77 lines (56 loc) · 4.4 KB
Raw

bert-small-uncased-whole-word-masking-squad-int8-0002

Use Case and High-Level Description

This is a small BERT-large like model distilled and quantized to INT8 on SQuAD v1.1 training set from larger BERT-large model (bert-large-uncased-whole-word-masking) provided by the Transformers library) and tuned on SQuAD v1.1 training set. The model performs question answering for English language; the input is a concatenated premise and question for the premise, and the output is the location of the answer to the question inside the premise. For details about the original model, check out BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, HuggingFace's Transformers: State-of-the-art Natural Language Processing.

Tokenization occurs using the BERT tokenizer (see the demo code for implementation details) and the enclosed vocab.txt dictionary file. Input is to be lower-cased before tokenizing.

Specification

Metric Value
GOps 23.9
MParams 41.1
Source framework PyTorch*

Despite number of parameters is the same as for FP32 version of the model the model occupied 4x times less memory due to it uses INT8 weights instead of FP32.

Accuracy

The quality metrics were calculated on the SQuAD v1.1 dataset ("dev" split). Maximum sequence length is 384, input is lower-cased.

Metric Value
F1 91.4%
Exact match (EM) 84.4%

Input

  1. Token IDs, name: input_ids, shape: [1x384]. Token IDs is sequence of integer values that is representing the tokenized premise and question. The sequence structure is as follows ([CLS], [SEP] and [PAD] should be replaced by corresponding token IDs as specified by the dictionary): [CLS] + tokenized question + [SEP] + tokenized premise of the question + [SEP] + ([PAD] tokens to pad to the maximum sequence length of 384)

  2. Input mask, name: attention_mask, shape: [1x384]. Input mask is a sequence of integer values representing the mask of valid values in the input. The values of this input are equal to:

    • 1 at positions corresponding to the [CLS] + tokenized question + [SEP] + tokenized premise of the question + [SEP] part of the input_ids (i.e. all positions except those containing the [PAD] tokens), and
    • 0 at all other positions

  3. Token types, name: token_type_ids, shape: [1x384]. Token types is sequence of integer values representing the segmentation of the input_ids into question and premise. The values are equal to:

    • 1 at positions corresponding to the tokenized premise of the question + [SEP] part of the input_ids, and
    • 0 at all other positions
  • [CLS] is a special symbol added in front of the question.
  • [SEP] is a special separator token inserted between the question and premise of the question.
  • [PAD] is a special token used to fill the rest of the input to get given input length (384).
  1. Position indexes, name: position_ids, shape: [1x384]. Position indexes is sequence of integer values from 0 to 383 representing the position index for each input token. The position_ids is always the same for any input tokens set

Output

The outputs of the net are two 1, 384 floating point-valued logit scores vectors that after soft-max operation are probabilities for start and end positions of answer in the premise of the question.

  1. Start position: name: output_s, shape: [1x384]. Start position is floating point-valued logit scores for start position.

  2. End position: name: output_e, shape: [1x384]. End position is floating point-valued logit scores for end position.

Legal Information

[*] Other names and brands may be claimed as the property of others.

The original bert-large-uncased-whole-word-masking model is taken from Transformers library, which is distributed under the Apache License, Version 2.0.