v5: target cluster range + caching + average linkage — app_v2.py

app_v2.py

@@ -15,7 +15,7 @@ from sample_extractor import (
     extract_stem, detect_onsets, classify_hits,
     cluster_hits, select_best, synthesize_from_cluster,
     sample_quality_score, export_midi, detect_bpm,
-    render_midi_with_samples, build_archive,
+    render_midi_with_samples, build_archive, cache_clear,
     DEMUCS_MODELS, DEMUCS_STEMS,
 )
 from synth_generator import generate_test_song
@@ -35,7 +35,7 @@ def audio_tuple(a, sr):
 
 def run_extraction(audio_in, stem_choice, demucs_model, demucs_shifts, demucs_overlap,
                    onset_mode, onset_delta, energy_db, pre_pad, min_dur, max_dur, min_gap,
-                   ncc_threshold, ncc_compare_ms,
+                   ncc_threshold, ncc_compare_ms, linkage, target_min, target_max,
                    do_synthesize, progress=gr.Progress()):
     if audio_in is None:
         return [None] * 8
@@ -74,9 +74,11 @@ def run_extraction(audio_in, stem_choice, demucs_model, demucs_shifts, demucs_ov
         progress(0.35, desc="Classifying...")
         hits = classify_hits(hits)
 
-        progress(0.45, desc=f"NCC clustering (threshold={ncc_threshold})...")
+        progress(0.45, desc=f"NCC clustering...")
         clusters = cluster_hits(hits, ncc_threshold=float(ncc_threshold),
-                                 max_compare_ms=float(ncc_compare_ms))
+                                 max_compare_ms=float(ncc_compare_ms),
+                                 target_min=int(target_min), target_max=int(target_max),
+                                 linkage=str(linkage))
 
         progress(0.65, desc="Scoring & selecting best...")
         select_best(clusters)
@@ -261,10 +263,19 @@ def build_app():
 
                 with gr.Accordion("🔗 Clustering", open=False):
                     with gr.Row():
-                        ncc_thresh = gr.Slider(0.5, 0.99, value=0.80, step=0.01,
-                                                label='NCC threshold (higher = stricter matching)')
+                        ncc_thresh = gr.Slider(0.3, 0.99, value=0.80, step=0.01,
+                                                label='NCC threshold (higher = stricter)')
                         ncc_ms = gr.Slider(50, 1000, value=200, step=50,
                                             label='Compare window (ms)')
+                        linkage_dd = gr.Dropdown(['average', 'complete', 'single'],
+                                                  value='average', label='Linkage')
+                    with gr.Row():
+                        target_min = gr.Number(value=0, label='Target min clusters (0 = use threshold)',
+                                               precision=0)
+                        target_max = gr.Number(value=0, label='Target max clusters (0 = use threshold)',
+                                               precision=0)
+                    gr.Markdown("*Set both target min/max > 0 to auto-search for the right threshold. "
+                                "Leave at 0 to use the NCC threshold directly.*")
 
                 with gr.Accordion("⚙️ Post-processing", open=False):
                     do_synth = gr.Checkbox(value=True, label='Synthesize optimal samples from clusters')
@@ -294,7 +305,7 @@ def build_app():
                     run_extraction,
                     [audio_in, stem_dd, demucs_model, demucs_shifts, demucs_overlap,
                      onset_mode, onset_delta, energy_db, pre_pad, min_dur, max_dur, min_gap,
-                     ncc_thresh, ncc_ms, do_synth],
+                     ncc_thresh, ncc_ms, linkage_dd, target_min, target_max, do_synth],
                     [stem_out, summary_md, rendered_out, sample_files,
                      midi_file, archive_file, status_txt, metrics_tbl])