Add enable_fp16 to face_app

adaface/face_id_to_ada_prompt.py

@@ -603,10 +603,7 @@ class Arc2Face_ID2AdaPrompt(FaceID2AdaPrompt):
         '''
         # Use the same model as ID2AdaPrompt does.
         # FaceAnalysis will try to find the ckpt in: models/insightface/models/antelopev2. 
-        # Note there's a second "model" in the path.        
-        # Note DON'T use CUDAExecutionProvider, as it will hang DDP training. 
-        # Seems when loading insightface onto the GPU, it will only reside on the first GPU. 
-        # Then the process on the second GPU has issue to communicate with insightface on the first GPU, causing hanging.
+        # Note there are two "models" in the path.        
         self.face_app = FaceAnalysis(name='antelopev2', root='models/insightface', 
                                             providers=['CPUExecutionProvider'])
         self.face_app.prepare(ctx_id=0, det_size=(512, 512))
@@ -661,7 +658,7 @@ class Arc2Face_ID2AdaPrompt(FaceID2AdaPrompt):
         else:
             device_id = device.index
             self.face_app = FaceAnalysis(name='antelopev2', root='models/insightface', 
-                                        providers=['CUDAExecutionProvider'],
+                                        providers=['CUDAExecutionProvider', {'enable_fp16': True}],
                                         provider_options=[{"device_id": str(device_id)}])
             self.face_app.prepare(ctx_id=device_id, det_size=(512, 512))
 
@@ -801,12 +798,12 @@ class ConsistentID_ID2AdaPrompt(FaceID2AdaPrompt):
 
         if str(device) == 'cpu':
             self.face_app = FaceAnalysis(name='buffalo_l', root='models/insightface', 
-                                        providers=['CPUExecutionProvider'])
+                                         providers=['CPUExecutionProvider'])
             self.face_app.prepare(ctx_id=0, det_size=(512, 512))
         else:
             device_id = device.index
             self.face_app = FaceAnalysis(name='buffalo_l', root='models/insightface', 
-                                        providers=['CUDAExecutionProvider'],
+                                        providers=['CUDAExecutionProvider', {'enable_fp16': True}],
                                         provider_options=[{"device_id": str(device_id)}])
             self.face_app.prepare(ctx_id=device_id, det_size=(512, 512))