CPU

app.py

@@ -1,4 +1,4 @@
-import spaces
+# import spaces
 import gradio as gr
 import torchaudio
 import pandas as pd
@@ -16,8 +16,8 @@ import torch
 from models.models import BiFormer
 
 
-DEVICE = torch.device('cuda')
-# DEVICE = torch.device('cpu')
+# DEVICE = torch.device('cuda')
+DEVICE = torch.device('cpu')
 
 # Configure logging
 logging.basicConfig(level=logging.INFO)
@@ -150,7 +150,7 @@ def get_top_emotion(probabilities):
     return f"{LABEL_TO_EMOTION[max_idx]} ({max(probabilities)*100:.1f}%)"
 
 
-@spaces.GPU
+# @spaces.GPU
 def process_audio(audio_path):
     """Main processing pipeline."""
     try:
@@ -312,6 +312,5 @@ def create_demo():
 
 
 if __name__ == "__main__":
-    print('START DEMO')
     demo = create_demo()
     demo.launch()

config.toml

@@ -54,7 +54,7 @@ text_column = "text"
 whisper_model = "base"
 
 # Указываем, где запускать Whisper: "cuda" (GPU) или "cpu"
-whisper_device = "cuda"
+whisper_device = "cpu"
 
 # Если для dev/test в CSV нет текста, нужно ли всё же вызывать Whisper?
 use_whisper_for_nontrain_if_no_text = true
@@ -120,7 +120,7 @@ audio_embedding_dim = 256  # размерность аудио-эмбеддин
 text_embedding_dim = 1024   # размерность текст-эмбеддинга
 emb_normalize = false  # нормализовать ли вектор L2-нормой
 max_tokens = 95         # ограничение на длину текста (токенов) при токенизации
-device = "cuda"          # "cuda" или "cpu", куда грузить модель
+device = "cpu"          # "cuda" или "cpu", куда грузить модель
 
 # audio_pooling = "mean"        # "mean", "cls", "max", "min", "last", "attention"
 # text_pooling = "cls"          # "mean", "cls", "max", "min", "last", "sum", "attention"

data_loading/feature_extractor.py

@@ -14,8 +14,8 @@ from transformers import (
 from data_loading.pretrained_extractors import EmotionModel, get_model_mamba, Mamba
 
 
-DEVICE = torch.device('cuda')
-# DEVICE = torch.device('cpu')
+# DEVICE = torch.device('cuda')
+DEVICE = torch.device('cpu')
 
 
 class PretrainedAudioEmbeddingExtractor:

data_loading/pretrained_extractors.py

@@ -14,8 +14,8 @@ from einops import rearrange, einsum
 from torch import Tensor
 from einops import rearrange
 
-DEVICE = torch.device('cuda')
-# DEVICE = torch.device('cpu')
+# DEVICE = torch.device('cuda')
+DEVICE = torch.device('cpu')
 
 ## Audio models