half bug fixed!

Model/trcaptionnet.py

@@ -40,11 +40,10 @@ class TRCaptionNet(nn.Module):
         # vision encoder
         self.vision_encoder, preprocess = clip.load(config["clip"], jit=False)
         self.vision_encoder.eval()
-        self.vision_encoder = self.vision_encoder.visual
+        self.vision_encoder = self.vision_encoder.visual.float()
         with torch.no_grad():
-            dummy_input_image = preprocess(Image.fromarray(numpy.zeros((512, 512, 3), dtype=numpy.uint8))).to(next(self.parameters()).device).half()
+            dummy_input_image = preprocess(Image.fromarray(numpy.zeros((512, 512, 3), dtype=numpy.uint8))).to(next(self.parameters()).device)
             encoder_output_size = self.vision_encoder(dummy_input_image.unsqueeze(0)).shape[-1]
-        self.vision_encoder = self.vision_encoder.float()
 
         # language decoder
         if not os.path.isfile(config["bert"]):