update

app.py

@@ -25,7 +25,7 @@ import numpy as np
 import requests
 from typing import Any, Optional, Tuple
 
-from rfm.data.dataset_types import Trajectory, ProgressSample, PreferenceSample
+from rfm.data.dataset_types import Trajectory, ProgressSample, PreferenceSample, SimilaritySample
 from rfm.evals.eval_utils import build_payload, post_batch_npy
 from rfm.evals.eval_viz_utils import create_combined_progress_success_plot, extract_frames
 from datasets import load_dataset as load_dataset_hf, get_dataset_config_names
@@ -368,7 +368,7 @@ def process_single_video(
         return None, f"Error processing video: {str(e)}"
 
 
-def process_dual_videos(
+def process_two_videos(
     video_a_path: str,
     video_b_path: str,
     task_text: str = "Complete the task",
@@ -376,7 +376,7 @@ def process_dual_videos(
     server_url: str = "",
     fps: float = 1.0,
 ) -> Tuple[Optional[str], Optional[str], Optional[str]]:
-    """Process two videos for preference or similarity prediction using eval server."""
+    """Process two videos for preference, similarity, or progress prediction using eval server."""
     if not server_url:
         return "Please provide a server URL and check connection first.", None, None
 
@@ -497,8 +497,44 @@ def process_dual_videos(
             else:
                 result_text += "Could not extract progress predictions from server response.\n"
 
-        else:  # similarity - not yet implemented in eval server response format
-            result_text = "Similarity prediction not yet supported in eval server response format."
+        elif prediction_type == "similarity":
+            # For similarity inference, we have two videos:
+            # - Video A as reference trajectory
+            # - Video B as similar trajectory
+            # diff_trajectory is None in inference mode (only need similarity between ref and sim)
+            
+            # Create SimilaritySample with Video A as ref and Video B as sim
+            similarity_sample = SimilaritySample(
+                ref_trajectory=trajectory_a,
+                sim_trajectory=trajectory_b,
+                diff_trajectory=None,  # None in inference mode
+                data_gen_strategy="demo",
+            )
+
+            # Build payload and send to server
+            files, sample_data = build_payload([similarity_sample])
+            response = post_batch_npy(server_url, files, sample_data, timeout_s=120.0)
+
+            # Process response - we only care about sim_score_ref_sim (similarity between Video A and Video B)
+            outputs_similarity = response.get("outputs_similarity", {})
+            sim_score_ref_sim = outputs_similarity.get("sim_score_ref_sim", [])
+
+            result_text = f"**Similarity Prediction:**\n"
+            if sim_score_ref_sim and len(sim_score_ref_sim) > 0:
+                sim_score = sim_score_ref_sim[0]
+                if sim_score is not None:
+                    result_text += f"- Similarity score (Video A vs Video B): {sim_score:.3f}\n"
+                    # Interpret similarity score (higher = more similar)
+                    if sim_score > 0.7:
+                        result_text += f"- Interpretation: High similarity - videos are very similar\n"
+                    elif sim_score > 0.4:
+                        result_text += f"- Interpretation: Moderate similarity - videos share some similarities\n"
+                    else:
+                        result_text += f"- Interpretation: Low similarity - videos are quite different\n"
+                else:
+                    result_text += "Could not extract similarity score from server response.\n"
+            else:
+                result_text += "Could not extract similarity prediction from server response.\n"
 
         # Return result text and both video paths
         return result_text, video_a_path, video_b_path
@@ -1292,10 +1328,10 @@ with demo:
         )
 
         analyze_dual_btn.click(
-            fn=process_dual_videos,
+            fn=process_two_videos,
             inputs=[video_a_input, video_b_input, task_text_dual, prediction_type, server_url_input, fps_input_dual],
             outputs=[result_text, video_a_display, video_b_display],
-            api_name="process_dual_videos",
+            api_name="process_two_videos",
         )
 
 

requirements.txt

@@ -26,7 +26,7 @@ watchfiles  # For file watching during development
 
 # RFM package (installed from git repository)
 # For local development, you can also install with: pip install -e ../ (from parent directory)
-git+https://github.com/aliang8/reward_fm.git@d0bfb225f0a8002ef301ea36a6eeadb7becc62d9
+git+https://github.com/aliang8/reward_fm.git@7fd45f3854d45aa297a9873c84e4dc663ef5519e
 
 # Make sure a newer version of gradio is installed
 gradio==4.44.0