background_inference.py

import os
import sys
import utils
import datasets
import eval_utils
from constants import DIALECTS_WITH_LABELS
from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_name = sys.argv[1]
commit_id = sys.argv[2]
inference_function = sys.argv[3]

utils.update_model_queue(
    repo_id=os.environ["PREDICTIONS_DATASET_NAME"],
    model_name=model_name,
    commit_id=commit_id,
    inference_function=inference_function,
    status="in_progress",
)

try:
    tokenizer = AutoTokenizer.from_pretrained(model_name, revision=commit_id)
    model = AutoModelForSequenceClassification.from_pretrained(
        model_name, revision=commit_id
    )

    # Load the dataset
    dataset_name = os.environ["DATASET_NAME"]
    dataset = datasets.load_dataset(dataset_name)["test"]

    sentences = dataset["sentence"]
    labels = {dialect: dataset[dialect] for dialect in DIALECTS_WITH_LABELS}

    predictions = []
    for i, sentence in enumerate(sentences):
        predictions.append(
            getattr(eval_utils, inference_function)(model, tokenizer, sentence)
        )
        print(
            f"Inference progress ({model_name}, {inference_function}): {round(100 * (i + 1) / len(sentences), 1)}%"
        )

    # Store the predictions in a private dataset
    utils.upload_predictions(
        os.environ["PREDICTIONS_DATASET_NAME"],
        predictions,
        model_name,
        commit_id,
        inference_function,
    )

    print(f"Inference completed!")

except Exception as e:
    print(f"An error occurred during inference of {model_name}: {e}")
    utils.update_model_queue(
        repo_id=os.environ["PREDICTIONS_DATASET_NAME"],
        model_name=model_name,
        commit_id=commit_id,
        inference_function=inference_function,
        status="failed (online)",
    )