Spaces:
Sleeping
Sleeping
File size: 3,917 Bytes
bdecbe9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 |
import streamlit as st
import os
import tempfile
import pandas as pd
from datasets import load_dataset
from transformers import WhisperProcessor, WhisperForConditionalGeneration
from peft import PeftModel
import torch
import librosa
import numpy as np
import evaluate
st.title("📊 Évaluation WER d'un modèle Whisper")
st.markdown("Ce Space permet d'évaluer la performance WER d'un modèle Whisper sur un dataset audio.")
# Section : Choix du modèle
st.subheader("1. Choix du modèle")
model_option = st.radio("Quel modèle veux-tu utiliser ?", (
"Whisper Large (baseline)",
"Whisper Large + LoRA (SimpleFrog/whisper_finetuned)",
"Whisper Large + LoRA + Post-processing (à venir)"
))
# Section : Lien du dataset
st.subheader("2. Chargement du dataset Hugging Face")
dataset_link = st.text_input("Lien du dataset (format: user/dataset_name)", value="SimpleFrog/mon_dataset")
hf_token = st.text_input("Token Hugging Face (si dataset privé)", type="password")
# Section : Bouton pour lancer l'évaluation
start_eval = st.button("🚀 Lancer l'évaluation WER")
if start_eval:
st.subheader("🔍 Traitement en cours...")
# 🔹 Télécharger dataset
with st.spinner("Chargement du dataset..."):
try:
dataset = load_dataset(dataset_link, split="test", token=hf_token)
except Exception as e:
st.error(f"Erreur lors du chargement du dataset : {e}")
st.stop()
# 🔹 Charger le modèle choisi
with st.spinner("Chargement du modèle..."):
base_model_name = "openai/whisper-large"
model = WhisperForConditionalGeneration.from_pretrained(base_model_name)
if "LoRA" in model_option:
model = PeftModel.from_pretrained(model, "SimpleFrog/whisper_finetuned", token=hf_token)
processor = WhisperProcessor.from_pretrained(base_model_name)
model.eval()
# 🔹 Préparer WER metric
wer_metric = evaluate.load("wer")
results = []
for example in dataset:
try:
audio_path = example["file_name"] # full path or relative path in AudioFolder
reference = example["text"]
# Load audio (we assume dataset is structured with 'file_name')
waveform, _ = librosa.load(audio_path, sr=16000)
waveform = np.expand_dims(waveform, axis=0)
inputs = processor(waveform, sampling_rate=16000, return_tensors="pt")
with torch.no_grad():
pred_ids = model.generate(input_features=inputs.input_features)
prediction = processor.batch_decode(pred_ids, skip_special_tokens=True)[0]
# 🔹 Nettoyage ponctuation pour WER "sans ponctuation"
def clean(text):
return ''.join([c for c in text.lower() if c.isalnum() or c.isspace()]).strip()
ref_clean = clean(reference)
pred_clean = clean(prediction)
wer = wer_metric.compute(predictions=[pred_clean], references=[ref_clean])
results.append({
"Fichier": audio_path,
"Référence": reference,
"Transcription": prediction,
"WER": round(wer, 4)
})
except Exception as e:
results.append({
"Fichier": example.get("file_name", "unknown"),
"Référence": "Erreur",
"Transcription": f"Erreur: {e}",
"WER": "-"
})
# 🔹 Afficher le tableau de résultats
df = pd.DataFrame(results)
st.subheader("📋 Résultats de la transcription")
st.dataframe(df)
mean_wer = df[df["WER"] != "-"]["WER"].mean()
st.markdown(f"### 🎯 WER moyen (sans ponctuation) : `{mean_wer:.3f}`")
# Bloc placeholder pour post-processing à venir
if "Post-processing" in model_option:
st.info("🛠️ Le post-processing sera ajouté prochainement ici...") |