Test on lengths

app.py

@@ -63,8 +63,6 @@ from moviepy import ImageSequenceClip
 from scipy.io import wavfile
 from avse_code import run_avse
 
-# Load face detector
-model = YOLO("yolov8n-face.pt").cuda()  # assumes CUDA available
 
 
 from decord import VideoReader, cpu
@@ -75,18 +73,18 @@ import spaces
 # Load model once globally
 #ckpt_path = "ckpts/ep215_0906.oat.ckpt"
 #model = AVSEModule.load_from_checkpoint(ckpt_path)
-avse_model = AVSEModule()
 #avse_state_dict = torch.load("ckpts/ep215_0906.oat.ckpt")
-avse_state_dict = torch.load("ckpts/ep220_0908.oat.ckpt")
-avse_model.load_state_dict(avse_state_dict, strict=True)
-avse_model.to("cuda")
-avse_model.eval()
 
 CHUNK_SIZE_AUDIO = 48000  # 3 sec at 16kHz
 CHUNK_SIZE_VIDEO = 75     # 25fps × 3 sec
 
 @spaces.GPU
 def run_avse_inference(video_path, audio_path):
+    avse_model = AVSEModule()
+    avse_state_dict = torch.load("ckpts/ep220_0908.oat.ckpt")
+    avse_model.load_state_dict(avse_state_dict, strict=True)
+    avse_model.to("cuda")
+    avse_model.eval()
     estimated = run_avse(video_path, audio_path)
     # Load audio
     #noisy, _ = sf.read(audio_path, dtype='float32')  # (N, )
@@ -164,6 +162,8 @@ def extract_resampled_audio(video_path, target_sr=16000):
 
 @spaces.GPU
 def yolo_detection(frame, verbose=False):
+    # Load face detector
+    model = YOLO("yolov8n-face.pt").cuda()  # assumes CUDA available
     return model(frame, verbose=verbose)[0]
 
 @spaces.GPU