schmidt-sebastian's picture
Initial upload of MediaPipe-TextClassifier-MobileBERT from MediaPipe Solutions
ed51b39
|
Raw History Blame Contribute Delete
7.73 kB
metadata
license: apache-2.0
library_name: litert
pipeline_tag: text-classification
language:
  - en
tags:
  - mediapipe
  - litert
  - tflite
  - text-classification
  - on-device
  - android
  - ios
  - web
  - python
  - flutter
  - nlp
  - mobilebert
  - sentiment-analysis
  - sst-2
  - dynamic-range
  - arxiv:2004.02984

MediaPipe Text Classifier (MobileBERT)

On-device English sentiment classification model (bert_classifier.tflite, 25.7 MB) trained on Stanford Sentiment Treebank (SST-2) movie reviews for the MediaPipe Text Classifier (TextClassifier) task.

Documentation: MediaPipe Text Classifier Docs

Samples: Android · iOS · Python Colab · Web · Flutter

Try live: Web Demo

MobileBERT sequence classifier fine-tuned on SST-2 (negative, positive), dynamic-range quantized with embedded BERT vocab.txt and labels.txt in TFLITE_METADATA (performing out-of-graph WordPiece tokenization into 128-token word_ids, type_ids, and mask inputs).

Files

File Precision Size SHA-256
bert_classifier.tflite dynamic-range 25.71 MB (25,707,538 B) 9b45012ab143d88d61e10ea501d6c8763f7202b86fa987711519d89bfa2a88b1

Download

pip install -U huggingface_hub
hf download litert-community/MediaPipe-TextClassifier-MobileBERT bert_classifier.tflite --local-dir .

Inputs, Outputs, and Model Specifications

bert_classifier.tflite

  • Embedded TFLITE_METADATA: name MobileBert text classifier · version 1 · author MediaPipe · packed associated files: vocab.txt, labels.txt
Direction Tensor Shape Type Notes
Input serving_default_input_word_ids:0 [1, 128] int32 ids — Tokenized ids of input text as concatenated query and passage.
Input serving_default_input_type_ids:0 [1, 128] int32 segment_ids — 0 for query and 1 for passage tokens.
Input serving_default_input_mask:0 [1, 128] int32 mask — Mask with 1 for real tokens and 0 for padding tokens.
Output StatefulPartitionedCall:0 [1, 2] float32 probability — Probabilities of labels respectively. — Labels: labels.txt

Use with MediaPipe Tasks

Python Quickstart

pip install -U mediapipe huggingface_hub
import mediapipe as mp
from huggingface_hub import hf_hub_download

model_path = hf_hub_download(repo_id="litert-community/MediaPipe-TextClassifier-MobileBERT", filename="bert_classifier.tflite")

BaseOptions = mp.tasks.BaseOptions
TextClassifier = mp.tasks.text.TextClassifier
TextClassifierOptions = mp.tasks.text.TextClassifierOptions

options = TextClassifierOptions(base_options=BaseOptions(model_asset_path=model_path))
with TextClassifier.create_from_options(options) as classifier:
    result = classifier.classify("An unforgettable and genuinely moving film!")
    for cat in result.classifications[0].categories:
        print(f"{cat.category_name}: {cat.score:.3f}")

Cross-Platform Guides & Samples

Latency Benchmarks

End-to-end MediaPipe Task pipeline average latency reported by the Text Classifier documentation:

Model Variant Device CPU Latency GPU Latency
MobileBERT Pixel 6 57.68 ms —

Intended Use, Limitations, and Fairness

  • Intended applications: Binary English sentiment analysis (positive vs. negative) on short text snippets such as reviews and feedback.
  • Out-of-scope & limitations:
    • English only: Not suitable for non-English text.
    • Domain shift: Trained on SST-2 movie reviews (https://nlp.stanford.edu/sentiment/index.html); accuracy may degrade on unrelated domains, sarcasm, or multi-sentence documents.
    • No standalone Google fairness PDF is published for this model on the MediaPipe documentation page.

Provenance

Every model binary in this repository is byte-identical (unmodified FlatBuffer weights and embedded TFLITE_METADATA bytes) to the official artifact served by Google from storage.googleapis.com/mediapipe-models.

Repo File Upstream URL GCS Generation Upstream Last-Modified Pinned Bucket Version
bert_classifier.tflite bert_classifier.tflite 1682480022043445 Wed, 26 Apr 2023 03:33:42 GMT /1/

License and Citation

  • License: apache-2.0 — Stated as Apache License, Version 2.0 (http://www.apache.org/licenses/LICENSE-2.0) in the embedded TFLITE_METADATA block of the .tflite model file(s).

Citation

@inproceedings{sun2020mobilebert,
  title     = {MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices},
  author    = {Sun, Zhiqing and Yu, Hongkun and Song, Xiaodan and Liu, Renjie and Yang, Yiming and Zhou, Denny},
  booktitle = {Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL)},
  year      = {2020},
  url       = {https://arxiv.org/abs/2004.02984}
}