Update modeling_mplug_owl2.py

modeling_mplug_owl2.py

@@ -93,7 +93,7 @@ def optimize_score_map_pytorch_cuda(c, seed=0, original_seed=20020, num_iteratio
 
     for _ in range(num_iterations):
         optimizer.zero_grad()
-        sum_log_diff = torch.sum(c * torch.log(torch.maximum(torch.sigmoid(initial_scores[:, None] - initial_scores), torch.tensor(1e-6, device=device))))
+        sum_log_diff = torch.sum(c * torch.log(torch.maximum(norm_cdf(initial_scores[:, None] - initial_scores), torch.tensor(1e-6, device=device))))
         sum_squares = torch.sum(initial_scores ** 2) / 2
 
         loss = -(sum_log_diff - sum_squares)
@@ -108,7 +108,7 @@ def optimize_score_map_pytorch_cuda(c, seed=0, original_seed=20020, num_iteratio
     
     # Reset the seed
     np.random.seed(original_seed)
-    return scaled_scores[-1]
+    return torch.tensor(scaled_scores[-1], device=device)
 
 def softmax(logits):
     # exp_logits = np.exp(logits - np.max(logits))
@@ -343,7 +343,7 @@ class MPLUGOwl2LlamaForCausalLM(LlamaForCausalLM, MPLUGOwl2MetaForCausalLM):
             return self.download_image(path)
         return Image.open(path).convert('RGB')
     
-    def score(self, image_path):
+    def score(self, image):
         prompt = "USER: <|image|> <|image|> Compared with the first image, what is your quality rating for second image? \nASSISTANT: The quality of the second image is"
         input_ids = tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).to(self.device)
         
@@ -352,7 +352,7 @@ class MPLUGOwl2LlamaForCausalLM(LlamaForCausalLM, MPLUGOwl2MetaForCausalLM):
         probabilities = []
         for index in self.anchor_indices:
             anchor_image = anchor_images[index]
-            image = self.load_image(image_path)
+            # image = self.load_image(image_path)
             images = [anchor_image, image]
             images = [expand2square(img, tuple(int(x*255) for x in self.image_processor.image_mean)) for img in images]
             image_tensor = self.image_processor.preprocess(images, return_tensors='pt')['pixel_values'].half().to(self.device)