Merge pull request #39 from bandi13/fix_memory

Process only 1k frames at a time to minimize memory footprint

refacer.py

@@ -173,6 +173,12 @@ class Refacer:
         if audio_stream is not None:
             self.video_has_audio = True
         
+    def reface_group(self, faces, frames, output):
+        with ThreadPoolExecutor(max_workers = self.use_num_cpus) as executor:
+            results = list(tqdm(executor.map(self.process_faces, frames), total=len(frames),desc="Processing frames"))
+            for result in results:
+                output.write(result)
+
     def reface(self, video_path, faces):
         self.__check_video_has_audio(video_path)
         output_video_path = os.path.join('out',Path(video_path).name)
@@ -199,15 +205,17 @@ class Refacer:
                     pbar.update()
                 else:
                     break
+                if (len(frames) > 1000):
+                    self.reface_group(faces,frames,output)
+                    frames=[]
+
             cap.release()
             pbar.close()
-        
-        with ThreadPoolExecutor(max_workers = self.use_num_cpus) as executor:
-            results = list(tqdm(executor.map(self.process_faces, frames), total=len(frames),desc="Processing frames"))
-            for result in results:
-                output.write(result)
-            output.release()
 
+        self.reface_group(faces,frames,output)
+        frames=[]
+        output.release()
+        
         return self.__convert_video(video_path,output_video_path)
     
     def __try_ffmpeg_encoder(self, vcodec):
@@ -248,4 +256,4 @@ class Refacer:
          #'h264_vaapi', #Intel HW acceleration
          #'h264_omx', #HW acceleration
          'libx264':'0' #No HW acceleration
-    }
+    }