Update src/streamlit_app.py

src/streamlit_app.py

@@ -1,33 +1,55 @@
-import os
 import torch
 import torchaudio
-import tempfile
+import os
 from transformers import WhisperProcessor, WhisperForConditionalGeneration
 from transformers import AutoTokenizer, AutoModelForSequenceClassification
 import streamlit as st
 
-os.environ["TRANSFORMERS_CACHE"] = "/app/cache"
-os.makedirs("/app/cache", exist_ok=True)
+def load_models():
+    whisper_processor = WhisperProcessor.from_pretrained("openai/whisper-tiny")
+    whisper_model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny")
+    text_model = AutoModelForSequenceClassification.from_pretrained("bert-base-uncased")
+    tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
+    return whisper_processor, whisper_model, text_model, tokenizer
 
-whisper_processor = WhisperProcessor.from_pretrained("openai/whisper-tiny")
-whisper_model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny")
-text_model = AutoModelForSequenceClassification.from_pretrained("bert-base-uncased")
-tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
+whisper_processor, whisper_model, text_model, tokenizer = load_models()
 
-def transcribe(audio_bytes):
-    with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp:
-        tmp.write(audio_bytes)
-        tmp_path = tmp.name
-    waveform, sample_rate = torchaudio.load(tmp_path)
-    input_features = whisper_processor(waveform.squeeze().numpy(), sampling_rate=sample_rate, return_tensors="pt").input_features
+def transcribe(audio_path):
+    waveform, sample_rate = torchaudio.load(audio_path)
+    input_features = whisper_processor(
+        waveform.squeeze().numpy(),
+        sampling_rate=sample_rate,
+        return_tensors="pt"
+    ).input_features
     predicted_ids = whisper_model.generate(input_features)
     transcription = whisper_processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
-    os.remove(tmp_path)
     return transcription
 
 def extract_text_features(text):
     inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
     outputs = text_model(**inputs)
-    return outputs.logits.argmax(dim=1).item()
+    predicted_class = outputs.logits.argmax(dim=1).item()
+    return "Hate Speech" if predicted_class == 1 else "Not Hate Speech"
+
+def predict(audio_file, text_input):
+    if not audio_file and not text_input:
+        return "Please provide either an audio file or some text."
+    if audio_file is not None:
+        audio_path = "temp_audio.wav"
+        with open(audio_path, "wb") as f:
+            f.write(audio_file.read())
+        transcribed_text = transcribe(audio_path)
+        prediction = extract_text_features(text_input or transcribed_text)
+        return f"Predicted: {prediction} \n\n(Transcribed: {transcribed_text})" if not text_input else f"Predicted: {prediction}"
+    elif text_input:
+        prediction = extract_text_features(text_input)
+        return f"Predicted: {prediction}"
+
+st.title("Hate Speech Detector")
+
+uploaded_audio = st.file_uploader("Upload Audio File (.mp3, .wav, .ogg, .flac, .opus)", type=["mp3", "wav", "ogg", "flac", "opus"])
+text_input = st.text_input("Or enter text:")
 
-def predict
+if st.button("Predict"):
+    result = predict(uploaded_audio, text_input)
+    st.success(result)