Sentiment Classification with ELECTRA-Small

Sentiment classification finetuned on Movie Review Dataset (IMDB English Dataset, NSMC Korean Dataset). Both English and Korean are supported.

Available models:

1. "Original" TorchScript ELECTRA-Small (53MB)
2. "Original" TFLite ELECTRA-Small (53MB)
3. FP16 post-training-quantized TFLite ELECTRA-Small (26MB)
4. "hybrid" (8-bits precision weights) post-training-quantized TFLite ELECTRA-Small (13MB)

Demo

Most of the assets are from Official Pytorch Android Code. (Tested with Galaxy S10)

📱 APK Download Link 📱

Build the demo app

./gradlew installDebug

Dependencies

1. Android

To convert the original model to tflite format, it has to use select TensorFlow Ops. It results in a bigger binary than the "normal" build but allows compatibility with Transformers architecture.

converter = tf.lite.TFLiteConverter.from_keras_model(model)
converter.target_spec.supported_ops = [tf.lite.OpsSet.SELECT_TF_OPS]

🚨 Using the transformers tflite model, you should build aar file by yourself. (Please check this documentation for tflite ops) 🚨

In this app, I used the same aar file provided from huggingface demo app. (The libs directory contains a custom build of aar.)

dependencies {
    implementation 'org.pytorch:pytorch_android:1.5.0'
    implementation 'org.pytorch:pytorch_android_torchvision:1.5.0'
    // implementation 'org.tensorflow:tensorflow-lite:2.1.0'
    // implementation 'org.tensorflow:tensorflow-lite-select-tf-ops:0.0.0-nightly'
    implementation(name: 'tensorflow-lite-with-select-tf-ops-0.0.0-nightly', ext: 'aar')
}

2. Python

• torch==1.5.0
• transformers==2.9.1
• tensorflow==2.1.0

🚨 Highly recommend to use tensorflow v2.1.0 instead of tensorflow v2.2.0. TF-Lite conversion is not working in tensorflow v2.2.0. (Related Issue) 🚨

Convert the model to TFLite or TorchScript

※ The models are already uploaded on huggingface s3. They will be automatically downloaded during build. If you want to download fp16 or 8bits model, uncomment the line in download.gradle.

🚨 TFLite conversion isn't working on CPU environment, working well with GPU environment. 🚨

You should specify the input shape(=max_seq_len) for model conversion.

# torchscript
$ python3 model_converter/{$TASK_NAME}/jit_compile.py --max_seq_len 40
# tflite (default)
$ python3 model_converter/{$TASK_NAME}/tflite_converter.py --max_seq_len 40
# tflite (fp16)
$ python3 model_converter/{$TASK_NAME}/tflite_converter.py --max_seq_len 40 --model fp16
# tflite (8bits)
$ python3 model_converter/{$TASK_NAME}/tflite_converter.py --max_seq_len 40 --model 8bits

More Details

1. Length & Padding

MAX_SEQ_LEN is set as 40 in this app. You may change this one by yourself.

• You should be cautious about the input shape when you converting the model(--max_seq_len option in python script)
• Also you need to change the MAX_SEQ_LEN in android source code.

private static final int MAX_SEQ_LEN = 40;

• In TFLite, dynamic input size is not supported! (Related Issue) So if the input shape doesn't match with max_seq_len, it crashes:( You should pad the input sequence for tflite model.
• But in torchscipt, even though we specified the input shape when converting the model, variable lengths are also possible. So I didn't pad the sequence for pytorch demo. If you want to pad the sequence for pytorch demo, please change the variable as below.

private static final boolean PAD_TO_MAX_LENGTH = true;

2. FP16 & 8Bits on TFLite

I've already uploaded fp16 and 8bits tflite model on huggingface s3. (English & Korean both)

If you want to use those models, uncomment the line in download.gradle as below. They will be automatically downloaded during gradle build.

task downloadLiteModel {
    def downloadFiles = [
        // "https://s3.amazonaws.com/models.huggingface.co/bert/monologg/koelectra-small-finetuned-sentiment/nsmc_small_fp16.tflite" : "nsmc_small_fp16.tflite",
        // "https://s3.amazonaws.com/models.huggingface.co/bert/monologg/koelectra-small-finetuned-sentiment/nsmc_small_8bits.tflite": "nsmc_small_8bits.tflite",
    ]
}

Also you need to change the MODEL_PATH on Activity.

// 1. fp16
private static final String MODEL_PATH = "imdb_small_fp16.tflite";
// 2. 8bits hybrid
private static final String MODEL_PATH = "imdb_small_8bits.tflite";

3. Slow inference when using TorchScript

At the first time running the inference using torchscript, the inference is quite slow. After the first pass, inference time comes back as normal.

It seems that the first time running the forward might do some preheating work. (Not sure about it...)

• [Related Issue 1], [Related Issue 2], [Related Issue 3], [Related Issue 4]

Reference

• ELECTRA
• KoELECTRA
• Huggingface Transformers
• Pytorch Android Demo
• Huggingface TFLite Android Demo
• TFLite Documentation
• TFLite Java API
• IMDB Dataset
• NSMC Dataset