Reduce audio row height (compact-audio CSS); sync dataset_loader

app.py

@@ -68,6 +68,10 @@ with gr.Blocks(
     .gradio-container { max-width: 1400px; }
     /* Results view: force all 4 spec images (Gaussian, Youtube, Ours, NN baseline) to same size */
     #results-column img { width: 700px !important; height: 280px !important; object-fit: contain; }
+    /* Reduce audio player row height (BG/FG/Mix) */
+    .compact-audio .gr-form { min-height: 0 !important; }
+    .compact-audio > div { min-height: 0 !important; max-height: 72px !important; }
+    .compact-audio audio { max-height: 48px !important; }
     """,
 ) as app:
     gr.Markdown("# NearestNeighbor Audio Demo")
@@ -103,9 +107,9 @@ with gr.Blocks(
                 nn_outputs.append(nn_p_md)
                 nn_img = gr.Image(label=f"NN{i+1}", show_label=True, height=480)
                 nn_outputs.append(nn_img)
-                nn_bg = gr.Audio(label="BG", show_label=True)
-                nn_fg = gr.Audio(label="FG", show_label=True)
-                nn_m = gr.Audio(label="Mix", show_label=True)
+                nn_bg = gr.Audio(label="BG", show_label=True, elem_classes=["compact-audio"])
+                nn_fg = gr.Audio(label="FG", show_label=True, elem_classes=["compact-audio"])
+                nn_m = gr.Audio(label="Mix", show_label=True, elem_classes=["compact-audio"])
                 nn_outputs.extend([nn_bg, nn_fg, nn_m])
 
     # ---- Results View: 3 prompts, each with 2 rows (Gaussian|Youtube, Ours|NN baseline) ----
@@ -122,26 +126,26 @@ with gr.Blocks(
                 with gr.Row():
                     with gr.Column():
                         res_outputs.append(gr.Image(label="Gaussian", **spec_size))
-                        res_outputs.append(gr.Audio(label="BG"))
-                        res_outputs.append(gr.Audio(label="FG"))
-                        res_outputs.append(gr.Audio(label="Mix"))
+                        res_outputs.append(gr.Audio(label="BG", elem_classes=["compact-audio"]))
+                        res_outputs.append(gr.Audio(label="FG", elem_classes=["compact-audio"]))
+                        res_outputs.append(gr.Audio(label="Mix", elem_classes=["compact-audio"]))
                     with gr.Column():
                         res_outputs.append(gr.Image(label="Youtube", **spec_size))
-                        res_outputs.append(gr.Audio(label="BG"))
-                        res_outputs.append(gr.Audio(label="FG"))
-                        res_outputs.append(gr.Audio(label="Mix"))
+                        res_outputs.append(gr.Audio(label="BG", elem_classes=["compact-audio"]))
+                        res_outputs.append(gr.Audio(label="FG", elem_classes=["compact-audio"]))
+                        res_outputs.append(gr.Audio(label="Mix", elem_classes=["compact-audio"]))
                 # Row 2: Ours | NN baseline (spec + BG/FG/Mix under each)
                 with gr.Row():
                     with gr.Column():
                         res_outputs.append(gr.Image(label="Ours", **spec_size))
-                        res_outputs.append(gr.Audio(label="BG"))
-                        res_outputs.append(gr.Audio(label="FG"))
-                        res_outputs.append(gr.Audio(label="Mix"))
+                        res_outputs.append(gr.Audio(label="BG", elem_classes=["compact-audio"]))
+                        res_outputs.append(gr.Audio(label="FG", elem_classes=["compact-audio"]))
+                        res_outputs.append(gr.Audio(label="Mix", elem_classes=["compact-audio"]))
                     with gr.Column():
                         res_outputs.append(gr.Image(label="NN baseline", **spec_size))
-                        res_outputs.append(gr.Audio(label="BG"))
-                        res_outputs.append(gr.Audio(label="FG"))
-                        res_outputs.append(gr.Audio(label="Mix"))
+                        res_outputs.append(gr.Audio(label="BG", elem_classes=["compact-audio"]))
+                        res_outputs.append(gr.Audio(label="FG", elem_classes=["compact-audio"]))
+                        res_outputs.append(gr.Audio(label="Mix", elem_classes=["compact-audio"]))
 
     def on_change(sid, view):
         use_dasheng = view in ("Nearest Neighbor (Dasheng)", "Results (Dasheng)")

dataset_loader.py

@@ -248,6 +248,24 @@ def get_results_demo_paths(bid: str, root_prefix: Optional[str] = None) -> dict:
         folders.sort(key=lambda x: x[0])
         return folders
 
+    def _match_dasheng_folder(folder_name: str, folders: list[tuple[str, str]]) -> Optional[tuple[str, str]]:
+        """Match prompt-derived folder_name to actual folder; allow truncated names (e.g. ..._the_autumn -> ..._the)."""
+        if not folder_name or not folders:
+            return None
+        # Exact match
+        for fn, fp in folders:
+            if fn == folder_name:
+                return (fn, fp)
+        # Folder may be truncated: actual fn is prefix of folder_name (e.g. fn="..._the", folder_name="..._the_autumn")
+        candidates = [(fn, fp) for fn, fp in folders if folder_name.startswith(fn)]
+        if candidates:
+            return max(candidates, key=lambda x: len(x[0]))
+        # Or folder_name is prefix of fn
+        candidates = [(fn, fp) for fn, fp in folders if fn.startswith(folder_name)]
+        if candidates:
+            return min(candidates, key=lambda x: len(x[0]))
+        return None
+
     baseline_folders = get_baseline_folders(baseline_inner, baseline_files)
     youtube_folders = get_youtube_folders()
 
@@ -267,12 +285,11 @@ def get_results_demo_paths(bid: str, root_prefix: Optional[str] = None) -> dict:
 
         bl_youtube = {"spec": None, "bg_wav": None, "fg_wav": None, "m_wav": None}
         if use_dasheng:
-            # Dasheng: match by prompt -> folder name (spaces to underscores)
             folder_name = prompt_text.replace(" ", "_") if prompt_text else ""
-            for fn, fp in youtube_folders:
-                if fn == folder_name:
-                    bl_youtube = _collect_block(youtube_files, fp)
-                    break
+            matched = _match_dasheng_folder(folder_name, youtube_folders)
+            if matched:
+                fn, fp = matched
+                bl_youtube = _collect_block(youtube_files, fp)
         else:
             for fn, fp in youtube_folders:
                 if fn.startswith(rel_prefix):
@@ -280,11 +297,26 @@ def get_results_demo_paths(bid: str, root_prefix: Optional[str] = None) -> dict:
                     break
 
         if use_dasheng:
-            # Dasheng: "ours" = prompt-named folder under inner
             folder_name = prompt_text.replace(" ", "_") if prompt_text else ""
-            ours_prefix = f"{inner}/{folder_name}"
-            nn_files = [f for f in files if f.startswith(ours_prefix + "/")]
-            ours_block = _collect_block(nn_files, ours_prefix)
+            # Ours: list prompt-named dirs under inner (exclude baseline, gaussian_baseline, youtube_noise_baseline)
+            skip = {"baseline", "youtube_noise_baseline", "gaussian_baseline"}
+            inner_dirs = set()
+            for f in files:
+                if not f.startswith(inner + "/"):
+                    continue
+                rest = f.replace(inner + "/", "", 1)
+                if "/" in rest:
+                    top = rest.split("/")[0]
+                    if top not in skip and not top.startswith("generated_baseline"):
+                        inner_dirs.add(top)
+            inner_folders = [(d, inner + "/" + d) for d in sorted(inner_dirs)]
+            ours_fn_fp = _match_dasheng_folder(folder_name, inner_folders)
+            if ours_fn_fp:
+                fn, fp = ours_fn_fp
+                nn_files = [f for f in files if f.startswith(fp + "/")]
+                ours_block = _collect_block(nn_files, fp)
+            else:
+                ours_block = {"spec": None, "bg_wav": None, "fg_wav": None, "m_wav": None}
         else:
             nn_files = [f for f in files if f.replace(inner + "/", "").startswith(rel_prefix)]
             ours_block = _collect_block(nn_files, inner + "/" + rel_prefix)