Feat: Remove the unnecessary calculation distance

We can direct use the similarities but sort descending as rank

src/ablation_experiment/validate_notebook.py

@@ -24,7 +24,7 @@ def compute_fiq_val_metrics_text_image_combinations(
     beta: float,
 ) -> pd.DataFrame:
     """
-    Compute validation metrics on FashionIQ dataset combining text and image distances.
+    Compute validation metrics on FashionIQ dataset combining text and image similarities.
 
     :param relative_val_dataset: FashionIQ validation dataset in relative mode
     :param blip_text_encoder: BLIP text encoder
@@ -33,14 +33,14 @@ def compute_fiq_val_metrics_text_image_combinations(
     :param image_index_features: validation image index features
     :param image_index_names: validation image index names
     :param combining_function: function that combines features
-    :param beta: beta value for the combination of text and image distances
+    :param beta: beta value for the combination of text and image similarities
     :return: the computed validation metrics
     """
-    all_text_distances = []
+    all_text_similarities = []
     results = []
     target_names = None
 
-    # Compute distances for individual text features
+    # Compute similarities for individual text features
     for text_features, text_names in zip(multiple_text_index_features, multiple_text_index_names):
         # Generate text predictions and normalize features
         predicted_text_features, target_names = generate_fiq_val_predictions(
@@ -55,12 +55,11 @@ def compute_fiq_val_metrics_text_image_combinations(
         text_features = F.normalize(text_features, dim=-1)
         predicted_text_features = F.normalize(predicted_text_features, dim=-1)
 
-        # Compute cosine similarity and convert to distance
+        # Compute cosine similarity
         cosine_similarities = torch.mm(predicted_text_features, text_features.T)
-        distances = 1 - cosine_similarities
-        all_text_distances.append(distances)
+        all_text_similarities.append(cosine_similarities)
 
-    # Normalize and compute distances for image features if available
+    # Normalize and compute similarities for image features if available
     if image_index_features is not None and len(image_index_features) > 0:
         predicted_image_features, _ = generate_fiq_val_predictions(
             blip_text_encoder,
@@ -71,17 +70,17 @@ def compute_fiq_val_metrics_text_image_combinations(
             no_print_output=True,
         )
 
-        # Normalize and compute distances
+        # Normalize and compute similarities
         image_index_features = F.normalize(image_index_features, dim=-1).float()
-        image_distances = 1 - predicted_image_features @ image_index_features.T
+        image_similarities = predicted_image_features @ image_index_features.T
     else:
-        image_distances = torch.zeros_like(all_text_distances[0])
+        image_similarities = torch.zeros_like(all_text_similarities[0])
 
-    # Merge text distances
-    merged_text_distances = torch.mean(torch.stack(all_text_distances), dim=0)
+    # Merge text similarities
+    merged_text_similarities = torch.mean(torch.stack(all_text_similarities), dim=0)
 
-    merged_distances = beta * merged_text_distances + (1 - beta) * image_distances
-    sorted_indices = torch.argsort(merged_distances, dim=-1).cpu()
+    merged_similarities = beta * merged_text_similarities + (1 - beta) * image_similarities
+    sorted_indices = torch.argsort(merged_similarities, dim=-1, descending=True).cpu()
     sorted_index_names = np.array(image_index_names if image_index_names else multiple_text_index_names[0])[
         sorted_indices]
     labels = torch.tensor(
@@ -110,7 +109,7 @@ def compute_fiq_val_metrics_text_image_combinations_clip(
     beta: float,
 ) -> pd.DataFrame:
     """
-    Compute validation metrics on FashionIQ dataset combining text and image distances.
+    Compute validation metrics on FashionIQ dataset combining text and image similarities.
 
     :param relative_val_dataset: FashionIQ validation dataset in relative mode
     :param clip_text_encoder: CLIP text encoder
@@ -120,14 +119,14 @@ def compute_fiq_val_metrics_text_image_combinations_clip(
     :param image_index_features: validation image index features
     :param image_index_names: validation image index names
     :param combining_function: function that combines features
-    :param beta: beta value for the combination of text and image distances
+    :param beta: beta value for the combination of text and image similarities
     :return: the computed validation metrics
     """
-    all_text_distances = []
+    all_text_similarities = []
     results = []
     target_names = None
 
-    # Compute distances for individual text features
+    # Compute similarities for individual text features
     for text_features, text_names in zip(multiple_text_index_features, multiple_text_index_names):
         # Generate text predictions and normalize features
         predicted_text_features, target_names = generate_fiq_val_predictions_clip(
@@ -143,12 +142,11 @@ def compute_fiq_val_metrics_text_image_combinations_clip(
         text_features = F.normalize(text_features, dim=-1)
         predicted_text_features = F.normalize(predicted_text_features, dim=-1)
 
-        # Compute cosine similarity and convert to distance
+        # Compute cosine similarity
         cosine_similarities = torch.mm(predicted_text_features, text_features.T)
-        distances = 1 - cosine_similarities
-        all_text_distances.append(distances)
+        all_text_similarities.append(cosine_similarities)
 
-    # Normalize and compute distances for image features if available
+    # Normalize and compute similarities for image features if available
     if image_index_features is not None and len(image_index_features) > 0:
         predicted_image_features, _ = generate_fiq_val_predictions_clip(
             clip_text_encoder,
@@ -160,17 +158,17 @@ def compute_fiq_val_metrics_text_image_combinations_clip(
             no_print_output=True,
         )
 
-        # Normalize and compute distances
+        # Normalize and compute similarities
         image_index_features = F.normalize(image_index_features, dim=-1).float()
-        image_distances = 1 - predicted_image_features @ image_index_features.T
+        image_similarities = predicted_image_features @ image_index_features.T
     else:
-        image_distances = torch.zeros_like(all_text_distances[0])
+        image_similarities = torch.zeros_like(all_text_similarities[0])
 
-    # Merge text distances
-    merged_text_distances = torch.mean(torch.stack(all_text_distances), dim=0)
+    # Merge text similarities
+    merged_text_similarities = torch.mean(torch.stack(all_text_similarities), dim=0)
 
-    merged_distances = beta * merged_text_distances + (1 - beta) * image_distances
-    sorted_indices = torch.argsort(merged_distances, dim=-1).cpu()
+    merged_similarities = beta * merged_text_similarities + (1 - beta) * image_similarities
+    sorted_indices = torch.argsort(merged_similarities, dim=-1, descending=True).cpu()
     sorted_index_names = np.array(
         image_index_names if image_index_names else multiple_text_index_names[0]
     )[sorted_indices]
@@ -208,7 +206,7 @@ def fiq_val_retrieval_text_image_combinations(
     :param blip_img_encoder: BLIP image model
     :param text_captions: text captions for the FashionIQ dataset
     :param preprocess: preprocess pipeline
-    :param beta: beta value for the combination of text and image distances
+    :param beta: beta value for the combination of text and image similarities
     :param cache: cache dictionary
     :return: DataFrame containing the retrieval metrics for each combination of text features
     """
@@ -332,7 +330,7 @@ def fiq_val_retrieval_text_image_combinations_clip(
     :param clip_tokenizer: CLIP tokenizer
     :param text_captions: text captions for the FashionIQ dataset
     :param preprocess: preprocess pipeline
-    :param beta: beta value for the combination of text and image distances
+    :param beta: beta value for the combination of text and image similarities
     :param cache: cache dictionary
     :return: DataFrame containing the retrieval metrics for each combination of text features
     """

src/ablation_experiment/validate_notebook_cirr.py

@@ -39,13 +39,13 @@ def compute_cirr_val_metrics_text_image_combinations(
     Returns:
         The computed validation metrics
     """
-    all_text_distances = []
+    all_text_similarities = []
     results = []
     reference_names = None
     target_names = None
     group_members = None
 
-    # Compute distances for individual text features
+    # Compute similarities for individual text features
     for text_features, text_names in zip(multiple_text_index_features, multiple_text_index_names):
         # Generate text predictions and normalize features
         predicted_text_features, reference_names, target_names, group_members = generate_cirr_val_predictions(
@@ -60,12 +60,11 @@ def compute_cirr_val_metrics_text_image_combinations(
         text_features = F.normalize(text_features, dim=-1)
         predicted_text_features = F.normalize(predicted_text_features, dim=-1)
 
-        # Compute cosine similarity and convert to distance
+        # Compute cosine similarity
         cosine_similarities = torch.mm(predicted_text_features, text_features.T)
-        distances = 1 - cosine_similarities
-        all_text_distances.append(distances)
+        all_text_similarities.append(cosine_similarities)
 
-    # Normalize and compute distances for image features if available
+    # Normalize and compute similarities for image features if available
     if image_index_features is not None and len(image_index_features) > 0:
         predicted_image_features, _, _, _ = generate_cirr_val_predictions(
             blip_text_encoder,
@@ -76,18 +75,18 @@ def compute_cirr_val_metrics_text_image_combinations(
             no_print_output=True,
         )
 
-        # Normalize and compute distances
+        # Normalize and compute similarities
         image_index_features = F.normalize(image_index_features, dim=-1).float()
-        image_distances = 1 - predicted_image_features @ image_index_features.T
+        image_similarities = predicted_image_features @ image_index_features.T
     else:
-        image_distances = torch.zeros_like(all_text_distances[0])
+        image_similarities = torch.zeros_like(all_text_similarities[0])
 
-    # Merge text distances
-    merged_text_distances = torch.mean(torch.stack(all_text_distances), dim=0)
+    # Merge text similarities
+    merged_text_similarities = torch.mean(torch.stack(all_text_similarities), dim=0)
 
-    merged_distances = beta * merged_text_distances + (1 - beta) * image_distances
+    merged_similarities = beta * merged_text_similarities + (1 - beta) * image_similarities
     # Sort the results
-    sorted_indices = torch.argsort(merged_distances, dim=-1).cpu()
+    sorted_indices = torch.argsort(merged_similarities, dim=-1, descending=True).cpu()
     sorted_index_names = np.array(
         image_index_names if image_index_names else multiple_text_index_names[0]
     )[sorted_indices]
@@ -172,13 +171,13 @@ def compute_cirr_val_metrics_text_image_combinations_clip(
     Returns:
         The computed validation metrics
     """
-    all_text_distances = []
+    all_text_similarities = []
     results = []
     reference_names = None
     target_names = None
     group_members = None
 
-    # Compute distances for individual text features
+    # Compute similarities for individual text features
     for text_features, text_names in zip(multiple_text_index_features, multiple_text_index_names):
         # Generate text predictions and normalize features
         predicted_text_features, reference_names, target_names, group_members = generate_cirr_val_predictions_clip(
@@ -194,12 +193,11 @@ def compute_cirr_val_metrics_text_image_combinations_clip(
         text_features = F.normalize(text_features, dim=-1)
         predicted_text_features = F.normalize(predicted_text_features, dim=-1)
 
-        # Compute cosine similarity and convert to distance
+        # Compute cosine similarity
         cosine_similarities = torch.mm(predicted_text_features, text_features.T)
-        distances = 1 - cosine_similarities
-        all_text_distances.append(distances)
+        all_text_similarities.append(cosine_similarities)
 
-    # Normalize and compute distances for image features if available
+    # Normalize and compute similarities for image features if available
     if image_index_features is not None and len(image_index_features) > 0:
         predicted_image_features, _, _, _ = generate_cirr_val_predictions_clip(
             clip_text_encoder,
@@ -211,18 +209,18 @@ def compute_cirr_val_metrics_text_image_combinations_clip(
             no_print_output=True,
         )
 
-        # Normalize and compute distances
+        # Normalize and compute similarities
         image_index_features = F.normalize(image_index_features, dim=-1).float()
-        image_distances = 1 - predicted_image_features @ image_index_features.T
+        image_similarities = predicted_image_features @ image_index_features.T
     else:
-        image_distances = torch.zeros_like(all_text_distances[0])
+        image_similarities = torch.zeros_like(all_text_similarities[0])
 
-    # Merge text distances
-    merged_text_distances = torch.mean(torch.stack(all_text_distances), dim=0)
+    # Merge text similarities
+    merged_text_similarities = torch.mean(torch.stack(all_text_similarities), dim=0)
 
-    merged_distances = beta * merged_text_distances + (1 - beta) * image_distances
+    merged_similarities = beta * merged_text_similarities + (1 - beta) * image_similarities
     # Sort the results
-    sorted_indices = torch.argsort(merged_distances, dim=-1).cpu()
+    sorted_indices = torch.argsort(merged_similarities, dim=-1, descending=True).cpu()
     sorted_index_names = np.array(
         image_index_names if image_index_names else multiple_text_index_names[0]
     )[sorted_indices]