functions/model.py

import math
import os

import pandas as pd
import torch
import whisper
from jiwer import wer

from zeno import (
    ZenoOptions,
    distill,
    metric,
    model,
    DistillReturn,
    ModelReturn,
    MetricReturn,
)


@model
def load_model(model_path):
    if "sst" in model_path:
        device = torch.device("cpu")
        model, decoder, utils = torch.hub.load(
            repo_or_dir="snakers4/silero-models",
            model="silero_stt",
            language="en",
            device=device,
        )
        (read_batch, _, _, prepare_model_input) = utils

        def pred(df, ops: ZenoOptions):
            files = [os.path.join(ops.data_path, f) for f in df[ops.data_column]]
            input = prepare_model_input(read_batch(files), device=device)
            return ModelReturn(model_output=[decoder(x.cpu()) for x in model(input)])

        return pred

    elif "whisper" in model_path:
        model = whisper.load_model("tiny")

        def pred(df, ops: ZenoOptions):
            files = [os.path.join(ops.data_path, f) for f in df[ops.data_column]]
            outs = []
            for f in files:
                outs.append(model.transcribe(f)["text"])
            return ModelReturn(model_output=outs)

        return pred


@distill
def country(df, ops: ZenoOptions):
    if df["birthplace"][0] == df["birthplace"][0]:
        return DistillReturn(distill_output=[df["birthplace"].str.split(", ")[-1][-1]])
    return DistillReturn(distill_output=[""] * len(df))


@distill
def wer_m(df, ops: ZenoOptions):
    return DistillReturn(
        distill_output=df.apply(
            lambda x: wer(x[ops.label_column], x[ops.output_column]), axis=1
        )
    )


@metric
def avg_wer(df, ops: ZenoOptions):
    avg = df[ops.distill_columns["wer_m"]].mean()
    if pd.isnull(avg) or math.isnan(avg):
        return MetricReturn(metric=0)
    return MetricReturn(metric=avg)