Add files using upload-large-folder tool

checkpoints/spatial_beats_ov1_local_spatial_v7j_ov123_exp/03_ov123_top4/val_predictions/epoch_0001_csv/valid__hm3d__00034-6imZUJGRUq4__000000-foa__132991__gt.csv

@@ -0,0 +1,34 @@
+frame_idx,frame_time_s,src_or_track_idx,class_idx,class_name,azimuth_deg,elevation_deg,distance_m,activity_prob
+0,0.0,0,41,singing,121.266,-42.349,2.338,1.0
+1,0.4041,0,41,singing,121.266,-42.349,2.338,1.0
+2,0.8083,0,41,singing,121.266,-42.349,2.338,1.0
+3,1.2124,0,41,singing,121.266,-42.349,2.338,1.0
+4,1.6166,0,41,singing,121.266,-42.349,2.338,1.0
+5,2.0207,0,41,singing,121.266,-42.349,2.338,1.0
+6,2.4248,0,41,singing,121.266,-42.349,2.338,1.0
+7,2.829,0,41,singing,121.266,-42.349,2.338,1.0
+8,3.2331,0,41,singing,121.266,-42.349,2.338,1.0
+9,3.6372,0,41,singing,121.266,-42.349,2.338,1.0
+10,4.0414,0,41,singing,121.266,-42.349,2.338,1.0
+11,4.4455,0,41,singing,121.266,-42.349,2.338,1.0
+12,4.8497,0,41,singing,121.266,-42.349,2.338,1.0
+13,5.2538,0,41,singing,121.266,-42.349,2.338,1.0
+14,5.6579,0,41,singing,121.266,-42.349,2.338,1.0
+15,6.0621,0,41,singing,121.266,-42.349,2.338,1.0
+16,6.4662,0,41,singing,121.266,-42.349,2.338,1.0
+17,6.8704,0,41,singing,121.266,-42.349,2.338,1.0
+18,7.2745,0,41,singing,121.266,-42.349,2.338,1.0
+19,7.6786,0,41,singing,121.266,-42.349,2.338,1.0
+20,8.0828,0,41,singing,121.266,-42.349,2.338,1.0
+21,8.4869,0,41,singing,121.266,-42.349,2.338,1.0
+22,8.8911,0,41,singing,121.266,-42.349,2.338,1.0
+23,9.2952,0,41,singing,121.266,-42.349,2.338,1.0
+24,9.6993,0,41,singing,121.266,-42.349,2.338,1.0
+25,10.1035,0,41,singing,121.266,-42.349,2.338,1.0
+26,10.5076,0,41,singing,121.266,-42.349,2.338,1.0
+27,10.9117,0,41,singing,121.266,-42.349,2.338,1.0
+28,11.3159,0,41,singing,121.266,-42.349,2.338,1.0
+29,11.72,0,41,singing,121.266,-42.349,2.338,1.0
+30,12.1242,0,41,singing,121.266,-42.349,2.338,1.0
+31,12.5283,0,41,singing,121.266,-42.349,2.338,1.0
+32,12.9324,0,41,singing,121.266,-42.349,2.338,1.0

checkpoints/spatial_beats_ov1_local_spatial_v7j_ov123_exp/03_ov123_top4/val_predictions/epoch_0001_csv/valid__hm3d__00034-6imZUJGRUq4__000032-foa__118401__gt.csv

@@ -0,0 +1,3 @@
+frame_idx,frame_time_s,src_or_track_idx,class_idx,class_name,azimuth_deg,elevation_deg,distance_m,activity_prob
+0,0.0,0,9,tool,134.861,-9.617,2.821,1.0
+1,0.412,0,9,tool,134.861,-9.617,2.821,1.0

checkpoints/spatial_beats_ov1_local_spatial_v7j_ov123_exp/03_ov123_top4/val_predictions/epoch_0001_csv/valid__hm3d__00034-6imZUJGRUq4__000048-foa__2050__gt.csv

@@ -0,0 +1,2 @@
+frame_idx,frame_time_s,src_or_track_idx,class_idx,class_name,azimuth_deg,elevation_deg,distance_m,activity_prob
+0,0.0,0,1,string_instrument,-41.991,6.244,1.138,1.0

checkpoints/spatial_beats_ov1_local_spatial_v7j_ov123_exp/03_ov123_top4/val_predictions/epoch_0001_csv/valid__hm3d__00034-6imZUJGRUq4__000064-foa__347902__pred.csv

@@ -0,0 +1,77 @@
+frame_idx,frame_time_s,src_or_track_idx,class_idx,class_name,azimuth_deg,elevation_deg,distance_m,activity_prob
+0,0.0,0,2,guitar,78.019,-31.913,2.959,0.9883
+1,0.3972,0,2,guitar,78.412,-31.649,2.974,0.9727
+2,0.7944,0,2,guitar,79.285,-30.843,3.034,0.5391
+3,1.1917,0,2,guitar,80.034,-30.1,3.019,0.0542
+4,1.5889,0,2,guitar,80.538,-29.535,2.959,0.0097
+5,1.9861,0,2,guitar,80.845,-29.198,2.871,0.0023
+6,2.3833,0,2,guitar,81.142,-28.607,2.797,0.0009
+7,2.7805,0,2,guitar,81.423,-28.247,2.724,0.0005
+8,3.1778,0,2,guitar,81.527,-28.004,2.68,0.0003
+9,3.575,0,2,guitar,81.835,-27.601,2.637,0.0002
+10,3.9722,0,2,guitar,81.913,-27.473,2.593,0.0001
+11,4.3694,0,2,guitar,82.002,-27.167,2.55,0.0001
+12,4.7666,0,2,guitar,82.192,-27.028,2.507,0.0001
+13,5.1639,0,2,guitar,82.235,-26.87,2.478,0.0001
+14,5.5611,0,2,guitar,82.279,-26.706,2.464,0.0001
+15,5.9583,0,2,guitar,82.38,-26.717,2.435,0.0001
+16,6.3555,0,2,guitar,82.626,-26.196,2.407,0.0
+17,6.7527,0,2,guitar,82.651,-26.104,2.393,0.0
+18,7.15,0,2,guitar,82.734,-25.923,2.364,0.0
+0,0.0,1,2,guitar,77.651,-31.146,2.945,0.9883
+1,0.3972,1,2,guitar,78.182,-30.561,2.959,0.9727
+2,0.7944,1,2,guitar,78.839,-29.916,3.034,0.5508
+3,1.1917,1,2,guitar,79.632,-29.032,3.034,0.0618
+4,1.5889,1,2,guitar,80.146,-28.427,2.974,0.0114
+5,1.9861,1,2,guitar,80.481,-27.888,2.9,0.0028
+6,2.3833,1,2,guitar,80.611,-27.538,2.827,0.0012
+7,2.7805,1,2,guitar,80.915,-27.156,2.768,0.0006
+8,3.1778,1,2,guitar,81.049,-26.889,2.724,0.0004
+9,3.575,1,2,guitar,81.239,-26.608,2.68,0.0003
+10,3.9722,1,2,guitar,81.401,-26.466,2.637,0.0002
+11,4.3694,1,2,guitar,81.479,-26.312,2.608,0.0001
+12,4.7666,1,2,guitar,81.711,-25.997,2.564,0.0001
+13,5.1639,1,2,guitar,81.634,-25.985,2.536,0.0001
+14,5.5611,1,2,guitar,81.72,-25.814,2.521,0.0001
+15,5.9583,1,2,guitar,81.862,-25.814,2.493,0.0001
+16,6.3555,1,2,guitar,82.015,-25.466,2.464,0.0001
+17,6.7527,1,2,guitar,82.002,-25.362,2.45,0.0
+18,7.15,1,2,guitar,82.165,-25.191,2.435,0.0
+0,0.0,2,2,guitar,77.498,-31.873,2.959,0.9844
+1,0.3972,2,2,guitar,77.984,-31.583,2.974,0.9688
+2,0.7944,2,2,guitar,78.815,-31.021,3.049,0.4395
+3,1.1917,2,2,guitar,79.462,-30.289,3.019,0.0337
+4,1.5889,2,2,guitar,79.851,-29.734,2.945,0.0057
+5,1.9861,2,2,guitar,80.101,-29.237,2.856,0.0013
+6,2.3833,2,2,guitar,80.224,-28.782,2.768,0.0006
+7,2.7805,2,2,guitar,80.489,-28.417,2.695,0.0003
+8,3.1778,2,2,guitar,80.521,-28.167,2.637,0.0002
+9,3.575,2,2,guitar,80.811,-27.918,2.593,0.0001
+10,3.9722,2,2,guitar,80.937,-27.454,2.55,0.0001
+11,4.3694,2,2,guitar,80.946,-27.311,2.507,0.0001
+12,4.7666,2,2,guitar,81.133,-27.166,2.478,0.0001
+13,5.1639,2,2,guitar,81.071,-27.09,2.435,0.0
+14,5.5611,2,2,guitar,81.133,-26.744,2.421,0.0
+15,5.9583,2,2,guitar,81.276,-26.76,2.393,0.0
+16,6.3555,2,2,guitar,81.342,-26.584,2.364,0.0
+17,6.7527,2,2,guitar,81.353,-26.399,2.35,0.0
+18,7.15,2,2,guitar,81.457,-26.406,2.322,0.0
+0,0.0,3,2,guitar,80.098,-32.192,2.959,0.9766
+1,0.3972,3,2,guitar,80.162,-31.832,2.959,0.9492
+2,0.7944,3,2,guitar,80.794,-31.194,3.004,0.3926
+3,1.1917,3,2,guitar,81.267,-30.391,2.974,0.0488
+4,1.5889,3,2,guitar,81.644,-29.769,2.915,0.0117
+5,1.9861,3,2,guitar,81.779,-29.351,2.827,0.0034
+6,2.3833,3,2,guitar,81.836,-29.018,2.768,0.0016
+7,2.7805,3,2,guitar,82.104,-28.491,2.71,0.0009
+8,3.1778,3,2,guitar,82.082,-28.394,2.666,0.0005
+9,3.575,3,2,guitar,82.368,-27.978,2.608,0.0004
+10,3.9722,3,2,guitar,82.464,-27.683,2.579,0.0003
+11,4.3694,3,2,guitar,82.51,-27.374,2.536,0.0002
+12,4.7666,3,2,guitar,82.608,-27.401,2.507,0.0002
+13,5.1639,3,2,guitar,82.604,-27.244,2.464,0.0001
+14,5.5611,3,2,guitar,82.655,-27.085,2.45,0.0001
+15,5.9583,3,2,guitar,82.792,-26.829,2.421,0.0001
+16,6.3555,3,2,guitar,82.903,-26.57,2.393,0.0001
+17,6.7527,3,2,guitar,82.988,-26.393,2.379,0.0001
+18,7.15,3,2,guitar,83.018,-26.302,2.35,0.0001

checkpoints/spatial_beats_ov1_local_spatial_v7j_ov123_exp/03_ov123_top4/val_predictions/epoch_0001_csv/valid__hm3d__00034-6imZUJGRUq4__000072-foa__60822__pred.csv

@@ -0,0 +1,129 @@
+frame_idx,frame_time_s,src_or_track_idx,class_idx,class_name,azimuth_deg,elevation_deg,distance_m,activity_prob
+0,0.0,0,11,war_sound,11.191,-42.083,1.174,0.1216
+1,0.3964,0,11,war_sound,10.795,-41.837,1.199,0.3418
+2,0.7929,0,11,war_sound,10.324,-41.213,1.212,0.5391
+3,1.1893,0,11,war_sound,10.042,-40.748,1.221,0.6484
+4,1.5858,0,11,war_sound,9.829,-40.457,1.223,0.6992
+5,1.9822,0,11,war_sound,9.502,-39.993,1.226,0.7188
+6,2.3787,0,11,war_sound,9.396,-39.68,1.226,0.7109
+7,2.7751,0,11,war_sound,9.265,-39.531,1.223,0.6758
+8,3.1716,0,11,war_sound,9.125,-39.227,1.226,0.625
+9,3.568,0,11,war_sound,8.997,-38.89,1.226,0.5547
+10,3.9645,0,11,war_sound,8.997,-38.89,1.226,0.4707
+11,4.3609,0,11,war_sound,8.755,-38.563,1.221,0.3789
+12,4.7574,0,11,war_sound,8.725,-38.528,1.221,0.2891
+13,5.1538,0,11,war_sound,8.609,-38.344,1.215,0.209
+14,5.5503,0,11,war_sound,8.453,-38.161,1.215,0.1533
+15,5.9467,0,11,war_sound,8.225,-38.137,1.212,0.1128
+16,6.3432,0,11,war_sound,8.271,-37.895,1.207,0.0747
+17,6.7396,0,11,war_sound,8.158,-37.66,1.201,0.0503
+18,7.1361,0,11,war_sound,8.085,-37.62,1.196,0.0352
+19,7.5325,0,11,war_sound,8.01,-37.58,1.193,0.0251
+20,7.929,0,11,war_sound,7.979,-37.491,1.188,0.017
+21,8.3254,0,11,war_sound,7.903,-37.24,1.182,0.0117
+22,8.7219,0,11,war_sound,7.788,-37.149,1.18,0.0086
+23,9.1183,0,11,war_sound,7.792,-37.099,1.172,0.0065
+24,9.5148,0,11,war_sound,7.796,-36.832,1.169,0.0052
+25,9.9112,0,11,war_sound,7.719,-36.84,1.166,0.0041
+26,10.3077,0,11,war_sound,7.68,-36.678,1.164,0.0032
+27,10.7041,0,11,war_sound,7.6,-36.684,1.15,0.0025
+28,11.1006,0,11,war_sound,7.56,-36.63,1.145,0.002
+29,11.497,0,11,war_sound,7.388,-36.47,1.142,0.0016
+30,11.8935,0,11,war_sound,7.389,-36.296,1.14,0.0013
+31,12.2899,0,11,war_sound,7.347,-36.122,1.137,0.0011
+0,0.0,1,11,war_sound,9.944,-41.433,1.26,0.1758
+1,0.3964,1,11,war_sound,9.419,-40.851,1.291,0.4414
+2,0.7929,1,11,war_sound,8.964,-40.057,1.305,0.6328
+3,1.1893,1,11,war_sound,8.718,-39.609,1.313,0.7148
+4,1.5858,1,11,war_sound,8.539,-39.297,1.319,0.7422
+5,1.9822,1,11,war_sound,8.277,-38.837,1.325,0.7344
+6,2.3787,1,11,war_sound,8.231,-38.683,1.325,0.707
+7,2.7751,1,11,war_sound,7.985,-38.388,1.319,0.6562
+8,3.1716,1,11,war_sound,7.985,-38.194,1.325,0.5859
+9,3.568,1,11,war_sound,7.825,-37.855,1.33,0.5039
+10,3.9645,1,11,war_sound,7.786,-37.858,1.325,0.4121
+11,4.3609,1,11,war_sound,7.592,-37.674,1.319,0.3184
+12,4.7574,1,11,war_sound,7.553,-37.282,1.319,0.2363
+13,5.1538,1,11,war_sound,7.477,-37.24,1.319,0.1689
+14,5.5503,1,11,war_sound,7.321,-37.049,1.313,0.1235
+15,5.9467,1,11,war_sound,7.165,-36.962,1.31,0.0913
+16,6.3432,1,11,war_sound,7.165,-36.912,1.305,0.0618
+17,6.7396,1,11,war_sound,7.045,-36.662,1.299,0.0435
+18,7.1361,1,11,war_sound,7.044,-36.767,1.296,0.0311
+19,7.5325,1,11,war_sound,7.003,-36.505,1.291,0.0226
+20,7.929,1,11,war_sound,6.961,-36.346,1.285,0.0164
+21,8.3254,1,11,war_sound,6.918,-36.345,1.282,0.012
+22,8.7219,1,11,war_sound,6.831,-36.346,1.279,0.0092
+23,9.1183,1,11,war_sound,6.787,-36.18,1.274,0.0071
+24,9.5148,1,11,war_sound,6.827,-36.007,1.268,0.0057
+25,9.9112,1,11,war_sound,6.74,-35.952,1.265,0.0046
+26,10.3077,1,11,war_sound,6.652,-35.781,1.26,0.0038
+27,10.7041,1,11,war_sound,6.605,-35.605,1.246,0.003
+28,11.1006,1,11,war_sound,6.515,-35.662,1.243,0.0025
+29,11.497,1,11,war_sound,6.419,-35.654,1.24,0.0021
+30,11.8935,1,11,war_sound,6.459,-35.467,1.237,0.0018
+31,12.2899,1,11,war_sound,6.41,-35.402,1.234,0.0015
+0,0.0,2,11,war_sound,13.083,-43.429,1.182,0.1348
+1,0.3964,2,11,war_sound,12.498,-42.745,1.207,0.3438
+2,0.7929,2,11,war_sound,11.944,-42.085,1.215,0.5234
+3,1.1893,2,11,war_sound,11.74,-41.778,1.221,0.6289
+4,1.5858,2,11,war_sound,11.514,-41.448,1.221,0.6836
+5,1.9822,2,11,war_sound,11.161,-40.984,1.223,0.707
+6,2.3787,2,11,war_sound,10.949,-40.665,1.218,0.7109
+7,2.7751,2,11,war_sound,10.823,-40.534,1.215,0.6875
+8,3.1716,2,11,war_sound,10.68,-40.381,1.21,0.6445
+9,3.568,2,11,war_sound,10.539,-40.23,1.21,0.5859
+10,3.9645,2,11,war_sound,10.48,-39.88,1.201,0.5078
+11,4.3609,2,11,war_sound,10.261,-39.737,1.199,0.4219
+12,4.7574,2,11,war_sound,10.142,-39.362,1.196,0.332
+13,5.1538,2,11,war_sound,10.022,-39.373,1.188,0.248
+14,5.5503,2,11,war_sound,9.822,-39.196,1.185,0.1826
+15,5.9467,2,11,war_sound,9.677,-39.174,1.174,0.1348
+16,6.3432,2,11,war_sound,9.611,-39.145,1.169,0.0903
+17,6.7396,2,11,war_sound,9.503,-38.92,1.161,0.061
+18,7.1361,2,11,war_sound,9.381,-38.528,1.156,0.0427
+19,7.5325,2,11,war_sound,9.353,-38.491,1.148,0.0293
+20,7.929,2,11,war_sound,9.296,-38.415,1.14,0.02
+21,8.3254,2,11,war_sound,9.184,-38.174,1.132,0.014
+22,8.7219,2,11,war_sound,9.124,-38.304,1.129,0.0103
+23,9.1183,2,11,war_sound,9.051,-38.054,1.121,0.0076
+24,9.5148,2,11,war_sound,9.106,-38.005,1.113,0.0059
+25,9.9112,2,11,war_sound,8.989,-37.751,1.108,0.0046
+26,10.3077,2,11,war_sound,8.924,-37.771,1.103,0.0036
+27,10.7041,2,11,war_sound,8.848,-37.618,1.09,0.0027
+28,11.1006,2,11,war_sound,8.77,-37.574,1.084,0.0023
+29,11.497,2,11,war_sound,8.701,-37.478,1.082,0.0018
+30,11.8935,2,11,war_sound,8.621,-37.317,1.074,0.0015
+31,12.2899,2,11,war_sound,8.585,-37.266,1.074,0.0012
+0,0.0,3,11,war_sound,11.085,-44.051,1.172,0.1318
+1,0.3964,3,11,war_sound,10.457,-43.344,1.215,0.3672
+2,0.7929,3,11,war_sound,9.944,-42.843,1.237,0.543
+3,1.1893,3,11,war_sound,9.653,-42.136,1.248,0.6211
+4,1.5858,3,11,war_sound,9.361,-41.803,1.257,0.6367
+5,1.9822,3,11,war_sound,9.106,-41.295,1.26,0.6094
+6,2.3787,3,11,war_sound,8.966,-41.132,1.257,0.5703
+7,2.7751,3,11,war_sound,8.828,-40.97,1.257,0.5117
+8,3.1716,3,11,war_sound,8.691,-40.613,1.257,0.4395
+9,3.568,3,11,war_sound,8.522,-40.429,1.257,0.3691
+10,3.9645,3,11,war_sound,8.565,-40.426,1.254,0.2988
+11,4.3609,3,11,war_sound,8.311,-40.246,1.248,0.2324
+12,4.7574,3,11,war_sound,8.184,-39.855,1.248,0.1787
+13,5.1538,3,11,war_sound,8.058,-39.863,1.243,0.1318
+14,5.5503,3,11,war_sound,7.978,-39.837,1.243,0.1011
+15,5.9467,3,11,war_sound,7.769,-39.614,1.237,0.0791
+16,6.3432,3,11,war_sound,7.726,-39.411,1.229,0.0574
+17,6.7396,3,11,war_sound,7.687,-39.378,1.223,0.0415
+18,7.1361,3,11,war_sound,7.603,-39.138,1.221,0.0311
+19,7.5325,3,11,war_sound,7.563,-39.102,1.212,0.0233
+20,7.929,3,11,war_sound,7.565,-38.849,1.207,0.0175
+21,8.3254,3,11,war_sound,7.435,-38.817,1.201,0.0128
+22,8.7219,3,11,war_sound,7.394,-38.846,1.201,0.01
+23,9.1183,3,11,war_sound,7.351,-38.807,1.193,0.0081
+24,9.5148,3,11,war_sound,7.351,-38.474,1.188,0.0065
+25,9.9112,3,11,war_sound,7.261,-38.322,1.18,0.0052
+26,10.3077,3,11,war_sound,7.171,-38.168,1.18,0.0042
+27,10.7041,3,11,war_sound,7.079,-38.126,1.166,0.0033
+28,11.1006,3,11,war_sound,7.079,-38.077,1.161,0.0027
+29,11.497,3,11,war_sound,6.937,-38.104,1.158,0.0023
+30,11.8935,3,11,war_sound,6.89,-37.753,1.15,0.0019
+31,12.2899,3,11,war_sound,6.841,-37.703,1.148,0.0016

checkpoints/spatial_beats_ov1_local_spatial_v7j_ov123_exp/03_ov123_top4/val_predictions/epoch_0001_csv/valid__hm3d__00034-6imZUJGRUq4__000088-foa__402628__gt.csv

@@ -0,0 +1,4 @@
+frame_idx,frame_time_s,src_or_track_idx,class_idx,class_name,azimuth_deg,elevation_deg,distance_m,activity_prob
+0,0.0,0,3,body_sound,-153.896,-52.022,0.906,1.0
+1,0.4159,0,3,body_sound,-153.896,-52.022,0.906,1.0
+2,0.8318,0,3,body_sound,-153.896,-52.022,0.906,1.0

checkpoints/spatial_beats_ov1_local_spatial_v7j_ov123_exp/03_ov123_top4/val_predictions/epoch_0001_csv/valid__hm3d__00056-kdw2Uapns3b__000004-foa__160791__pred.csv

@@ -0,0 +1,69 @@
+frame_idx,frame_time_s,src_or_track_idx,class_idx,class_name,azimuth_deg,elevation_deg,distance_m,activity_prob
+0,0.0,0,11,war_sound,176.802,-31.065,3.987,0.7617
+1,0.3919,0,11,war_sound,178.23,-31.069,3.972,0.8477
+2,0.7838,0,11,war_sound,179.029,-31.361,3.972,0.8594
+3,1.1756,0,11,war_sound,179.7,-31.416,3.957,0.8594
+4,1.5675,0,11,war_sound,-179.897,-31.012,3.957,0.8438
+5,1.9594,0,11,war_sound,-179.459,-31.539,3.941,0.8086
+6,2.3513,0,11,war_sound,-179.555,-31.528,3.941,0.7422
+7,2.7432,0,11,war_sound,-179.312,-31.355,3.941,0.6289
+8,3.1351,0,11,war_sound,-179.39,-31.668,3.957,0.4727
+9,3.5269,0,11,war_sound,-179.314,-31.52,3.957,0.2852
+10,3.9188,0,11,war_sound,-179.001,-31.219,3.957,0.1543
+11,4.3107,0,11,war_sound,-178.998,-31.197,3.957,0.0771
+12,4.7026,0,11,war_sound,-179.092,-30.839,3.941,0.0374
+13,5.0945,0,11,war_sound,-178.809,-31.179,3.926,0.0188
+14,5.4864,0,11,war_sound,-178.397,-31.312,3.911,0.0107
+15,5.8782,0,11,war_sound,-177.756,-31.078,3.896,0.0069
+16,6.2701,0,11,war_sound,-178.058,-31.085,3.865,0.0041
+0,0.0,1,11,war_sound,177.822,-29.955,4.049,0.7578
+1,0.3919,1,11,war_sound,179.75,-30.068,4.018,0.8633
+2,0.7838,1,11,war_sound,-179.143,-30.576,4.018,0.8867
+3,1.1756,1,11,war_sound,-177.891,-30.489,4.003,0.8906
+4,1.5675,1,11,war_sound,-177.365,-30.073,4.003,0.875
+5,1.9594,1,11,war_sound,-176.89,-30.663,4.003,0.8359
+6,2.3513,1,11,war_sound,-176.465,-30.768,4.003,0.7578
+7,2.7432,1,11,war_sound,-176.36,-30.736,4.003,0.6289
+8,3.1351,1,11,war_sound,-176.025,-30.877,4.018,0.4512
+9,3.5269,1,11,war_sound,-176.038,-30.715,4.018,0.2559
+10,3.9188,1,11,war_sound,-175.837,-30.703,4.018,0.1357
+11,4.3107,1,11,war_sound,-175.461,-30.656,4.003,0.0674
+12,4.7026,1,11,war_sound,-175.409,-30.472,4.003,0.0337
+13,5.0945,1,11,war_sound,-175.056,-30.778,3.972,0.0175
+14,5.4864,1,11,war_sound,-174.644,-30.728,3.957,0.0107
+15,5.8782,1,11,war_sound,-173.935,-30.822,3.941,0.0069
+16,6.2701,1,11,war_sound,-174.062,-30.565,3.911,0.0042
+0,0.0,2,11,war_sound,177.152,-34.475,4.049,0.6953
+1,0.3919,2,11,war_sound,178.852,-34.064,4.018,0.8125
+2,0.7838,2,11,war_sound,-179.948,-34.414,4.018,0.8438
+3,1.1756,2,11,war_sound,-178.911,-34.631,4.003,0.8477
+4,1.5675,2,11,war_sound,-178.091,-34.331,3.987,0.8438
+5,1.9594,2,11,war_sound,-177.433,-34.832,3.987,0.8203
+6,2.3513,2,11,war_sound,-176.879,-35.034,3.972,0.7695
+7,2.7432,2,11,war_sound,-176.476,-35.109,3.972,0.6797
+8,3.1351,2,11,war_sound,-175.608,-35.582,3.972,0.543
+9,3.5269,2,11,war_sound,-175.162,-35.675,3.972,0.3496
+10,3.9188,2,11,war_sound,-174.312,-35.602,3.987,0.1973
+11,4.3107,2,11,war_sound,-173.569,-36.303,3.972,0.0981
+12,4.7026,2,11,war_sound,-173.092,-36.022,3.972,0.0466
+13,5.0945,2,11,war_sound,-171.959,-36.895,3.941,0.0229
+14,5.4864,2,11,war_sound,-170.831,-37.133,3.926,0.0128
+15,5.8782,2,11,war_sound,-169.746,-37.29,3.911,0.0078
+16,6.2701,2,11,war_sound,-169.753,-37.308,3.88,0.0046
+0,0.0,3,11,war_sound,176.355,-31.689,4.11,0.6992
+1,0.3919,3,11,war_sound,178.979,-31.573,4.08,0.8867
+2,0.7838,3,11,war_sound,-179.076,-31.935,4.049,0.9297
+3,1.1756,3,11,war_sound,-177.725,-32.019,4.018,0.9336
+4,1.5675,3,11,war_sound,-177.027,-31.692,4.003,0.9219
+5,1.9594,3,11,war_sound,-176.65,-32.322,4.003,0.8789
+6,2.3513,3,11,war_sound,-176.294,-32.471,4.018,0.7891
+7,2.7432,3,11,war_sound,-176.397,-32.64,4.018,0.6289
+8,3.1351,3,11,war_sound,-176.284,-33.063,4.018,0.4297
+9,3.5269,3,11,war_sound,-176.467,-32.998,4.018,0.2363
+10,3.9188,3,11,war_sound,-176.028,-32.803,4.018,0.1216
+11,4.3107,3,11,war_sound,-175.65,-32.991,4.018,0.061
+12,4.7026,3,11,war_sound,-175.742,-32.688,4.003,0.0311
+13,5.0945,3,11,war_sound,-175.139,-33.232,3.972,0.0159
+14,5.4864,3,11,war_sound,-174.603,-33.199,3.941,0.0097
+15,5.8782,3,11,war_sound,-173.664,-33.271,3.926,0.0063
+16,6.2701,3,11,war_sound,-173.989,-33.087,3.896,0.0038

checkpoints/spatial_beats_ov1_local_spatial_v7j_ov123_exp/03_ov123_top4/val_predictions/epoch_0001_csv/valid__hm3d__00056-kdw2Uapns3b__000012-foa__278193__gt.csv

@@ -0,0 +1,18 @@
+frame_idx,frame_time_s,src_or_track_idx,class_idx,class_name,azimuth_deg,elevation_deg,distance_m,activity_prob
+1,0.4002,0,43,writing,49.464,-45.574,2.133,1.0
+2,0.8004,0,43,writing,49.464,-45.574,2.133,1.0
+3,1.2006,0,43,writing,49.464,-45.574,2.133,1.0
+4,1.6008,0,43,writing,49.464,-45.574,2.133,1.0
+5,2.001,0,43,writing,49.464,-45.574,2.133,1.0
+6,2.4011,0,43,writing,49.464,-45.574,2.133,1.0
+7,2.8013,0,43,writing,49.464,-45.574,2.133,1.0
+8,3.2015,0,43,writing,49.464,-45.574,2.133,1.0
+9,3.6017,0,43,writing,49.464,-45.574,2.133,1.0
+10,4.0019,0,43,writing,49.464,-45.574,2.133,1.0
+11,4.4021,0,43,writing,49.464,-45.574,2.133,1.0
+12,4.8023,0,43,writing,49.464,-45.574,2.133,1.0
+13,5.2025,0,43,writing,49.464,-45.574,2.133,1.0
+14,5.6027,0,43,writing,49.464,-45.574,2.133,1.0
+15,6.0029,0,43,writing,49.464,-45.574,2.133,1.0
+16,6.403,0,43,writing,49.464,-45.574,2.133,1.0
+17,6.8032,0,43,writing,49.464,-45.574,2.133,1.0

checkpoints/spatial_beats_ov1_local_spatial_v7j_ov123_exp/03_ov123_top4/val_predictions/epoch_0001_csv/valid__ov2_000011__gt.csv

@@ -0,0 +1,55 @@
+frame_idx,frame_time_s,src_or_track_idx,class_idx,class_name,azimuth_deg,elevation_deg,distance_m,activity_prob
+0,0.0,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+1,0.4,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+2,0.8,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+3,1.2,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+4,1.6,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+5,2.0,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+6,2.4,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+7,2.8,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+8,3.2,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+9,3.6,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+10,4.0,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+11,4.4,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+12,4.8,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+13,5.2,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+14,5.6,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+15,6.0,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+16,6.4,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+17,6.8,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+18,7.2,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+19,7.6,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+20,8.0,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+21,8.4,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+22,8.8,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+23,9.2,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+24,9.6,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+25,10.0,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+26,10.4,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+27,10.8,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+28,11.2,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+29,11.6,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+30,12.0,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+31,12.4,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+32,12.8,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+33,13.2,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+34,13.6,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+35,14.0,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+36,14.4,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+37,14.8,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+38,15.2,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+39,15.6,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+40,16.0,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+41,16.4,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+42,16.8,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+43,17.2,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+44,17.6,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+45,18.0,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+46,18.4,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+47,18.8,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+48,19.2,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+49,19.6,0,23,telephone_alarm,-59.911,-6.962,4.73,1.0
+22,8.8,1,13,breathing,-50.319,10.656,4.931,1.0
+23,9.2,1,13,breathing,-50.319,10.656,4.931,1.0
+24,9.6,1,13,breathing,-50.319,10.656,4.931,1.0
+25,10.0,1,13,breathing,-50.319,10.656,4.931,1.0

checkpoints/spatial_beats_ov1_local_spatial_v7j_ov123_exp/03_ov123_top4/val_predictions/epoch_0001_csv/valid__ov2_000065__pred.csv

@@ -0,0 +1,65 @@
+frame_idx,frame_time_s,src_or_track_idx,class_idx,class_name,azimuth_deg,elevation_deg,distance_m,activity_prob
+0,0.0,0,39,clock,-85.299,-10.221,3.034,0.7852
+1,0.4099,0,39,clock,-85.36,-10.511,3.019,0.8438
+2,0.8197,0,39,clock,-85.349,-10.704,3.004,0.8516
+3,1.2296,0,39,clock,-85.41,-10.824,3.004,0.8398
+4,1.6395,0,39,clock,-85.379,-10.883,3.019,0.8203
+5,2.0493,0,39,clock,-85.318,-11.001,3.004,0.793
+6,2.4592,0,39,clock,-85.318,-10.941,3.019,0.7656
+7,2.869,0,39,clock,-85.247,-10.794,3.019,0.7344
+8,3.2789,0,39,clock,-85.463,-10.197,3.019,0.7148
+9,3.6888,0,39,clock,-86.724,-8.789,3.019,0.6836
+10,4.0986,0,39,clock,-89.433,-6.25,2.989,0.6406
+11,4.5085,0,39,clock,-93.785,-3.464,2.989,0.5664
+12,4.9184,0,39,clock,-98.955,-1.392,3.019,0.4043
+13,5.3282,0,30,door,-103.081,-0.228,3.078,0.1768
+14,5.7381,0,30,door,-105.571,0.332,3.108,0.0549
+15,6.1479,0,30,door,-107.309,0.773,3.108,0.017
+0,0.0,1,39,clock,-86.191,-8.998,3.078,0.5781
+1,0.4099,1,39,clock,-86.378,-9.304,3.063,0.7031
+2,0.8197,1,39,clock,-86.484,-9.465,3.063,0.7266
+3,1.2296,1,39,clock,-86.591,-9.525,3.063,0.7188
+4,1.6395,1,39,clock,-86.621,-9.525,3.063,0.6953
+5,2.0493,1,39,clock,-86.591,-9.565,3.063,0.6641
+6,2.4592,1,39,clock,-86.623,-9.546,3.078,0.6328
+7,2.869,1,39,clock,-86.654,-9.286,3.093,0.6055
+8,3.2789,1,39,clock,-86.841,-8.784,3.093,0.582
+9,3.6888,1,39,clock,-87.777,-7.74,3.093,0.5586
+10,4.0986,1,39,clock,-89.706,-5.781,3.078,0.5352
+11,4.5085,1,39,clock,-93.151,-3.435,3.078,0.5078
+12,4.9184,1,30,door,-98.182,-1.249,3.108,0.4453
+13,5.3282,1,30,door,-103.03,0.133,3.153,0.3027
+14,5.7381,1,30,door,-106.356,0.822,3.213,0.1387
+15,6.1479,1,30,door,-108.637,1.298,3.243,0.0454
+0,0.0,2,39,clock,-86.635,-9.256,3.183,0.6562
+1,0.4099,2,39,clock,-86.73,-9.552,3.168,0.6875
+2,0.8197,2,39,clock,-86.729,-9.698,3.168,0.6719
+3,1.2296,2,39,clock,-86.777,-9.761,3.168,0.6367
+4,1.6395,2,39,clock,-86.745,-9.698,3.183,0.5938
+5,2.0493,2,39,clock,-86.73,-9.678,3.168,0.5469
+6,2.4592,2,39,clock,-86.864,-9.512,3.183,0.5078
+7,2.869,2,39,clock,-87.217,-9.0,3.168,0.4824
+8,3.2789,2,39,clock,-88.193,-7.947,3.168,0.4668
+9,3.6888,2,39,clock,-91.074,-5.882,3.153,0.4551
+10,4.0986,2,30,door,-95.458,-3.672,3.153,0.4395
+11,4.5085,2,30,door,-100.305,-2.091,3.183,0.3926
+12,4.9184,2,30,door,-105.217,-1.086,3.228,0.2578
+13,5.3282,2,30,door,-108.316,-0.519,3.273,0.105
+14,5.7381,2,30,door,-110.136,-0.181,3.288,0.0374
+15,6.1479,2,30,door,-111.525,0.134,3.288,0.0132
+0,0.0,3,38,camera,-86.549,-8.481,3.168,0.8789
+1,0.4099,3,38,camera,-86.442,-8.689,3.153,0.9414
+2,0.8197,3,38,camera,-86.245,-8.722,3.153,0.9531
+3,1.2296,3,38,camera,-86.298,-8.633,3.168,0.957
+4,1.6395,3,38,camera,-86.642,-8.188,3.183,0.9531
+5,2.0493,3,38,camera,-87.995,-7.231,3.183,0.9492
+6,2.4592,3,38,camera,-90.54,-5.825,3.198,0.9375
+7,2.869,3,38,camera,-93.637,-4.373,3.213,0.9141
+8,3.2789,3,30,door,-97.042,-3.067,3.243,0.8789
+9,3.6888,3,30,door,-101.275,-1.766,3.273,0.8086
+10,4.0986,3,30,door,-104.937,-0.723,3.303,0.6992
+11,4.5085,3,30,door,-107.977,-0.005,3.333,0.5352
+12,4.9184,3,30,door,-110.65,0.616,3.363,0.3438
+13,5.3282,3,30,door,-112.895,1.196,3.394,0.1797
+14,5.7381,3,30,door,-114.336,1.502,3.394,0.0967
+15,6.1479,3,30,door,-115.564,1.859,3.409,0.0481

checkpoints/spatial_beats_ov1_local_spatial_v7j_ov123_exp/03_ov123_top4/val_predictions/epoch_0001_csv/valid__ov2_000076__gt.csv

@@ -0,0 +1,55 @@
+frame_idx,frame_time_s,src_or_track_idx,class_idx,class_name,azimuth_deg,elevation_deg,distance_m,activity_prob
+0,0.0,0,16,speech,-66.274,-28.619,2.21,1.0
+1,0.4,0,16,speech,-66.274,-28.619,2.21,1.0
+2,0.8,0,16,speech,-66.274,-28.619,2.21,1.0
+3,1.2,0,16,speech,-66.274,-28.619,2.21,1.0
+4,1.6,0,16,speech,-66.274,-28.619,2.21,1.0
+5,2.0,0,16,speech,-66.274,-28.619,2.21,1.0
+6,2.4,0,16,speech,-66.274,-28.619,2.21,1.0
+7,2.8,0,16,speech,-66.274,-28.619,2.21,1.0
+8,3.2,0,16,speech,-66.274,-28.619,2.21,1.0
+9,3.6,0,16,speech,-66.274,-28.619,2.21,1.0
+10,4.0,0,16,speech,-66.274,-28.619,2.21,1.0
+11,4.4,0,16,speech,-66.274,-28.619,2.21,1.0
+12,4.8,0,16,speech,-66.274,-28.619,2.21,1.0
+13,5.2,0,16,speech,-66.274,-28.619,2.21,1.0
+14,5.6,0,16,speech,-66.274,-28.619,2.21,1.0
+15,6.0,0,16,speech,-66.274,-28.619,2.21,1.0
+16,6.4,0,16,speech,-66.274,-28.619,2.21,1.0
+17,6.8,0,16,speech,-66.274,-28.619,2.21,1.0
+18,7.2,0,16,speech,-66.274,-28.619,2.21,1.0
+19,7.6,0,16,speech,-66.274,-28.619,2.21,1.0
+20,8.0,0,16,speech,-66.274,-28.619,2.21,1.0
+21,8.4,0,16,speech,-66.274,-28.619,2.21,1.0
+22,8.8,0,16,speech,-66.274,-28.619,2.21,1.0
+23,9.2,0,16,speech,-66.274,-28.619,2.21,1.0
+24,9.6,0,16,speech,-66.274,-28.619,2.21,1.0
+25,10.0,0,16,speech,-66.274,-28.619,2.21,1.0
+26,10.4,0,16,speech,-66.274,-28.619,2.21,1.0
+27,10.8,0,16,speech,-66.274,-28.619,2.21,1.0
+28,11.2,0,16,speech,-66.274,-28.619,2.21,1.0
+29,11.6,0,16,speech,-66.274,-28.619,2.21,1.0
+30,12.0,0,16,speech,-66.274,-28.619,2.21,1.0
+31,12.4,0,16,speech,-66.274,-28.619,2.21,1.0
+32,12.8,0,16,speech,-66.274,-28.619,2.21,1.0
+33,13.2,0,16,speech,-66.274,-28.619,2.21,1.0
+34,13.6,0,16,speech,-66.274,-28.619,2.21,1.0
+35,14.0,0,16,speech,-66.274,-28.619,2.21,1.0
+36,14.4,0,16,speech,-66.274,-28.619,2.21,1.0
+37,14.8,0,16,speech,-66.274,-28.619,2.21,1.0
+38,15.2,0,16,speech,-66.274,-28.619,2.21,1.0
+39,15.6,0,16,speech,-66.274,-28.619,2.21,1.0
+40,16.0,0,16,speech,-66.274,-28.619,2.21,1.0
+41,16.4,0,16,speech,-66.274,-28.619,2.21,1.0
+42,16.8,0,16,speech,-66.274,-28.619,2.21,1.0
+43,17.2,0,16,speech,-66.274,-28.619,2.21,1.0
+44,17.6,0,16,speech,-66.274,-28.619,2.21,1.0
+45,18.0,0,16,speech,-66.274,-28.619,2.21,1.0
+46,18.4,0,16,speech,-66.274,-28.619,2.21,1.0
+47,18.8,0,16,speech,-66.274,-28.619,2.21,1.0
+48,19.2,0,16,speech,-66.274,-28.619,2.21,1.0
+49,19.6,0,16,speech,-66.274,-28.619,2.21,1.0
+11,4.4,1,43,writing,157.445,-19.401,1.894,1.0
+12,4.8,1,43,writing,157.445,-19.401,1.894,1.0
+13,5.2,1,43,writing,157.445,-19.401,1.894,1.0
+14,5.6,1,43,writing,157.445,-19.401,1.894,1.0

checkpoints/spatial_beats_ov1_local_spatial_v7j_ov123_exp/03_ov123_top4/val_predictions/epoch_0001_csv/valid__ov2_000076__pred.csv

@@ -0,0 +1,201 @@
+frame_idx,frame_time_s,src_or_track_idx,class_idx,class_name,azimuth_deg,elevation_deg,distance_m,activity_prob
+0,0.0,0,13,breathing,-61.489,-32.474,1.689,0.5039
+1,0.4,0,13,breathing,-61.976,-32.936,1.695,0.7461
+2,0.8,0,13,breathing,-62.082,-33.075,1.701,0.832
+3,1.2,0,13,breathing,-62.176,-33.416,1.701,0.8672
+4,1.6,0,13,breathing,-61.998,-33.373,1.701,0.8828
+5,2.0,0,13,breathing,-61.916,-33.527,1.701,0.8945
+6,2.4,0,13,breathing,-61.916,-33.527,1.695,0.9023
+7,2.8,0,13,breathing,-61.916,-33.666,1.695,0.9062
+8,3.2,0,13,breathing,-61.916,-33.666,1.695,0.9102
+9,3.6,0,13,breathing,-61.916,-33.666,1.695,0.9141
+10,4.0,0,13,breathing,-61.916,-33.804,1.695,0.9141
+11,4.4,0,13,breathing,-62.01,-33.722,1.695,0.9141
+12,4.8,0,13,breathing,-61.916,-33.666,1.689,0.9141
+13,5.2,0,13,breathing,-61.916,-33.804,1.689,0.9141
+14,5.6,0,13,breathing,-61.916,-33.804,1.689,0.9141
+15,6.0,0,13,breathing,-62.186,-33.766,1.689,0.9141
+16,6.4,0,13,breathing,-61.916,-33.942,1.682,0.9141
+17,6.8,0,13,breathing,-62.092,-33.847,1.682,0.9141
+18,7.2,0,13,breathing,-62.092,-33.985,1.676,0.9102
+19,7.6,0,13,breathing,-61.998,-34.067,1.676,0.9102
+20,8.0,0,13,breathing,-62.082,-34.194,1.676,0.9062
+21,8.4,0,13,breathing,-61.998,-34.067,1.67,0.9062
+22,8.8,0,13,breathing,-62.082,-34.331,1.67,0.9062
+23,9.2,0,13,breathing,-62.082,-34.331,1.67,0.9062
+24,9.6,0,13,breathing,-62.082,-34.331,1.663,0.9023
+25,10.0,0,13,breathing,-62.082,-34.331,1.67,0.9023
+26,10.4,0,13,breathing,-62.26,-34.237,1.663,0.9023
+27,10.8,0,13,breathing,-62.26,-34.375,1.663,0.9023
+28,11.2,0,13,breathing,-62.439,-34.419,1.663,0.9023
+29,11.6,0,13,breathing,-62.439,-34.281,1.663,0.8984
+30,12.0,0,13,breathing,-62.439,-34.419,1.657,0.8945
+31,12.4,0,13,breathing,-62.439,-34.419,1.657,0.8984
+32,12.8,0,13,breathing,-62.799,-34.368,1.651,0.8984
+33,13.2,0,13,breathing,-63.071,-34.571,1.651,0.8945
+34,13.6,0,13,breathing,-63.253,-34.475,1.651,0.8945
+35,14.0,0,13,breathing,-63.435,-34.518,1.644,0.8945
+36,14.4,0,13,breathing,-63.619,-34.668,1.644,0.8906
+37,14.8,0,13,breathing,-63.804,-34.818,1.638,0.8906
+38,15.2,0,13,breathing,-63.99,-34.861,1.632,0.8906
+39,15.6,0,13,breathing,-64.28,-35.253,1.625,0.8828
+40,16.0,0,13,breathing,-65.041,-35.283,1.625,0.8789
+41,16.4,0,13,breathing,-65.247,-35.408,1.625,0.875
+42,16.8,0,13,breathing,-65.96,-35.611,1.613,0.8672
+43,17.2,0,13,breathing,-66.295,-36.021,1.619,0.8555
+44,17.6,0,13,breathing,-67.172,-36.374,1.613,0.8438
+45,18.0,0,13,breathing,-67.751,-36.547,1.613,0.8242
+46,18.4,0,13,breathing,-67.938,-36.915,1.607,0.8086
+47,18.8,0,13,breathing,-68.308,-36.929,1.607,0.793
+48,19.2,0,13,breathing,-68.962,-37.242,1.607,0.7734
+49,19.6,0,13,breathing,-68.901,-37.328,1.6,0.7539
+0,0.0,1,13,breathing,-61.277,-31.69,1.676,0.7461
+1,0.4,1,13,breathing,-61.834,-32.138,1.682,0.8906
+2,0.8,1,13,breathing,-61.928,-32.484,1.682,0.9336
+3,1.2,1,13,breathing,-62.031,-32.486,1.689,0.9492
+4,1.6,1,13,breathing,-62.205,-32.528,1.689,0.957
+5,2.0,1,13,breathing,-62.031,-32.764,1.682,0.9609
+6,2.4,1,13,breathing,-62.031,-32.764,1.682,0.9648
+7,2.8,1,13,breathing,-62.042,-32.704,1.676,0.9688
+8,3.2,1,13,breathing,-61.858,-32.722,1.682,0.9688
+9,3.6,1,13,breathing,-62.031,-32.903,1.682,0.9688
+10,4.0,1,13,breathing,-62.214,-32.883,1.676,0.9727
+11,4.4,1,13,breathing,-62.214,-32.883,1.676,0.9727
+12,4.8,1,13,breathing,-62.214,-32.883,1.676,0.9727
+13,5.2,1,13,breathing,-62.042,-32.979,1.676,0.9727
+14,5.6,1,13,breathing,-62.214,-33.02,1.676,0.9727
+15,6.0,1,13,breathing,-62.214,-33.02,1.67,0.9727
+16,6.4,1,13,breathing,-62.214,-33.02,1.67,0.9727
+17,6.8,1,13,breathing,-62.214,-33.02,1.663,0.9727
+18,7.2,1,13,breathing,-62.214,-33.157,1.663,0.9727
+19,7.6,1,13,breathing,-62.214,-33.157,1.663,0.9727
+20,8.0,1,13,breathing,-62.387,-33.199,1.663,0.9727
+21,8.4,1,13,breathing,-62.214,-33.157,1.657,0.9727
+22,8.8,1,13,breathing,-62.387,-33.335,1.657,0.9727
+23,9.2,1,13,breathing,-62.387,-33.471,1.651,0.9727
+24,9.6,1,13,breathing,-62.387,-33.471,1.651,0.9727
+25,10.0,1,13,breathing,-62.387,-33.471,1.651,0.9727
+26,10.4,1,13,breathing,-62.387,-33.471,1.644,0.9727
+27,10.8,1,13,breathing,-62.387,-33.471,1.644,0.9727
+28,11.2,1,13,breathing,-62.56,-33.649,1.644,0.9727
+29,11.6,1,13,breathing,-62.56,-33.513,1.644,0.9727
+30,12.0,1,13,breathing,-62.56,-33.513,1.638,0.9727
+31,12.4,1,13,breathing,-62.56,-33.649,1.638,0.9727
+32,12.8,1,13,breathing,-62.913,-33.394,1.632,0.9727
+33,13.2,1,13,breathing,-62.913,-33.53,1.632,0.9727
+34,13.6,1,13,breathing,-63.086,-33.571,1.625,0.9727
+35,14.0,1,13,breathing,-63.086,-33.571,1.625,0.9727
+36,14.4,1,13,breathing,-63.26,-33.611,1.619,0.9688
+37,14.8,1,13,breathing,-63.259,-33.949,1.613,0.9688
+38,15.2,1,13,breathing,-63.435,-33.787,1.607,0.9688
+39,15.6,1,13,breathing,-63.612,-34.166,1.6,0.9648
+40,16.0,1,13,breathing,-64.228,-34.103,1.594,0.9609
+41,16.4,1,13,breathing,-64.144,-34.152,1.588,0.957
+42,16.8,1,13,breathing,-64.413,-34.212,1.576,0.9492
+43,17.2,1,13,breathing,-64.601,-34.46,1.576,0.9375
+44,17.6,1,13,breathing,-65.147,-34.58,1.569,0.9258
+45,18.0,1,13,breathing,-65.26,-34.781,1.563,0.9102
+46,18.4,1,13,breathing,-65.275,-34.996,1.551,0.8945
+47,18.8,1,13,breathing,-65.283,-35.104,1.551,0.8828
+48,19.2,1,13,breathing,-65.772,-35.146,1.551,0.8633
+49,19.6,1,13,breathing,-65.772,-35.006,1.545,0.8516
+0,0.0,2,13,breathing,-61.39,-31.056,1.689,0.6797
+1,0.4,2,13,breathing,-61.783,-31.637,1.695,0.8359
+2,0.8,2,13,breathing,-61.975,-31.919,1.695,0.8828
+3,1.2,2,13,breathing,-61.892,-31.942,1.695,0.9023
+4,1.6,2,13,breathing,-61.796,-32.169,1.695,0.9141
+5,2.0,2,13,breathing,-61.713,-32.19,1.695,0.918
+6,2.4,2,13,breathing,-61.713,-32.335,1.689,0.9219
+7,2.8,2,13,breathing,-61.535,-32.292,1.689,0.9258
+8,3.2,2,13,breathing,-61.631,-32.211,1.689,0.9258
+9,3.6,2,13,breathing,-61.713,-32.479,1.689,0.9258
+10,4.0,2,13,breathing,-61.713,-32.479,1.689,0.9258
+11,4.4,2,13,breathing,-61.809,-32.399,1.689,0.9258
+12,4.8,2,13,breathing,-61.809,-32.399,1.682,0.9258
+13,5.2,2,13,breathing,-61.535,-32.58,1.682,0.9258
+14,5.6,2,13,breathing,-61.809,-32.399,1.682,0.9258
+15,6.0,2,13,breathing,-61.809,-32.543,1.676,0.9258
+16,6.4,2,13,breathing,-61.809,-32.543,1.676,0.9219
+17,6.8,2,13,breathing,-61.892,-32.667,1.676,0.9219
+18,7.2,2,13,breathing,-61.713,-32.767,1.67,0.9219
+19,7.6,2,13,breathing,-61.617,-32.848,1.67,0.918
+20,8.0,2,13,breathing,-61.796,-32.892,1.67,0.918
+21,8.4,2,13,breathing,-61.713,-32.767,1.663,0.918
+22,8.8,2,13,breathing,-61.892,-32.811,1.663,0.9141
+23,9.2,2,13,breathing,-61.796,-33.036,1.663,0.9141
+24,9.6,2,13,breathing,-61.892,-32.954,1.657,0.9141
+25,10.0,2,13,breathing,-61.892,-32.954,1.657,0.9141
+26,10.4,2,13,breathing,-61.976,-32.936,1.657,0.9102
+27,10.8,2,13,breathing,-61.975,-33.079,1.657,0.9102
+28,11.2,2,13,breathing,-61.976,-33.222,1.657,0.9102
+29,11.6,2,13,breathing,-61.976,-32.936,1.651,0.9062
+30,12.0,2,13,breathing,-62.156,-33.123,1.651,0.9062
+31,12.4,2,13,breathing,-62.071,-32.997,1.644,0.9062
+32,12.8,2,13,breathing,-62.431,-33.084,1.644,0.9062
+33,13.2,2,13,breathing,-62.431,-33.084,1.644,0.9062
+34,13.6,2,13,breathing,-62.613,-33.27,1.644,0.9023
+35,14.0,2,13,breathing,-62.613,-33.127,1.638,0.9023
+36,14.4,2,13,breathing,-62.792,-33.275,1.638,0.8984
+37,14.8,2,13,breathing,-62.883,-33.44,1.632,0.8984
+38,15.2,2,13,breathing,-63.068,-33.377,1.625,0.8945
+39,15.6,2,13,breathing,-63.157,-33.611,1.619,0.8906
+40,16.0,2,13,breathing,-63.621,-33.718,1.619,0.8867
+41,16.4,2,13,breathing,-63.714,-33.739,1.619,0.8789
+42,16.8,2,13,breathing,-63.997,-33.622,1.607,0.8711
+43,17.2,2,13,breathing,-64.096,-33.86,1.607,0.8633
+44,17.6,2,13,breathing,-64.578,-34.076,1.607,0.8477
+45,18.0,2,13,breathing,-64.782,-34.047,1.6,0.832
+46,18.4,2,13,breathing,-64.885,-34.328,1.594,0.8203
+47,18.8,2,13,breathing,-64.99,-34.314,1.594,0.8047
+48,19.2,2,13,breathing,-65.302,-34.457,1.594,0.7852
+49,19.6,2,13,breathing,-65.401,-34.329,1.588,0.7695
+0,0.0,3,13,breathing,-63.637,-36.285,1.676,0.4766
+1,0.4,3,13,breathing,-64.418,-36.53,1.689,0.8164
+2,0.8,3,13,breathing,-64.695,-36.693,1.689,0.918
+3,1.2,3,13,breathing,-64.879,-36.644,1.689,0.9531
+4,1.6,3,13,breathing,-64.977,-36.666,1.689,0.9688
+5,2.0,3,13,breathing,-64.873,-36.805,1.689,0.9766
+6,2.4,3,13,breathing,-64.977,-36.942,1.682,0.9805
+7,2.8,3,13,breathing,-64.879,-37.058,1.676,0.9844
+8,3.2,3,13,breathing,-64.879,-36.92,1.676,0.9883
+9,3.6,3,13,breathing,-65.172,-36.986,1.676,0.9883
+10,4.0,3,13,breathing,-65.376,-37.146,1.676,0.9883
+11,4.4,3,13,breathing,-65.376,-37.146,1.676,0.9922
+12,4.8,3,13,breathing,-65.583,-37.168,1.676,0.9922
+13,5.2,3,13,breathing,-65.683,-37.328,1.676,0.9922
+14,5.6,3,13,breathing,-65.892,-37.351,1.676,0.9922
+15,6.0,3,13,breathing,-66.194,-37.416,1.676,0.9922
+16,6.4,3,13,breathing,-66.206,-37.535,1.67,0.9922
+17,6.8,3,13,breathing,-66.537,-37.84,1.67,0.9922
+18,7.2,3,13,breathing,-66.846,-37.904,1.67,0.9922
+19,7.6,3,13,breathing,-66.772,-38.126,1.67,0.9961
+20,8.0,3,13,breathing,-67.205,-38.195,1.67,0.9961
+21,8.4,3,13,breathing,-67.451,-38.487,1.663,0.9961
+22,8.8,3,13,breathing,-67.77,-38.551,1.67,0.9961
+23,9.2,3,13,breathing,-67.918,-38.824,1.663,0.9961
+24,9.6,3,13,breathing,-68.155,-38.994,1.663,0.9961
+25,10.0,3,13,breathing,-68.749,-39.072,1.663,0.9961
+26,10.4,3,13,breathing,-68.911,-39.354,1.663,0.9961
+27,10.8,3,13,breathing,-69.585,-39.335,1.663,0.9961
+28,11.2,3,13,breathing,-70.017,-39.938,1.657,0.9961
+29,11.6,3,13,breathing,-70.429,-39.98,1.657,0.9961
+30,12.0,3,13,breathing,-70.934,-40.051,1.657,0.9961
+31,12.4,3,13,breathing,-71.525,-40.396,1.651,0.9961
+32,12.8,3,13,breathing,-72.294,-40.505,1.651,0.9961
+33,13.2,3,13,breathing,-72.754,-40.844,1.644,0.9961
+34,13.6,3,13,breathing,-73.453,-41.343,1.651,0.9961
+35,14.0,3,13,breathing,-74.194,-41.564,1.644,0.9922
+36,14.4,3,13,breathing,-74.896,-41.93,1.644,0.9922
+37,14.8,3,13,breathing,-75.758,-42.608,1.638,0.9922
+38,15.2,3,13,breathing,-76.735,-42.859,1.632,0.9922
+39,15.6,3,13,breathing,-78.257,-44.081,1.625,0.9922
+40,16.0,3,13,breathing,-79.821,-44.387,1.625,0.9883
+41,16.4,3,13,breathing,-80.643,-44.94,1.625,0.9883
+42,16.8,3,13,breathing,-82.416,-45.748,1.619,0.9844
+43,17.2,3,13,breathing,-83.481,-46.684,1.619,0.9805
+44,17.6,3,13,breathing,-85.533,-47.372,1.619,0.9727
+45,18.0,3,13,breathing,-86.876,-48.197,1.619,0.9688
+46,18.4,3,13,breathing,-87.407,-48.445,1.619,0.9609
+47,18.8,3,13,breathing,-88.061,-48.872,1.625,0.957
+48,19.2,3,13,breathing,-89.771,-49.398,1.632,0.9492
+49,19.6,3,13,breathing,-89.933,-49.812,1.625,0.9414

checkpoints/spatial_beats_ov1_local_spatial_v7j_ov123_exp/03_ov123_top4/val_predictions/epoch_0001_csv/valid__ov2_000099__gt.csv

@@ -0,0 +1,40 @@
+frame_idx,frame_time_s,src_or_track_idx,class_idx,class_name,azimuth_deg,elevation_deg,distance_m,activity_prob
+0,0.0,0,37,zipper,80.687,15.8,3.062,1.0
+1,0.3972,0,37,zipper,80.687,15.8,3.062,1.0
+2,0.7943,0,37,zipper,80.687,15.8,3.062,1.0
+3,1.1915,0,37,zipper,80.687,15.8,3.062,1.0
+4,1.5887,0,37,zipper,80.687,15.8,3.062,1.0
+5,1.9858,0,37,zipper,80.687,15.8,3.062,1.0
+6,2.383,0,37,zipper,80.687,15.8,3.062,1.0
+7,2.7802,0,37,zipper,80.687,15.8,3.062,1.0
+8,3.1773,0,37,zipper,80.687,15.8,3.062,1.0
+9,3.5745,0,37,zipper,80.687,15.8,3.062,1.0
+10,3.9717,0,37,zipper,80.687,15.8,3.062,1.0
+11,4.3688,0,37,zipper,80.687,15.8,3.062,1.0
+12,4.766,0,37,zipper,80.687,15.8,3.062,1.0
+13,5.1631,0,37,zipper,80.687,15.8,3.062,1.0
+14,5.5603,0,37,zipper,80.687,15.8,3.062,1.0
+15,5.9575,0,37,zipper,80.687,15.8,3.062,1.0
+16,6.3546,0,37,zipper,80.687,15.8,3.062,1.0
+17,6.7518,0,37,zipper,80.687,15.8,3.062,1.0
+18,7.149,0,37,zipper,80.687,15.8,3.062,1.0
+19,7.5461,0,37,zipper,80.687,15.8,3.062,1.0
+4,1.5887,1,22,train,-83.499,-24.228,1.804,1.0
+5,1.9858,1,22,train,-83.499,-24.228,1.804,1.0
+6,2.383,1,22,train,-83.499,-24.228,1.804,1.0
+7,2.7802,1,22,train,-83.499,-24.228,1.804,1.0
+8,3.1773,1,22,train,-83.499,-24.228,1.804,1.0
+9,3.5745,1,22,train,-83.499,-24.228,1.804,1.0
+10,3.9717,1,22,train,-83.499,-24.228,1.804,1.0
+11,4.3688,1,22,train,-83.499,-24.228,1.804,1.0
+12,4.766,1,22,train,-83.499,-24.228,1.804,1.0
+13,5.1631,1,22,train,-83.499,-24.228,1.804,1.0
+14,5.5603,1,22,train,-83.499,-24.228,1.804,1.0
+15,5.9575,1,22,train,-83.499,-24.228,1.804,1.0
+16,6.3546,1,22,train,-83.499,-24.228,1.804,1.0
+17,6.7518,1,22,train,-83.499,-24.228,1.804,1.0
+18,7.149,1,22,train,-83.499,-24.228,1.804,1.0
+19,7.5461,1,22,train,-83.499,-24.228,1.804,1.0
+20,7.9433,1,22,train,-83.499,-24.228,1.804,1.0
+21,8.3405,1,22,train,-83.499,-24.228,1.804,1.0
+22,8.7376,1,22,train,-83.499,-24.228,1.804,1.0

checkpoints/spatial_beats_ov1_local_spatial_v7j_ov123_exp/03_ov123_top4/val_predictions/epoch_0001_csv/valid__ov2_000119__pred.csv

@@ -0,0 +1,201 @@
+frame_idx,frame_time_s,src_or_track_idx,class_idx,class_name,azimuth_deg,elevation_deg,distance_m,activity_prob
+0,0.0,0,55,aircraft,-16.401,-19.656,3.819,0.793
+1,0.4,0,55,aircraft,-15.906,-19.959,3.804,0.793
+2,0.8,0,55,aircraft,-15.514,-19.909,3.804,0.7812
+3,1.2,0,55,aircraft,-15.222,-20.017,3.804,0.7695
+4,1.6,0,55,aircraft,-15.131,-19.942,3.789,0.7578
+5,2.0,0,55,aircraft,-15.034,-19.991,3.789,0.7461
+6,2.4,0,55,aircraft,-14.763,-20.014,3.789,0.7383
+7,2.8,0,55,aircraft,-14.578,-19.906,3.789,0.7305
+8,3.2,0,55,aircraft,-14.668,-19.816,3.773,0.7227
+9,3.6,0,27,animal,-14.581,-19.864,3.773,0.7148
+10,4.0,0,55,aircraft,-14.581,-19.864,3.789,0.7031
+11,4.4,0,55,aircraft,-14.578,-19.824,3.773,0.6953
+12,4.8,0,55,aircraft,-14.488,-19.666,3.773,0.6875
+13,5.2,0,55,aircraft,-14.488,-19.666,3.773,0.6836
+14,5.6,0,55,aircraft,-14.491,-19.623,3.773,0.6758
+15,6.0,0,55,aircraft,-14.672,-19.691,3.773,0.668
+16,6.4,0,55,aircraft,-14.398,-19.508,3.773,0.6602
+17,6.8,0,55,aircraft,-14.491,-19.623,3.773,0.6562
+18,7.2,0,27,animal,-14.491,-19.623,3.773,0.6484
+19,7.6,0,55,aircraft,-14.4,-19.547,3.773,0.6406
+20,8.0,0,55,aircraft,-14.307,-19.35,3.773,0.6328
+21,8.4,0,55,aircraft,-14.309,-19.471,3.773,0.6289
+22,8.8,0,55,aircraft,-14.218,-19.478,3.773,0.6211
+23,9.2,0,55,aircraft,-14.309,-19.387,3.773,0.6172
+24,9.6,0,55,aircraft,-14.218,-19.395,3.773,0.6094
+25,10.0,0,55,aircraft,-14.218,-19.478,3.773,0.6055
+26,10.4,0,55,aircraft,-14.309,-19.387,3.773,0.5977
+27,10.8,0,55,aircraft,-14.309,-19.304,3.773,0.5938
+28,11.2,0,55,aircraft,-14.127,-19.281,3.773,0.5898
+29,11.6,0,55,aircraft,-14.309,-19.22,3.773,0.5859
+30,12.0,0,55,aircraft,-14.309,-19.22,3.773,0.5781
+31,12.4,0,55,aircraft,-14.217,-19.191,3.773,0.5742
+32,12.8,0,55,aircraft,-14.398,-19.011,3.773,0.5742
+33,13.2,0,55,aircraft,-14.491,-19.122,3.773,0.5664
+34,13.6,0,55,aircraft,-14.491,-19.039,3.773,0.5625
+35,14.0,0,55,aircraft,-14.307,-18.934,3.773,0.5586
+36,14.4,0,55,aircraft,-14.581,-18.948,3.773,0.5586
+37,14.8,0,55,aircraft,-14.491,-18.871,3.773,0.5547
+38,15.2,0,55,aircraft,-14.491,-18.871,3.773,0.5508
+39,15.6,0,55,aircraft,-14.491,-18.703,3.758,0.5469
+40,16.0,0,55,aircraft,-14.585,-18.729,3.758,0.543
+41,16.4,0,55,aircraft,-14.585,-18.729,3.758,0.543
+42,16.8,0,55,aircraft,-14.585,-18.56,3.758,0.543
+43,17.2,0,55,aircraft,-14.494,-18.397,3.758,0.543
+44,17.6,0,55,aircraft,-14.585,-18.135,3.743,0.543
+45,18.0,0,55,aircraft,-14.494,-18.057,3.743,0.543
+46,18.4,0,55,aircraft,-14.497,-18.001,3.727,0.5391
+47,18.8,0,55,aircraft,-14.589,-17.908,3.727,0.5391
+48,19.2,0,55,aircraft,-14.5,-17.77,3.727,0.5352
+49,19.6,0,55,aircraft,-14.5,-17.77,3.712,0.5195
+0,0.0,1,55,aircraft,-16.649,-18.834,3.804,0.8906
+1,0.4,1,55,aircraft,-16.192,-19.084,3.789,0.9062
+2,0.8,1,27,animal,-15.922,-19.116,3.789,0.9023
+3,1.2,1,27,animal,-15.658,-19.148,3.789,0.8984
+4,1.6,1,55,aircraft,-15.489,-19.056,3.789,0.8945
+5,2.0,1,27,animal,-15.4,-19.179,3.789,0.8906
+6,2.4,1,55,aircraft,-15.161,-19.273,3.789,0.8867
+7,2.8,1,27,animal,-14.995,-19.107,3.789,0.8828
+8,3.2,1,27,animal,-15.075,-19.1,3.773,0.8789
+9,3.6,1,27,animal,-15.001,-19.139,3.773,0.875
+10,4.0,1,27,animal,-15.001,-19.139,3.789,0.8672
+11,4.4,1,55,aircraft,-14.995,-19.107,3.789,0.8633
+12,4.8,1,27,animal,-14.915,-18.966,3.789,0.8594
+13,5.2,1,27,animal,-14.831,-18.869,3.789,0.8555
+14,5.6,1,27,animal,-14.836,-18.899,3.789,0.8516
+15,6.0,1,55,aircraft,-14.995,-18.886,3.789,0.8477
+16,6.4,1,27,animal,-14.915,-18.893,3.789,0.8438
+17,6.8,1,27,animal,-14.836,-18.825,3.789,0.8398
+18,7.2,1,27,animal,-14.915,-18.893,3.789,0.8359
+19,7.6,1,27,animal,-14.756,-18.906,3.789,0.832
+20,8.0,1,27,animal,-14.756,-18.758,3.789,0.8281
+21,8.4,1,55,aircraft,-14.756,-18.832,3.789,0.8203
+22,8.8,1,27,animal,-14.756,-18.832,3.789,0.8164
+23,9.2,1,27,animal,-14.673,-18.661,3.789,0.8164
+24,9.6,1,55,aircraft,-14.597,-18.771,3.789,0.8125
+25,10.0,1,27,animal,-14.676,-18.764,3.789,0.8086
+26,10.4,1,55,aircraft,-14.593,-18.668,3.789,0.8047
+27,10.8,1,27,animal,-14.676,-18.69,3.789,0.8008
+28,11.2,1,55,aircraft,-14.514,-18.674,3.804,0.7969
+29,11.6,1,55,aircraft,-14.676,-18.616,3.789,0.793
+30,12.0,1,55,aircraft,-14.676,-18.616,3.804,0.7891
+31,12.4,1,55,aircraft,-14.673,-18.514,3.804,0.7891
+32,12.8,1,55,aircraft,-14.915,-18.523,3.804,0.7852
+33,13.2,1,55,aircraft,-14.836,-18.529,3.789,0.7812
+34,13.6,1,55,aircraft,-14.836,-18.529,3.804,0.7812
+35,14.0,1,55,aircraft,-14.756,-18.462,3.804,0.7773
+36,14.4,1,55,aircraft,-14.915,-18.449,3.804,0.7734
+37,14.8,1,55,aircraft,-14.756,-18.387,3.804,0.7734
+38,15.2,1,55,aircraft,-14.836,-18.307,3.789,0.7695
+39,15.6,1,55,aircraft,-14.836,-18.158,3.804,0.7656
+40,16.0,1,55,aircraft,-15.001,-18.172,3.789,0.7656
+41,16.4,1,55,aircraft,-15.001,-18.172,3.789,0.7617
+42,16.8,1,55,aircraft,-14.921,-17.954,3.789,0.7617
+43,17.2,1,55,aircraft,-14.76,-17.966,3.789,0.7617
+44,17.6,1,55,aircraft,-15.087,-17.74,3.789,0.7578
+45,18.0,1,55,aircraft,-14.926,-17.677,3.789,0.7578
+46,18.4,1,55,aircraft,-14.926,-17.525,3.773,0.7539
+47,18.8,1,55,aircraft,-14.931,-17.472,3.773,0.75
+48,19.2,1,55,aircraft,-15.018,-17.412,3.773,0.7422
+49,19.6,1,55,aircraft,-14.85,-17.249,3.773,0.7305
+0,0.0,2,55,aircraft,-17.142,-18.923,3.819,0.7578
+1,0.4,2,55,aircraft,-16.666,-19.178,3.804,0.7461
+2,0.8,2,55,aircraft,-16.385,-19.134,3.804,0.7305
+3,1.2,2,55,aircraft,-16.04,-19.361,3.804,0.707
+4,1.6,2,55,aircraft,-15.863,-19.189,3.789,0.6875
+5,2.0,2,55,aircraft,-15.781,-19.273,3.789,0.6719
+6,2.4,2,55,aircraft,-15.616,-19.287,3.789,0.6562
+7,2.8,2,55,aircraft,-15.36,-19.199,3.773,0.6406
+8,3.2,2,55,aircraft,-15.442,-19.116,3.773,0.6289
+9,3.6,2,55,aircraft,-15.278,-19.053,3.773,0.6172
+10,4.0,2,55,aircraft,-15.285,-19.163,3.773,0.6055
+11,4.4,2,55,aircraft,-15.278,-19.053,3.773,0.5938
+12,4.8,2,55,aircraft,-15.195,-18.908,3.773,0.582
+13,5.2,2,55,aircraft,-15.113,-18.915,3.758,0.5742
+14,5.6,2,55,aircraft,-15.202,-18.939,3.758,0.5664
+15,6.0,2,55,aircraft,-15.285,-18.933,3.758,0.5547
+16,6.4,2,55,aircraft,-15.113,-18.838,3.758,0.543
+17,6.8,2,55,aircraft,-15.12,-18.793,3.758,0.5352
+18,7.2,2,55,aircraft,-15.12,-18.869,3.758,0.5312
+19,7.6,2,55,aircraft,-14.954,-18.729,3.758,0.5195
+20,8.0,2,55,aircraft,-14.954,-18.729,3.758,0.5117
+21,8.4,2,55,aircraft,-15.037,-18.722,3.758,0.5039
+22,8.8,2,55,aircraft,-14.87,-18.736,3.758,0.4961
+23,9.2,2,55,aircraft,-14.954,-18.652,3.758,0.4902
+24,9.6,2,55,aircraft,-14.87,-18.581,3.758,0.4824
+25,10.0,2,55,aircraft,-14.87,-18.581,3.758,0.4785
+26,10.4,2,55,aircraft,-14.783,-18.405,3.758,0.4707
+27,10.8,2,55,aircraft,-14.87,-18.504,3.758,0.4648
+28,11.2,2,55,aircraft,-14.87,-18.504,3.758,0.459
+29,11.6,2,55,aircraft,-14.87,-18.427,3.758,0.4551
+30,12.0,2,55,aircraft,-14.865,-18.321,3.758,0.4492
+31,12.4,2,55,aircraft,-14.865,-18.321,3.758,0.4453
+32,12.8,2,55,aircraft,-15.037,-18.336,3.758,0.4414
+33,13.2,2,55,aircraft,-15.037,-18.336,3.743,0.4355
+34,13.6,2,55,aircraft,-15.037,-18.259,3.758,0.4297
+35,14.0,2,55,aircraft,-14.87,-18.272,3.758,0.4277
+36,14.4,2,55,aircraft,-15.037,-18.181,3.743,0.4238
+37,14.8,2,55,aircraft,-14.87,-18.117,3.743,0.4199
+38,15.2,2,55,aircraft,-14.954,-18.032,3.743,0.416
+39,15.6,2,55,aircraft,-14.787,-17.968,3.743,0.4121
+40,16.0,2,55,aircraft,-14.954,-17.799,3.727,0.4102
+41,16.4,2,55,aircraft,-14.87,-17.805,3.727,0.4082
+42,16.8,2,55,aircraft,-14.876,-17.752,3.727,0.4102
+43,17.2,2,55,aircraft,-14.792,-17.602,3.727,0.4082
+44,17.6,2,55,aircraft,-14.792,-17.444,3.712,0.4102
+45,18.0,2,55,aircraft,-14.708,-17.372,3.712,0.4102
+46,18.4,2,55,aircraft,-14.713,-17.315,3.697,0.4102
+47,18.8,2,55,aircraft,-14.713,-17.236,3.697,0.4082
+48,19.2,2,55,aircraft,-14.544,-17.009,3.682,0.4062
+49,19.6,2,55,aircraft,-14.544,-16.93,3.682,0.3926
+0,0.0,3,55,aircraft,-16.427,-23.745,3.712,0.9102
+1,0.4,3,55,aircraft,-16.172,-23.989,3.697,0.9531
+2,0.8,3,55,aircraft,-15.785,-23.913,3.667,0.9609
+3,1.2,3,55,aircraft,-15.247,-23.992,3.651,0.9648
+4,1.6,3,55,aircraft,-14.983,-23.733,3.636,0.9609
+5,2.0,3,55,aircraft,-14.375,-23.685,3.621,0.9609
+6,2.4,3,55,aircraft,-13.617,-23.491,3.606,0.957
+7,2.8,3,55,aircraft,-13.123,-23.277,3.606,0.957
+8,3.2,3,55,aircraft,-12.804,-22.907,3.59,0.9531
+9,3.6,3,55,aircraft,-12.165,-22.711,3.59,0.9492
+10,4.0,3,55,aircraft,-11.701,-22.504,3.59,0.9453
+11,4.4,3,55,aircraft,-11.407,-22.428,3.575,0.9453
+12,4.8,3,55,aircraft,-11.039,-22.219,3.56,0.9414
+13,5.2,3,55,aircraft,-10.679,-22.014,3.56,0.9414
+14,5.6,3,55,aircraft,-10.462,-21.914,3.545,0.9375
+15,6.0,3,55,aircraft,-10.269,-21.84,3.545,0.9336
+16,6.4,3,55,aircraft,-9.849,-21.645,3.53,0.9336
+17,6.8,3,55,aircraft,-9.733,-21.652,3.515,0.9297
+18,7.2,3,55,aircraft,-9.513,-21.449,3.515,0.9297
+19,7.6,3,55,aircraft,-9.284,-21.327,3.499,0.9258
+20,8.0,3,55,aircraft,-9.146,-21.123,3.499,0.9258
+21,8.4,3,55,aircraft,-8.909,-20.898,3.484,0.9219
+22,8.8,3,55,aircraft,-8.72,-20.908,3.484,0.9219
+23,9.2,3,55,aircraft,-8.644,-20.778,3.469,0.918
+24,9.6,3,55,aircraft,-8.493,-20.652,3.469,0.918
+25,10.0,3,55,aircraft,-8.373,-20.556,3.454,0.918
+26,10.4,3,55,aircraft,-8.259,-20.427,3.439,0.918
+27,10.8,3,55,aircraft,-8.259,-20.293,3.439,0.918
+28,11.2,3,55,aircraft,-8.066,-20.203,3.439,0.918
+29,11.6,3,55,aircraft,-8.109,-19.897,3.424,0.9141
+30,12.0,3,55,aircraft,-7.953,-19.671,3.409,0.918
+31,12.4,3,55,aircraft,-7.803,-19.543,3.409,0.918
+32,12.8,3,55,aircraft,-7.882,-19.637,3.394,0.918
+33,13.2,3,55,aircraft,-7.92,-19.5,3.378,0.918
+34,13.6,3,55,aircraft,-7.769,-19.371,3.378,0.918
+35,14.0,3,55,aircraft,-7.658,-19.199,3.378,0.918
+36,14.4,3,55,aircraft,-7.658,-19.062,3.363,0.918
+37,14.8,3,55,aircraft,-7.47,-19.028,3.363,0.918
+38,15.2,3,55,aircraft,-7.317,-18.827,3.348,0.918
+39,15.6,3,55,aircraft,-7.087,-18.628,3.333,0.918
+40,16.0,3,55,aircraft,-7.241,-18.508,3.318,0.918
+41,16.4,3,55,aircraft,-7.125,-18.442,3.318,0.9141
+42,16.8,3,55,aircraft,-6.931,-18.262,3.303,0.9141
+43,17.2,3,55,aircraft,-6.66,-18.131,3.303,0.9141
+44,17.6,3,55,aircraft,-6.657,-17.869,3.288,0.9102
+45,18.0,3,55,aircraft,-6.458,-17.897,3.273,0.9062
+46,18.4,3,55,aircraft,-6.336,-17.779,3.273,0.9062
+47,18.8,3,55,aircraft,-6.217,-17.638,3.273,0.9023
+48,19.2,3,55,aircraft,-6.093,-17.517,3.258,0.8984
+49,19.6,3,55,aircraft,-6.087,-17.464,3.258,0.8906

checkpoints/spatial_beats_ov1_local_spatial_v7j_ov123_exp/03_ov123_top4/val_predictions/epoch_0001_csv/valid__ov2_000156__pred.csv

@@ -0,0 +1,201 @@
+frame_idx,frame_time_s,src_or_track_idx,class_idx,class_name,azimuth_deg,elevation_deg,distance_m,activity_prob
+0,0.0,0,35,insect,-92.579,-19.456,3.896,0.8086
+1,0.4,0,35,insect,-92.913,-19.244,3.911,0.8359
+2,0.8,0,35,insect,-92.945,-19.131,3.911,0.8398
+3,1.2,0,35,insect,-93.152,-18.937,3.911,0.8438
+4,1.6,0,35,insect,-93.309,-18.708,3.911,0.8398
+5,2.0,0,35,insect,-93.058,-18.938,3.911,0.8398
+6,2.4,0,35,insect,-93.261,-18.673,3.896,0.8359
+7,2.8,0,35,insect,-93.258,-19.056,3.911,0.832
+8,3.2,0,35,insect,-93.112,-19.138,3.896,0.8281
+9,3.6,0,35,insect,-93.624,-18.899,3.896,0.8242
+10,4.0,0,35,insect,-93.208,-18.677,3.896,0.8203
+11,4.4,0,35,insect,-93.261,-18.559,3.896,0.8125
+12,4.8,0,35,insect,-93.88,-18.664,3.896,0.8086
+13,5.2,0,35,insect,-93.801,-18.629,3.896,0.8086
+14,5.6,0,35,insect,-93.351,-19.099,3.896,0.8008
+15,6.0,0,35,insect,-93.226,-18.599,3.88,0.7969
+16,6.4,0,35,insect,-93.786,-18.939,3.88,0.793
+17,6.8,0,35,insect,-93.287,-18.869,3.88,0.7891
+18,7.2,0,35,insect,-93.321,-18.597,3.865,0.7852
+19,7.6,0,35,insect,-93.334,-18.716,3.865,0.7773
+20,8.0,0,35,insect,-93.062,-18.642,3.865,0.7734
+21,8.4,0,35,insect,-92.869,-18.529,3.85,0.7695
+22,8.8,0,35,insect,-93.076,-18.488,3.85,0.7656
+23,9.2,0,35,insect,-92.632,-18.455,3.85,0.7656
+24,9.6,0,35,insect,-92.837,-18.761,3.85,0.7617
+25,10.0,0,35,insect,-92.375,-18.459,3.834,0.7617
+26,10.4,0,35,insect,-92.869,-18.529,3.819,0.7578
+27,10.8,0,35,insect,-92.429,-18.151,3.819,0.7578
+28,11.2,0,35,insect,-91.979,-18.045,3.819,0.7539
+29,11.6,0,35,insect,-91.798,-17.926,3.804,0.7539
+30,12.0,0,35,insect,-91.133,-17.668,3.804,0.75
+31,12.4,0,35,insect,-90.93,-17.338,3.789,0.7539
+32,12.8,0,35,insect,-90.887,-17.494,3.789,0.75
+33,13.2,0,35,insect,-90.926,-17.153,3.773,0.7461
+34,13.6,0,35,insect,-90.438,-17.059,3.758,0.7461
+35,14.0,0,35,insect,-90.517,-16.676,3.758,0.7422
+36,14.4,0,35,insect,-89.977,-16.677,3.743,0.7344
+37,14.8,0,35,insect,-89.863,-16.632,3.727,0.7266
+38,15.2,0,35,insect,-89.798,-16.435,3.712,0.7109
+39,15.6,0,35,insect,-89.395,-15.866,3.682,0.6953
+40,16.0,0,35,insect,-89.134,-15.632,3.667,0.6758
+41,16.4,0,35,insect,-88.471,-15.387,3.651,0.6523
+42,16.8,0,35,insect,-88.862,-15.282,3.636,0.6289
+43,17.2,0,35,insect,-88.632,-14.999,3.621,0.5977
+44,17.6,0,35,insect,-88.703,-14.457,3.59,0.5547
+45,18.0,0,35,insect,-88.527,-14.561,3.575,0.5234
+46,18.4,0,35,insect,-88.097,-14.558,3.56,0.5
+47,18.8,0,35,insect,-88.295,-14.348,3.545,0.4824
+48,19.2,0,35,insect,-88.028,-14.133,3.53,0.4609
+49,19.6,0,35,insect,-87.886,-14.027,3.53,0.4551
+0,0.0,1,35,insect,-90.976,-22.752,4.003,0.8945
+1,0.4,1,35,insect,-91.486,-22.089,4.003,0.9258
+2,0.8,1,35,insect,-91.505,-21.682,4.003,0.9336
+3,1.2,1,35,insect,-91.79,-21.656,4.003,0.9336
+4,1.6,1,35,insect,-91.693,-21.142,4.003,0.9375
+5,2.0,1,35,insect,-91.164,-21.347,4.003,0.9336
+6,2.4,1,35,insect,-91.447,-21.206,4.003,0.9336
+7,2.8,1,35,insect,-91.486,-21.522,4.018,0.9336
+8,3.2,1,35,insect,-91.229,-21.5,4.018,0.9336
+9,3.6,1,35,insect,-91.7,-21.179,4.018,0.9297
+10,4.0,1,35,insect,-91.18,-21.157,4.018,0.9297
+11,4.4,1,35,insect,-91.212,-21.007,4.018,0.9297
+12,4.8,1,35,insect,-91.831,-21.063,4.018,0.9258
+13,5.2,1,35,insect,-91.691,-21.038,4.018,0.9258
+14,5.6,1,35,insect,-90.899,-21.388,4.018,0.9219
+15,6.0,1,35,insect,-91.006,-20.982,4.018,0.9219
+16,6.4,1,35,insect,-91.312,-21.156,4.018,0.918
+17,6.8,1,35,insect,-90.85,-21.274,4.003,0.918
+18,7.2,1,35,insect,-91.015,-21.158,4.003,0.9141
+19,7.6,1,35,insect,-90.991,-21.106,4.003,0.9141
+20,8.0,1,35,insect,-90.564,-21.25,4.003,0.9102
+21,8.4,1,35,insect,-90.265,-21.02,4.003,0.9102
+22,8.8,1,35,insect,-90.575,-20.992,4.003,0.9062
+23,9.2,1,35,insect,-90.116,-21.02,4.003,0.9062
+24,9.6,1,35,insect,-90.333,-21.34,4.003,0.9023
+25,10.0,1,35,insect,-89.764,-21.135,3.987,0.9023
+26,10.4,1,35,insect,-90.258,-20.993,3.987,0.9023
+27,10.8,1,35,insect,-89.702,-20.904,3.987,0.8984
+28,11.2,1,35,insect,-89.228,-20.727,3.987,0.8945
+29,11.6,1,35,insect,-88.997,-20.669,3.972,0.8945
+30,12.0,1,35,insect,-88.086,-20.459,3.972,0.8906
+31,12.4,1,35,insect,-88.153,-20.173,3.972,0.8867
+32,12.8,1,35,insect,-88.039,-20.202,3.972,0.8828
+33,13.2,1,35,insect,-87.894,-20.002,3.957,0.8789
+34,13.6,1,35,insect,-87.582,-19.915,3.957,0.875
+35,14.0,1,35,insect,-87.49,-19.538,3.941,0.8672
+36,14.4,1,35,insect,-86.947,-19.434,3.941,0.8555
+37,14.8,1,35,insect,-86.681,-19.444,3.941,0.8438
+38,15.2,1,35,insect,-86.456,-19.279,3.911,0.8164
+39,15.6,1,35,insect,-85.787,-18.923,3.911,0.793
+40,16.0,1,35,insect,-85.604,-18.65,3.896,0.7656
+41,16.4,1,35,insect,-84.885,-18.404,3.88,0.7305
+42,16.8,1,35,insect,-84.893,-18.367,3.88,0.6914
+43,17.2,1,35,insect,-84.474,-17.982,3.865,0.6484
+44,17.6,1,35,insect,-84.277,-17.385,3.85,0.5898
+45,18.0,1,35,insect,-84.033,-17.489,3.834,0.543
+46,18.4,1,35,insect,-83.484,-17.472,3.819,0.5156
+47,18.8,1,35,insect,-83.362,-17.245,3.804,0.4883
+48,19.2,1,35,insect,-83.027,-17.009,3.789,0.459
+49,19.6,1,35,insect,-82.722,-16.912,3.789,0.4512
+0,0.0,2,35,insect,-91.976,-19.607,3.972,0.8203
+1,0.4,2,35,insect,-92.456,-19.299,3.987,0.832
+2,0.8,2,35,insect,-93.016,-19.04,3.987,0.8281
+3,1.2,2,35,insect,-93.383,-19.121,3.972,0.8281
+4,1.6,2,35,insect,-93.524,-18.866,3.972,0.8242
+5,2.0,2,35,insect,-93.031,-18.875,3.972,0.8164
+6,2.4,2,35,insect,-93.506,-18.7,3.972,0.8125
+7,2.8,2,35,insect,-93.702,-19.214,3.987,0.8086
+8,3.2,2,35,insect,-93.523,-19.218,3.972,0.8047
+9,3.6,2,35,insect,-94.152,-19.166,3.972,0.7969
+10,4.0,2,35,insect,-93.9,-18.913,3.972,0.793
+11,4.4,2,35,insect,-93.969,-18.736,3.972,0.7891
+12,4.8,2,35,insect,-94.74,-18.934,3.972,0.7812
+13,5.2,2,35,insect,-94.62,-18.897,3.972,0.7773
+14,5.6,2,35,insect,-93.973,-19.299,3.972,0.7734
+15,6.0,2,35,insect,-94.203,-18.819,3.972,0.7656
+16,6.4,2,35,insect,-94.679,-18.985,3.972,0.7578
+17,6.8,2,35,insect,-94.282,-19.253,3.957,0.7578
+18,7.2,2,35,insect,-94.476,-19.03,3.957,0.75
+19,7.6,2,35,insect,-94.508,-19.039,3.957,0.7461
+20,8.0,2,35,insect,-94.122,-19.218,3.957,0.7383
+21,8.4,2,35,insect,-94.34,-18.952,3.941,0.7344
+22,8.8,2,35,insect,-94.362,-18.911,3.941,0.7305
+23,9.2,2,35,insect,-93.977,-18.961,3.941,0.7266
+24,9.6,2,35,insect,-94.545,-19.429,3.941,0.7227
+25,10.0,2,35,insect,-93.831,-19.094,3.926,0.7188
+26,10.4,2,35,insect,-94.508,-19.039,3.926,0.7188
+27,10.8,2,35,insect,-94.049,-18.959,3.926,0.7148
+28,11.2,2,35,insect,-93.415,-18.706,3.926,0.7109
+29,11.6,2,35,insect,-93.54,-18.708,3.911,0.7109
+30,12.0,2,35,insect,-92.761,-18.633,3.911,0.707
+31,12.4,2,35,insect,-92.684,-18.244,3.911,0.707
+32,12.8,2,35,insect,-92.457,-18.419,3.911,0.707
+33,13.2,2,35,insect,-92.551,-18.077,3.896,0.7031
+34,13.6,2,35,insect,-92.117,-18.21,3.896,0.6992
+35,14.0,2,35,insect,-92.235,-17.908,3.88,0.6953
+36,14.4,2,35,insect,-91.406,-17.841,3.865,0.6875
+37,14.8,2,35,insect,-91.437,-17.878,3.865,0.6836
+38,15.2,2,35,insect,-91.336,-17.715,3.85,0.6719
+39,15.6,2,35,insect,-90.742,-17.209,3.834,0.6562
+40,16.0,2,35,insect,-90.379,-16.85,3.819,0.6406
+41,16.4,2,35,insect,-89.448,-16.624,3.804,0.6172
+42,16.8,2,35,insect,-89.612,-16.574,3.789,0.5938
+43,17.2,2,35,insect,-89.13,-16.3,3.773,0.5664
+44,17.6,2,35,insect,-88.793,-15.651,3.758,0.5234
+45,18.0,2,35,insect,-88.435,-15.711,3.727,0.4902
+46,18.4,2,35,insect,-87.769,-15.705,3.727,0.4668
+47,18.8,2,35,insect,-87.869,-15.397,3.712,0.4492
+48,19.2,2,35,insect,-87.482,-15.263,3.697,0.4258
+49,19.6,2,35,insect,-87.258,-15.148,3.697,0.4199
+0,0.0,3,35,insect,-101.232,-10.541,3.85,0.9336
+1,0.4,3,35,insect,-100.479,-10.492,3.834,0.9688
+2,0.8,3,35,insect,-99.844,-10.152,3.819,0.9766
+3,1.2,3,35,insect,-99.707,-10.201,3.819,0.9805
+4,1.6,3,35,insect,-99.357,-9.904,3.804,0.9805
+5,2.0,3,35,insect,-98.722,-9.965,3.804,0.9844
+6,2.4,3,35,insect,-98.531,-9.552,3.789,0.9844
+7,2.8,3,35,insect,-98.267,-9.932,3.789,0.9844
+8,3.2,3,35,insect,-97.942,-9.833,3.773,0.9844
+9,3.6,3,35,insect,-98.078,-9.563,3.773,0.9805
+10,4.0,3,35,insect,-97.716,-9.571,3.773,0.9805
+11,4.4,3,35,insect,-97.574,-9.455,3.758,0.9805
+12,4.8,3,35,insect,-97.843,-9.361,3.758,0.9805
+13,5.2,3,35,insect,-97.574,-9.279,3.743,0.9805
+14,5.6,3,35,insect,-97.125,-9.58,3.743,0.9805
+15,6.0,3,35,insect,-96.992,-9.306,3.743,0.9805
+16,6.4,3,35,insect,-97.125,-9.361,3.727,0.9766
+17,6.8,3,35,insect,-96.814,-9.484,3.727,0.9766
+18,7.2,3,35,insect,-96.86,-9.295,3.712,0.9766
+19,7.6,3,35,insect,-96.684,-9.212,3.712,0.9766
+20,8.0,3,35,insect,-96.209,-9.194,3.697,0.9766
+21,8.4,3,35,insect,-96.34,-9.149,3.697,0.9727
+22,8.8,3,35,insect,-96.209,-9.065,3.682,0.9727
+23,9.2,3,35,insect,-95.954,-9.015,3.682,0.9727
+24,9.6,3,35,insect,-96.263,-9.167,3.682,0.9688
+25,10.0,3,35,insect,-95.839,-9.035,3.667,0.9688
+26,10.4,3,35,insect,-96.097,-9.031,3.667,0.9688
+27,10.8,3,35,insect,-95.897,-8.803,3.667,0.9648
+28,11.2,3,35,insect,-95.702,-8.704,3.651,0.9609
+29,11.6,3,35,insect,-95.408,-8.543,3.651,0.957
+30,12.0,3,35,insect,-94.992,-8.368,3.636,0.9531
+31,12.4,3,35,insect,-94.976,-8.192,3.636,0.9492
+32,12.8,3,35,insect,-95.111,-8.144,3.636,0.9453
+33,13.2,3,35,insect,-95.163,-7.938,3.621,0.9375
+34,13.6,3,35,insect,-95.065,-7.89,3.621,0.9336
+35,14.0,3,35,insect,-95.176,-7.648,3.606,0.9258
+36,14.4,3,35,insect,-95.02,-7.605,3.606,0.9102
+37,14.8,3,35,insect,-95.108,-7.528,3.59,0.8984
+38,15.2,3,35,insect,-95.052,-7.369,3.59,0.875
+39,15.6,3,35,insect,-95.036,-7.098,3.575,0.8477
+40,16.0,3,35,insect,-95.201,-6.867,3.56,0.8242
+41,16.4,3,35,insect,-94.821,-6.873,3.56,0.7969
+42,16.8,3,35,insect,-95.135,-6.755,3.545,0.7617
+43,17.2,3,35,insect,-95.031,-6.644,3.53,0.7266
+44,17.6,3,35,insect,-95.156,-6.382,3.515,0.6797
+45,18.0,3,35,insect,-95.308,-6.38,3.515,0.6406
+46,18.4,3,35,insect,-94.966,-6.346,3.499,0.6133
+47,18.8,3,35,insect,-95.069,-6.265,3.484,0.5938
+48,19.2,3,35,insect,-94.993,-6.266,3.484,0.5664
+49,19.6,3,35,insect,-94.924,-6.071,3.484,0.5586

checkpoints/spatial_beats_ov1_local_spatial_v7j_ov123_exp/03_ov123_top4/val_predictions/epoch_0001_csv/valid__ov3_000079__pred.csv

@@ -0,0 +1,201 @@
+frame_idx,frame_time_s,src_or_track_idx,class_idx,class_name,azimuth_deg,elevation_deg,distance_m,activity_prob
+0,0.0,0,48,appliance,-140.389,-13.446,3.273,0.6445
+1,0.4,0,48,appliance,-140.194,-13.8,3.288,0.7656
+2,0.8,0,48,appliance,-140.232,-13.977,3.288,0.8008
+3,1.2,0,48,appliance,-140.119,-14.094,3.288,0.8164
+4,1.6,0,48,appliance,-140.119,-14.094,3.288,0.8203
+5,2.0,0,48,appliance,-140.119,-14.233,3.288,0.8203
+6,2.4,0,48,appliance,-140.119,-14.302,3.288,0.8203
+7,2.8,0,48,appliance,-139.929,-14.356,3.288,0.8203
+8,3.2,0,48,appliance,-140.119,-14.302,3.288,0.8164
+9,3.6,0,48,appliance,-140.232,-14.394,3.288,0.8125
+10,4.0,0,48,appliance,-140.119,-14.371,3.273,0.8086
+11,4.4,0,48,appliance,-140.119,-14.371,3.273,0.8047
+12,4.8,0,38,camera,-140.119,-14.302,3.273,0.8008
+13,5.2,0,38,camera,-140.119,-14.302,3.273,0.7969
+14,5.6,0,38,camera,-140.119,-14.302,3.273,0.7891
+15,6.0,0,38,camera,-140.119,-14.302,3.273,0.7852
+16,6.4,0,38,camera,-140.232,-14.325,3.258,0.7773
+17,6.8,0,38,camera,-140.232,-14.255,3.258,0.7734
+18,7.2,0,38,camera,-140.232,-14.325,3.258,0.7617
+19,7.6,0,38,camera,-140.346,-14.278,3.258,0.7539
+20,8.0,0,38,camera,-140.346,-14.208,3.258,0.7461
+21,8.4,0,38,camera,-140.46,-14.161,3.243,0.7383
+22,8.8,0,38,camera,-140.46,-14.091,3.243,0.7305
+23,9.2,0,38,camera,-140.575,-14.044,3.228,0.7266
+24,9.6,0,38,camera,-140.575,-14.044,3.228,0.7188
+25,10.0,0,38,camera,-140.575,-13.904,3.213,0.7148
+26,10.4,0,38,camera,-140.575,-13.834,3.198,0.707
+27,10.8,0,38,camera,-140.575,-13.764,3.198,0.6992
+28,11.2,0,38,camera,-140.69,-13.716,3.183,0.6953
+29,11.6,0,38,camera,-140.616,-13.579,3.168,0.6914
+30,12.0,0,38,camera,-140.616,-13.437,3.168,0.6875
+31,12.4,0,38,camera,-140.616,-13.296,3.138,0.6836
+32,12.8,0,38,camera,-140.616,-13.225,3.138,0.6836
+33,13.2,0,38,camera,-140.541,-13.227,3.108,0.6797
+34,13.6,0,38,camera,-140.583,-13.034,3.093,0.6797
+35,14.0,0,38,camera,-140.507,-12.89,3.078,0.6758
+36,14.4,0,38,camera,-140.43,-12.744,3.049,0.6758
+37,14.8,0,38,camera,-140.711,-12.659,3.019,0.6719
+38,15.2,0,38,camera,-140.477,-12.504,2.959,0.668
+39,15.6,0,38,camera,-140.688,-12.334,2.915,0.6641
+40,16.0,0,38,camera,-140.859,-12.218,2.871,0.6602
+41,16.4,0,38,camera,-140.874,-12.169,2.827,0.6562
+42,16.8,0,38,camera,-140.89,-11.956,2.797,0.6523
+43,17.2,0,38,camera,-141.124,-11.844,2.739,0.6406
+44,17.6,0,38,camera,-141.201,-11.797,2.68,0.625
+45,18.0,0,38,camera,-141.34,-11.704,2.651,0.6133
+46,18.4,0,38,camera,-141.456,-11.645,2.637,0.6016
+47,18.8,0,38,camera,-141.429,-11.695,2.608,0.5898
+48,19.2,0,38,camera,-141.582,-11.64,2.579,0.5742
+49,19.6,0,38,camera,-141.461,-11.577,2.579,0.5586
+0,0.0,1,48,appliance,-140.787,-12.734,3.303,0.7305
+1,0.4,1,48,appliance,-140.774,-12.911,3.318,0.832
+2,0.8,1,48,appliance,-140.732,-13.175,3.318,0.8633
+3,1.2,1,48,appliance,-140.616,-13.296,3.318,0.875
+4,1.6,1,48,appliance,-140.5,-13.275,3.318,0.8789
+5,2.0,1,48,appliance,-140.5,-13.416,3.318,0.8828
+6,2.4,1,48,appliance,-140.5,-13.416,3.318,0.8828
+7,2.8,1,48,appliance,-140.5,-13.487,3.318,0.8828
+8,3.2,1,48,appliance,-140.5,-13.487,3.318,0.8828
+9,3.6,1,48,appliance,-140.5,-13.557,3.318,0.8789
+10,4.0,1,48,appliance,-140.5,-13.557,3.318,0.8789
+11,4.4,1,48,appliance,-140.385,-13.536,3.318,0.875
+12,4.8,1,48,appliance,-140.575,-13.483,3.318,0.8711
+13,5.2,1,48,appliance,-140.5,-13.557,3.318,0.8672
+14,5.6,1,48,appliance,-140.5,-13.557,3.318,0.8633
+15,6.0,1,48,appliance,-140.385,-13.536,3.318,0.8594
+16,6.4,1,48,appliance,-140.385,-13.536,3.318,0.8555
+17,6.8,1,48,appliance,-140.5,-13.557,3.318,0.8516
+18,7.2,1,48,appliance,-140.575,-13.483,3.318,0.8438
+19,7.6,1,48,appliance,-140.5,-13.487,3.318,0.8359
+20,8.0,1,48,appliance,-140.616,-13.508,3.318,0.832
+21,8.4,1,48,appliance,-140.616,-13.508,3.318,0.8242
+22,8.8,1,48,appliance,-140.805,-13.385,3.318,0.8164
+23,9.2,1,48,appliance,-140.616,-13.437,3.318,0.8125
+24,9.6,1,48,appliance,-140.805,-13.385,3.318,0.8047
+25,10.0,1,48,appliance,-140.805,-13.315,3.318,0.7969
+26,10.4,1,48,appliance,-140.921,-13.265,3.303,0.793
+27,10.8,1,48,appliance,-140.921,-13.265,3.303,0.7852
+28,11.2,1,48,appliance,-140.921,-13.194,3.303,0.7773
+29,11.6,1,48,appliance,-140.848,-13.054,3.303,0.7734
+30,12.0,1,48,appliance,-140.732,-13.033,3.303,0.7656
+31,12.4,1,48,appliance,-140.732,-12.891,3.288,0.7617
+32,12.8,1,48,appliance,-140.921,-12.84,3.288,0.7578
+33,13.2,1,48,appliance,-140.921,-12.769,3.273,0.75
+34,13.6,1,48,appliance,-140.848,-12.697,3.273,0.7461
+35,14.0,1,48,appliance,-140.848,-12.554,3.273,0.7383
+36,14.4,1,48,appliance,-140.848,-12.411,3.258,0.7344
+37,14.8,1,48,appliance,-140.891,-12.428,3.243,0.7227
+38,15.2,1,48,appliance,-140.818,-12.208,3.213,0.7109
+39,15.6,1,48,appliance,-140.862,-12.077,3.198,0.6992
+40,16.0,1,48,appliance,-140.906,-12.018,3.183,0.6875
+41,16.4,1,48,appliance,-140.755,-11.933,3.153,0.668
+42,16.8,1,48,appliance,-140.722,-11.714,3.123,0.6523
+43,17.2,1,48,appliance,-140.812,-11.585,3.093,0.6289
+44,17.6,1,48,appliance,-141.033,-11.394,3.049,0.5977
+45,18.0,1,48,appliance,-140.922,-11.305,3.019,0.5703
+46,18.4,1,48,appliance,-141.234,-11.193,3.004,0.5508
+47,18.8,1,48,appliance,-141.072,-11.171,2.989,0.5312
+48,19.2,1,48,appliance,-141.093,-11.009,2.959,0.5117
+49,19.6,1,48,appliance,-140.872,-10.975,2.959,0.4941
+0,0.0,2,48,appliance,-141.861,-12.341,3.258,0.6523
+1,0.4,2,48,appliance,-141.755,-12.679,3.258,0.7266
+2,0.8,2,48,appliance,-141.703,-12.807,3.273,0.7461
+3,1.2,2,48,appliance,-141.582,-12.86,3.273,0.75
+4,1.6,2,48,appliance,-141.582,-12.86,3.273,0.75
+5,2.0,2,48,appliance,-141.582,-13.007,3.273,0.7422
+6,2.4,2,48,appliance,-141.703,-13.102,3.273,0.7383
+7,2.8,2,48,appliance,-141.703,-13.175,3.273,0.7305
+8,3.2,2,48,appliance,-141.703,-13.102,3.273,0.7227
+9,3.6,2,48,appliance,-141.703,-13.175,3.273,0.7148
+10,4.0,2,48,appliance,-141.703,-13.175,3.273,0.707
+11,4.4,2,48,appliance,-141.703,-13.175,3.273,0.6992
+12,4.8,2,48,appliance,-141.703,-13.175,3.258,0.6875
+13,5.2,2,48,appliance,-141.703,-13.175,3.258,0.6836
+14,5.6,2,48,appliance,-141.633,-13.251,3.258,0.6719
+15,6.0,2,48,appliance,-141.703,-13.175,3.258,0.6602
+16,6.4,2,48,appliance,-141.633,-13.251,3.258,0.6523
+17,6.8,2,48,appliance,-141.948,-13.144,3.258,0.6406
+18,7.2,2,48,appliance,-141.825,-13.123,3.258,0.6289
+19,7.6,2,48,appliance,-141.755,-13.199,3.258,0.6172
+20,8.0,2,48,appliance,-141.878,-13.146,3.258,0.6055
+21,8.4,2,48,appliance,-141.878,-13.146,3.243,0.5938
+22,8.8,2,48,appliance,-142.193,-12.964,3.243,0.582
+23,9.2,2,48,appliance,-142.001,-13.093,3.243,0.5742
+24,9.6,2,48,appliance,-142.001,-13.019,3.243,0.5625
+25,10.0,2,48,appliance,-142.193,-12.89,3.228,0.5547
+26,10.4,2,48,appliance,-142.125,-12.891,3.228,0.5469
+27,10.8,2,48,appliance,-142.001,-12.87,3.228,0.5352
+28,11.2,2,48,appliance,-142.125,-12.816,3.228,0.5273
+29,11.6,2,48,appliance,-142.249,-12.612,3.213,0.5234
+30,12.0,2,48,appliance,-142.249,-12.537,3.213,0.5156
+31,12.4,2,48,appliance,-142.125,-12.442,3.198,0.5117
+32,12.8,2,48,appliance,-142.249,-12.387,3.183,0.5078
+33,13.2,2,48,appliance,-142.181,-12.385,3.183,0.5
+34,13.6,2,48,appliance,-142.306,-12.178,3.168,0.498
+35,14.0,2,48,appliance,-142.111,-12.078,3.153,0.4941
+36,14.4,2,48,appliance,-142.041,-11.92,3.138,0.4922
+37,14.8,2,10,machine,-142.224,-11.954,3.108,0.4883
+38,15.2,2,10,machine,-142.082,-11.708,3.078,0.4863
+39,15.6,2,10,machine,-142.198,-11.574,3.049,0.4824
+40,16.0,2,10,machine,-142.257,-11.505,3.019,0.4805
+41,16.4,2,10,machine,-142.245,-11.345,2.989,0.4766
+42,16.8,2,10,machine,-142.293,-11.272,2.945,0.4727
+43,17.2,2,10,machine,-142.312,-11.148,2.915,0.4668
+44,17.6,2,10,machine,-142.509,-11.044,2.871,0.457
+45,18.0,2,10,machine,-142.647,-10.971,2.827,0.4473
+46,18.4,2,10,machine,-142.904,-10.913,2.812,0.4414
+47,18.8,2,10,machine,-142.595,-10.91,2.797,0.4297
+48,19.2,2,10,machine,-142.896,-10.853,2.768,0.4238
+49,19.6,2,10,machine,-142.778,-10.834,2.768,0.4082
+0,0.0,3,48,appliance,-144.149,-13.1,3.093,0.6445
+1,0.4,3,48,appliance,-143.951,-13.385,3.093,0.832
+2,0.8,3,48,appliance,-143.887,-13.568,3.108,0.8906
+3,1.2,3,48,appliance,-143.653,-13.666,3.093,0.9102
+4,1.6,3,48,appliance,-143.536,-13.715,3.093,0.9219
+5,2.0,3,48,appliance,-143.536,-13.852,3.093,0.9297
+6,2.4,3,48,appliance,-143.536,-13.92,3.078,0.9336
+7,2.8,3,5,water,-143.653,-14.008,3.078,0.9336
+8,3.2,3,5,water,-143.536,-13.92,3.078,0.9375
+9,3.6,3,48,appliance,-143.536,-13.988,3.063,0.9375
+10,4.0,3,5,water,-143.536,-13.92,3.063,0.9375
+11,4.4,3,5,water,-143.536,-13.92,3.063,0.9375
+12,4.8,3,38,camera,-143.707,-13.866,3.049,0.9336
+13,5.2,3,38,camera,-143.653,-13.872,3.049,0.9336
+14,5.6,3,38,camera,-143.653,-13.803,3.034,0.9336
+15,6.0,3,38,camera,-143.48,-13.858,3.034,0.9297
+16,6.4,3,38,camera,-143.597,-13.809,3.019,0.9297
+17,6.8,3,38,camera,-143.715,-13.76,3.004,0.9258
+18,7.2,3,38,camera,-143.541,-13.676,3.004,0.9258
+19,7.6,3,38,camera,-143.778,-13.716,2.989,0.9219
+20,8.0,3,38,camera,-143.722,-13.652,2.974,0.9219
+21,8.4,3,38,camera,-143.786,-13.607,2.959,0.918
+22,8.8,3,38,camera,-143.73,-13.54,2.945,0.9141
+23,9.2,3,38,camera,-143.917,-13.514,2.93,0.9141
+24,9.6,3,38,camera,-143.861,-13.518,2.915,0.9102
+25,10.0,3,38,camera,-144.052,-13.346,2.886,0.9062
+26,10.4,3,38,camera,-143.938,-13.353,2.871,0.9062
+27,10.8,3,38,camera,-144.007,-13.229,2.841,0.9023
+28,11.2,3,38,camera,-144.148,-13.276,2.827,0.9023
+29,11.6,3,38,camera,-144.162,-13.074,2.812,0.8945
+30,12.0,3,38,camera,-144.118,-13.023,2.783,0.8945
+31,12.4,3,14,laughter,-144.072,-12.97,2.739,0.8906
+32,12.8,3,14,laughter,-144.345,-12.853,2.724,0.8867
+33,13.2,3,14,laughter,-144.302,-12.958,2.695,0.8828
+34,13.6,3,14,laughter,-144.258,-12.817,2.68,0.8789
+35,14.0,3,14,laughter,-144.421,-12.693,2.651,0.875
+36,14.4,3,14,laughter,-144.27,-12.746,2.622,0.8633
+37,14.8,3,14,laughter,-144.593,-12.727,2.608,0.8555
+38,15.2,3,14,laughter,-144.53,-12.624,2.564,0.8438
+39,15.6,3,14,laughter,-144.86,-12.659,2.536,0.8281
+40,16.0,3,14,laughter,-144.891,-12.635,2.507,0.8164
+41,16.4,3,14,laughter,-144.853,-12.758,2.493,0.793
+42,16.8,3,14,laughter,-145.291,-12.646,2.478,0.7734
+43,17.2,3,14,laughter,-145.389,-12.632,2.464,0.7422
+44,17.6,3,14,laughter,-145.48,-12.739,2.435,0.7031
+45,18.0,3,14,laughter,-145.854,-12.699,2.421,0.668
+46,18.4,3,14,laughter,-145.911,-12.785,2.421,0.6406
+47,18.8,3,14,laughter,-145.768,-12.902,2.407,0.6172
+48,19.2,3,14,laughter,-145.892,-12.893,2.407,0.5898
+49,19.6,3,14,laughter,-145.685,-12.803,2.407,0.5703

checkpoints/spatial_beats_ov1_local_spatial_v7j_ov123_exp/03_ov123_top4/val_predictions/epoch_0001_csv/valid__ov3_000112__pred.csv

@@ -0,0 +1,201 @@
+frame_idx,frame_time_s,src_or_track_idx,class_idx,class_name,azimuth_deg,elevation_deg,distance_m,activity_prob
+0,0.0,0,34,home_sound,-114.752,47.18,1.108,0.6406
+1,0.4,0,34,home_sound,-114.485,46.845,1.134,0.832
+2,0.8,0,34,home_sound,-114.063,46.887,1.142,0.8867
+3,1.2,0,34,home_sound,-113.929,46.891,1.148,0.9062
+4,1.6,0,34,home_sound,-113.929,47.044,1.148,0.918
+5,2.0,0,34,home_sound,-113.629,47.261,1.148,0.9219
+6,2.4,0,34,home_sound,-113.629,47.412,1.145,0.9258
+7,2.8,0,34,home_sound,-113.479,47.444,1.148,0.9297
+8,3.2,0,34,home_sound,-113.349,47.295,1.148,0.9297
+9,3.6,0,34,home_sound,-113.349,47.295,1.148,0.9336
+10,4.0,0,34,home_sound,-113.22,47.295,1.145,0.9336
+11,4.4,0,34,home_sound,-113.22,47.444,1.145,0.9336
+12,4.8,0,34,home_sound,-113.22,47.295,1.148,0.9336
+13,5.2,0,34,home_sound,-113.07,47.476,1.145,0.9336
+14,5.6,0,34,home_sound,-112.921,47.359,1.148,0.9336
+15,6.0,0,34,home_sound,-112.818,47.179,1.145,0.9336
+16,6.4,0,34,home_sound,-112.794,47.359,1.148,0.9336
+17,6.8,0,34,home_sound,-112.944,47.327,1.142,0.9336
+18,7.2,0,34,home_sound,-112.669,47.21,1.145,0.9297
+19,7.6,0,34,home_sound,-112.794,47.359,1.142,0.9297
+20,8.0,0,34,home_sound,-112.645,47.39,1.145,0.9297
+21,8.4,0,34,home_sound,-112.345,47.452,1.142,0.9297
+22,8.8,0,34,home_sound,-112.371,47.272,1.145,0.9258
+23,9.2,0,34,home_sound,-112.345,47.452,1.145,0.9258
+24,9.6,0,34,home_sound,-112.194,47.483,1.145,0.9258
+25,10.0,0,34,home_sound,-112.043,47.513,1.145,0.9258
+26,10.4,0,34,home_sound,-111.741,47.425,1.148,0.9219
+27,10.8,0,34,home_sound,-111.892,47.543,1.145,0.9219
+28,11.2,0,34,home_sound,-111.741,47.574,1.148,0.9219
+29,11.6,0,34,home_sound,-111.437,47.485,1.15,0.918
+30,12.0,0,34,home_sound,-111.284,47.364,1.153,0.918
+31,12.4,0,34,home_sound,-111.055,47.558,1.15,0.918
+32,12.8,0,34,home_sound,-110.862,47.271,1.153,0.918
+33,13.2,0,34,home_sound,-110.825,47.452,1.153,0.9141
+34,13.6,0,34,home_sound,-110.672,47.481,1.158,0.9141
+35,14.0,0,34,home_sound,-110.323,47.722,1.153,0.9102
+36,14.4,0,34,home_sound,-110.245,47.736,1.161,0.9062
+37,14.8,0,34,home_sound,-109.969,47.661,1.164,0.9023
+38,15.2,0,34,home_sound,-109.608,47.904,1.166,0.8984
+39,15.6,0,34,home_sound,-109.081,48.022,1.172,0.8906
+40,16.0,0,34,home_sound,-109.193,47.876,1.177,0.8906
+41,16.4,0,34,home_sound,-108.654,48.149,1.177,0.8828
+42,16.8,0,34,home_sound,-108.296,48.419,1.18,0.875
+43,17.2,0,34,home_sound,-107.843,48.551,1.185,0.8672
+44,17.6,0,34,home_sound,-107.507,48.884,1.191,0.8555
+45,18.0,0,34,home_sound,-107.11,49.17,1.196,0.8438
+46,18.4,0,34,home_sound,-106.79,49.447,1.199,0.832
+47,18.8,0,34,home_sound,-105.998,49.794,1.204,0.8203
+48,19.2,0,34,home_sound,-105.959,50.115,1.207,0.8047
+49,19.6,0,34,home_sound,-105.482,50.181,1.207,0.7891
+0,0.0,1,34,home_sound,-115.006,48.218,1.092,0.5664
+1,0.4,1,34,home_sound,-114.585,47.868,1.126,0.7852
+2,0.8,1,34,home_sound,-114.305,47.866,1.134,0.8555
+3,1.2,1,34,home_sound,-114.014,47.899,1.142,0.8828
+4,1.6,1,34,home_sound,-113.878,47.897,1.145,0.8984
+5,2.0,1,34,home_sound,-113.86,48.233,1.148,0.9102
+6,2.4,1,34,home_sound,-113.572,48.262,1.148,0.9141
+7,2.8,1,34,home_sound,-113.572,48.114,1.15,0.918
+8,3.2,1,34,home_sound,-113.265,48.18,1.15,0.9219
+9,3.6,1,34,home_sound,-113.418,48.295,1.153,0.9258
+10,4.0,1,34,home_sound,-113.286,48.143,1.153,0.9258
+11,4.4,1,34,home_sound,-113.286,48.291,1.153,0.9258
+12,4.8,1,34,home_sound,-113.133,48.176,1.156,0.9297
+13,5.2,1,34,home_sound,-112.85,48.204,1.153,0.9297
+14,5.6,1,34,home_sound,-112.697,48.088,1.158,0.9297
+15,6.0,1,34,home_sound,-112.722,48.052,1.156,0.9297
+16,6.4,1,34,home_sound,-112.569,48.084,1.158,0.9297
+17,6.8,1,34,home_sound,-112.569,48.084,1.156,0.9297
+18,7.2,1,34,home_sound,-112.569,47.936,1.158,0.9297
+19,7.6,1,34,home_sound,-112.389,48.299,1.156,0.9297
+20,8.0,1,34,home_sound,-112.235,48.183,1.158,0.9297
+21,8.4,1,34,home_sound,-112.08,48.214,1.156,0.9297
+22,8.8,1,34,home_sound,-112.109,48.177,1.158,0.9297
+23,9.2,1,34,home_sound,-111.956,48.06,1.158,0.9297
+24,9.6,1,34,home_sound,-111.926,48.245,1.158,0.9297
+25,10.0,1,34,home_sound,-111.57,48.137,1.161,0.9258
+26,10.4,1,34,home_sound,-111.459,48.188,1.164,0.9258
+27,10.8,1,34,home_sound,-111.337,48.182,1.161,0.9258
+28,11.2,1,34,home_sound,-111.103,48.227,1.164,0.9258
+29,11.6,1,34,home_sound,-110.791,48.138,1.166,0.9219
+30,12.0,1,34,home_sound,-110.674,48.339,1.169,0.9219
+31,12.4,1,34,home_sound,-110.399,48.36,1.166,0.9219
+32,12.8,1,34,home_sound,-110.241,48.24,1.169,0.918
+33,13.2,1,34,home_sound,-110.041,48.306,1.169,0.918
+34,13.6,1,34,home_sound,-110.041,48.306,1.174,0.918
+35,14.0,1,34,home_sound,-109.482,48.255,1.172,0.9141
+36,14.4,1,34,home_sound,-109.596,48.415,1.177,0.9102
+37,14.8,1,34,home_sound,-109.225,48.51,1.18,0.9062
+38,15.2,1,34,home_sound,-108.765,48.618,1.185,0.9023
+39,15.6,1,34,home_sound,-108.213,48.74,1.188,0.8984
+40,16.0,1,34,home_sound,-108.351,48.933,1.193,0.8945
+41,16.4,1,34,home_sound,-107.782,49.211,1.196,0.8867
+42,16.8,1,34,home_sound,-107.311,49.352,1.196,0.8789
+43,17.2,1,34,home_sound,-106.931,49.826,1.201,0.8711
+44,17.6,1,34,home_sound,-106.552,50.192,1.207,0.8555
+45,18.0,1,34,home_sound,-106.12,50.493,1.207,0.8438
+46,18.4,1,34,home_sound,-105.673,50.639,1.212,0.832
+47,18.8,1,34,home_sound,-104.821,51.0,1.215,0.8203
+48,19.2,1,34,home_sound,-104.844,51.504,1.218,0.8008
+49,19.6,1,34,home_sound,-104.34,51.406,1.218,0.7852
+0,0.0,2,34,home_sound,-113.944,49.314,1.21,0.6914
+1,0.4,2,34,home_sound,-113.506,48.976,1.229,0.8359
+2,0.8,2,34,home_sound,-113.363,49.122,1.232,0.875
+3,1.2,2,34,home_sound,-113.222,49.112,1.234,0.8906
+4,1.6,2,34,home_sound,-113.222,49.264,1.234,0.8945
+5,2.0,2,34,home_sound,-112.838,49.305,1.232,0.8945
+6,2.4,2,34,home_sound,-112.756,49.473,1.229,0.8984
+7,2.8,2,34,home_sound,-112.812,49.652,1.232,0.8984
+8,3.2,2,34,home_sound,-112.593,49.356,1.232,0.8945
+9,3.6,2,34,home_sound,-112.593,49.506,1.232,0.8945
+10,4.0,2,34,home_sound,-112.674,49.639,1.229,0.8945
+11,4.4,2,34,home_sound,-112.538,49.478,1.226,0.8906
+12,4.8,2,34,home_sound,-112.457,49.494,1.229,0.8906
+13,5.2,2,34,home_sound,-112.457,49.494,1.229,0.8906
+14,5.6,2,34,home_sound,-112.16,49.366,1.232,0.8867
+15,6.0,2,34,home_sound,-112.19,49.322,1.229,0.8828
+16,6.4,2,34,home_sound,-112.109,49.338,1.232,0.8828
+17,6.8,2,34,home_sound,-112.404,49.466,1.229,0.8789
+18,7.2,2,34,home_sound,-111.866,49.238,1.229,0.8789
+19,7.6,2,34,home_sound,-112.079,49.531,1.226,0.875
+20,8.0,2,34,home_sound,-111.785,49.254,1.229,0.8711
+21,8.4,2,34,home_sound,-111.541,49.302,1.226,0.8672
+22,8.8,2,34,home_sound,-111.785,49.403,1.229,0.8633
+23,9.2,2,34,home_sound,-111.541,49.302,1.229,0.8633
+24,9.6,2,34,home_sound,-111.67,49.464,1.232,0.8594
+25,10.0,2,34,home_sound,-111.378,49.334,1.232,0.8594
+26,10.4,2,34,home_sound,-111.132,49.381,1.234,0.8555
+27,10.8,2,34,home_sound,-111.296,49.35,1.234,0.8516
+28,11.2,2,34,home_sound,-110.968,49.412,1.234,0.8516
+29,11.6,2,34,home_sound,-110.556,49.34,1.24,0.8477
+30,12.0,2,34,home_sound,-110.391,49.371,1.24,0.8438
+31,12.4,2,34,home_sound,-110.225,49.401,1.24,0.8438
+32,12.8,2,34,home_sound,-110.186,49.222,1.243,0.8438
+33,13.2,2,34,home_sound,-110.059,49.281,1.243,0.8359
+34,13.6,2,34,home_sound,-109.892,49.311,1.248,0.832
+35,14.0,2,34,home_sound,-109.475,49.233,1.248,0.8281
+36,14.4,2,34,home_sound,-109.425,49.429,1.257,0.8242
+37,14.8,2,34,home_sound,-109.12,49.518,1.26,0.8203
+38,15.2,2,34,home_sound,-108.55,49.651,1.268,0.8164
+39,15.6,2,34,home_sound,-107.995,49.97,1.274,0.8086
+40,16.0,2,34,home_sound,-107.907,49.675,1.282,0.8008
+41,16.4,2,34,home_sound,-107.511,49.969,1.285,0.793
+42,16.8,2,34,home_sound,-106.9,50.101,1.288,0.7852
+43,17.2,2,34,home_sound,-106.494,50.439,1.299,0.7773
+44,17.6,2,34,home_sound,-105.877,50.653,1.308,0.7617
+45,18.0,2,34,home_sound,-105.317,50.979,1.313,0.75
+46,18.4,2,34,home_sound,-105.146,51.464,1.319,0.7383
+47,18.8,2,34,home_sound,-103.934,51.666,1.325,0.7305
+48,19.2,2,34,home_sound,-103.878,52.148,1.336,0.7109
+49,19.6,2,34,home_sound,-103.456,51.925,1.33,0.6953
+0,0.0,3,34,home_sound,-122.397,54.007,1.092,0.6211
+1,0.4,3,34,home_sound,-122.486,53.624,1.121,0.9023
+2,0.8,3,34,home_sound,-122.442,53.688,1.129,0.9609
+3,1.2,3,34,home_sound,-122.242,53.538,1.132,0.9766
+4,1.6,3,34,home_sound,-122.085,53.585,1.134,0.9844
+5,2.0,3,34,home_sound,-121.926,53.763,1.137,0.9883
+6,2.4,3,34,home_sound,-121.965,53.832,1.137,0.9922
+7,2.8,3,34,home_sound,-121.844,54.08,1.14,0.9922
+8,3.2,3,34,home_sound,-121.682,53.864,1.142,0.9922
+9,3.6,3,34,home_sound,-121.62,53.989,1.148,0.9922
+10,4.0,3,34,home_sound,-121.557,53.982,1.148,0.9961
+11,4.4,3,34,home_sound,-121.658,53.927,1.148,0.9961
+12,4.8,3,34,home_sound,-121.531,54.046,1.156,0.9961
+13,5.2,3,34,home_sound,-121.402,54.301,1.158,0.9961
+14,5.6,3,34,home_sound,-121.167,54.209,1.164,0.9961
+15,6.0,3,34,home_sound,-120.896,54.177,1.164,0.9961
+16,6.4,3,34,home_sound,-120.861,54.381,1.166,0.9961
+17,6.8,3,34,home_sound,-120.998,54.539,1.166,0.9961
+18,7.2,3,34,home_sound,-120.683,54.715,1.172,0.9961
+19,7.6,3,34,home_sound,-120.644,54.927,1.169,0.9961
+20,8.0,3,34,home_sound,-120.423,55.192,1.174,0.9961
+21,8.4,3,34,home_sound,-120.303,55.548,1.174,0.9961
+22,8.8,3,34,home_sound,-120.379,55.274,1.18,0.9961
+23,9.2,3,34,home_sound,-119.852,55.51,1.18,0.9961
+24,9.6,3,34,home_sound,-119.908,55.963,1.182,0.9961
+25,10.0,3,34,home_sound,-119.565,56.132,1.188,0.9961
+26,10.4,3,34,home_sound,-119.506,56.367,1.191,0.9961
+27,10.8,3,34,home_sound,-119.221,56.561,1.191,0.9961
+28,11.2,3,34,home_sound,-118.942,56.735,1.191,0.9961
+29,11.6,3,34,home_sound,-118.348,56.987,1.196,0.9961
+30,12.0,3,34,home_sound,-118.065,57.407,1.199,0.9961
+31,12.4,3,34,home_sound,-117.543,57.64,1.199,0.9961
+32,12.8,3,34,home_sound,-117.543,57.64,1.201,0.9961
+33,13.2,3,34,home_sound,-116.898,58.024,1.201,0.9922
+34,13.6,3,34,home_sound,-116.79,58.286,1.21,0.9922
+35,14.0,3,34,home_sound,-115.877,58.601,1.21,0.9922
+36,14.4,3,34,home_sound,-115.627,59.025,1.215,0.9922
+37,14.8,3,34,home_sound,-115.115,59.501,1.221,0.9922
+38,15.2,3,34,home_sound,-113.806,60.139,1.229,0.9922
+39,15.6,3,34,home_sound,-112.437,61.317,1.234,0.9883
+40,16.0,3,34,home_sound,-112.224,61.488,1.243,0.9883
+41,16.4,3,34,home_sound,-110.37,62.343,1.251,0.9844
+42,16.8,3,34,home_sound,-108.673,63.3,1.254,0.9805
+43,17.2,3,34,home_sound,-106.887,64.273,1.265,0.9766
+44,17.6,3,34,home_sound,-104.225,65.652,1.279,0.9688
+45,18.0,3,34,home_sound,-101.886,66.654,1.291,0.9609
+46,18.4,3,34,home_sound,-100.163,67.602,1.299,0.9531
+47,18.8,3,34,home_sound,-96.413,68.202,1.31,0.9492
+48,19.2,3,34,home_sound,-93.262,69.619,1.325,0.9375
+49,19.6,3,34,home_sound,-92.556,69.302,1.325,0.9336

checkpoints/spatial_beats_ov1_local_spatial_v7j_ov123_exp/03_ov123_top4/val_predictions/epoch_0001_csv/valid__ov3_000149__pred.csv

@@ -0,0 +1,201 @@
+frame_idx,frame_time_s,src_or_track_idx,class_idx,class_name,azimuth_deg,elevation_deg,distance_m,activity_prob
+0,0.0,0,45,rain,145.62,-25.37,2.915,0.9297
+1,0.4,0,45,rain,145.69,-25.575,2.9,0.957
+2,0.8,0,45,rain,145.585,-25.848,2.9,0.9609
+3,1.2,0,45,rain,145.368,-26.018,2.9,0.9609
+4,1.6,0,45,rain,145.204,-26.134,2.9,0.9609
+5,2.0,0,45,rain,145.317,-26.229,2.9,0.957
+6,2.4,0,45,rain,145.154,-26.344,2.886,0.957
+7,2.8,0,45,rain,144.878,-26.363,2.886,0.9531
+8,3.2,0,45,rain,144.992,-26.458,2.886,0.9492
+9,3.6,0,45,rain,144.992,-26.458,2.886,0.9492
+10,4.0,0,45,rain,144.83,-26.412,2.886,0.9453
+11,4.4,0,45,rain,144.83,-26.571,2.886,0.9414
+12,4.8,0,45,rain,144.83,-26.412,2.886,0.9375
+13,5.2,0,45,rain,144.83,-26.571,2.886,0.9375
+14,5.6,0,45,rain,144.83,-26.571,2.886,0.9336
+15,6.0,0,45,rain,144.83,-26.571,2.886,0.9297
+16,6.4,0,45,rain,144.83,-26.729,2.886,0.9258
+17,6.8,0,45,rain,144.83,-26.571,2.886,0.9219
+18,7.2,0,45,rain,144.669,-26.525,2.871,0.918
+19,7.6,0,45,rain,144.669,-26.683,2.886,0.9141
+20,8.0,0,45,rain,144.782,-26.618,2.886,0.9102
+21,8.4,0,45,rain,144.669,-26.683,2.886,0.9062
+22,8.8,0,45,rain,144.896,-26.554,2.886,0.9023
+23,9.2,0,45,rain,144.896,-26.554,2.886,0.8984
+24,9.6,0,45,rain,144.622,-26.573,2.886,0.8945
+25,10.0,0,45,rain,144.735,-26.509,2.886,0.8906
+26,10.4,0,45,rain,144.735,-26.509,2.886,0.8867
+27,10.8,0,45,rain,144.848,-26.445,2.886,0.8789
+28,11.2,0,45,rain,144.848,-26.602,2.886,0.875
+29,11.6,0,45,rain,144.848,-26.445,2.886,0.8711
+30,12.0,0,45,rain,144.848,-26.445,2.886,0.8672
+31,12.4,0,45,rain,144.848,-26.445,2.886,0.8633
+32,12.8,0,45,rain,144.96,-26.382,2.886,0.8594
+33,13.2,0,45,rain,144.96,-26.382,2.886,0.8555
+34,13.6,0,45,rain,144.913,-26.274,2.886,0.8516
+35,14.0,0,45,rain,145.072,-26.162,2.886,0.8438
+36,14.4,0,45,rain,145.182,-26.099,2.886,0.8398
+37,14.8,0,45,rain,145.024,-26.056,2.886,0.8359
+38,15.2,0,45,rain,145.182,-25.943,2.871,0.8281
+39,15.6,0,45,rain,145.182,-25.943,2.871,0.8242
+40,16.0,0,45,rain,145.134,-25.681,2.871,0.8203
+41,16.4,0,45,rain,145.182,-25.629,2.886,0.8164
+42,16.8,0,45,rain,145.293,-25.41,2.886,0.8125
+43,17.2,0,45,rain,145.134,-25.21,2.886,0.8086
+44,17.6,0,45,rain,145.293,-25.094,2.886,0.8086
+45,18.0,0,45,rain,145.182,-24.837,2.871,0.8047
+46,18.4,0,45,rain,145.182,-24.518,2.871,0.8047
+47,18.8,0,45,rain,144.96,-24.477,2.871,0.8008
+48,19.2,0,45,rain,145.12,-24.195,2.871,0.7969
+49,19.6,0,45,rain,145.008,-24.092,2.871,0.7812
+0,0.0,1,45,rain,143.259,-23.577,2.93,0.6406
+1,0.4,1,45,rain,143.005,-23.372,2.915,0.6484
+2,0.8,1,45,rain,142.798,-23.396,2.9,0.6523
+3,1.2,1,45,rain,142.759,-23.377,2.9,0.6484
+4,1.6,1,45,rain,142.883,-23.318,2.9,0.6445
+5,2.0,1,45,rain,142.883,-23.318,2.9,0.6406
+6,2.4,1,45,rain,142.719,-23.359,2.886,0.6328
+7,2.8,1,45,rain,142.595,-23.419,2.9,0.6289
+8,3.2,1,45,rain,142.595,-23.245,2.9,0.6211
+9,3.6,1,45,rain,142.719,-23.186,2.9,0.6133
+10,4.0,1,45,rain,142.719,-23.099,2.9,0.6094
+11,4.4,1,45,rain,142.843,-23.04,2.9,0.6016
+12,4.8,1,45,rain,142.843,-22.953,2.9,0.5938
+13,5.2,1,45,rain,142.843,-22.866,2.9,0.5898
+14,5.6,1,45,rain,143.007,-22.736,2.9,0.5781
+15,6.0,1,45,rain,143.007,-22.736,2.9,0.5703
+16,6.4,1,45,rain,143.007,-22.736,2.9,0.5625
+17,6.8,1,45,rain,143.171,-22.605,2.9,0.5547
+18,7.2,1,45,rain,143.007,-22.474,2.9,0.5508
+19,7.6,1,45,rain,143.13,-22.329,2.9,0.5391
+20,8.0,1,45,rain,143.13,-22.241,2.9,0.5312
+21,8.4,1,45,rain,143.13,-22.241,2.9,0.5234
+22,8.8,1,45,rain,143.253,-22.184,2.9,0.5156
+23,9.2,1,45,rain,143.294,-22.02,2.9,0.5078
+24,9.6,1,45,rain,143.416,-21.875,2.9,0.5039
+25,10.0,1,45,rain,143.416,-21.875,2.9,0.5
+26,10.4,1,45,rain,143.416,-21.699,2.9,0.4922
+27,10.8,1,45,rain,143.538,-21.554,2.9,0.4844
+28,11.2,1,45,rain,143.702,-21.596,2.9,0.4805
+29,11.6,1,45,rain,143.702,-21.33,2.9,0.4746
+30,12.0,1,45,rain,143.823,-21.098,2.9,0.4707
+31,12.4,1,45,rain,143.867,-21.015,2.9,0.4688
+32,12.8,1,45,rain,144.107,-20.905,2.9,0.4648
+33,13.2,1,45,rain,144.107,-20.905,2.9,0.4629
+34,13.6,1,45,rain,144.107,-20.638,2.915,0.4609
+35,14.0,1,45,rain,144.439,-20.538,2.915,0.4629
+36,14.4,1,45,rain,144.439,-20.268,2.9,0.4648
+37,14.8,1,45,rain,144.439,-20.088,2.9,0.4668
+38,15.2,1,45,rain,144.654,-19.947,2.9,0.4727
+39,15.6,1,45,rain,144.654,-19.673,2.9,0.4824
+40,16.0,1,45,rain,145.058,-19.422,2.9,0.4883
+41,16.4,1,45,rain,145.161,-19.235,2.9,0.4961
+42,16.8,1,45,rain,145.161,-18.864,2.9,0.5078
+43,17.2,1,45,rain,145.213,-18.766,2.9,0.5234
+44,17.6,1,45,rain,145.437,-18.421,2.886,0.5352
+45,18.0,1,45,rain,145.491,-18.224,2.886,0.5469
+46,18.4,1,45,rain,145.251,-18.041,2.886,0.5547
+47,18.8,1,45,rain,145.48,-17.874,2.886,0.5547
+48,19.2,1,45,rain,145.536,-17.669,2.871,0.5586
+49,19.6,1,45,rain,145.414,-17.623,2.871,0.543
+0,0.0,2,45,rain,143.912,-24.349,2.959,0.7344
+1,0.4,2,45,rain,143.692,-24.488,2.945,0.7148
+2,0.8,2,45,rain,143.607,-24.443,2.93,0.6836
+3,1.2,2,45,rain,143.448,-24.399,2.915,0.6562
+4,1.6,2,45,rain,143.407,-24.295,2.915,0.6367
+5,2.0,2,45,rain,143.407,-24.295,2.915,0.6094
+6,2.4,2,45,rain,143.248,-24.251,2.9,0.5977
+7,2.8,2,45,rain,143.248,-24.087,2.9,0.582
+8,3.2,2,45,rain,143.248,-23.922,2.9,0.5664
+9,3.6,2,45,rain,143.248,-23.922,2.9,0.5547
+10,4.0,2,45,rain,143.248,-23.922,2.9,0.543
+11,4.4,2,45,rain,143.366,-23.617,2.9,0.5352
+12,4.8,2,45,rain,143.366,-23.617,2.9,0.5195
+13,5.2,2,45,rain,143.366,-23.534,2.9,0.5117
+14,5.6,2,45,rain,143.407,-23.552,2.886,0.5
+15,6.0,2,45,rain,143.524,-23.412,2.886,0.4922
+16,6.4,2,45,rain,143.524,-23.329,2.886,0.4805
+17,6.8,2,45,rain,143.683,-23.288,2.886,0.4727
+18,7.2,2,45,rain,143.524,-23.162,2.886,0.4648
+19,7.6,2,45,rain,143.683,-23.121,2.886,0.4512
+20,8.0,2,45,rain,143.641,-22.939,2.886,0.4434
+21,8.4,2,45,rain,143.799,-22.981,2.886,0.4316
+22,8.8,2,45,rain,143.915,-22.757,2.886,0.4238
+23,9.2,2,45,rain,143.799,-22.814,2.886,0.416
+24,9.6,2,45,rain,143.799,-22.646,2.886,0.4102
+25,10.0,2,45,rain,144.074,-22.547,2.886,0.4043
+26,10.4,2,45,rain,144.074,-22.379,2.886,0.3965
+27,10.8,2,45,rain,144.189,-22.323,2.886,0.3887
+28,11.2,2,45,rain,144.348,-22.28,2.886,0.3828
+29,11.6,2,45,rain,144.348,-22.111,2.886,0.375
+30,12.0,2,45,rain,144.348,-21.942,2.886,0.3691
+31,12.4,2,45,rain,144.348,-21.858,2.886,0.3652
+32,12.8,2,45,rain,144.622,-21.842,2.886,0.3613
+33,13.2,2,45,rain,144.622,-21.757,2.886,0.3574
+34,13.6,2,45,rain,144.782,-21.627,2.886,0.3516
+35,14.0,2,45,rain,144.896,-21.402,2.886,0.3516
+36,14.4,2,45,rain,144.943,-21.324,2.871,0.3477
+37,14.8,2,45,rain,145.056,-21.099,2.886,0.3477
+38,15.2,2,45,rain,145.105,-20.933,2.871,0.3496
+39,15.6,2,45,rain,145.105,-20.675,2.871,0.3516
+40,16.0,2,45,rain,145.542,-20.522,2.871,0.3535
+41,16.4,2,45,rain,145.481,-20.317,2.871,0.3574
+42,16.8,2,45,rain,145.646,-20.002,2.871,0.3652
+43,17.2,2,45,rain,145.646,-19.826,2.856,0.373
+44,17.6,2,45,rain,145.751,-19.558,2.856,0.3867
+45,18.0,2,45,rain,145.804,-19.377,2.856,0.3965
+46,18.4,2,45,rain,145.69,-19.158,2.856,0.4043
+47,18.8,2,45,rain,145.629,-19.025,2.841,0.4062
+48,19.2,2,45,rain,145.798,-18.693,2.841,0.4141
+49,19.6,2,45,rain,145.566,-18.704,2.841,0.3984
+0,0.0,3,45,rain,142.631,-22.002,2.915,0.8828
+1,0.4,3,45,rain,142.789,-21.385,2.871,0.9102
+2,0.8,3,45,rain,143.017,-20.791,2.841,0.9062
+3,1.2,3,45,rain,143.168,-20.332,2.827,0.8945
+4,1.6,3,45,rain,143.358,-20.038,2.812,0.8867
+5,2.0,3,45,rain,143.624,-19.773,2.797,0.875
+6,2.4,3,45,rain,143.705,-19.604,2.783,0.8633
+7,2.8,3,45,rain,143.705,-19.349,2.783,0.8555
+8,3.2,3,45,rain,143.86,-19.042,2.768,0.8477
+9,3.6,3,45,rain,144.015,-18.819,2.768,0.8359
+10,4.0,3,45,rain,144.058,-18.728,2.768,0.8281
+11,4.4,3,45,rain,144.215,-18.589,2.753,0.8203
+12,4.8,3,45,rain,144.643,-18.435,2.753,0.8164
+13,5.2,3,45,rain,144.576,-18.3,2.753,0.8086
+14,5.6,3,45,rain,144.735,-18.246,2.739,0.8047
+15,6.0,3,45,rain,144.782,-18.15,2.739,0.7969
+16,6.4,3,45,rain,144.782,-17.972,2.739,0.793
+17,6.8,3,45,rain,144.943,-17.827,2.739,0.7891
+18,7.2,3,45,rain,144.992,-17.728,2.724,0.7852
+19,7.6,3,45,rain,145.154,-17.671,2.724,0.7812
+20,8.0,3,45,rain,145.041,-17.628,2.724,0.7773
+21,8.4,3,45,rain,145.204,-17.479,2.724,0.7734
+22,8.8,3,45,rain,145.204,-17.389,2.724,0.7695
+23,9.2,3,45,rain,145.419,-17.318,2.724,0.7695
+24,9.6,3,45,rain,145.419,-17.227,2.724,0.7695
+25,10.0,3,45,rain,145.471,-17.214,2.724,0.7695
+26,10.4,3,45,rain,145.471,-17.03,2.71,0.7695
+27,10.8,3,45,rain,145.637,-16.969,2.71,0.7695
+28,11.2,3,45,rain,145.69,-17.048,2.71,0.7695
+29,11.6,3,45,rain,145.858,-16.801,2.71,0.7695
+30,12.0,3,45,rain,145.629,-16.707,2.71,0.7734
+31,12.4,3,45,rain,145.682,-16.597,2.71,0.7773
+32,12.8,3,45,rain,145.968,-16.487,2.695,0.7773
+33,13.2,3,45,rain,146.023,-16.47,2.695,0.7812
+34,13.6,3,45,rain,146.08,-16.358,2.695,0.7812
+35,14.0,3,45,rain,146.136,-16.34,2.695,0.7852
+36,14.4,3,45,rain,146.194,-16.225,2.695,0.7891
+37,14.8,3,45,rain,146.077,-16.175,2.68,0.793
+38,15.2,3,45,rain,146.31,-16.09,2.68,0.7969
+39,15.6,3,45,rain,146.251,-15.92,2.68,0.8008
+40,16.0,3,45,rain,146.547,-15.904,2.666,0.8008
+41,16.4,3,45,rain,146.489,-15.831,2.666,0.8008
+42,16.8,3,45,rain,146.55,-15.81,2.666,0.8047
+43,17.2,3,45,rain,146.43,-15.757,2.666,0.8047
+44,17.6,3,45,rain,146.674,-15.564,2.666,0.8008
+45,18.0,3,45,rain,146.737,-15.54,2.651,0.8008
+46,18.4,3,45,rain,146.616,-15.485,2.651,0.7969
+47,18.8,3,45,rain,146.557,-15.509,2.651,0.793
+48,19.2,3,45,rain,146.743,-15.334,2.651,0.7891
+49,19.6,3,45,rain,146.497,-15.429,2.651,0.7773

checkpoints/spatial_beats_ov1_local_spatial_v7j_ov123_exp/03_ov123_top4/val_predictions/epoch_0001_csv/valid__ov3_000159__gt.csv

@@ -0,0 +1,45 @@
+frame_idx,frame_time_s,src_or_track_idx,class_idx,class_name,azimuth_deg,elevation_deg,distance_m,activity_prob
+6,2.4128,0,19,vehicle,-32.065,-0.311,2.365,1.0
+7,2.815,0,19,vehicle,-32.065,-0.311,2.365,1.0
+8,3.2171,0,19,vehicle,-32.065,-0.311,2.365,1.0
+9,3.6193,0,19,vehicle,-32.065,-0.311,2.365,1.0
+10,4.0214,0,19,vehicle,-32.065,-0.311,2.365,1.0
+11,4.4235,0,19,vehicle,-32.065,-0.311,2.365,1.0
+12,4.8257,0,19,vehicle,-32.065,-0.311,2.365,1.0
+13,5.2278,0,19,vehicle,-32.065,-0.311,2.365,1.0
+14,5.63,0,19,vehicle,-32.065,-0.311,2.365,1.0
+15,6.0321,0,19,vehicle,-32.065,-0.311,2.365,1.0
+16,6.4342,0,19,vehicle,-32.065,-0.311,2.365,1.0
+17,6.8364,0,19,vehicle,-32.065,-0.311,2.365,1.0
+18,7.2385,0,19,vehicle,-32.065,-0.311,2.365,1.0
+19,7.6407,0,19,vehicle,-32.065,-0.311,2.365,1.0
+20,8.0428,0,19,vehicle,-32.065,-0.311,2.365,1.0
+21,8.4449,0,19,vehicle,-32.065,-0.311,2.365,1.0
+22,8.8471,0,19,vehicle,-32.065,-0.311,2.365,1.0
+23,9.2492,0,19,vehicle,-32.065,-0.311,2.365,1.0
+21,8.4449,1,36,typing,95.334,-20.364,4.225,1.0
+22,8.8471,1,36,typing,95.334,-20.364,4.225,1.0
+23,9.2492,1,36,typing,95.334,-20.364,4.225,1.0
+24,9.6514,1,36,typing,95.334,-20.364,4.225,1.0
+25,10.0535,1,36,typing,95.334,-20.364,4.225,1.0
+26,10.4556,1,36,typing,95.334,-20.364,4.225,1.0
+27,10.8578,1,36,typing,95.334,-20.364,4.225,1.0
+28,11.2599,1,36,typing,95.334,-20.364,4.225,1.0
+29,11.662,1,36,typing,95.334,-20.364,4.225,1.0
+30,12.0642,1,36,typing,95.334,-20.364,4.225,1.0
+31,12.4663,1,36,typing,95.334,-20.364,4.225,1.0
+32,12.8685,1,36,typing,95.334,-20.364,4.225,1.0
+33,13.2706,1,36,typing,95.334,-20.364,4.225,1.0
+34,13.6727,1,36,typing,95.334,-20.364,4.225,1.0
+22,8.8471,2,42,tearing,151.596,-2.896,1.779,1.0
+23,9.2492,2,42,tearing,151.596,-2.896,1.779,1.0
+28,11.2599,3,48,appliance,108.664,2.827,4.402,1.0
+29,11.662,3,48,appliance,108.664,2.827,4.402,1.0
+30,12.0642,3,48,appliance,108.664,2.827,4.402,1.0
+31,12.4663,3,48,appliance,108.664,2.827,4.402,1.0
+32,12.8685,3,48,appliance,108.664,2.827,4.402,1.0
+33,13.2706,3,48,appliance,108.664,2.827,4.402,1.0
+34,13.6727,3,48,appliance,108.664,2.827,4.402,1.0
+35,14.0749,3,48,appliance,108.664,2.827,4.402,1.0
+36,14.477,3,48,appliance,108.664,2.827,4.402,1.0
+37,14.8792,3,48,appliance,108.664,2.827,4.402,1.0

docs/0413.md

@@ -0,0 +1,512 @@
+# 0413 当前 Spatial-BEATs 基线总结
+
+本文档记录 2026-04-13 时点仓库内当前可复现的 `Spatial-BEATs` 基线状态。  
+重点不是回顾全部试错，而是明确：
+
+- 当前应该以哪个 checkpoint / preset 为基线
+- 当前模型架构到底是什么
+- 当前训练和验证流程如何工作
+- 这条线已经证明了什么，没证明什么
+- 后续继续改时，哪些地方不要再误改
+
+当前本地快照提交：
+
+```text
+b782333 snapshot: restore 0410 local_spatial baseline state
+```
+
+## 1. 当前基线对象
+
+当前恢复并确认的基线是：
+
+```text
+train_spatial_beats.py --preset ov1_local_spatial
+```
+
+对应的历史参考 checkpoint 是：
+
+```text
+checkpoints/spatial_beats_ov1_local_spatial_run1/best.pt
+```
+
+这条线对应的设计目标是：
+
+- 用 `W` 通道的 BEATs 路径提供语义时间序列
+- 用 `WXYZ + IVxyz` 的 local spatial branch 提供局部空间时间序列
+- 两者在 token 时间轴上对齐后相加融合
+- 用一个单路 fused temporal token 序列同时做：
+  - clip-level class
+  - direction
+  - distance
+  - 最终 `llm_spatial_tokens`
+
+这条线不是：
+
+- `semantic_aux_classifier` 双头版本
+- 原始 Kaldi fbank W semantic special path 版本
+- class/spatial 分别从不同 token 读出的版本
+
+这些后续尝试都已经从当前基线中移除。
+
+## 2. 当前模型架构
+
+当前 `ov1_local_spatial` 的主路径如下：
+
+```text
+FOA waveform [B, 4, T]
+  -> SpatialBEATsPreprocessor
+  -> foa_feat [B, 7, T_f, F]
+     channels = [W, X, Y, Z, IVx, IVy, IVz]
+
+  -> extract_patch_tokens()
+     base path:
+       W_logmel -> original single-channel BEATs patch embedding
+     spatial delta path:
+       7ch foa_feat -> SpatialDeltaPatchAdapter -> delta_patch_tokens
+     patch_tokens = base_patch_tokens + delta_patch_tokens
+
+  -> BEATs trunk
+  -> frequency_pool
+  -> TemporalResampler(target_token_rate = 2.5 Hz)
+  -> ShallowTemporalReadout
+  -> semantic_embeddings [B, T_s, D]
+
+  -> LocalSpatialEncoder(foa_feat) -> local_patch_rate_tokens [B, T_f, D_s]
+  -> LocalSpatialResampler -> local_spatial_tokens [B, T_s, D_s]
+  -> local_spatial_proj: Linear(D_s, D)
+  -> fused_embeddings = LayerNorm(semantic_embeddings + local_update)
+
+  -> LocalSpatialPredictionHeads(fused_embeddings)
+     -> class pooled token
+     -> spatial pooled token
+     -> pred_class_logits [B, 65]
+     -> pred_direction [B, 3]
+     -> pred_distance [B, 1]
+
+  -> SpatialTokenProjector(fused_embeddings)
+     -> llm_spatial_tokens [B, T_s, d_llm]
+```
+
+### 2.1 FOA 前端
+
+代码入口：
+
+- [spatial_modules.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_modules.py)
+- [spatial_beats.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_beats.py)
+
+当前前端行为：
+
+- 输入波形按 DCASE 存储顺序 `[W, Y, Z, X]`
+- `SpatialBEATsPreprocessor` 内部先重排到 `[W, X, Y, Z]`
+- 计算：
+  - `WXYZ logmel`
+  - `IVx, IVy, IVz`
+- 输出 `foa_feat [B, 7, T_f, F]`
+
+这里仍然是当前基线的一部分。  
+也就是说，当前基线不是完全复用原始 `BEATs.py preprocess()`，而是：
+
+- 用自实现 STFT/mel 前端
+- 再用 BEATs 统计量做归一化
+
+这点非常重要，因为后续很多分类迁移问题都和这里有关。
+
+### 2.2 Patch token 构造
+
+当前 patch 输入仍然是“W base + 7ch delta”风格：
+
+- `patch_embedding.proj` 只接收 `W_logmel`
+- `SpatialDeltaPatchAdapter` 从完整 `foa_feat` 生成 patch-level residual update
+- 最终送入 trunk 的是：
+
+```text
+patch_tokens = base_patch_tokens + delta_patch_tokens
+```
+
+当前 `ov1_local_spatial` preset 中：
+
+- `train_patch_embedding_in_stage1 = False`
+- `train_spatial_adapter_in_stage1 = False`
+- `patch_adapter_residual_alpha_init = 0.0`
+- `patch_adapter_out_proj_scale_init = 0.0`
+
+也就是说，在这条恢复后的 run1 基线里：
+
+- trunk 冻结
+- patch embedding 冻结
+- patch delta adapter 也不训练
+
+真正训练的是后面的 local spatial 分支和 fused readout。
+
+### 2.3 BEATs trunk 路径
+
+当前 trunk 保持的是原 BEATs 主干：
+
+- `layer_norm`
+- `post_extract_proj`
+- `encoder.pos_conv`
+- `encoder.layers.*`
+- `encoder.layer_norm`
+
+预训练初始化来自：
+
+```text
+pretrain_ckpt/BEATs_iter3_plus_AS2M.pt/BEATs_iter3_plus_AS2M.pt
+```
+
+加载逻辑在：
+
+- [spatial_beats.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_beats.py)
+
+当前日志行为应类似：
+
+```text
+[SpatialBEATs] Reusing 201 compatible BEATs keys
+[SpatialBEATs] Reusing original single-channel BEATs patch embedding
+```
+
+### 2.4 Local spatial 分支
+
+当前 spatial branch 是当前基线的核心增量，结构为：
+
+- `LocalSpatialEncoder`
+  - 2D CNN 提取局部多通道空间模式
+  - 频率池化
+  - temporal Transformer
+- `TemporalResampler`
+  - 把 local spatial token 序列重采样到和 semantic token 同样的 `T_s`
+- `local_spatial_proj`
+  - `Linear(D_s, D)`，小尺度初始化
+- `local_spatial_fusion_norm`
+  - `LayerNorm(semantic + local_update)`
+
+代码位置：
+
+- [spatial_modules.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_modules.py)
+- [spatial_beats.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_beats.py)
+
+### 2.5 Prediction heads
+
+当前 `LocalSpatialPredictionHeads` 是 0410 那条线真正使用的版本：
+
+- 输入只有 `fused_tokens`
+- class 和 spatial 都从同一个 fused sequence 上做 attention pooling
+- 不是 class 从 semantic token 读、spatial 从 fused token 读的后改版本
+
+结构：
+
+```text
+fused_tokens [B, T_s, D]
+  -> class_score -> class attention pooling -> class_token
+  -> spatial_score -> spatial attention pooling -> spatial_token
+
+class_token   -> class_head -> pred_class_logits
+spatial_token -> direction_head -> pred_direction (normalized)
+spatial_token -> distance_head  -> pred_distance (softplus)
+```
+
+初始化细节：
+
+- `class_score.weight/bias = 0`
+- `spatial_score.weight/bias = 0`
+
+所以初始 attention pooling 等价于均匀 mean-pool。  
+这就是为什么它可以比较平滑地接入之前的 class head checkpoint。
+
+## 3. 当前 supervision 与 loss
+
+当前 `ov1_local_spatial` 使用：
+
+```text
+cfg.loss.supervision_mode = "mono_ast"
+```
+
+对应 loss 计算在：
+
+- [spatial_loss.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_loss.py)
+
+当前这条线不是 Hungarian slot matching，而是单源 clip-level supervision：
+
+- 每个样本必须只有一个有效源
+- 从第一个有效源读取：
+  - class
+  - azimuth
+  - elevation
+  - distance
+
+loss 定义：
+
+```text
+loss_cls_aux = cross_entropy(pred_class_logits, cls_target)
+loss_direction = 1 - cos(pred_direction, gt_direction)
+loss_dist = smooth_l1(pred_distance, gt_distance)
+
+loss_total =
+    lambda_cls_aux * loss_cls_aux
+  + lambda_direction * loss_direction
+  + lambda_dist * loss_dist
+```
+
+当前 `ov1_local_spatial` 默认权重：
+
+```text
+lambda_cls_aux = 1.0
+lambda_direction = 12.0
+lambda_dist = 2.0
+```
+
+valid metrics：
+
+- `class_acc`
+- `azi_mae_deg`
+- `ele_mae_deg`
+- `dist_mae`
+
+### 3.1 Active window mask
+
+虽然是单源 clip-level readout，这条线仍然会构建弱时间窗 mask：
+
+- `build_primary_source_window_mask()`
+- 根据 source start/end time 映射到 token 轴
+- 作为 `active_window_mask` 提供给 `LocalSpatialPredictionHeads`
+
+这意味着 pooled class/spatial token 默认更关注标注的 active window，而不是整段所有 token。
+
+## 4. 当前训练配置
+
+当前恢复后的 `ov1_local_spatial` preset 在：
+
+- [train_spatial_beats.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/train_spatial_beats.py)
+
+关键配置：
+
+```text
+batch_size = 8
+num_workers = 4
+num_epochs = 20
+learning_rate = 1e-4
+weight_decay = 0.05
+
+freeze_trunk_in_stage1 = True
+unfreeze_full_trunk = False
+train_patch_embedding_in_stage1 = False
+train_spatial_adapter_in_stage1 = False
+freeze_projector_by_default = True
+
+dataset.max_clip_duration_seconds = 20.0
+dataset.crop_mode = "start"
+
+best_metric_name = "azi_mae_deg"
+minimize_best_metric = True
+class_finetuned_ckpt = checkpoints/beats_ov1_event_cls_head_only/best.pt
+```
+
+当前最重要的初始化来源有两个：
+
+1. `BEATs pretrained trunk`
+2. `checkpoints/beats_ov1_event_cls_head_only/best.pt`
+
+第二个 checkpoint 的加载策略是：
+
+```text
+beats.patch_embedding.weight -> patch_embedding.proj.weight
+beats.patch_embedding.bias   -> patch_embedding.proj.bias
+beats.layer_norm.*           -> layer_norm.*
+beats.post_extract_proj.*    -> post_extract_proj.*
+beats.encoder.*              -> encoder.*
+classifier.weight            -> local_spatial_prediction_heads.class_head.weight
+classifier.bias              -> local_spatial_prediction_heads.class_head.bias
+```
+
+所以当前 `ov1_local_spatial` 的初始分类能力，并不是随机的，而是：
+
+- 先来自 W-channel 事件分类 baseline 的 class head
+- 再接到 fused temporal readout 上
+
+## 5. 当前 checkpoint 已证明的结果
+
+参考 checkpoint：
+
+```text
+checkpoints/spatial_beats_ov1_local_spatial_run1/best.pt
+```
+
+这条 run 的已知最佳结果是：
+
+```text
+best epoch = 7
+best metric = azi_mae_deg
+
+train:
+  loss_total      = 4.0454
+  loss_cls_aux    = 1.3232
+  loss_direction  = 0.1966
+  loss_dist       = 0.1816
+  class_acc       = 0.6184
+  azi_mae_deg     = 24.68
+  ele_mae_deg     = 10.07
+  dist_mae        = 0.465
+
+valid:
+  loss_total      = 5.1220
+  loss_cls_aux    = 2.5258
+  loss_direction  = 0.1856
+  loss_dist       = 0.1847
+  class_acc       = 0.3985
+  azi_mae_deg     = 23.52
+  ele_mae_deg     = 9.65
+  dist_mae        = 0.475
+```
+
+更关键的是 `epoch_0000` 的行为：
+
+```text
+train class_acc ≈ 0.246
+val   class_acc ≈ 0.244
+val   azi_mae   ≈ 71.49
+val   ele_mae   ≈ 71.54
+```
+
+这说明 run1 的初始状态是：
+
+- class 明显高于随机
+- 空间一开始很差
+
+这正是后续 0413 恢复工作所要回到的行为。
+
+## 6. 0413 已排除掉的错误分支
+
+为了避免后续再把这条基线改坏，下面这些都已经确认不是当前基线的一部分：
+
+### 6.1 不是 `semantic_aux_classifier` 版本
+
+后面曾尝试：
+
+- 在训练期额外挂一个完全独立的 semantic auxiliary classifier
+- 让 class loss 不再从 fused token 读
+
+结果：
+
+- 这条线改变了初始化行为
+- 会让 `Epoch 0` class 接近随机
+- 与原始 run1 不一致
+
+现在已经从当前基线彻底移除。
+
+### 6.2 不是 `use_original_beats_semantic_frontend_for_local_spatial` 版本
+
+后面曾尝试：
+
+- 让 local_spatial 路径里的 semantic branch 改走原始 W-only Kaldi fbank + BEATs
+
+结果：
+
+- 这条线会明显改变 local_spatial 的初始化分布
+- 也是导致后续 “空间一开始很好、class 一开始很差” 的原因之一
+
+现在已经从当前基线移除。
+
+### 6.3 不是 class/spatial 分开读不同 token 的版本
+
+后面曾尝试：
+
+- class 从 `semantic_tokens` 池化
+- spatial 从 `fused_tokens` 池化
+
+这同样改变了 0410 run1 的行为。  
+当前已经恢复成：
+
+- class 和 spatial 都从 `fused_tokens` 读
+
+## 7. 当前仓库里额外存在但不属于这条基线的内容
+
+当前仓库除了 `ov1_local_spatial` 之外，还保留了一些后续研发分支：
+
+- `ov1_ast`
+- `ov1_pretrunk_ast_*`
+- `ov123_local_spatial_slot`
+- `ov123_local_spatial_track`
+- `ov123_local_spatial_accdoa`
+
+这些分支仍然在代码里，但它们不是当前这份文档关注的“恢复后的单源 local_spatial 基线”。
+
+后续如果继续研究 frame-level / slot-level 多源建模，应当：
+
+- 把它们视为新的实验线
+- 不要再直接污染 `ov1_local_spatial`
+
+## 8. 当前推荐的复现命令
+
+恢复后的基线复现命令：
+
+```bash
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 torchrun --nproc_per_node=8 --master_port=29531 \
+  train_spatial_beats.py \
+  --preset ov1_local_spatial \
+  --distributed \
+  --batch-size 8 \
+  --num-workers 24 \
+  --num-epochs 20 \
+  --output-dir checkpoints/spatial_beats_ov1_local_spatial_run1
+```
+
+如果只想做行为检查，先跑 1 个 epoch 就够：
+
+```bash
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 torchrun --nproc_per_node=8 --master_port=29531 \
+  train_spatial_beats.py \
+  --preset ov1_local_spatial \
+  --distributed \
+  --batch-size 8 \
+  --num-workers 24 \
+  --num-epochs 1 \
+  --output-dir checkpoints/spatial_beats_ov1_local_spatial_repro_check
+```
+
+判断是否仍在 run1 轨道上的最直接标准：
+
+- `Epoch 0 class_acc` 应该明显高于随机
+- `Epoch 0 azi/ele` 应该仍然很差
+
+而不是：
+
+- `Epoch 0 class` 接近 0
+- `Epoch 0 spatial` 一开始就很好
+
+## 9. 目前这条线的结论
+
+当前 `ov1_local_spatial` 基线已经证明：
+
+1. 单路 fused token 架构是能学到空间的。
+2. 这条线能在不动 trunk 的前提下，把 azimuth / elevation / distance 拉下来。
+3. 但它的 class 保持能力有限，最佳 valid `class_acc` 约 `0.40`，明显低于纯 W-channel BEATs 分类 baseline。
+4. 所以这条线当前更适合作为：
+   - “可工作空间基线”
+   - 后续 frame-level spatial 扩展的 warm-start
+
+它还没有证明：
+
+1. class 和 spatial 可以同时都做到强泛化。
+2. fused token 已经足够好，可以直接无脑接 LLM。
+
+## 10. 后续继续迭代时的建议
+
+如果以后继续改，不要再直接破坏 `ov1_local_spatial` 这条基线。更稳的方式是：
+
+1. 保留 `ov1_local_spatial` 不动，作为 frozen baseline。
+2. 新实验单独开新 preset。
+3. 新实验优先从：
+   - `checkpoints/spatial_beats_ov1_local_spatial_run1/best.pt`
+   - 或 `DEFAULT_OV1_LOCAL_SPATIAL_INIT`
+   warm-start。
+4. 任何涉及下面这些点的修改，都应视为“新架构”，不要再说和 run1 一样：
+   - semantic aux classifier
+   - 原始 W-BEATs special frontend
+   - class/spatial 分离 readout
+   - 新的 frame-level supervision path
+
+当前最安全的角色定位是：
+
+```text
+ov1_local_spatial = 已恢复并可复现的 0410 单源 local-spatial baseline
+```

docs/0416.md

@@ -0,0 +1,215 @@
+# 0416 Spatial-BEATs 实验记录
+
+本文档记录 2026-04-16 对话期间的所有改动、实验和结果。
+
+## 1. Bug 修复
+
+### 1.1 方位角坐标系修复（val_predictions 显示 bug）
+
+**问题**：val_predictions jsonl 文件中 `gt_azimuth_deg` 是 DCASE 标准 `[-180°, 180°]`，但 `pred_azimuth_deg` 被 `torch.remainder(..., 360.0)` 转成了 `[0°, 360°]`。导致 GT=-168° 和 Pred=192° 看起来差 360° 但实际是同一个方向。
+
+**影响**：仅影响 val_predictions 文件的可读性。**训练 loss 和 epoch log metrics 不受影响**（`_circular_distance_deg` 对两种坐标系都能正确计算圆周距离）。
+
+**修复**（`spatial_loss.py`）：
+- `_azi_ele_deg_from_direction_vector`：去掉 `torch.remainder(..., 360.0)`，`atan2` 直接返回 `[-180°, 180°]`
+- 新增 `_to_dcase_azimuth()` 工具函数
+- `build_validation_examples`、`build_pretrunk_ast_validation_examples` 中 pred_azi 使用 `_to_dcase_azimuth` 转换
+
+### 1.2 Vocabulary 标签修复（65→63 类）
+
+**问题 1**：`female_singing`(171例) 和 `male_singing`(104例) 与 `singing` 是父子关系，模型在 65 类 softmax 下完全区分不了（test accuracy=0%）。
+
+**问题 2**：`string_instrument` 类中有 **1644 个样本**（22.8%）是 Hi-hat/Crash_cymbal/Cymbal 打击乐，被错误映射。这些应该归入 `percussion`。
+
+**修复**（`fix_vocabulary_and_manifests.py`）：
+- `female_singing` + `male_singing` → 合并到 `singing`（65→63 类）
+- `string_instrument` 中 `mono_primary_label` 为 Hi-hat/Crash_cymbal/Cymbal 的样本 → 重标为 `percussion`
+- 对 ov1/ov2/ov3 manifest 全部原地修复，备份为 `.bak_20260416`
+- `final_vocabulary.csv` 重编号为连续的 1~63
+
+**受影响文件**：
+| 文件 | 变更 |
+|---|---|
+| `final_vocabulary.csv` | 65→63 类，去掉 female_singing/male_singing |
+| `ov1_foa.jsonl` | 460 条 singing 合并 + 1644 条 cymbal 修正 |
+| `ov2_foa.jsonl` / `ov3_foa.jsonl` | 无变动（这些类别未出现） |
+
+**代码 default 值更新**：
+| 文件 | 变更 |
+|---|---|
+| `spatial_beats.py` | `source_num_classes: 65 → 63` |
+| `spatial_dataset.py` | `num_classes: 65 → 63` |
+| `spatial_modules.py` | 6 处函数参数默认值 `65 → 63` |
+| `spatial_atst.py` | `source_num_classes: 65 → 63` |
+
+### 1.3 load_checkpoint strict 改为 non-strict
+
+**问题**：stage1→stage2 resume 时，如果两阶段 model config 不完全一致（比如 stage1 无 semantic_anchor 但 stage2 有），`strict=True` 会报错 missing key。
+
+**修复**（`train_spatial_beats.py`）：`load_checkpoint` 改用 `strict=False`，missing/unexpected key 打印 warning 不报错。
+
+## 2. v2 Test 集完整评测
+
+编写了 `eval_spatial_beats.py` 评测脚本，在 ov1 **完整 test 集（1800 样本）** 上评测了 v2 stage2 best.pt。
+
+**checkpoint**：`checkpoints/spatial_beats_ov1_local_spatial_v2_exp/02_spatial/best.pt`
+
+| 指标 | 值 |
+|---|---|
+| class_acc | **43.83%** |
+| azi_mae | **19.87°** |
+| ele_mae | 8.66° |
+| dist_mae | 0.554 m |
+| SELD F1 | 0.3139 |
+| SELD LR | 0.6956 |
+| SELD LE | 8.64° |
+| **SELD Score ↓** | **0.6027** |
+
+### 63 类映射后（post-hoc）
+
+| 指标 | 65类 | 63类映射 |
+|---|---|---|
+| class_acc | 43.83% | **44.44%** (+0.6%) |
+| 空间指标 | 不变 | 不变 |
+
+提升来自 singing 子类合并（+11 个正确样本）。
+
+### Per-class 分析亮点
+
+- **Accuracy=0 的 4 个类**：female_singing, male_singing, tape, wood
+- **Accuracy≥0.7 的 5 个类**：bird(70.6%), guitar(70.9%), thunderstorm(76%), knock(77.8%), car(100%)
+- **最大混淆来源**：父子标签层级冲突（guitar↔keyboard_instrument, wind_instrument→musical_instrument, male_singing→singing）
+
+## 3. bypass / purify 实验结果
+
+两个新架构实验的 stage1 都**全 trunk 解冻**（12 层），目标是用更激进的策略达到更高 class_acc。
+
+| 实验 | Stage1 策略 | Stage1 class_acc | Stage2 SELD |
+|---|---|---|---|
+| bypass | 全解冻 + bypass_local_fusion + 零空间 | ~25% ❌ | 0.743 |
+| purify | 全解冻 + freeze_local_spatial + 零空间 | ~25% ❌ | 0.816 |
+| **v2（对照）** | top-2 + semantic_anchor + 空间多任务 | **~56%** ✅ | **0.603** |
+
+**结论**：全 trunk 解冻 + SpatialBEATs 架构 = 崩。top-2 解冻更稳定。
+
+## 4. 分类能力损失分析
+
+### 纯 BEATs 分类实验（无空间任务）
+
+| 解冻策略 | val_acc |
+|---|---|
+| head_only（trunk 全冻） | 62.6% |
+| top-8（层 4-11） | **69.1%** |
+| full（全 12 层） | 70.0% |
+
+### SpatialBEATs 分类损失链
+
+```
+纯 BEATs 62.6%（trunk 全冻）
+  → SpatialBEATs v2 stage1: 56%  (-6.6%: CNN噪声 + 多任务干扰 + 只解冻top-2)
+  → SpatialBEATs v2 stage2: 44%  (-12%: λ_dir=12 空间梯度冲击语义)
+```
+
+### 关键发现：v4 的 trunk 起点
+
+v4 config 中 `class_finetuned_ckpt` 之前指向的是 `head_only/best.pt`（62.6%），其 trunk 权重和原始 BEATs **完全相同**（head_only 全冻 trunk 训练）。已修正为指向 `02_full/best.pt`（70%），获取 FSD50K 适配过的 trunk。
+
+## 5. 失败实验
+
+### v3 / v3ws（bypass + top-8/top-4）
+- v3 epoch1: cls=0.116 ❌
+- v3ws epoch1: cls=0.080 ❌
+- 原因：bypass 模式 + SpatialBEATs = DDP 不稳定
+
+### v3b / v3bws（freeze_local_spatial + top-8/top-4）
+- v3b 15 epoch: cls=35.67%，azi=89.66°（空间几乎随机）
+- v3bws: 无效果
+- 原因：多变量同时改动（top-8 + freeze_local_spatial + ddp_find_unused + 无 anchor），不如 v2 的组合
+
+## 6. 新增架构：Frame-Level Track Supervision
+
+### 动机
+
+当前 `mono_ast` 是 clip-level 预测（attention pool → 1 class + 1 direction），无法：
+1. 输出 DCASE 格式逐帧检测
+2. 扩展到 ov2/ov3 多源
+3. 约束 trunk 逐帧表征质量
+
+### 实现
+
+在 `readout_scheme="local_spatial"` 下新增可选 `FrameTrack` 分支，与 clip-level head **并行运行**：
+
+```
+fused_spatial_embeddings [B, T_s, D]
+  ├── attention pool → clip-level mono_ast prediction (已有)
+  └── SourceQueryDecoder → [B, K, T_s, D] → FrameTrack prediction (新增)
+```
+
+**完全复用已有代码**：`SourceQueryDecoder`、`FrameTrackPredictionHeads`、`compute_frame_track_losses` 一行都没改。
+
+**控制开关**：
+- `SpatialBEATsConfig.enable_frame_track: bool = False`
+- `SpatialLossConfig.enable_frame_track_loss: bool = False`
+- 默认关闭，所有现有实验零影响
+
+**改动文件**：
+| 文件 | 改动 |
+|---|---|
+| `spatial_beats.py` | Config flag + __init__ 创建 head + forward 产出 |
+| `spatial_loss.py` | Config flag |
+| `train_spatial_beats.py` | run_train_step 追加 loss + validate 追加 examples + preset |
+| `spatial_modules.py` | **不改** |
+
+**新 preset**：`ov1_local_spatial_v4f_spatial`
+
+## 7. 当前实验矩阵
+
+### 正在跑
+
+| 实验 | 状态 | 配置要点 |
+|---|---|---|
+| v4 stage1 | 运行中 | v2 架构复刻 + 63 类 + **70% trunk init** |
+
+### 等 v4 stage1 结束后
+
+| 实验 | 脚本 | 配置要点 |
+|---|---|---|
+| v4 stage2 | `run_ov1_v4.sh` | v2 复刻（λ_dir=12, anchor=0.5） |
+| v4g stage2 | `run_ov1_v4g.sh` | 温和版（λ_dir=6, λ_cls=2, anchor=1.5） |
+| v4f stage2 | `run_ov1_v4f.sh` | v4 + 并行 frame-level track head |
+
+### 预期效果
+
+| 实验 | 预期 class_acc | 预期 azi_mae | 新能力 |
+|---|---|---|---|
+| v4 stage2 | ~44%（同 v2） | ~20° | baseline |
+| v4g stage2 | **~50%+** | ~25-30° | 分类更好，空间稍差 |
+| v4f stage2 | ~44% + frame metrics | ~20° | DCASE 逐帧输出 |
+
+## 8. 新增文件清单
+
+| 文件 | 用途 |
+|---|---|
+| `eval_spatial_beats.py` | 独立评测脚本，支持所有 preset |
+| `fix_vocabulary_and_manifests.py` | 一次性 vocab+manifest 修复脚本 |
+| `run_ov1_v3.sh` | v3 实验（bypass，已失败） |
+| `run_ov1_v3ws.sh` | v3ws 实验（bypass+warmstart，已失败） |
+| `run_ov1_v3b.sh` | v3b 实验（freeze_local_spatial，效果差） |
+| `run_ov1_v3bws.sh` | v3bws 实验（同上+warmstart） |
+| `run_ov1_v4.sh` | v4 实验（v2 复刻 + 63 类 + 70% trunk） |
+| `run_ov1_v4g.sh` | v4g 温和空间版 |
+| `run_ov1_v4f.sh` | v4f frame-level track 版 |
+
+## 9. 关键经验总结
+
+1. **不要全解冻 trunk**：纯 BEATs 全解冻 OK（70%），但 SpatialBEATs 架构下全解冻必崩（25%）。top-2 是当前唯一验证过的安全策略。
+
+2. **bypass_local_fusion 不可用**：在 DDP 训练下导致不稳定，即使只是 top-8 解冻也会崩。freeze_local_spatial 稍好但仍差。
+
+3. **Semantic anchor 有效**：v2 的 anchor（λ=0.5）让 stage2 class_acc 从 25%（无 anchor 的 kaldi_spatial）保到 44%。
+
+4. **标签质量是分类瓶颈**：65 类中父子层级冲突（singing/female_singing、musical_instrument/string_instrument）和标签映射 bug（cymbal→string_instrument）贡献了大量"假错误"。
+
+5. **trunk 初始化很重要**：v4 之前一直用的是 `head_only/best.pt`（trunk ≡ 原始 BEATs），现在改为 `02_full/best.pt`（trunk 已适配 FSD50K，70%），预期 stage1 起点更高。
+
+6. **一次只改一个变量**：v3b/v3bws 同时改了 5 个变量导致无法诊断，v4 只改了 trunk init 这一项。

docs/0422.md

@@ -0,0 +1,569 @@
+# 2026-04-22 记录：v7h 基线、v8 融合升级与当前诊断
+
+本文档记录 2026-04-22 这轮关于 `v7h -> v8` 的设计、问答、结果和代码修改。
+
+相关旧文档：
+- `docs/0422_v7h_v7j.md`：主要记录 `v7h / v7j` 的诊断和尝试
+
+本文聚焦：
+- 当前最可用版本 `v7h` 到底是什么
+- 为什么要做 `v8`
+- `v8` 的结构改了什么，没改什么
+- 目前 `epoch0/epoch1` 的真实现象
+- 后续应该继续看什么
+
+## 1. 当前结论
+
+截至本轮：
+
+- 当前最可用的稳定基线仍然是 `v7h`
+- `v8` 不是坏形态，值得继续训练
+- `v8` 当前的主要问题不是 duplicate 崩坏，而是 `ov2/ov3` 上的 class binding 还不稳，尤其 `ov3` 仍有明显 class collapse
+- `v8` 现在还处于 two-stage 的 stage 1，前 3 个 epoch 本来就不训 `dir/dist`，所以 `F20 / LE_CD / oazi` 在 `epoch0-2` 不适合过早下结论
+
+## 2. 关键问答与结论
+
+### 2.1 BEATs 适不适合做逐帧、多源预测
+
+结论：
+
+- 原始 `BEATs` 官方下游头更偏 `clip-level audio tagging`
+- 但 `BEATs` 作为语义 backbone 是适合逐帧任务的
+- 真正不够的是“直接拿官方 head 做 strong-label / 多源 SELD”
+
+对当前项目更准确的判断是：
+
+- `BEATs` 适合作为语义分支
+- `local spatial branch` 适合作为空间分支
+- 真正的难点在于：
+  - 如何融合语义/空间表征
+  - 如何让 query decoder 在同一时刻分辨多个源
+  - 如何让 per-frame class / DoA 监督不互相拖累
+
+### 2.2 现在 4 个 track 到底在预测什么
+
+当前 `local_spatial_track` 路线的输出是：
+
+- 每帧固定产出 `K=4` 组 candidate tracks
+- 每条 track 在每帧预测：
+  - `activity`
+  - `class`
+  - `direction`
+  - `distance`
+
+它不是“每帧一定有 4 个真实声源”，而是“每帧最多从 4 个候选槽位里激活若干条”。
+
+结构链路是：
+
+```text
+fused_embeddings [B, T_s, D]
+  -> SourceQueryDecoder
+     -> track_latents [B, K, D]
+     -> track_time_features [B, K, T_s, D]
+  -> FrameTrackPredictionHeads
+     -> pred_activity
+     -> pred_class_logits
+     -> pred_direction
+     -> pred_distance
+```
+
+### 2.3 `oracle_cls / oracle_azi / oracle_ele` 是什么
+
+这几个量是 `track` 头的 oracle 诊断指标，不是官方 DCASE metric。
+
+定义：
+
+- 先在 GT-active 的 frame-source 对上做 matching
+- matching 时不走 activity threshold，只看 head 本身
+- 在 matched pair 上统计：
+  - `oracle_cls`：class 是否对
+  - `oracle_azi`：azimuth MAE
+  - `oracle_ele`：elevation MAE
+
+用途：
+
+- 看 class/DoA head 本身有没有学到
+- 不包含最终检测误差，不等于 `F20`
+
+### 2.4 要不要先训练一个独立的 per-frame class-only BEATs
+
+结论：不建议单独另起一个 class-only 模型作为主路线。
+
+原因：
+
+- 最终任务的 class 是从
+  `fused -> source_query_decoder -> track_time_features -> class_head`
+  这条路径读出来的
+- 单独训一个普通 per-frame classifier，学不到 query/track binding
+
+更合理的做法是：
+
+- 保留同一套 `track-query` 架构
+- 先做 `activity + class` warmup
+- 再打开 `dir/dist`
+
+这就是当前 `v8` 的 two-stage 设计。
+
+### 2.5 v7 阶段关于初始化和解冻的结论
+
+已经确认的结论：
+
+- `source_query_decoder` 和 `activity_head` 随机初始化是合理的
+- `frame_track_prediction_heads.class/direction/distance` 可以从旧的 `local_spatial_prediction_heads` 迁移同语义权重
+- 如果 `v7` 只在 `ov1` 上学过，到了 `ov123` 不建议永远冻结 trunk
+- 当前合理策略是：
+  - 继续从已有可用 checkpoint 热启动
+  - 解冻 trunk 顶部少量层，例如 top-4
+
+## 3. 当前最可用基线：v7h
+
+`v7h` 继承自 `v7f_ov123_top4`，是当前最可用的稳定基线。
+
+核心训练设定：
+
+- `readout_scheme = local_spatial_track`
+- `enable_clip_aux_head = False`
+- `ov1:ov2:ov3 = 1:3:3`
+- 去掉 focal BCE
+- `frame_activity_pos_weight = 3.0`
+- class-cost warmup 采用 `1 + 2` epoch
+- trunk 顶部 4 层解冻
+
+相关配置可参考：
+- `train_spatial_beats.py:1465`
+- `docs/0422_v7h_v7j.md`
+
+历史最好结果记录：
+
+```text
+v7h ep3 val:
+  F20=0.246
+  LE_CD=34.3°
+  LR_CD=0.504
+  SELD=0.608
+```
+
+这也是当前继续做 `v8` 的热启动来源。
+
+## 4. v8 的设计目标
+
+### 4.1 为什么要做 v8
+
+对 `v7h` 的判断是：
+
+- 前端不要乱改
+- `source_query_decoder` 和 `frame_track_prediction_heads` 也先不要乱改
+- 真正可能偏弱的是 fused token 的构造方式
+
+`v7h` 的融合只有：
+
+```text
+fused = LayerNorm(semantic_embeddings + local_update)
+```
+
+这隐含假设：
+
+- semantic token 和 spatial token 已经天然对齐
+- 只要相加，query decoder 就能自动学会利用空间信息
+
+这个假设偏强，因此设计 `v8`：
+
+- 前端完全不改
+- 只升级 semantic/spatial 的 fusion
+
+### 4.2 v8 设计原则
+
+- 以 semantic token 为主骨架
+- spatial token 不直接硬加，而是通过 cross-attention 注入
+- 新模块要能安全从 `v7h` 热启动
+
+因此 `v8` 的 fusion 采用：
+
+```text
+semantic <- spatial cross-attention (2 layers)
+  + gated direct spatial residual
+  + same final LayerNorm
+```
+
+## 5. v8 新模型架构
+
+### 5.1 整体链路
+
+`v8` 只改 fused 部分，其他都和 `v7h` 一致：
+
+```text
+FOA waveform [B, 4, T]
+  -> SpatialBEATsPreprocessor
+  -> patch embedding + BEATs trunk
+  -> frequency_pool
+  -> TemporalResampler (2.5 Hz)
+  -> temporal_readout
+  -> semantic_embeddings [B, T_s, D]
+
+  -> LocalSpatialEncoder
+  -> local_spatial_resampler
+  -> local_spatial_proj
+  -> local_update [B, T_s, D]
+
+  -> LocalSpatialCrossFuser (new in v8)
+     semantic <- spatial cross-attn x 2
+     + gated spatial residual
+  -> local_spatial_fusion_norm
+  -> fused_embeddings [B, T_s, D]
+
+  -> SourceQueryDecoder
+  -> FrameTrackPredictionHeads
+```
+
+### 5.2 代码落点
+
+新增或修改位置：
+
+- `spatial_modules.py:1723-1842`
+  - `LocalSpatialCrossFusionBlock`
+  - `LocalSpatialCrossFuser`
+- `spatial_beats.py:173-182`
+  - 新增 fusion 配置字段
+- `spatial_beats.py:1104-1118`
+  - `build_local_spatial_fusion()` 改为支持 `cross_attn_gated`
+- `train_spatial_beats.py:1536-1549`
+  - 新增 `make_ov1_local_spatial_v8_ov123_top4_config()`
+- `run_ov1_v8_ov123_top4.sh:1-68`
+  - 新的启动脚本
+
+### 5.3 v8 的训练可学习参数
+
+为了保证 `v8` 新融合模块真的训练到，代码里做了两处接线：
+
+- `train_spatial_beats.py:2528-2535`
+  - `local_spatial_fuser` 被加入 `always_train_prefixes`
+- `train_spatial_beats.py:2646-2653`
+  - `local_spatial_fuser.` 被加入 `_SPATIAL_PREFIXES`
+
+也就是说：
+
+- 新 fuser 走 `spatial_lr` 组
+- 不会被当成 trunk 或 head 漏掉
+
+## 6. v8 的两阶段训练
+
+### 6.1 设计原因
+
+直接在新融合结构上同时训练 `class + direction + distance`，早期很容易出现：
+
+- class 还没稳定
+- DoA 噪声先进入 matching
+- assignment 被错误的 spatial cost 带偏
+
+所以 `v8` 采用 two-stage：
+
+- stage 1：先训 `activity + class`
+- stage 2：再恢复 `dir/dist`
+
+### 6.2 具体配置
+
+`v8` preset 中：
+
+```text
+frame_spatial_loss_warmup_epochs = 3
+frame_spatial_loss_warmup_scale = 0.0
+```
+
+对应代码：
+
+- `train_spatial_beats.py:1547-1548`
+- `train_spatial_beats.py:4079-4102`
+
+含义：
+
+- `epoch 0-2`
+  - `lambda_frame_direction = 0`
+  - `lambda_frame_distance = 0`
+  - `frame_match_dir_cost_weight = 0`
+  - `frame_match_dist_cost_weight = 0`
+- `epoch 3+`
+  - 自动恢复完整方向/距离监督
+
+## 7. v8 热启动方式
+
+默认脚本：
+
+```text
+./run_ov1_v8_ov123_top4.sh
+```
+
+默认行为：
+
+- 从 `v7h` 的 `best.pt` 热启动
+- 只新增 `local_spatial_fuser.*` 参数
+- 其余 trunk / local_spatial / source_query_decoder / frame-track heads 继承已有权重
+- 训练比例继续保持 `ov1:ov2:ov3 = 1:3:3`
+
+脚本位置：
+
+- `run_ov1_v8_ov123_top4.sh:16-37`
+
+## 8. v8 当前结果
+
+### 8.1 epoch 0 日志
+
+用户记录：
+
+```text
+[Epoch 0] train:
+  loss=1.3445  activity=0.3099  cls_aux=1.0346
+  direction=0.4527  dist=0.4894
+  act↑=0.887  act↓=0.136  sep=0.751
+  ocls=0.786  oazi=45.2°  oele=18.4°
+
+[Epoch 0] val:
+  loss=2.1304  activity=0.2310  cls_aux=1.8995
+  direction=0.4377  dist=0.4337
+  act↑=0.925  act↓=0.092  sep=0.832
+  ocls=0.652  oazi=46.9°  oele=20.1°
+  ER20=1.047 F20=0.096 LE_CD=51.8° LR_CD=0.486 SELD=0.688
+```
+
+### 8.2 epoch 0 CSV 诊断
+
+`epoch_0000_csv` 的诊断结论：
+
+- 不是 `v7j` 那种 3-4 条轨全亮的崩坏
+- `activity` 分离已经很好
+- 当前主要瓶颈是 `class + angle`，不是 duplicate
+
+粗统计：
+
+```text
+TP=207  FP=2147  FN=1844
+P=0.088  R=0.101  F1=0.094
+
+FP breakdown:
+  class_wrong = 1211
+  angle_wrong = 911
+  duplicate   = 25
+
+active_hist:
+  0 active: 185
+  1 active: 748
+  2 active: 683
+  3 active: 80
+
+no_gt_frames=286
+no_gt_with_pred=119
+ratio=0.416
+
+mean_maxprob_gt   = 0.972
+mean_maxprob_nogt = 0.400
+```
+
+解释：
+
+- `act↑/act↓/sep` 很强，说明 activity head 已经把有源/无源分开了
+- 但 `class_wrong` 和 `angle_wrong` 很高
+- `F20` 低并不意外，因为 stage 1 根本还没训 `dir/dist`
+
+### 8.3 epoch 0 的 ov1 / ov2 / ov3 形态
+
+代表样本观察：
+
+`ov1`
+
+```text
+mean_act_by_track = {0: 0.88, 1: 0.007, 2: 0.001, 3: 0.0}
+active_tracks_per_frame_hist = {1: 37}
+top_pred_classes = singing(37)
+```
+
+说明单源样本已经能稳定只亮一条轨。
+
+`ov2`
+
+```text
+mean_act_by_track = {0: 0.999, 1: 0.745, 2: 0.133, 3: 0.0}
+active_tracks_per_frame_hist = {1: 5, 2: 45}
+top_pred_classes = frog(50), bird(37), tool(8)
+```
+
+说明双源样本已经在尝试输出两条轨。
+
+`ov3`
+
+```text
+mean_act_by_track = {0: 0.977, 1: 0.519, 2: 0.036, 3: 0.0}
+active_tracks_per_frame_hist = {1: 17, 2: 32}
+top_pred_classes = tool(81)
+```
+
+说明三源样本里第二条轨开始亮，但有明显 class collapse。
+
+## 9. v8 的 epoch 1：重点看 ov2 / ov3
+
+第二个 epoch 之后，重点检查了 `ov23`。
+
+### 9.1 active track 数量变化
+
+在“有预测的帧”上，平均亮起的轨数：
+
+```text
+epoch0:
+  ov2 = 1.578
+  ov3 = 1.767
+
+epoch1:
+  ov2 = 1.736
+  ov3 = 1.942
+```
+
+这说明：
+
+- `ov2/ov3` 上第二条轨更积极了
+- 模型更愿意在 overlap 场景输出多轨
+
+这本身不是坏事，但如果 class/DoA 没跟上，就会先表现为 FP 上升。
+
+### 9.2 代表样本对比
+
+`ov2` 代表样本 `valid__ov2_000000__pred.csv`
+
+`epoch0`
+
+```text
+mean_act_by_track = {0: 0.999, 1: 0.745, 2: 0.133, 3: 0.0}
+active_hist = {1: 5, 2: 45}
+top_classes = frog(50), bird(37), tool(8)
+```
+
+`epoch1`
+
+```text
+mean_act_by_track = {0: 0.996, 1: 0.312, 2: 0.625, 3: 0.339}
+active_hist = {1: 11, 2: 39}
+top_classes = frog(50), wind(39)
+```
+
+解释：
+
+- 第二条活跃轨从 `track1` 转向 `track2`
+- 类别也从 `bird` 变成了 `wind`
+- 说明 query 责任在重排，但 class binding 还不稳定
+
+`ov3` 代表样本 `valid__ov3_000004__pred.csv`
+
+`epoch0`
+
+```text
+mean_act_by_track = {0: 0.977, 1: 0.519, 2: 0.036, 3: 0.0}
+active_hist = {1: 17, 2: 32}
+top_classes = tool(81)
+```
+
+`epoch1`
+
+```text
+mean_act_by_track = {0: 0.966, 1: 0.665, 2: 0.065, 3: 0.127}
+active_hist = {1: 11, 2: 39}
+top_classes = tool(89)
+```
+
+解释：
+
+- 第二条轨更稳定地亮了
+- 但 class collapse 没缓解，反而更统一地塌成 `tool`
+
+因此当前对 `epoch1` 的判断是：
+
+- `ov2`：有变化，但不能算明显变好
+- `ov3`：暂时没有变好，仍然是当前最大问题点
+
+## 10. 当前诊断
+
+截至目前，对 `v8` 的判断是：
+
+### 10.1 已经证明的事
+
+- `v8` 没有走向 `v7j` 那种 duplicate 崩坏
+- `activity` 分离做得比担心中要好
+- `ov2/ov3` 上的多轨输出能力确实在长出来
+
+### 10.2 还没解决的事
+
+- `ov2/ov3` 的 class binding 还不稳
+- `ov3` 仍然存在明显 class collapse
+- stage 1 期间不训 `dir/dist`，所以 `F20 / LE_CD / oazi` 现在还不能作为最终判断
+
+### 10.3 当前最值得关注的信号
+
+- `epoch 3+` 进入 stage 2 后：
+  - `ocls` 是否继续上升
+  - `oazi` 是否明显下降
+  - `F20` 是否出现拐点
+- `ov3` 的 top predicted classes 是否开始从单一 `tool` 分裂成多个类
+- `ov23` 的平均 active tracks 是否继续上涨过快
+
+## 11. 本轮实际代码修改
+
+本轮已经落地的修改：
+
+### 11.1 新增 v8 融合模块
+
+- `spatial_modules.py:1723-1842`
+  - `LocalSpatialCrossFusionBlock`
+  - `LocalSpatialCrossFuser`
+
+### 11.2 新增 fusion 配置
+
+- `spatial_beats.py:173-182`
+  - `local_spatial_fusion_mode`
+  - `local_spatial_fusion_layers`
+  - `local_spatial_fusion_heads`
+  - `local_spatial_fusion_dropout`
+  - `local_spatial_fusion_gate_bias`
+  - `local_spatial_fusion_direct_gate_bias`
+
+### 11.3 改 fused token 构造
+
+- `spatial_beats.py:1104-1118`
+  - 从 `semantic + local_update`
+  - 改成支持 `local_spatial_fuser(...)`
+
+### 11.4 新增 v8 preset
+
+- `train_spatial_beats.py:1536-1549`
+  - 继承 `v7h`
+  - 打开 `cross_attn_gated`
+  - 打开 two-stage spatial warmup
+  - 输出目录改为 `v8_ov123_exp/03_ov123_top4`
+
+### 11.5 训练侧接线
+
+- `train_spatial_beats.py:2528-2535`
+  - `local_spatial_fuser` 加入 `always_train_prefixes`
+- `train_spatial_beats.py:2646-2653`
+  - `local_spatial_fuser.` 加入 `_SPATIAL_PREFIXES`
+
+### 11.6 新增脚本
+
+- `run_ov1_v8_ov123_top4.sh:1-68`
+
+## 12. 下一步建议
+
+当前建议不改结构，先继续训练 `v8`：
+
+1. 至少跑到 `epoch 3` 之后，确认 stage 2 开启后的趋势
+2. 优先看 `ov3` 是否开始摆脱 `tool` collapse
+3. 如果 `epoch 3-5` 之后仍然：
+   - `ov3` 继续单类塌缩
+   - `ocls` 不升
+   - `oazi` 不降
+   - `F20` 没有明显抬升
+   再考虑下一轮结构修改
+
+当前最合理的工作顺序是：
+
+- 先把 `v8` 跑穿 stage 1 / stage 2
+- 再根据 `ov23` 的 class collapse 是否缓解，决定下一轮改：
+  - query decoder
+  - matching
+  - finer token rate
+  - 或额外的 class-preserving auxiliary

docs/0423.md

@@ -0,0 +1,969 @@
+# 2026-04-23 — v9_ov123_top4：class-first 针对性修复（A→D→E→B→C→F）
+
+## 0. 背景与本轮任务
+
+v8 / v8a 在 `03_ov123_top4` 上完成了 10+ 个 epoch 的训练，整体 spatial 指标比 v7h 有小幅改善，但 `class_ok` 一直卡在 **~45%** 左右。用户要求：
+
+1. 专注 frame 级预测，不引入任何 clip 级监督；
+2. 分析 v8 / v8a 的 val CSV（epoch 9 / epoch 11），找出 class 准确率上不去的真正瓶颈；
+3. 在不破坏现有代码框架和训练逻辑的前提下，按顺序落地 Fix A→D→E→B→C→F；
+4. 全部作为 v9 新 preset + 新脚本，从 v8a `best.pt` 热启动（保证 epoch0 前向输出与 v8a 完全相同）。
+
+本文档记录 2026-04-23 这一轮的 **所有代码修改细节**，以及每个修改背后的 CSV 诊断证据。
+
+相关旧文档：
+- `docs/0422.md`：v7h → v8 架构升级（cross-attention fusion）
+- `docs/0422_v7h_v7j.md`：v7h / v7j / v7i 的 per-frame 多源诊断与 class-weighted CE 的引入
+- `docs/0421.md`：v7f → ov123 per-frame 扩展与 frame-track CSV dump
+
+---
+
+## 1. CSV 诊断：v8 ep9 / v8a ep11 的错误模式
+
+### 1.1 整体准确率（按 "activity ≥ 0.5 的 active track 里 DOA 最近的一条" 取 class）
+
+| 指标 | v8 ep9 | v8a ep11 |
+|---|---|---|
+| overall `cls_ok` | 933/2051 = **45.5%** | 934/2051 = **45.5%** |
+| DOA≤20° 命中内 `cls_ok` | 48.1% | 51.9% |
+| **Oracle cls**（无视 activity，取 DOA 最近 track 的 class） | **38.7%** | **46.4%** |
+
+两个关键数字几乎相等：`cls_ok(nearest active) ≈ oracle_cls`。这说明：
+
+> **activity 选的 track 和 DOA 选的 track 语义上一致**，K=4 的 track binding 是 ok 的；真正错的是 **"那条被选中的 track 自己的 class_logits 就预测错了"**。
+
+换句话说：**瓶颈不在 matching，不在 activity，而在 class head 本身的输出分布**。
+
+### 1.2 按 ov 分组
+
+v8 ep9：
+```
+ov2 : gt= 682  DOA_ok=51.6%  cls_ok=37.1%  any_track_cls_ok=38.9%  cls&DOA=23.9%
+ov3 : gt=1158  DOA_ok=42.5%  cls_ok=47.1%  any_track_cls_ok=59.6%  cls&DOA=18.0%
+other: gt= 211  DOA_ok=97.6%  cls_ok=64.0%  any_track_cls_ok=64.0%  cls&DOA=63.0%
+```
+
+v8a ep11 类似，ov3 略退：
+```
+ov2 :  cls_ok=44.0%  cls&DOA=33.7%
+ov3 :  cls_ok=43.1%  cls&DOA=14.2%
+other: cls_ok=64.0%  cls&DOA=63.0%
+```
+
+`other`（近似 ov1 单源）是 64%，ov2 是 37-44%，ov3 降到 43% 且 `cls&DOA` 只有 14%。**多源重叠帧的 class 比单源帧低 20 个点**，这是 ov3 demixing 失败的直接证据。
+
+### 1.3 每类错误模式（Oracle：无视 activity 取 DOA 最近）
+
+```
+aircraft  (n=100): speech(50%)  human_vocalization(50%)   → 永远 0% 正确
+vehicle   (n= 68): machine(74%) aircraft(13%) train(10%)  → 永远 0% 正确
+frog      (n= 50): bird(98~100%)                           → 永远预测成 bird
+speech    (n= 53): human_vocalization(94%) breathing(6%)   → 永远 0% 正确
+crackle   (n= 51): rain(53%)    machine(41%)               → 永远 0% 正确
+wind      (n= 16): fire(50~88%) vehicle/wood               → 永远 0% 正确
+drawer_cab(n= 50): tool(74~84%)                            → 永远 0% 正确
+tape      (n= 50): human_vocalization(88%) typing(46%)     → 永远 0% 正确
+knock     (n= 36): home_sound(64%) human_vocalization      → 永远 0% 正确
+train     (n=115): train(43%) crushing(37%) vehicle(15%)   → ~10%
+```
+
+错误全都是 **同父类 "sibling collapse"**：
+
+- `aircraft / vehicle / train` 同属 transportation；
+- `speech / human_vocalization / breathing / laughter` 同属 human-voice；
+- `frog / bird / insect` 同属 animal-vocal；
+- `drawer_cabinet / tool / home_sound / door` 同属 indoor-mechanical。
+
+### 1.4 Label bug 复核：frog→bird 是不是数据映射错了？
+
+查了 `build_ov{1,23}_foa_dataset.py`、`final_vocabulary.csv`、`spatial_dataset.py:_resolve_class_index`，以及 jsonl 里具体样本：
+
+```
+ov1 train frog: 113    ov1 train bird: 3270    →  1:29
+ov1 valid frog:   9    ov1 valid bird:   56
+ov1 test  frog:   3    ov1 test  bird:   51
+ov2 train frog:1078    ov2 train bird: 1255
+ov3 train frog:1384    ov3 train bird: 1432
+```
+
+**frog 标签完全正确**，不是 label bug。100% 预测成 bird 的根因是 ov1 训练集里 bird 比 frog 多 **29 倍**，纯粹的 class imbalance；其他两个 manifest (ov2/ov3) 已经平衡，但 ov1 训练数据量占比仍然不小，整体 prior 偏向 bird。
+
+## 2. 六个 Fix 的整体设计原则
+
+1. **全部 additive**：v9 只 **增加** 新 config 字段与新子模块，从不删除或覆盖 v8a 已有的路径；
+2. **全部 zero-init 或 identity-init**：新参数在 ckpt 加载时贡献 0（可验证 `v9(v8a.pt) - v8a(v8a.pt)` 前向最大 abs diff = 0）；
+3. **strict=False 兼容**：v9 的 18 个新参数在 v8a ckpt 中缺失，走现有 `_load_spatial_init_checkpoint` / `load_state_dict(..., strict=False)` 流程自然初始化；
+4. **不改 forward signature 的兼容性**：`FrameTrackPredictionHeads.forward` 新增的几个关键字参数默认 `None`，旧调用链（如 v7 / v8 非-track readout）不受影响；
+5. **所有 LR / schedule 默认行为等价于 v8a**：`class_head_lr_scale=1.0` 时走原 3-group 快速路径，`frame_class_ontology_smoothing=0.0` 时走原 `F.cross_entropy`。
+
+## 3. 文件级修改清单
+
+| 文件 | 修改内容 |
+|---|---|
+| `spatial_loss.py` | 新增 `frame_class_ontology_smoothing`、`frame_class_ontology_groups` 字段；CE 分支支持本体软标签 |
+| `spatial_modules.py` | `FrameTrackPredictionHeads` 新增 6 个构造参数 + 新增 `ClassHeadSpectralDemixer` 模块 |
+| `spatial_beats.py` | `SpatialBEATsConfig` 新增 8 个字段；两处 `FrameTrackPredictionHeads` 构造传入新参数；forward 两处 call 传入 `pre_pool_features`；新增 `_derive_pre_pool_time_mask` helper |
+| `train_spatial_beats.py` | 新增 `class_head_lr_scale`、`class_head_freeze_during_ramp_epochs`、`class_head_lr_scale_during_ramp` 字段；`build_optimizer` 拆出 `cls_head` group；epoch loop 动态写 cls_head LR；新增 `_V9_CLASS_WEIGHTS`、`_V9_ONTOLOGY_GROUPS`、`make_ov1_local_spatial_v9_ov123_top4_config`；preset dispatch 和 `--preset` choices 注册 `v9_ov123_top4` |
+| `run_ov1_v9_ov123_top4.sh` | 新增启动脚本（chmod +x） |
+
+---
+
+## 4. Fix A —— frog/稀有类的诊断（无代码改动）
+
+Fix A 是诊断性质的，没有代码落地。结论：frog 不是 label bug，是 ov1 train 的 29:1 imbalance。解决方案被折进 Fix D：
+
+- `bird` 权重：1.0 → **0.6**（压 catch-all）
+- `frog` 权重：1.0 → **3.0**（提 minority）
+- `insect` 权重：从 v7I 的 4.0 → **1.0**（v8/v8a 已经 100%，不需要加权）
+
+## 5. Fix D —— 重新设计 class 权重 `_V9_CLASS_WEIGHTS`
+
+### 5.1 数据证据
+
+v7I 的权重表给了 aircraft / vehicle / insect **4×** 权重：
+
+```
+v8 ep9:    aircraft 0%,  vehicle 0%,  insect 100%
+v8a ep11:  aircraft 0%,  vehicle 0%,  insect 100%,  printer 100%→59%（退化！）
+```
+
+- aircraft / vehicle 的 4× 没有任何效果，它们的声学特征与 speech / machine 真实接近；
+- 4× 让误分类（aircraft→speech）的 loss 放大 4 倍，模型为了降 loss 把 **speech 的预测分布也拉保守**，导致 speech 也掉到 0%（双输）；
+- v8a ep11 还有一个异常：printer 从 v8 ep9 的 100% 掉到 59%，原因见 Fix E。
+
+### 5.2 设计规则
+
+- **"catch-all" 类**（GT frame 上被当万金油预测的类）→ 权重下调：
+  - `human_vocalization` 0.4（被 speech / tape / typing / knock 当靶子）
+  - `bird` 0.6（frog 几乎全部坍缩到 bird）
+  - `machine` 0.5（vehicle / crackle 坍缩到 machine）
+  - `rain` 0.5（crackle / singing 的 collapse 目标）
+  - `breathing` 0.5（speech 94% 错分到这里）
+  - `home_sound` 0.6（knock / printer 的 collapse）
+- **易被 collapse 的稀有类** → 权重上调（但不过分，≤ 3×）：
+  - `frog` 3.0
+  - `crackle` 2.0、`tape` 2.0、`knock` 2.0、`drawer_cabinet` 2.0、`speech` 2.0
+  - `aircraft` / `vehicle` / `train` 回到 1.0-1.5（4× 已证实无效且伤 sibling）
+- **过度自信的稳健类** → 略压：
+  - `singing` 0.7、`printer` 0.7（在 ov123 任务里吸收错 FP）
+
+### 5.3 代码改动
+
+位置：`train_spatial_beats.py`（紧挨 `_V7I_CLASS_WEIGHTS` 之后）
+
+```python
+_V9_CLASS_WEIGHTS: List[float] = [
+    # 0  wind_instrument    1  string_instrument  2  guitar             3  body_sound
+    1.0,                    1.0,                  1.0,                  1.0,
+    # 4  drum               5  water              6  human_vocalization  7  keyboard_instrument
+    1.0,                    1.0,                  0.4,                  1.0,
+    # 8  bird               9  tool              10  machine            11  war_sound
+    0.6,                    1.0,                  0.5,                  1.0,
+    # 12 metal_clink       13  breathing         14  laughter           15  percussion
+    1.0,                    0.5,                  1.0,                  1.0,
+    # 16 speech            17  bell              18  dog                19  vehicle
+    2.0,                    1.0,                  1.0,                  1.5,
+    # 20 alarm             21  footsteps         22  train              23  telephone_alarm
+    1.0,                    1.0,                  1.5,                  1.0,
+    # 24 glass             25  wind              26  kitchenware        27  animal
+    1.0,                    1.5,                  1.0,                  1.0,
+    # 28 musical_instrument 29 thunderstorm      30  door               31  male_speech
+    1.0,                    1.0,                  1.0,                  1.0,
+    # 32 female_speech      33 cat               34  home_sound         35  insect
+    1.0,                    1.0,                  0.6,                  1.0,
+    # 36 typing             37 zipper            38  camera             39  clock
+    1.0,                    1.0,                  1.0,                  1.0,
+    # 40 fire               41 singing           42  tearing            43  writing
+    1.0,                    0.7,                  1.0,                  1.0,
+    # 44 car                45 rain              46  scratch            47  gong
+    1.0,                    0.5,                  1.0,                  1.0,
+    # 48 appliance          49 paper             50  drawer_cabinet     51  ocean
+    1.0,                    1.0,                  2.0,                  1.0,
+    # 52 knock              53 crackle           54  finger_snapping    55  aircraft
+    2.0,                    2.0,                  1.0,                  1.0,
+    # 56 crushing           57 printer           58  tape               59  wood
+    1.0,                    0.7,                  2.0,                  1.0,
+    # 60 crack              61 cooking           62  frog
+    1.0,                    1.0,                  3.0,
+]
+assert len(_V9_CLASS_WEIGHTS) == 63
+```
+
+v9 preset 里：`cfg.loss.frame_class_loss_weights = list(_V9_CLASS_WEIGHTS)`。
+
+## 6. Fix E —— class head LR 单独分组 + DOA ramp 期冻结
+
+### 6.1 数据证据
+
+v8a ep5 vs ep11（stage 2 DOA ramp 打开前后）：
+
+```
+ep5  ov3: ocls=44.4%  printer=100%
+ep11 ov3: ocls=44.4%  printer=59%     ← class 被 DOA 梯度扰动掉了
+```
+
+打开 dir/dist 的 Hungarian cost 和 loss 之后，class binding 被扰动。传统做法是整体降 LR，但那会连带把 trunk / decoder 的学习也拖慢。更好的做法是 **单独把 class_head 从 optimizer 拉出来，DOA ramp 期间冻结**。
+
+### 6.2 代码改动
+
+#### 6.2.1 新增 config 字段（`train_spatial_beats.py`）
+
+```python
+    trunk_lr_scale: float = 1.0
+    spatial_lr_scale: float = 1.0
+    # v9: isolated LR multiplier for the class_head inside
+    # frame_track_prediction_heads. When < 1.0 the class head is put in its
+    # own param group with lr = base_lr * class_head_lr_scale.  Used during
+    # DOA ramp (stage 2) to prevent class binding from being perturbed by
+    # the newly-unlocked dir/dist gradients.  1.0 = legacy behaviour.
+    class_head_lr_scale: float = 1.0
+    # Optional epoch-range override that further scales the class head LR
+    # specifically during the DOA ramp.  When set, between
+    # frame_spatial_loss_warmup_epochs and frame_spatial_loss_warmup_epochs
+    # + class_head_freeze_during_ramp_epochs the class head LR is set to
+    # class_head_lr_scale_during_ramp (defaults to 0.0 = frozen).  After the
+    # ramp window the LR returns to class_head_lr_scale.
+    class_head_freeze_during_ramp_epochs: int = 0
+    class_head_lr_scale_during_ramp: float = 0.0
+```
+
+#### 6.2.2 `build_optimizer` 拆出 cls_head group
+
+- 新增 `_CLASS_HEAD_PREFIXES`：
+  ```python
+  _CLASS_HEAD_PREFIXES = (
+      "frame_track_prediction_heads.class_head.",
+      "frame_track_prediction_heads.class_head_mlp.",
+      "frame_track_prediction_heads.class_head_demixer.",
+  )
+  ```
+- fast path 条件从 `trunk_scale == 1.0 and spatial_scale == 1.0` 改为再加 `and cls_head_scale == 1.0`；
+- 当 `cls_head_scale != 1.0` 时，匹配 `_CLASS_HEAD_PREFIXES` 的参数从 head_params 抽出，放入 `cls_head_params`；
+- 每个 param group 增加 `group_name` 字段（`"trunk"` / `"spatial"` / `"head"` / `"cls_head"`），便于 epoch loop 按 name 定位。
+
+关键代码片段：
+```python
+for name, param in model.named_parameters():
+    if not param.requires_grad:
+        continue
+    if name.startswith(_CLASS_HEAD_PREFIXES) and cls_head_scale != 1.0:
+        cls_head_params.append(param)
+    elif name.startswith(_TRUNK_PREFIXES):
+        trunk_params.append(param)
+    elif name.startswith(_SPATIAL_PREFIXES):
+        spatial_params.append(param)
+    else:
+        head_params.append(param)
+
+param_groups = []
+if trunk_params:   param_groups.append({"params": trunk_params,   "lr": base_lr * trunk_scale,   "weight_decay": wd, "group_name": "trunk"})
+if spatial_params: param_groups.append({"params": spatial_params, "lr": base_lr * spatial_scale, "weight_decay": wd, "group_name": "spatial"})
+if head_params:    param_groups.append({"params": head_params,    "lr": base_lr,                 "weight_decay": wd, "group_name": "head"})
+if cls_head_params:param_groups.append({"params": cls_head_params,"lr": base_lr * cls_head_scale, "weight_decay": wd, "group_name": "cls_head"})
+```
+
+#### 6.2.3 epoch loop 动态写 cls_head LR
+
+紧跟 spatial loss schedule 之后（在 `_log(f"[Epoch {epoch}] start")` 之前）：
+
+```python
+_cls_ramp_len = int(train_cfg.class_head_freeze_during_ramp_epochs)
+if _cls_ramp_len > 0 and _sp_warmup > 0 and train_cfg.class_head_lr_scale != 1.0:
+    in_ramp = _sp_warmup <= epoch < _sp_warmup + _cls_ramp_len
+    if in_ramp:
+        _cls_scale = train_cfg.class_head_lr_scale_during_ramp
+    else:
+        _cls_scale = train_cfg.class_head_lr_scale
+    for _g in optimizer.param_groups:
+        if _g.get("group_name") == "cls_head":
+            _g["lr"] = train_cfg.learning_rate * _cls_scale
+    _log(
+        f"[Epoch {epoch}] cls_head_lr scale={_cls_scale:.3f}  "
+        f"lr={train_cfg.learning_rate * _cls_scale:.2e}  "
+        f"(ramp_window={_sp_warmup}..{_sp_warmup + _cls_ramp_len - 1})"
+    )
+```
+
+#### 6.2.4 v9 preset 设定
+
+```python
+cfg.class_head_lr_scale = 0.3
+cfg.class_head_freeze_during_ramp_epochs = 4
+cfg.class_head_lr_scale_during_ramp = 0.0
+```
+
+v9 的 `SPATIAL_LR=1.5e-5`，各阶段 cls_head LR：
+
+| 阶段 | epoch | lambda_dir | cls_head_lr |
+|---|---|---|---|
+| stage 1（class-only warmup） | 0-2 | 0.0 | 4.5e-6 |
+| stage 2（DOA ramp，cls_head 冻结） | 3-6 | ramp 0→1 | **0.0** |
+| stage 3（全放开） | 7+ | 1.0 | 4.5e-6 |
+
+## 7. Fix B —— 本体论（hierarchical）标签平滑
+
+### 7.1 核心思想
+
+v8/v8a 的 class 错误 **≥70% 是 sibling collapse**（同 AudioSet 父类）。Hard CE 把 `frog→bird` 和 `frog→aircraft` 一视同仁地惩罚满 log loss，这不合理：
+
+- 对下游 LLM 而言 `frog↔bird` 混淆可以靠语义上下文恢复；
+- `aircraft↔speech` 这种跨域错误则完全荒谬。
+
+希望 loss 的惩罚强度 **匹配错误的 "语义距离"**。最简单的实现：在同父类内部做 label smoothing，跨父类保持硬 CE。
+
+### 7.2 代码改动
+
+#### 7.2.1 `SpatialLossConfig` 新字段（`spatial_loss.py`）
+
+```python
+    # v9 hierarchical label smoothing for the frame-track class head.
+    # When frame_class_ontology_smoothing > 0, the CE target becomes a soft
+    # label distribution:
+    #   target[c_gt] = 1 - eps
+    #   target[c_sib] = eps / |siblings|  (for each sibling in same ontology
+    #                                       group as c_gt; excludes c_gt itself)
+    #   target[c_other] = 0
+    # ...
+    frame_class_ontology_smoothing: float = 0.0
+    # Parallel list of sibling groups.  Each entry is a list of class indices
+    # belonging to the same AudioSet ontology parent.  A class may appear in
+    # only one group.  Empty list = no hierarchical smoothing.
+    frame_class_ontology_groups: List[List[int]] = None
+```
+
+`__post_init__` 把 `None` 替换为 `[]` 以防意外。
+
+#### 7.2.2 CE 分支改写（`spatial_loss.py:compute_frame_track_losses`）
+
+原有分支：
+```python
+loss_class = F.cross_entropy(
+    class_logits_flat, class_target_flat, weight=_cls_weights
+)
+```
+
+改为：
+```python
+eps_onto = float(config.frame_class_ontology_smoothing)
+onto_groups = config.frame_class_ontology_groups
+if eps_onto > 0.0 and onto_groups:
+    num_classes = class_logits_flat.size(-1)
+    # Build a [C, C] soft-target "mixing" table on first use and
+    # cache it on the config object to avoid per-batch rebuild.
+    if (
+        not hasattr(config, "_onto_mixing_table")
+        or config._onto_mixing_table is None
+        or config._onto_mixing_table.shape[0] != num_classes
+        or config._onto_mixing_table.dtype != class_logits_flat.dtype
+        or config._onto_mixing_table.device != device
+    ):
+        table = torch.zeros((num_classes, num_classes), dtype=class_logits_flat.dtype, device=device)
+        table.fill_diagonal_(1.0)
+        for group in onto_groups:
+            members = [int(c) for c in group if 0 <= int(c) < num_classes]
+            if len(members) < 2:
+                continue
+            for c in members:
+                siblings = [s for s in members if s != c]
+                table[c].zero_()
+                table[c, c] = 1.0 - eps_onto
+                sib_mass = eps_onto / len(siblings)
+                for s in siblings:
+                    table[c, s] = sib_mass
+        config._onto_mixing_table = table
+    soft_target = config._onto_mixing_table[class_target_flat]
+    log_probs = F.log_softmax(class_logits_flat, dim=-1)
+    if _cls_weights is not None:
+        sample_w = _cls_weights[class_target_flat]
+        per_sample_loss = -(soft_target * log_probs).sum(dim=-1)
+        loss_class = (per_sample_loss * sample_w).sum() / sample_w.sum().clamp_min(1e-8)
+    else:
+        loss_class = -(soft_target * log_probs).sum(dim=-1).mean)
+else:
+    loss_class = F.cross_entropy(class_logits_flat, class_target_flat, weight=_cls_weights)
+```
+
+实现要点：
+- **mixing table 缓存**：表存在 `config` 对象上（不是模块），第一次构建并缓存，之后按形状/device/dtype 重用，开销可忽略；
+- **per-class weight 兼容**：当同时启用 ontology smoothing 和 class weights 时，正类权重照常生效（按 GT hard class idx 加权每样本 loss）；
+- **不在 group 里的类**：table 默认对角线 = 1，所以未列入任何 group 的类自然退化为 hard one-hot，零影响；
+- **完全开关**：`frame_class_ontology_smoothing=0` 或 `frame_class_ontology_groups=[]` 都会走原 `F.cross_entropy`，v8/v8a 训练复现不受影响。
+
+### 7.3 Ontology groups（`_V9_ONTOLOGY_GROUPS`）
+
+位置：`train_spatial_beats.py`，v9 preset 之前。
+
+```python
+_V9_ONTOLOGY_GROUPS: List[List[int]] = [
+    # transportation: aircraft, vehicle, train, car
+    [55, 19, 22, 44],
+    # human voice (non-singing): speech, human_vocalization, male_speech,
+    # female_speech, breathing, laughter
+    [16, 6, 31, 32, 13, 14],
+    # animal vocal: bird, frog, insect, dog, cat, animal
+    [8, 62, 35, 18, 33, 27],
+    # indoor mechanical + appliances: tool, machine, appliance, printer,
+    # home_sound, door, drawer_cabinet, kitchenware, camera, clock, typing,
+    # zipper, tape, cooking
+    [9, 10, 48, 57, 34, 30, 50, 26, 38, 39, 36, 37, 58, 61],
+    # percussive / impact: knock, footsteps, crack, crackle, crushing,
+    # scratch, finger_snapping, tearing, writing, paper
+    [52, 21, 60, 53, 56, 46, 54, 42, 43, 49],
+    # weather / water / ambience: wind, rain, thunderstorm, ocean, water,
+    # fire, glass, metal_clink, wood
+    [25, 45, 29, 51, 5, 40, 24, 12, 59],
+    # musical instruments: wind_instrument, string_instrument, guitar, drum,
+    # keyboard_instrument, percussion, musical_instrument, gong, bell,
+    # singing
+    [0, 1, 2, 4, 7, 15, 28, 47, 17, 41],
+    # alarms / signals: alarm, telephone_alarm, war_sound
+    [20, 23, 11],
+]
+```
+
+**8 组，覆盖 62/63 个 class**（只有 `body_sound` 没归组，走 hard CE）。每个 class 仅出现在一个组中。
+
+v9 preset：
+```python
+cfg.loss.frame_class_ontology_smoothing = 0.1
+cfg.loss.frame_class_ontology_groups = [list(g) for g in _V9_ONTOLOGY_GROUPS]
+```
+
+## 8. Fix C —— 频谱级 demixing cross-attention
+
+### 8.1 动机
+
+- `track_time_features[B, K, T_s, D]` 是 `SourceQueryDecoder` 从 `fused_embeddings[B, T_s, D]` 里 decode 出来的；`fused_embeddings` 在 `frequency_pool` 之后，**频率维已经被池化掉了**；
+- DOA 能在多源重叠帧 demix（IV 通道物理上就编码了方向），但 class 没有等价物理解混通路；
+- 需要让每个 track latent 能"回看"池化前的 trunk 输出，从 F_p 个频率 token 里挑自己负责的那部分。
+
+### 8.2 模型结构
+
+- 输入：
+  - `track_time_features: [B, K, T_s, D]`（已经过 `input_norm`）
+  - `pre_pool_features: [B, T_p * F_p, D]`（BEATs trunk 输出，无 task tokens，未经 frequency_pool）
+  - `pre_pool_grid_size: (T_p, F_p)`
+  - `pre_pool_time_mask: [B, T_p]`（True = 有效时间步）
+- 时间对齐：frame `t ∈ [0, T_s)` → trunk 时间步 `t_p = round(t * T_p / T_s)`，clip 到 `[0, T_p-1]`；
+- KV 构造：`kv_grid[:, t_p, :, :]` → `[B, T_s, F_p, D]`，然后在 K 轴 expand 到 `[B, K, T_s, F_p, D]`，flatten 成 `[B*K*T_s, F_p, D]`；
+- Query：`track_time_features` reshape 成 `[B*K*T_s, 1, D]`；
+- 经 1 层 `nn.MultiheadAttention`（`num_heads=8`, `dropout=0.1`，`batch_first=True`），再 `out_proj(Linear 768→768)`，乘以标量 `gate`，加回 class head 的 `class_input`。
+
+### 8.3 初始化策略（关键）
+
+- `out_proj.weight` 全零、`out_proj.bias` 全零 → **加载时 demixer 输出 = 0**，class_logits 与 v8a 完全相同；
+- `gate = 1e-2`（**不是 0**！）→ 前向 = `gate * 0 = 0`（身份保证），但 `∂L/∂out_proj.weight = gate × ...` **非零**，梯度从 step 0 就能流进 demixer 的 attention 权重；这是关键的"gradient-warmup trick"，否则两端 zero 会把 demixer 永久冻在 0。
+
+### 8.4 `ClassHeadSpectralDemixer` 源码（`spatial_modules.py`）
+
+```python
+class ClassHeadSpectralDemixer(nn.Module):
+    def __init__(self, embed_dim=768, num_layers=1, num_heads=8, dropout=0.1):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_layers = max(1, int(num_layers))
+        self.kv_norm = nn.LayerNorm(embed_dim)
+        self.q_norm = nn.LayerNorm(embed_dim)
+        self.layers = nn.ModuleList([
+            nn.MultiheadAttention(embed_dim=embed_dim, num_heads=num_heads,
+                                  dropout=dropout, batch_first=True)
+            for _ in range(self.num_layers)
+        ])
+        self.out_proj = nn.Linear(embed_dim, embed_dim)
+        nn.init.zeros_(self.out_proj.weight)
+        nn.init.zeros_(self.out_proj.bias)
+        self.gate = nn.Parameter(torch.full((1,), 1e-2))
+
+    def forward(self, track_time_features, pre_pool_features,
+                pre_pool_grid_size, pre_pool_time_mask=None):
+        B, K, T_s, D = track_time_features.shape
+        T_p, F_p = int(pre_pool_grid_size[0]), int(pre_pool_grid_size[1])
+        if pre_pool_features.size(-1) != D:
+            raise ValueError(...)
+        expected = T_p * F_p
+        if pre_pool_features.size(1) != expected:
+            # Fall back gracefully — demixer is additive & zero-gated.
+            return track_time_features.new_zeros(track_time_features.shape)
+        kv_grid = pre_pool_features.view(B, T_p, F_p, D)
+        if T_s > 0 and T_p > 0:
+            time_idx = torch.arange(T_s, device=kv_grid.device).float() * (T_p / max(1, T_s))
+            time_idx = time_idx.round().clamp_(0, T_p - 1).long()
+        else:
+            time_idx = torch.zeros((T_s,), dtype=torch.long, device=kv_grid.device)
+        kv_per_frame = kv_grid[:, time_idx, :, :]
+        kv_per_frame = kv_per_frame.unsqueeze(1).expand(B, K, T_s, F_p, D).contiguous()
+        kv_flat = kv_per_frame.view(B * K * T_s, F_p, D)
+        kv_flat = self.kv_norm(kv_flat)
+        q_flat = track_time_features.reshape(B * K * T_s, 1, D)
+        q_flat = self.q_norm(q_flat)
+        key_padding_mask = None
+        if pre_pool_time_mask is not None:
+            per_frame_valid = pre_pool_time_mask[:, time_idx]
+            per_frame_valid = per_frame_valid.unsqueeze(1).expand(B, K, T_s).reshape(-1)
+            if not per_frame_valid.all():
+                ignore = ~per_frame_valid
+                key_padding_mask = ignore.unsqueeze(1).expand(-1, F_p).contiguous()
+        attn_out = q_flat
+        for layer in self.layers:
+            attn_out, _ = layer(attn_out, kv_flat, kv_flat,
+                                key_padding_mask=key_padding_mask, need_weights=False)
+        residual = self.out_proj(attn_out).view(B, K, T_s, D)
+        return residual * self.gate
+```
+
+### 8.5 `_derive_pre_pool_time_mask` helper（`spatial_beats.py`）
+
+紧跟 `_build_patch_padding_mask` 之后新增：
+
+```python
+def _derive_pre_pool_time_mask(
+    self,
+    patch_padding_mask: Optional[Tensor],
+    grid_size: Tuple[int, int],
+) -> Optional[Tensor]:
+    """Return a [B, T_p] boolean mask where True marks *valid* trunk time
+    steps.  Used by the v9 class-head spectral demixer to ignore padded
+    tail frames."""
+    if patch_padding_mask is None:
+        return None
+    t_p, f_p = grid_size
+    B = patch_padding_mask.size(0)
+    pad_grid = patch_padding_mask.view(B, t_p, f_p)
+    time_valid = ~pad_grid.all(dim=-1)
+    return time_valid
+```
+
+`patch_padding_mask` 的语义是 `True = padded`，time-valid 是 "某时间步还有非 padded 频率位置" → `~pad_grid.all(dim=-1)`。
+
+### 8.6 forward 两处 call 更新（`spatial_beats.py`）
+
+两处调用 `self.frame_track_prediction_heads(...)` 都扩展：
+
+```python
+_pre_pool_time_mask = self._derive_pre_pool_time_mask(
+    patch_padding_mask=patch_padding_mask,
+    grid_size=grid_size,
+)
+frame_track_prediction_output = self.frame_track_prediction_heads(
+    track_time_features=track_time_features,
+    track_latents=track_latents,
+    pre_pool_features=encoder_memory,
+    pre_pool_grid_size=grid_size,
+    pre_pool_time_mask=_pre_pool_time_mask,
+)
+```
+
+- 第 1 处在 `readout_scheme == "local_spatial"` 分支下 frame-track parallel 路径；
+- 第 2 处在 `readout_scheme == "local_spatial_track"` 分支下纯 per-frame 路径（v9 走这里）。
+
+两处都用到的上下文变量：
+- `encoder_memory`：`self.encode_patches(...)` 的输出，`[B, T_p*F_p, D]`，**已经过 trunk 但未 frequency_pool**，正好是 demixer 需要的 pre_pool features；
+- `grid_size`：`self.extract_patch_tokens(...)` 返回的 `(T_p, F_p)`；
+- `patch_padding_mask`：`self._build_patch_padding_mask(...)` 返回。
+
+## 9. Fix F —— 2-layer MLP 残差
+
+### 9.1 动机
+
+当前 class head 是 `nn.Linear(768, 63)`，对多源混合 token 的表达能力可能不够。加一个 2-layer MLP 残差 branch：
+
+```
+class_logits = class_head(x) + gate * class_head_mlp(x)
+class_head_mlp = Linear(768, 1536) → GELU → Dropout → LayerNorm → Linear(1536, 63)
+```
+
+### 9.2 初始化策略
+
+与 Fix C 的 demixer 同构：
+- `class_head_mlp[-1].weight / bias` 全零 → 残差输出 = 0；
+- `class_head_mlp_gate = 1e-2` → 前向仍然 = 0，但梯度可流进 MLP 最后一层（非零）；
+- 经过 1-2 个 step MLP 最后一层有非零权重后，前一层（GELU 前）也开始获得梯度。
+
+### 9.3 `FrameTrackPredictionHeads` 重构（`spatial_modules.py`）
+
+构造签名扩展：
+```python
+def __init__(
+    self,
+    embed_dim: int = 768,
+    num_classes: int = 63,
+    dropout: float = 0.1,
+    use_class_head_mlp_residual: bool = False,
+    class_head_mlp_hidden_multiplier: int = 2,
+    class_head_mlp_dropout: float = 0.1,
+    use_class_head_demixer: bool = False,
+    class_head_demixer_layers: int = 1,
+    class_head_demixer_heads: int = 8,
+    class_head_demixer_dropout: float = 0.1,
+) -> None:
+```
+
+构造体内新增：
+```python
+self.use_class_head_mlp_residual = bool(use_class_head_mlp_residual)
+if self.use_class_head_mlp_residual:
+    hidden = embed_dim * max(1, int(class_head_mlp_hidden_multiplier))
+    self.class_head_mlp = nn.Sequential(
+        nn.Linear(embed_dim, hidden),
+        nn.GELU(),
+        nn.Dropout(class_head_mlp_dropout),
+        nn.LayerNorm(hidden),
+        nn.Linear(hidden, num_classes),
+    )
+    nn.init.zeros_(self.class_head_mlp[-1].weight)
+    nn.init.zeros_(self.class_head_mlp[-1].bias)
+    self.class_head_mlp_gate = nn.Parameter(torch.full((1,), 1e-2))
+else:
+    self.class_head_mlp = None
+    self.class_head_mlp_gate = None
+
+self.use_class_head_demixer = bool(use_class_head_demixer)
+if self.use_class_head_demixer:
+    self.class_head_demixer = ClassHeadSpectralDemixer(
+        embed_dim=embed_dim,
+        num_layers=class_head_demixer_layers,
+        num_heads=class_head_demixer_heads,
+        dropout=class_head_demixer_dropout,
+    )
+else:
+    self.class_head_demixer = None
+```
+
+forward 改写：
+```python
+def forward(
+    self,
+    track_time_features: Tensor,
+    track_latents: Tensor,
+    pre_pool_features: Optional[Tensor] = None,
+    pre_pool_grid_size: Optional[Tuple[int, int]] = None,
+    pre_pool_time_mask: Optional[Tensor] = None,
+) -> FrameTrackPredictionOutput:
+    ...
+    x = self.input_norm(track_time_features)
+    activity = self.activity_head(x).squeeze(-1)
+    class_input = x
+    if (
+        self.class_head_demixer is not None
+        and pre_pool_features is not None
+        and pre_pool_grid_size is not None
+    ):
+        demix_residual = self.class_head_demixer(
+            track_time_features=x,
+            pre_pool_features=pre_pool_features,
+            pre_pool_grid_size=pre_pool_grid_size,
+            pre_pool_time_mask=pre_pool_time_mask,
+        )
+        class_input = class_input + demix_residual
+    class_logits = self.class_head(class_input)
+    if self.class_head_mlp is not None and self.class_head_mlp_gate is not None:
+        class_logits = class_logits + self.class_head_mlp_gate * self.class_head_mlp(class_input)
+    direction = F.normalize(self.direction_head(x), dim=-1)
+    distance = F.softplus(self.distance_head(x)).squeeze(-1)
+    ...
+```
+
+**关键细节**：
+- demixer 的 residual 加到 `class_input` 上（即 class_head 的输入），而不是加到 logits 上；这样 demixer 得到的"纠偏"信号先经过 `class_head` 的共享投影再生成 logits；
+- MLP 分支也看 `class_input`（包含 demixer residual），这样 demixer 的信息同样能进入 MLP 分支；
+- `direction_head` 和 `distance_head` 的输入 `x` 不加 demixer（demixer 是 class-specific），保持 DOA head 对 v8a ckpt 的完全等价。
+
+### 9.4 `SpatialBEATsConfig` 新字段（`spatial_beats.py`）
+
+```python
+self.frame_track_dropout: float = 0.1
+self.frame_accdoa_hidden_dim: int = 256
+self.frame_accdoa_dropout: float = 0.1
+
+# v9: optional zero-initialised MLP residual branch inside
+# FrameTrackPredictionHeads.  ...
+self.use_class_head_mlp_residual: bool = False
+self.class_head_mlp_hidden_multiplier: int = 2
+self.class_head_mlp_dropout: float = 0.1
+# v9: optional spectral demixing cross-attention branch.  ...
+self.use_class_head_demixer: bool = False
+self.class_head_demixer_layers: int = 1
+self.class_head_demixer_heads: int = 8
+self.class_head_demixer_dropout: float = 0.1
+```
+
+### 9.5 两处构造都传入新参数
+
+`spatial_beats.py` 两处 `FrameTrackPredictionHeads(...)` 调用都改为：
+
+```python
+self.frame_track_prediction_heads = FrameTrackPredictionHeads(
+    embed_dim=cfg.encoder_embed_dim,
+    num_classes=cfg.source_num_classes,
+    dropout=cfg.frame_track_dropout,
+    use_class_head_mlp_residual=cfg.use_class_head_mlp_residual,
+    class_head_mlp_hidden_multiplier=cfg.class_head_mlp_hidden_multiplier,
+    class_head_mlp_dropout=cfg.class_head_mlp_dropout,
+    use_class_head_demixer=cfg.use_class_head_demixer,
+    class_head_demixer_layers=cfg.class_head_demixer_layers,
+    class_head_demixer_heads=cfg.class_head_demixer_heads,
+    class_head_demixer_dropout=cfg.class_head_demixer_dropout,
+)
+```
+
+## 10. v9 preset 与启动脚本
+
+### 10.1 `make_ov1_local_spatial_v9_ov123_top4_config`（`train_spatial_beats.py`）
+
+位置：v7k 系列的最后一个 preset 之后，`v3: top-8 unfreeze` 注释分割线之前。
+
+```python
+def make_ov1_local_spatial_v9_ov123_top4_config(
+    ov1_manifest_path: str = DEFAULT_OV1_MANIFEST,
+    ov2_manifest_path: str = DEFAULT_OV2_MANIFEST,
+    ov3_manifest_path: str = DEFAULT_OV3_MANIFEST,
+) -> TrainSpatialBEATsConfig:
+    """v9 = v8a + class-first cleanup (fixes A..F).
+
+    Inherits v8a (cross-attn fusion + segment matching + 4-epoch DOA ramp),
+    applies:
+      - _V9_CLASS_WEIGHTS (suppress catch-all classes, boost frog/crackle/tape)
+      - ontology-aware label smoothing (eps=0.1)
+      - class head residual MLP + spectral demixer (both zero-init)
+      - class_head_lr_scale=0.3 with full freeze during the 4-epoch DOA ramp
+
+    Frontend / trunk / source_query_decoder / activity / dir / dist heads
+    are unchanged.  Hot-start from v8a best.pt works with strict=False.
+    """
+    cfg = make_ov1_local_spatial_v8a_ov123_top4_config(
+        ov1_manifest_path=ov1_manifest_path,
+        ov2_manifest_path=ov2_manifest_path,
+        ov3_manifest_path=ov3_manifest_path,
+    )
+
+    # (D) Re-balanced class weights driven by v8/v8a CSV confusion analysis.
+    cfg.loss.frame_class_loss_weights = list(_V9_CLASS_WEIGHTS)
+
+    # (B) Hierarchical (ontology-aware) label smoothing.
+    cfg.loss.frame_class_ontology_smoothing = 0.1
+    cfg.loss.frame_class_ontology_groups = [list(g) for g in _V9_ONTOLOGY_GROUPS]
+
+    # (F) Zero-gated MLP residual on the class head.
+    cfg.model.use_class_head_mlp_residual = True
+    cfg.model.class_head_mlp_hidden_multiplier = 2
+    cfg.model.class_head_mlp_dropout = 0.1
+
+    # (C) Zero-gated spectral demixing cross-attention on the class head.
+    cfg.model.use_class_head_demixer = True
+    cfg.model.class_head_demixer_layers = 1
+    cfg.model.class_head_demixer_heads = 8
+    cfg.model.class_head_demixer_dropout = 0.1
+
+    # (E) Class head gets its own LR group.
+    cfg.class_head_lr_scale = 0.3
+    cfg.class_head_freeze_during_ramp_epochs = 4
+    cfg.class_head_lr_scale_during_ramp = 0.0
+
+    cfg.num_epochs = 12
+    cfg.output_dir = "checkpoints/spatial_beats_ov1_local_spatial_v9_ov123_exp/03_ov123_top4"
+    return cfg
+```
+
+### 10.2 preset 注册
+
+`train_spatial_beats.py` 中两处：
+
+1. 在 `args.preset == "ov1_local_spatial_v8a_ov123_top4"` 分支之后，新增：
+   ```python
+   elif args.preset == "ov1_local_spatial_v9_ov123_top4":
+       cfg = make_ov1_local_spatial_v9_ov123_top4_config(
+           ov1_manifest_path=args.ov1_manifest,
+           ov2_manifest_path=args.ov2_manifest,
+           ov3_manifest_path=args.ov3_manifest,
+       )
+   ```
+2. argparse `--preset` 的 `choices=(...)` 列表里在 `"ov1_local_spatial_v8a_ov123_top4"` 之后加入 `"ov1_local_spatial_v9_ov123_top4"`。
+
+### 10.3 `run_ov1_v9_ov123_top4.sh`
+
+新建文件（chmod +x）：
+
+```bash
+#!/usr/bin/env bash
+set -euo pipefail
+
+# v9_ov123_top4: v8a + class-first cleanup (Fix A..F from docs/0423.md analysis)
+# 所有 fix 都是 additive + zero-init，从 v8a best.pt 热启动 epoch-0 输出与 v8a 完全相同
+
+GPUS="${GPUS:-8}"
+BATCH_SIZE="${BATCH_SIZE:-8}"
+NUM_WORKERS="${NUM_WORKERS:-8}"
+SPATIAL_EPOCHS="${SPATIAL_EPOCHS:-12}"
+SPATIAL_LR="${SPATIAL_LR:-1.5e-5}"
+AMP="${AMP:-fp32}"
+
+OV1_MANIFEST="${OV1_MANIFEST:-/apdcephfs_cq10/.../ov1_foa.jsonl}"
+OV2_MANIFEST="${OV2_MANIFEST:-/apdcephfs_cq10/.../ov2_foa.jsonl}"
+OV3_MANIFEST="${OV3_MANIFEST:-/apdcephfs_cq10/.../ov3_foa.jsonl}"
+
+RESUME_CKPT="${RESUME_CKPT:-checkpoints/spatial_beats_ov1_local_spatial_v8a_ov123_exp/03_ov123_top4/best.pt}"
+OUT_DIR="${OUT_DIR:-checkpoints/spatial_beats_ov1_local_spatial_v9_ov123_exp/03_ov123_top4}"
+
+torchrun --nproc_per_node="${GPUS}" --master-port="${MASTER_PORT:-29557}" train_spatial_beats.py \
+  --preset ov1_local_spatial_v9_ov123_top4 \
+  --resume "${RESUME_CKPT}" \
+  --output-dir "${OUT_DIR}" \
+  --ov1-manifest "${OV1_MANIFEST}" \
+  --ov2-manifest "${OV2_MANIFEST}" \
+  --ov3-manifest "${OV3_MANIFEST}" \
+  --batch-size "${BATCH_SIZE}" \
+  --num-workers "${NUM_WORKERS}" \
+  --num-epochs "${SPATIAL_EPOCHS}" \
+  --learning-rate "${SPATIAL_LR}" \
+  --amp "${AMP}" \
+  --no-resume-optimizer \
+  --reset-epoch-on-resume \
+  --reset-best-on-resume
+```
+
+---
+
+## 11. 正确性验证（已完成）
+
+### 11.1 语法检查
+
+四个 py 文件（`train_spatial_beats.py` / `spatial_loss.py` / `spatial_beats.py` / `spatial_modules.py`）+ `run_ov1_v9_ov123_top4.sh` 全部通过 ast.parse / bash -n。
+
+### 11.2 模块级单测
+
+`FrameTrackPredictionHeads` + `ClassHeadSpectralDemixer`：
+- 开启所有 v9 选项后，**max abs diff = 0.00e+00** 相对于无 v9 选项的同一模型；
+- 带 padding mask 时 `pred_class_logits` 无 NaN；
+- 打开 MLP gate / demixer gate + 轻微扰动权重后，logits 差异 ~1e-2（符合预期）。
+
+### 11.3 v9 vs v8a 端到端前向恒等
+
+```python
+torch.manual_seed(42)
+m_v8a = SpatialBEATs(make_ov1_local_spatial_v8a_ov123_top4_config().model).eval()
+torch.manual_seed(42)
+m_v9  = SpatialBEATs(make_ov1_local_spatial_v9_ov123_top4_config().model).eval()
+
+# 两个模型都 load v8a/best.pt，strict=False
+# 同一 waveform 输入
+max_abs_diff = (o_v8a - o_v9).abs().max()
+# 实测：0.00e+00
+```
+
+**v9 加载 v8a ckpt 后的 `pred_class_logits` 与 v8a 模型加载同一 ckpt 的输出逐元素完全相等**（0.00e+00）。
+
+### 11.4 v8a ckpt 加载统计
+
+```
+ckpt keys            = 425
+loadable (v9 shapes) = 425
+v9 missing params    = 18   ← class_head_mlp.* (7 个) + class_head_demixer.* (11 个)
+ckpt unexpected      = 0
+```
+
+18 个新参数通过 `strict=False` 默认初始化（按 Fix C / F 的 zero-out-proj + tiny-gate 策略）。
+
+### 11.5 梯度流
+
+假 CE loss + backward：
+- `class_head.weight`：grad_norm ≈ 9.06（正常）
+- `class_head_mlp.4.weight`（MLP 最后一层，zero-init）：grad_norm ≈ 1.32e-1（**非零，因为 gate = 1e-2**）
+- `class_head_mlp.0.weight`（MLP 第一层）：grad_norm = 0（正常，要等最后一层非零后才有梯度，1 step 内开解）
+- `class_head_demixer.out_proj.weight`：grad_norm ≈ 1.60e-2（**非零**）
+- `class_head_demixer.layers.0.in_proj_weight`：grad_norm = 0（同理）
+
+### 11.6 Optimizer 分组
+
+```
+group 0: name=trunk    lr=3.00e-06  n_params=238
+group 1: name=spatial  lr=9.00e-06  n_params= 84
+group 2: name=head     lr=3.00e-05  n_params= 91
+group 3: name=cls_head lr=9.00e-06  n_params= 19
+```
+
+（上例 base_lr=3e-5；v9 启动脚本里 SPATIAL_LR=1.5e-5，对应 cls_head lr = 4.5e-6。）
+
+## 12. 关键训练时间线（SPATIAL_EPOCHS=12）
+
+继承 v8a 的 `frame_spatial_loss_warmup_epochs=3`、`frame_spatial_loss_ramp_epochs=4`，叠加 v9 新的 cls_head LR 调��：
+
+| epoch | lambda_dir / dist | dir/dist match cost | cls_head lr | 说明 |
+|---|---|---|---|---|
+| 0-2 | 0.0 | 0.0 | 4.5e-6 | Stage 1：class-only warmup，cls_head 低 LR 继续微调（但本来就 v8a 延续） |
+| 3 | ramp 0.25 | 0.25 | **0.0** | Stage 2 起点，cls_head 冻结 |
+| 4 | ramp 0.50 | 0.50 | **0.0** | cls_head 冻结 |
+| 5 | ramp 0.75 | 0.75 | **0.0** | cls_head 冻结 |
+| 6 | ramp 1.00 | 1.00 | **0.0** | cls_head 冻结，DOA 全开 |
+| 7-11 | 1.0 | 1.0 | 4.5e-6 | Stage 3：全部放开，cls_head 低 LR 微调 |
+
+## 13. 验证清单 / 观察优先级
+
+启动后按以下顺序观察指标：
+
+1. **epoch 0 validation metrics 是否与 v8a 的最后一个 epoch 等价**
+   - 预期 val loss ≈ v8a best val loss（因为前向 = v8a）
+   - 若不等价，说明 hot-start 出问题；
+2. **epoch 0-2（stage 1）**：
+   - `ocls`（oracle class acc）是否比 v8a ep2 同期高？关键指标，主要受 Fix D + B + C + F 驱动；
+   - 若 Fix D 生效：aircraft / vehicle 不再 0%，frog 不再 100%→bird；
+   - 若 Fix B 生效：smooth CE loss 比 hard CE 降得更快（从 1 - eps 开始）；
+   - 若 Fix C/F 生效：每 epoch 训完后 `class_head_mlp_gate` / `class_head_demixer.gate` 应从 0.01 慢慢增长；
+3. **epoch 3-6（stage 2，cls_head 冻结）**：
+   - DOA 指标改善（`LE_CD` 下降，`oazi` 下降），`ocls` **不应退化**（cls_head LR=0，梯度不回传）；
+   - 若 `ocls` 退化：可能是 fusion/trunk 的梯度通过 demixer 影响了 class_head 输入分布，此时需要考虑把 demixer 也一起冻；
+4. **epoch 7+（stage 3，全开）**：
+   - `F20` 是否突破 v8a 的上限（v7h 基线 0.246，v8a ~ 0.22-0.25）；
+   - ov3 class_ok 是否从 43% 往上走；
+   - 若 ov2/ov3 class_ok 都有改善但 ov3 仍然落后：Fix C 的频谱 demixing 还需要再加层数或 heads。
+
+## 14. 代码 diff 总览
+
+### 14.1 `spatial_loss.py`
+- 新增 2 个 `SpatialLossConfig` 字段：`frame_class_ontology_smoothing`、`frame_class_ontology_groups`；
+- `__post_init__` 处理 `frame_class_ontology_groups = None`；
+- `compute_frame_track_losses` 里 CE 分支支持 soft target + class weight 复合。
+
+### 14.2 `spatial_modules.py`
+- `FrameTrackPredictionHeads.__init__` 新增 6 个 kwargs；
+- `FrameTrackPredictionHeads.forward` 新增 3 个 Optional kwargs（`pre_pool_features` / `pre_pool_grid_size` / `pre_pool_time_mask`），内部整合 demixer + MLP residual；
+- 新增 `ClassHeadSpectralDemixer` 模块。
+
+### 14.3 `spatial_beats.py`
+- `SpatialBEATsConfig.__init__` 新增 8 个 `self.*` 字段；
+- 两处 `FrameTrackPredictionHeads(...)` 构造传入全部新参数；
+- 两处 `self.frame_track_prediction_heads(...)` 调用传入 pre-pool 参数；
+- 新增 `SpatialBEATs._derive_pre_pool_time_mask` 实例方法。
+
+### 14.4 `train_spatial_beats.py`
+- `TrainSpatialBEATsConfig` 新增 3 个字段：`class_head_lr_scale`、`class_head_freeze_during_ramp_epochs`、`class_head_lr_scale_during_ramp`；
+- 新增常量 `_V9_CLASS_WEIGHTS`（63 长度）+ `_V9_ONTOLOGY_GROUPS`（8 组，覆盖 62/63 class）；
+- 新增 preset 函数 `make_ov1_local_spatial_v9_ov123_top4_config`；
+- `build_optimizer` 新增 `_CLASS_HEAD_PREFIXES`、`cls_head_params` 分组、`group_name` 标签、fast-path 条件更新；
+- epoch loop 动态写 cls_head LR（紧跟 spatial loss schedule 后）；
+- preset dispatch 和 `--preset` argparse choices 注册 `ov1_local_spatial_v9_ov123_top4`。
+
+### 14.5 `run_ov1_v9_ov123_top4.sh`
+- 新建 shell 脚本，默认 8 GPU / BS=8 / 12 epoch / LR=1.5e-5 / fp32 / master_port=29557；
+- 默认 RESUME_CKPT 指向 v8a best.pt；
+- 复用 `--no-resume-optimizer --reset-epoch-on-resume --reset-best-on-resume` 三开关。
+
+## 15. 关键文件与行数
+
+| 文件 | 新增代码主要位置 |
+|---|---|
+| `spatial_loss.py` | `SpatialLossConfig` 新增字段紧跟 `frame_class_loss_weights` 之后；CE 分支改写在 `compute_frame_track_losses` 内 |
+| `spatial_modules.py` | `FrameTrackPredictionHeads` 整块重写；`ClassHeadSpectralDemixer` 定义在 `FrameTrackPredictionHeads` 之后 |
+| `spatial_beats.py` | `SpatialBEATsConfig.__init__` 的 v9 字段紧跟 `frame_accdoa_dropout`；`_derive_pre_pool_time_mask` 紧跟 `_build_patch_padding_mask` |
+| `train_spatial_beats.py` | `_V9_CLASS_WEIGHTS` 紧跟 `_V7I_CLASS_WEIGHTS` 之后；`_V9_ONTOLOGY_GROUPS` 和 `make_ov1_local_spatial_v9_ov123_top4_config` 紧跟 v7k 系列结束处；`class_head_*_scale` 字段紧跟 `spatial_lr_scale` 之后 |
+
+## 16. 后续计划
+
+如果 v9 ep3 `ocls` 仍不超过 v8a 同期，下一步（v9a/v9b）候选：
+
+1. **demixer 加层**：`class_head_demixer_layers = 2` 或增加 heads 到 16；
+2. **demixer 冻结共进退**：DOA ramp 期把 `class_head_demixer.*` 也当作 cls_head group 的一部分（目前 `_CLASS_HEAD_PREFIXES` 已经包含它，在 ramp 期也会冻结——已经生效，但需要验证效果）；
+3. **ontology smoothing eps 提升**：0.1 → 0.2，进一步容忍 sibling collapse；
+4. **source_query_decoder 加正交正则**（前一版建议）：`L_orth = ||Q Q^T - I||_F^2`，λ=0.01；
+5. **重新审视 matching**：如果 cls 瓶颈解了但 F20 仍不涨，回到 `docs/0422.md` 的 track-dead / duplicate 诊断。
+
+但当前 v9 已经把 "class head 本身" 这一块做得比较彻底，应该先跑满 12 epoch 再决定。

docs/0424.md

@@ -0,0 +1,269 @@
+# 2026-04-24 — `v9_real_balanced_10hz` real dump 诊断
+
+本文档记录 `v9_real_balanced_10hz` 在真实数据 dump 上的直接 CSV 诊断结果，不依赖训练日志里的 aggregate `val_metrics`。
+
+相关文件：
+- dump 目录：`checkpoints/v9_10hz_eval_dump/v9_real_balanced_10hz`
+- 统计脚本：[scripts/analyze_csv_dump.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/scripts/analyze_csv_dump.py)
+
+## 1. 背景
+
+用户的核心问题不是 “aggregate 指标多少”，而是：
+
+1. `real_ov1 / real_ov2 / real_ov3` 到底差在哪；
+2. 是类错、角度错，还是对的 track 没被最终输出；
+3. `10Hz + real mix` 到底有没有把模型带坏。
+
+为避免继续靠 `oracle_* / F20 / LE_CD` 猜，我们直接分析了导出的 `__pred.csv / __gt.csv`。
+
+## 2. 统计脚本
+
+新增脚本：
+
+```bash
+python3 scripts/analyze_csv_dump.py \
+  --dump-dir checkpoints/v9_10hz_eval_dump/v9_real_balanced_10hz
+```
+
+默认行为：
+- 不做阈值，直接分析 raw `__pred.csv` 的全部 track
+
+分析阈值后的最终输出：
+
+```bash
+python3 scripts/analyze_csv_dump.py \
+  --dump-dir checkpoints/v9_10hz_eval_dump/v9_real_balanced_10hz \
+  --threshold 0.5 \
+  --threshold-sweep 0.3 0.4 0.6
+```
+
+脚本会输出两类统计：
+
+### 2.1 GT-side
+
+对每个 GT source/frame 看：
+
+- `hit_cls_and_angle`
+  含义：至少存在一个同类预测，且最佳角误差 `<=20°`
+- `class_right_angle_wrong`
+  含义：存在同类预测，但最佳角误差 `>20°`
+- `no_same_class_pred_but_other_preds_exist`
+  含义：这帧模型有别的 active 预测，但没有任何同类预测
+- `no_pred_in_frame`
+  含义：这帧一个 active 预测都没有
+
+### 2.2 Pred-side
+
+对 threshold 后的预测看：
+
+- `matched_tp`
+  含义：能和某个 GT 做同类且 `<=20°` 的匹配
+- `same_class_angle_wrong_fp`
+  含义：有同类 GT，但角度没进 `20°`
+- `wrong_class_or_spurious_fp`
+  含义：没有任何同类 GT
+
+## 3. 一个重要事实：dump 里的 `pred.csv` 不是最终输出
+
+这次 `v9_real_balanced_10hz` 的 `pred.csv` 保存的是 **每帧 4 条 raw track 输出**，不是已经过 `activity_prob>=0.5` 筛选后的最终预测。
+
+这非常重要，因为它允许我们把问题拆成两层：
+
+1. **raw 4-track 里有没有可用候选**
+2. **过阈值以后，最终留下来的到底是什么**
+
+## 4. Raw 4-track 结果
+
+命令：
+
+```bash
+python3 scripts/analyze_csv_dump.py \
+  --dump-dir checkpoints/v9_10hz_eval_dump/v9_real_balanced_10hz
+```
+
+核心结果：
+
+### 4.1 `real_ov1`
+
+- `avg_gt/frame = 1.00`
+- `avg_pred/frame = 4.00`
+- `hit_cls_and_angle = 54.2%`
+- `class_right_angle_wrong = 45.8%`
+- `no_same_class_pred_but_other_preds_exist = 0.0%`
+
+解释：
+- raw 4 条里，**每个 GT 都能找到同类候选**
+- 其中一半以上角度也已经进了 `20°`
+- 所以单源 real 上，raw 候选并不差
+
+### 4.2 `real_ov2`
+
+- `avg_gt/frame = 1.98`
+- `avg_pred/frame = 4.00`
+- `hit_cls_and_angle = 37.4%`
+- `class_right_angle_wrong = 56.9%`
+- `no_same_class_pred_but_other_preds_exist = 5.7%`
+
+解释：
+- raw 4 条里，大多数 GT 还是能找到同类候选
+- 但**主问题已经是角度本身错**
+
+### 4.3 `real_ov3`
+
+- `avg_gt/frame = 2.88`
+- `avg_pred/frame = 4.00`
+- `hit_cls_and_angle = 33.9%`
+- `class_right_angle_wrong = 41.6%`
+- `no_same_class_pred_but_other_preds_exist = 24.5%`
+
+解释：
+- 即使给满 4 条 raw 候选，仍有 `24.5%` 的 GT 找不到任何同类预测
+- 所以 `real_ov3` 从 raw 层面就已经有明显 `class/source binding` 问题
+
+## 5. `activity>=0.5` 后的结果
+
+命令：
+
+```bash
+python3 scripts/analyze_csv_dump.py \
+  --dump-dir checkpoints/v9_10hz_eval_dump/v9_real_balanced_10hz \
+  --threshold 0.5
+```
+
+### 5.1 `real_ov1`
+
+- `avg_gt/frame = 1.00`
+- `avg_pred/frame = 1.14`
+- `hit_cls_and_angle = 22.4%`
+- `class_right_angle_wrong = 40.2%`
+- `no_same_class_pred_but_other_preds_exist = 37.3%`
+
+解释：
+- raw 4-track 时，同类候选是 `100%` 存在的
+- 过阈值后，`37.3%` 的 GT 直接变成 “这帧有别的 active 预测，但没有同类预测”
+- 这说明 `real_ov1` 的主问题不是 “不会预测”，而是：
+  - 对的 track 没被留住
+  - 或被别的错 track 抢走
+  - 也就是 **decode / track ranking / calibration** 问题
+
+### 5.2 `real_ov2`
+
+- `avg_gt/frame = 1.98`
+- `avg_pred/frame = 1.93`
+- `hit_cls_and_angle = 17.9%`
+- `class_right_angle_wrong = 73.9%`
+- `no_same_class_pred_but_other_preds_exist = 8.2%`
+
+解释：
+- 轨数和 GT 基本对齐，不是明显少报
+- 主要失败项是 **同类有了，但角度错**
+- 所以 `real_ov2` 主问题不是 threshold，也不是主要类错，而是 **角度本身错**
+
+### 5.3 `real_ov3`
+
+- `avg_gt/frame = 2.88`
+- `avg_pred/frame = 1.82`
+- `247` 帧里有 `161` 帧是 `pred < gt`
+- `hit_cls_and_angle = 26.3%`
+- `class_right_angle_wrong = 30.2%`
+- `no_same_class_pred_but_other_preds_exist = 43.5%`
+
+解释：
+- 这里同时有三件事：
+  1. **active 轨数不够**
+  2. **同类 track 经常找不到**
+  3. 即使找到了，同类里也有不少角度不对
+
+因此 `real_ov3` 是：
+- `binding` 错
+- `decode` 少亮轨
+- `angle` 也错
+
+三件事叠在一起。
+
+## 6. Threshold sweep：是不是纯阈值问题
+
+命令：
+
+```bash
+python3 scripts/analyze_csv_dump.py \
+  --dump-dir checkpoints/v9_10hz_eval_dump/v9_real_balanced_10hz \
+  --threshold 0.5 \
+  --threshold-sweep 0.3 0.4 0.6
+```
+
+### 6.1 `real_ov1`
+
+从 `0.3 -> 0.6`：
+
+- `avg_pred/frame` 几乎不变：`1.14 -> 1.05`
+- `hit` 基本只在高阈值 `0.6` 时下降
+- `no_same_cls` 一直很高：`37.3% -> 43.4%`
+
+结论：
+- **不是简单阈值问题**
+- 更像是对的 track 本来就没排到最终输出前面
+
+### 6.2 `real_ov2`
+
+从 `0.3 -> 0.6`：
+
+- `avg_pred/frame` 几乎不变：`1.99 -> 1.91`
+- `hit` 几乎不变：`17.9% -> 17.8%`
+- `class_right_angle_wrong` 一直卡在 `73%` 左右
+
+结论：
+- **几乎完全不是 threshold 问题**
+- 就是 **角度头本身错**
+
+### 6.3 `real_ov3`
+
+从 `0.5 -> 0.3`：
+
+- `avg_pred/frame` 从 `1.82 -> 2.15`
+- `under_frames` 从 `161 -> 140`
+- `no_same_cls` 从 `43.5% -> 34.7%`
+- 但 `hit` 只从 `26.3% -> 27.0%`
+
+结论：
+- 降阈值确实能缓一点 “少亮轨”
+- 但收益有限，根问题还在
+- 所以 `real_ov3` 不是纯阈值问题
+
+## 7. 最终结论
+
+### 7.1 每个 split 的主问题
+
+- `real_ov1`
+  - 主问题：**decode / track ranking / calibration**
+  - 证据：raw 4-track 同类候选 `100%` 存在，但阈值后大量 GT 找不到同类 track
+
+- `real_ov2`
+  - 主问题：**角度本身错**
+  - 证据：轨数基本对，class 也不是主要问题，但 `73.9%` 变成 “同类有了但角度错”
+
+- `real_ov3`
+  - 主问题：**binding 错 + 少亮轨 + 角度错**
+  - 证据：
+    - raw 层 already `24.5%` 找不到同类
+    - threshold 后 `avg_pred/frame = 1.82 < 2.88`
+    - 同时还有 `30.2%` 的同类角度错
+
+### 7.2 这次问题不是一句 “real 很差” 能说清的
+
+更准确的说法应该是：
+
+- `real_ov1`：内部候选还可以，但最终输出选坏了
+- `real_ov2`：主要是 spatial regression 错
+- `real_ov3`：multi-source 下 source binding 和 spatial 一起掉了
+
+### 7.3 这也解释了为什么只看 aggregate 指标会误判
+
+同一个 `LE_CD / F20` 很差，背后可能是三种完全不同的失败机制：
+
+- 对的候选存在，但没被输出
+- 同类 track 有了，但角度偏得很远
+- raw 4-track 里就没把 source 绑定出来
+
+所以以后继续看 real dump 时，应该固定用 `scripts/analyze_csv_dump.py`，先把问题拆成这三类，再决定改哪一层。
+

docs/SPATIAL_AUDIO_FRAMEWORKS_ANALYSIS.md

@@ -0,0 +1,724 @@
+# Comprehensive Analysis: Spatial Audio Frameworks & Alternative Architectures in Spatial-BEATs Codebase
+
+*Last Updated: 2026-04-27*
+*Analysis Scope: Complete codebase survey for spatial audio frameworks, alternative SELD approaches, and experimental architectures*
+
+---
+
+## Executive Summary
+
+This codebase implements **Spatial-BEATs**, a spatial sound event localization and detection (SELD) system built on top of the BEATs audio pre-training model. The implementation includes:
+
+1. **Three parallel multi-source supervision routes** (Routes A/B/C) for per-frame spatial prediction
+2. **Multiple alternative spatial architectures** already implemented and compared
+3. **Strong references to existing SELD frameworks** from the DCASE challenge series
+4. **An experimental series (v11) exploring architectural improvements** through spectral demixing and paradigm shifts
+
+All architectures coexist through conditional compilation (no destructive changes), allowing side-by-side comparison.
+
+---
+
+## Part 1: Existing Spatial Audio Frameworks Referenced
+
+### 1.1 **Spatial-AST** (Primary Inspiration)
+- **References**: Mentioned extensively throughout codebase
+- **Location**: `.gitignore:9` explicitly lists "Spatial-AST" as a protected directory
+- **Role**: Foundational work for task-token-based spatial audio processing
+- **Influence in code**:
+  - `PreTrunkASTPredictionHeads`: Direct implementation of Spatial-AST task-token architecture (lines 1177-1237 in `spatial_modules.py`)
+  - Pre-trunk supervision strategy with separate task tokens for distance, DoA, and class
+  - Single-source readout design adapted for multi-source scenarios
+  
+**Key Characteristic**: Task tokens injected **before** the trunk transformer (pre-trunk), requiring separate head predictions after trunk passage
+
+---
+
+### 1.2 **DCASE Challenge SELD Baseline**
+- **References**: Extensively cited throughout codebase
+- **Location**: 
+  - `spatial_loss.py` lines 3079-3300+ (SELDMetricsAccumulator, OfficialDCASESELDMetrics)
+  - Multiple DCASE-style FOA channel ordering conventions documented
+  - Official evaluation metrics implemented: ER20, F20, LE_CD, LR_CD, SELD_score
+
+**DCASE Metrics Computed**:
+```
+SELD = (ER + (1-F) + LE/180 + (1-LR)) / 4
+Where:
+- ER: Error Rate (precision/recall weighted)
+- F: F-score
+- LE: Localization Error (great-circle distance)
+- LR: Localization Recall
+```
+
+**Route C directly borrows DCASE paradigm**:
+- Per-class activity-coupled direction (ACCDOA: Activity-Coupled Cartesian Direction of Arrival)
+- No explicit source-to-slot matching required
+- Simple and stable for ov2/ov3 data constraints (same-class overlap ~ 0)
+
+**Connection to DCASE Code**:
+```python
+# Line 3304 references official implementation:
+# https://github.com/sharathadavanne/seld-dcase2023/blob/master/SELD_evaluation_metrics.py
+```
+
+---
+
+### 1.3 **EINV2 (Event Independent Network V2)** 
+- **References**: 
+  - `docs/spatial_beats_ov123_frame_routes.md` line 6, 36
+  - `run_ov123_local_spatial_track.sh` line 4
+  
+**EINV2 Paradigm Borrowed**:
+- K learnable track queries per clip (K=4 in current implementation)
+- Clip-level Hungarian matching for track-to-source binding
+- Per-track temporal attention to capture source continuity across frames
+- **Implemented as Route B** in this codebase
+
+**Route B Architecture** (`local_spatial_track`):
+```
+SourceQueryDecoder:
+  ├─ Stage 1 (track-level): K queries → TransformerDecoder → [B, K, D] track latents
+  └─ Stage 2 (per-frame): Expand with temporal positional embeddings → [B, K, T_s, D]
+       ↓
+FrameTrackPredictionHeads:
+  ├─ Activity prediction: [B, K, T_s, 1]
+  ├─ Class prediction: [B, K, T_s, num_classes]
+  ├─ Direction prediction (L2-norm): [B, K, T_s, 3]
+  └─ Distance prediction (softplus): [B, K, T_s, 1]
+```
+
+**Key Innovation from v9→v11a/b**: Added `ClassHeadSpectralDemixer` to break the bottleneck where multiple sources were compressed into a single D-dim vector after frequency pooling
+
+---
+
+## Part 2: Three Parallel Multi-Source Supervision Routes
+
+All three routes coexist through conditional compilation (no conflicts):
+
+### 2.1 **Route A: `local_spatial_slot`** (Per-frame K-slot assignment)
+
+**Architecture**:
+```python
+class FrameSlotHead(nn.Module):
+    """Route A — per-frame K-slot head for multi-source supervision."""
+    # Per time step independently predicts K slots
+    # Each slot: class, direction (L2-norm), distance (softplus)
+    # Per-step Hungarian matching for source assignment
+```
+
+**Supervision Strategy**:
+- `compute_frame_slot_losses()` (line 2573+ in `spatial_loss.py`)
+- Per-frame independent slot allocation (no temporal continuity assumption)
+- Per-step Hungarian matching to assign ground-truth sources to K slots
+- Loss weighted: `w_act + w_cls + w_dir + w_dist` per (batch, time, slot)
+
+**Loss Configuration** (Route A typical):
+```python
+lambda_frame_activity = 1.0
+lambda_frame_class = 1.0
+lambda_frame_direction = 4.0
+lambda_frame_distance = 1.0
+lambda_clip_aux = 0.1
+frame_num_slots = 4
+```
+
+**When to Use**: Scenarios with frequent source entry/exit and short trajectories
+
+**Shell Script**: `run_ov123_local_spatial_slot.sh`
+
+---
+
+### 2.2 **Route B: `local_spatial_track`** (K track queries with temporal self-attention)
+
+**Architecture** (Already documented above):
+- Two-stage Transformer decoder (`SourceQueryDecoder`)
+- Learnable source queries with temporal positional embeddings
+- Clip-level Hungarian matching (not per-step)
+- Strong temporal coherence assumption
+
+**Supervision Strategy**:
+- `compute_frame_track_losses()` (line 2682+ in `spatial_loss.py`)
+- Clip-level Hungarian: matched track k* ↔ gt source n*
+- Per-matched-track supervision across entire time window
+- Unmatched tracks supervised as "inactive" across all frames
+
+**Loss Configuration** (Route B typical):
+```python
+lambda_frame_activity = 1.0
+lambda_frame_class = 1.0
+lambda_frame_direction = 4.0
+lambda_frame_distance = 1.0
+lambda_clip_aux = 0.1
+frame_num_slots = 4  # via num_queries in SourceQueryDecoder
+```
+
+**When to Use**: Continuous source trajectories, inter-frame identity persistence important
+
+**Shell Script**: `run_ov123_local_spatial_track.sh`
+
+**EINV2 Connection**: This route is explicitly designed as Spatial-BEATs' interpretation of EINV2's event-independent tracking paradigm
+
+---
+
+### 2.3 **Route C: `local_spatial_accdoa`** (Per-class ACCDOA vector field)
+
+**Architecture**:
+```python
+class ACCDOAHeads(nn.Module):
+    """Route C — per-class per-frame ACCDOA head.
+    
+    For each (batch, time, class) emits a 3D vector v whose:
+    - Magnitude (||v||) encodes class activity at that frame
+    - Direction (v/||v||) encodes DoA (azimuth, elevation)
+    - Distance predicted separately per (batch, time, class)
+    
+    Output:
+        pred_accdoa: [B, T_s, num_classes, 3]
+        pred_distance: [B, T_s, num_classes]
+    """
+```
+
+**Key Properties**:
+- **No explicit matching** required (no Hungarian, no query assignment)
+- **Per-class decomposition**: Each class has its own spatial slot → eliminates binding ambiguity
+- **Activity-coupled**: Class activity directly from vector magnitude (||v||)
+- **Assumes**: No same-class overlap within a frame (true for ov2/ov3)
+
+**Supervision Strategy**:
+- `compute_frame_accdoa_losses()` (line 2857+ in `spatial_loss.py`)
+- Target construction (lines 2803-2854):
+  ```python
+  # For each valid source at (b, t, class):
+  accdoa_target[b, t, cls, :] = unit_direction_vector
+  distance_target[b, t, cls] = distance_value
+  # Otherwise: accdoa_target = zero vector, distance masked out
+  ```
+- Loss = MSE(predicted_accdoa, target_accdoa) across valid time steps
+- Distance = smooth_l1 loss only for active (b,t,c) locations
+
+**Loss Configuration** (Route C - ACCDOA specific):
+```python
+lambda_frame_activity = 4.0  # ACCDOA MSE dominates
+lambda_frame_class = 0.0      # No separate class CE (implied by per-class slot)
+lambda_frame_direction = 0.0  # No separate direction (in ACCDOA MSE already)
+lambda_frame_distance = 1.0
+lambda_clip_aux = 0.1
+frame_accdoa_activity_threshold = 0.5  # For inference: ||v|| > 0.5 → active
+```
+
+**When to Use**: 
+- No same-class overlap constraint (ov2/ov3 data)
+- Simpler topology, less query binding complexity
+- Per-class interpretation more natural for sound events
+
+**Shell Script**: `run_ov123_local_spatial_accdoa.sh`
+
+**DCASE Connection**: Directly based on DCASE SELD challenge baseline (Activity-Coupled Cartesian DoA)
+
+---
+
+## Part 3: Historical Architecture Evolution (v7 → v11)
+
+### 3.1 **v7 Series: Early Frame-Level Approaches**
+
+**v7 baseline**: Clip-level single-source `LocalSpatialPredictionHeads`
+- Works for ov1 (single source per clip)
+- Can't handle multi-source ov2/ov3
+
+**v7→v8 jump**: Introduction of per-frame supervision with per-track queries
+
+---
+
+### 3.2 **v9 Series: Frame-Track with Class-Head Spectral Demixer (Current Baseline)**
+
+**v9 Core Innovations**:
+1. **Symmetric attention through class head spectral demixer** (`ClassHeadSpectralDemixer`)
+   - Input: Per-track per-frame features `[B, K, T_s, D]`
+   - Problem: Multiple sources compressed into single D-vector after frequency pooling
+   - Solution: For class head, attend back to BEATs trunk **pre-pool** tokens `[B, T_p*F_p, D]` with class-specific cross-attention
+   - Gate mechanism: `out = (class_head(x) + gate * demixer_mlp(x))`
+   - Zero-initialized: `demixer_mlp` output layer starts at 0; gate starts at 1e-2
+   - Guarantee: epoch-0 is equivalent to non-demixer baseline
+
+2. **Class-weighted CE loss** (`frame_class_loss_weights = _V9_CLASS_WEIGHTS`)
+   - Per-class weight matrix to handle dataset imbalance
+   - Located: `train_spatial_beats.py` lines 1674-1769
+
+3. **Direction/Distance heads remain bottlenecked**:
+   - Still use only post-pooled `track_time_features`
+   - **This is the root cause of real_ov2 failures** (73.9% of predictions have right class but wrong angle)
+
+**v9 Loss Weights**:
+```python
+lambda_frame_activity = 1.0
+lambda_frame_class = 1.0
+lambda_frame_direction = 4.0  # Weighted 4x over activity
+lambda_frame_distance = 1.0
+```
+
+**Configuration Factory**: `make_ov1_local_spatial_v9_ov123_top4_config()`
+
+---
+
+### 3.3 **v10 Series: Phase-Wise Training (Activity Re-balancing)**
+
+**v10 Phase-1 (`v10_phase1_cls`)**: Pure classification refinement
+- Freeze all spatial prediction sub-heads (direction_head, distance_head)
+- Train only class head + new `num_active_head` (how many sources active at each frame)
+- Rationale: v9's class recall peaked early (ep3) then dropped; spatial kept improving
+- Hyperparameters:
+  ```python
+  lambda_frame_direction = 0.0   # Freeze
+  lambda_frame_distance = 0.0    # Freeze
+  lambda_frame_activity = 0.5    # Weakened to not drag class around
+  lambda_frame_num_active = 0.5  # New: multi-source count CE
+  base_lr = 7.5e-6               # Halved from v9's 1.5e-5
+  ```
+
+**v10b (`v10b_phase1_activity`)**: Activity re-balancing on top of v10 phase-1
+- Re-enable spatial heads, but with different activity weighting
+- Diagnosed: v10 phase-1's lambda_frame_activity=0.5 was too weak for ov3
+- Fix:
+  ```python
+  lambda_frame_activity = 1.0    # Restored
+  lambda_frame_num_active = 0.8  # Tuned
+  ```
+
+**Root Finding**: Activity supervision strength varies by split (ov1/ov2/ov3 imbalances)
+
+**Shell Scripts**: 
+- `run_ov1_v10_phase1_cls.sh`
+- `run_ov1_v10b_phase1_activity.sh`
+
+---
+
+### 3.4 **v11 Series: Architectural Refinements (2026-04-27)**
+
+Four independent experiments to isolate failure modes:
+
+#### **v11a: Symmetric Spectral Demixer for DOA**
+- **Problem Diagnosed**: v9 added `ClassHeadSpectralDemixer` for class head only
+- **Real_ov2 Symptom**: 73.9% of predictions have `class_right_angle_wrong`
+- **Root Cause**: direction/distance heads still see only post-pooled vectors
+- **Solution**: Add symmetric `spatial_head_demixer` for direction and distance heads
+  ```python
+  # In FrameTrackPredictionHeads.__init__:
+  self.spatial_head_demixer = ClassHeadSpectralDemixer(...)
+  # Attend to BEATs trunk pre-pool tokens just like class_demixer
+  ```
+- **Initialization**: Zero-gated (epoch-0 identical to v9)
+- **Configuration Factory**: `make_ov1_local_spatial_v11a_ov123_top4_config()`
+- **Shell Script**: `run_ov1_v11a_ov123_top4.sh`
+- **Expected Outcome**: real_ov2 angle errors should decrease significantly
+
+#### **v11b: DOA Demixer with LocalSpatial Pre-Pool KV**
+- **Question**: v11a demixer attends to BEATs mono fbank pre-pool (no direction info)
+- **Hypothesis**: If demixer attends to LocalSpatial's 7-channel (4-FOA + 3-IV) pre-pool, might be better
+- **Implementation**:
+  ```python
+  LocalSpatialEncoder.forward(return_pre_pool=True)
+  # Returns 4D CNN feature before frequency pooling: [B, D_s, T_f, F_cnn]
+  # Reshape to [B, T_f*F_cnn, D_s] + grid info
+  # Project with local_spatial_pre_pool_proj: Linear(D_s → D=768)
+  # Pass to spatial_head_demixer as alternative KV source
+  ```
+- **Hyperparameters**:
+  - `spatial_demixer_use_local_spatial_kv = True`
+  - Demixer gate still 1e-2, output layer still 0
+- **Configuration Factory**: `make_ov1_local_spatial_v11b_ov123_top4_config()`
+- **Shell Script**: `run_ov1_v11b_ov123_top4.sh`
+- **Expected Comparison**:
+  - v11b >> v11a → Physical IV signal crucial for DOA (BEATs pre-pool insufficient)
+  - v11b ≈ v11a → Already enough spatial context in fuser (IV mixed back in)
+  - v11b < v11a → New KV too noisy or proj under-trained
+
+#### **v11c: Paradigm Shift to ACCDOA**
+- **Problem Addressed**: real_ov3 24.5% raw GT layer without same-class candidates
+- **Root Cause**: K-track binding (Hungarian matching) failure for ov3
+- **Hypothesis**: Query-binding stage is bottleneck, not head improvements
+- **Solution**: Replace entire Route B topology with Route C (ACCDOA)
+  - No queries, no Hungarian, per-class vector slots
+  - Each class inherently has its own "slot" → binding non-issue
+- **Initialization**: From ov1 local_spatial warmup checkpoint (strict=False)
+  - Reuses: BEATs trunk + LocalSpatialEncoder + fusion stack
+  - Replaces: SourceQueryDecoder + FrameTrackPredictionHeads → ACCDOAHeads
+- **Hyperparameters**:
+  ```python
+  lambda_frame_activity = 4.0  # ACCDOA MSE dominates
+  lambda_frame_class = 0.0
+  lambda_frame_direction = 0.0
+  lambda_frame_distance = 1.0
+  num_epochs = 24
+  learning_rate = 3e-5  # Conservative (not tuned on multi-source)
+  ```
+- **Configuration Factory**: `make_ov1_local_spatial_v11c_ov123_accdoa()`
+- **Shell Script**: `run_ov1_v11c_ov123_accdoa.sh`
+- **Expected Outcome**:
+  - real_ov3 `no_same_class_pred_but_other_preds_exist` drops → query binding proven as bottleneck
+  - real_ov2 may improve (per-class vector naturally separates sources)
+  - sim_ov1 class accuracy may drop slightly (activity-coupling with DOA reduces class signal)
+
+#### **v11d: Decode-Time Activity Calibration**
+- **Problem**: real_ov1 loses 37% of same-class candidates post-activity-thresholding
+- **Root Cause**: Threshold (0.5) vs. actual activity probability distribution mismatch
+- **Solution**: Post-hoc decode recalibration (no retraining)
+  ```python
+  # Three decode strategies:
+  1. threshold: Fixed thr in {0.3, 0.4, 0.5, 0.6}
+  2. topk_hat: Per-frame top-K̂ by activity, K̂ = v10's num_active_head argmax
+  3. topk_hat_min: (K̂ membership) AND (thr > min_thr)
+  ```
+- **Tool**: `scripts/calibrate_activity.py` (new, stdlib-only)
+- **Inputs**: Pre-computed `__pred.csv` and `__gt.csv` with `activity_prob` and `num_active_pred` columns
+- **Outputs**: Per-split optimal decode config
+- **Property**: Pure post-processing, no model change, reproducible for all checkpoints
+- **Expected Outcome**: real_ov1 no_same_class drops significantly; real_ov2/ov3 limited gains (their problems are not ranking)
+
+**v11 Diagnostic Summary**:
+| Split | Symptom | Candidate Root Cause | Experiment |
+|-------|---------|----------------------|------------|
+| real_ov1 | 37% drop post-thresholding | Ranking/calibration | v11d |
+| real_ov2 | 73.9% class_right_angle_wrong | DOA head bottleneck | v11a/b |
+| real_ov3 | 24.5% no_same_class_candidate | Query binding | v11c |
+
+---
+
+## Part 4: Key Implementation Patterns
+
+### 4.1 **Shared Preprocessing & Fusion Stack**
+
+```python
+# All routes use identical front-end:
+FOA waveform (16kHz, 4ch)
+  ↓
+SpatialBEATsPreprocessor (mel-spectrogram + log)
+  ↓
+SpatialPatchEmbedding (patch tokenization)
+  ↓
+BEATs TransformerEncoder (frozen or LoRA)
+  ↓
+FrequencyPool (mean across frequency)
+  ↓
+TemporalResampler (→ 2.5 Hz)
+  ↓
+LocalSpatialEncoder (FOA-specific spatial encoding)
+  ↓
+LocalSpatialFusion (semantic + spatial projection)
+  ↓
+fused_spatial_embeddings: [B, T_s, D] at 2.5 Hz
+  ↓
+Route-specific readout head (A/B/C/original)
+```
+
+**Fusion Mechanism**:
+```python
+local_spatial_out = LocalSpatialEncoder(foa_features)
+fused = LayerNorm(semantic_seq + LocalSpatialProjector(local_spatial_out))
+```
+
+**Shared Clip Auxiliary Head** (all routes):
+- `LocalSpatialPredictionHeads` runs in parallel
+- Clip-level single-source supervisions
+- Weight: `lambda_clip_aux = 0.1`
+- Provides stable classification signal for all routes
+
+---
+
+### 4.2 **Spatial Loss Dispatch (spatial_loss.py)**
+
+**Conditional Loss Computation**:
+```python
+if supervision_mode == "local_spatial_slot":
+    loss_out = compute_frame_slot_losses(
+        prediction=model_output.frame_slot_prediction_output,
+        batch=batch, cfg=loss_config)
+elif supervision_mode == "local_spatial_track":
+    loss_out = compute_frame_track_losses(...)
+elif supervision_mode == "local_spatial_accdoa":
+    loss_out = compute_frame_accdoa_losses(...)
+```
+
+**Loss Output Fields** (reused via semantic mapping):
+```python
+@dataclass
+class SpatialLossOutput:
+    loss_total: Tensor  # Weighted sum of components
+    loss_activity: Tensor  # BCE / ACCDOA magnitude / etc.
+    loss_cls: Tensor  # Per-source class CE
+    loss_dir: Tensor  # 1 - cos(direction)
+    loss_dist: Tensor  # smooth_l1(distance)
+    loss_cls_aux: Tensor  # Track/slot matched class (route B/A specific)
+    loss_temp: Tensor  # Clip aux head loss × 0.1
+    # No changes to dataclass structure; repurpose fields per route
+```
+
+---
+
+### 4.3 **Checkpoint Hot-Starting Strategy**
+
+**Typical Flow** (for routes A/B/C):
+1. Initialize from `ov1_local_spatial_run1/best.pt` (ov1 single-source baseline)
+   - Carries: BEATs trunk, LocalSpatialEncoder, fusion stack
+   - Frozen trunk ensures warm-start stability
+2. Load with `strict=False` to skip incompatible heads (Route B/C have different head topology)
+3. New parameters initialized with:
+   - LayerNorm: default
+   - Linear: trunc_normal_(std=2e-5) for light initialization
+   - **Spectral demixer gate**: 1e-2 (near-zero residual)
+   - **Spectral demixer output**: zeros (bit-equivalent to baseline at epoch 0)
+
+**Route C Special Case** (v11c):
+- v9 best.pt not compatible (no ACCDOAHeads)
+- Hot-start from ov1 local_spatial instead
+- Trade-off: Less pre-training, but topology matches by default
+
+---
+
+## Part 5: Experimental Configurations Summary
+
+### Available Presets:
+```python
+# Legacy single-source routes (unchanged):
+"ov1_ast"                 # Spatial-AST style pre-trunk
+"ov1_local_spatial"       # Original ov1 baseline
+"ov1_pretrunk_ast"        # Pre-trunk task tokens
+
+# Multi-source frame-level routes (new):
+"ov123_local_spatial_slot"      # Route A
+"ov123_local_spatial_track"     # Route B (EINV2-style)
+"ov123_local_spatial_accdoa"    # Route C (DCASE SELD style)
+
+# v9 series (current production):
+"ov1_local_spatial_v9_ov123_top4"
+
+# v10 series (phase-wise):
+"ov1_local_spatial_v10_phase1_cls"
+"ov1_local_spatial_v10b_phase1_activity"
+
+# v11 series (architectural refinements):
+"ov1_local_spatial_v11a_ov123_top4"  # v9 + spatial demixer
+"ov1_local_spatial_v11b_ov123_top4"  # v11a + local spatial KV
+"ov1_local_spatial_v11c_ov123_accdoa" # ACCDOA paradigm
+```
+
+### Loss Weight Comparison Table:
+
+| Preset | `lambda_activity` | `lambda_class` | `lambda_direction` | `lambda_distance` | `lambda_clip_aux` | Notes |
+|--------|-----|-----|-----|-----|-----|---|
+| ov123_local_spatial_slot | 1.0 | 1.0 | 4.0 | 1.0 | 0.1 | Route A (DETR-style) |
+| ov123_local_spatial_track | 1.0 | 1.0 | 4.0 | 1.0 | 0.1 | Route B (EINV2-style) |
+| ov123_local_spatial_accdoa | 4.0 | 0.0 | 0.0 | 1.0 | 0.1 | Route C (DCASE style) |
+| v9_ov123_top4 | 1.0 | 1.0 | 4.0 | 1.0 | 0.1 | + class spectral demixer |
+| v11a_ov123_top4 | 1.0 | 1.0 | 4.0 | 1.0 | 0.1 | v9 + spatial demixer |
+| v11b_ov123_top4 | 1.0 | 1.0 | 4.0 | 1.0 | 0.1 | v11a + local spatial KV |
+| v11c_accdoa | 4.0 | 0.0 | 0.0 | 1.0 | 0.1 | ACCDOA with 24 ep, lr=3e-5 |
+
+---
+
+## Part 6: Code Organization Reference
+
+### Key Files & Line Ranges:
+
+**`spatial_modules.py`**:
+- Lines 1177-1237: `PreTrunkASTPredictionHeads` (Spatial-AST single-source)
+- Lines 1240-1310: `FixedSlotReadout` (auxiliary multi-source readout)
+- Lines 1467-1482: `FrameACCDOAPredictionOutput` (dataclass, Route C)
+- Lines 1484-1568: `FrameSlotHead` (Route A)
+- Lines 1569-1684: `SourceQueryDecoder` (Route B, EINV2-style)
+- Lines 1685-2130: `FrameTrackPredictionHeads` (Route B heads)
+- Lines 2132-2198: `ACCDOAHeads` (Route C)
+- Lines 2226-2350+: `FrameWisePredictionHeads` (legacy ov1 frame-wise)
+
+**`spatial_loss.py`**:
+- Lines 2573-2650: `compute_frame_slot_losses()` (Route A)
+- Lines 2682-2750: `compute_frame_track_losses()` (Route B)
+- Lines 2803-2854: `_build_accdoa_targets()` (Route C target construction)
+- Lines 2857-2945: `compute_frame_accdoa_losses()` (Route C)
+- Lines 3079-3300: `SELDMetricsAccumulator` + `OfficialDCASESELDMetrics` (DCASE)
+- Lines 3528-3580: `OfficialDCASEMetricsAccumulator` (Route B evaluation)
+
+**`spatial_beats.py`**:
+- Lines 347-362: Spatial-AST-style supervision documentation
+- Lines 673: `PreTrunkASTPredictionHeads` instantiation
+- Lines 760-810: `ACCDOAHeads` and route selection logic
+- Lines 1119: Spatial-AST task token encoding in trunk
+- Lines 1260-1350: Single-source Spatial-AST-style output building
+
+**`train_spatial_beats.py`**:
+- Lines 570-650: `make_ov1_ast_config()` (Spatial-AST factory)
+- Lines 675-750: BAT/Spatial-AST-style warmup
+- Lines 760-910: `make_ov1_local_spatial_*()` factories
+- Lines 2228-2280: `make_ov1_local_spatial_v9_ov123_top4_config()` (v9 baseline)
+- Lines 2286-2390: `make_ov1_local_spatial_v11a_ov123_top4_config()` (v11a with demixer)
+- Lines 2394-2545: v10 phase-1 and v10b configs
+- Lines 2987-3011: `make_ov123_local_spatial_accdoa_config()` (Route C)
+
+**Documentation**:
+- `docs/spatial_beats_ov123_frame_routes.md`: Complete Route A/B/C design spec (36KB)
+- `docs/0427_v11_series.md`: v11 experiments diagnostic (15KB)
+- `docs/spatial_beats_design_guide.md`: Spatial-AST comparison
+- `docs/spatial_beats_implementation_spec.md`: Implementation details
+
+---
+
+## Part 7: Research References & Connections
+
+### Explicit Code References:
+1. **BEATs** (foundational):
+   - https://arxiv.org/abs/2212.09058
+   - https://github.com/microsoft/unilm/tree/master/beats
+
+2. **DCASE SELD Evaluation**:
+   - https://github.com/sharathadavanne/seld-dcase2023/blob/master/SELD_evaluation_metrics.py
+   - Metrics: ER20, F20, LE_CD, LR_CD, SELD_score
+
+3. **SELD Metrics**:
+   - Reference implementation: scipy.optimize.linear_sum_assignment (Hungarian)
+   - Great-circle distance formula for localization error
+   - DCASE gt/pred row format: `[frame_idx, class_id, azi_deg, ele_deg, distance_m]`
+
+4. **AudioMAE** (inspiration for BEATs pre-training):
+   - Modified with binaural + IPD front-end in early Spatial-AST work
+
+### Implicit References:
+- **DETR** (Detection Transformer): Influences Route A slot-based design
+- **Transformer Decoders**: Standard PyTorch modules used in Route B (SourceQueryDecoder)
+- **FairSeq**: Referenced in code headers for fairness/best-practice attribution
+
+---
+
+## Part 8: Validation & Evaluation Framework
+
+### Metrics Computed Across All Routes:
+
+**Per-epoch validation**:
+- `class_acc`: Matched-source class top-1 accuracy
+- `azi_mae_deg`: Azimuth mean absolute error
+- `ele_mae_deg`: Elevation mean absolute error
+- `dist_mae_m`: Distance mean absolute error
+- `activity_f1`: Per-frame source activity F1-score
+- `num_active_mae`: Mean absolute error in number of active sources
+
+**Official DCASE metrics** (when applicable):
+- `ER`: Error Rate (0=perfect, 1=worst)
+- `F`: F-score (0=worst, 1=perfect)
+- `LE_CD`: Localization Error in degrees (0=perfect)
+- `LR_CD`: Localization Recall (0=worst, 1=perfect)
+- `SELD_score`: Joint score = (ER + (1-F) + LE/180 + (1-LR)) / 4
+
+**Best metric strategy**:
+- `best_metric_name = "class_acc"` (primary)
+- `minimize_best_metric = False` (maximize accuracy)
+- Fallback: direction error (azimuth MAE)
+
+### Validation Examples:
+All routes produce per-sample JSONL dumps for offline analysis:
+```python
+build_frame_{slot,track,accdoa}_validation_examples()
+# Outputs: per-sample predictions vs ground-truth for manual inspection
+```
+
+---
+
+## Part 9: Practical Comparison Guide
+
+### When to Use Each Route:
+
+**Route A (Slot)**: 
+- ✓ Frequent source entry/exit
+- ✓ Short trajectories
+- ✓ Minimal temporal coherence assumption
+- ✗ Heavy Hungarian per-frame (compute cost)
+
+**Route B (Track - EINV2 style)**:
+- ✓ Continuous source trajectories
+- ✓ Strong temporal coherence
+- ✓ Identity persistence across frames important
+- ✓ Current production baseline (v9)
+- ✗ Query binding complexity in crowded ov3 scenarios
+
+**Route C (ACCDOA)**:
+- ✓ Simple, no matching required
+- ✓ Per-class decomposition naturally separates sources
+- ✓ Assumes no same-class overlap (ov2/ov3 satisfied)
+- ✓ More interpretable output
+- ✗ Activity and DOA coupled (mag/direction trade-off)
+- ✗ Slightly lower ov1 class accuracy (known trade-off)
+
+### Architecture Progression for Model Development:
+
+1. **Baseline** → Start with v9 (current production)
+2. **Diagnosis** → Run v11a (is DOA the bottleneck?)
+3. **Refinement** → Based on v11a results, pick v11b or v11c
+4. **Post-hoc Tuning** → v11d (activity calibration, if needed)
+
+---
+
+## Part 10: Future Investigation Directions
+
+Based on current code exploration:
+
+1. **Hybrid Routes**:
+   - Could combine Route B's temporal coherence with Route C's per-class decomposition
+   - E.g., per-class track queries (K track queries × num_classes classes)
+
+2. **Demixer Expansion**:
+   - v11a/b only address class/direction heads; what about other bottlenecks?
+   - Could add demixer to activity head for multi-source activity separation
+
+3. **Loss Reweighting**:
+   - Current `lambda_frame_direction=4.0` fixed; could be data/split adaptive
+   - v10's phase-wise training shows promise; more sophisticated scheduling?
+
+4. **Temporal Aggregation**:
+   - All routes do per-frame independent supervision; temporal smoothing loss could help
+   - Route B's temporal decoders unused for loss computation (only forward pass)
+
+5. **Multi-Modal Fusion**:
+   - Current: FOA only; could inject synthetic room geometry or source priors
+   - BEATs backbone frozen; LoRA tuning on spatial-specific parameters
+
+---
+
+## Appendix: Common Command Reference
+
+```bash
+# Route A training
+./run_ov123_local_spatial_slot.sh
+
+# Route B training (v9 baseline with enhancements)
+./run_ov1_v9_ov123_top4.sh
+./run_ov1_v11a_ov123_top4.sh  # + spatial demixer
+./run_ov1_v11b_ov123_top4.sh  # + local spatial KV
+
+# Route C training (ACCDOA)
+./run_ov123_local_spatial_accdoa.sh
+./run_ov1_v11c_ov123_accdoa.sh
+
+# v11d: Post-hoc activity calibration
+python3 scripts/calibrate_activity.py --dump-dir <csv_dir>
+
+# Evaluation on real splits
+python3 scripts/eval_v7k_real_valid.py \
+  --ckpt <ckpt_path> \
+  --dump-pred-dir <dump_dir> \
+  --dump-splits real_ov1,real_ov2,real_ov3
+```
+
+---
+
+## Summary of Findings
+
+This codebase represents a comprehensive, well-engineered exploration of spatial audio architectures:
+
+1. **Three coexisting paradigms** (Routes A/B/C) allow direct comparison
+2. **Strong DCASE SELD foundation** with official evaluation metrics
+3. **Spatial-AST influence** visible in task-token and pre-trunk designs
+4. **EINV2 adaptation** in Route B's track-query approach
+5. **Systematic v11 experiments** to isolate architectural failure modes
+6. **Zero-initialized residuals** enable safe architectural extension without breaking baselines
+7. **Per-split diagnostics** guide targeted improvements
+8. **Checkpoint reuse strategy** allows efficient transfer across related tasks
+
+The codebase prioritizes **safety** (strict=False loading, zero-init demixers) and **interpretability** (per-route metrics, JSONL dumps) while maintaining experimental rigor (fixed random seeds, official DCASE evaluation).
+

docs/SPATIAL_FRAMEWORKS_QUICK_REFERENCE.md

@@ -0,0 +1,192 @@
+# Spatial Audio Frameworks — Quick Reference Card
+
+## Frameworks Found in Codebase
+
+### 1. **Spatial-AST** (Pre-trunk Task Tokens)
+- **File**: `spatial_modules.py` lines 1177-1237
+- **Class**: `PreTrunkASTPredictionHeads`
+- **Strategy**: Task tokens (distance, DoA, class) injected **before** trunk
+- **Single-source only**: Yes (ov1)
+- **Status**: Legacy (can coexist with multi-source routes)
+
+### 2. **DCASE SELD Baseline** (Per-Class ACCDOA)
+- **File**: `spatial_modules.py` lines 2132-2198 (`ACCDOAHeads`)
+- **File**: `spatial_loss.py` lines 2803-2945
+- **Strategy**: Activity-Coupled Cartesian DoA vector field per class
+- **No matching required**: Yes (implicit per-class slots)
+- **Status**: Route C (production alternative)
+- **Metrics**: ER20, F20, LE_CD, LR_CD, SELD_score
+
+### 3. **EINV2** (Event Independent Network)
+- **File**: `spatial_modules.py` lines 1569-2130
+- **Classes**: `SourceQueryDecoder` + `FrameTrackPredictionHeads`
+- **Strategy**: K learnable track queries with clip-level Hungarian
+- **Temporal coherence**: Yes (strong)
+- **Status**: Route B (current production baseline v9)
+- **Variants**: v9 (+ class demixer), v11a (+ spatial demixer), v11b (+ local spatial KV)
+
+## Three Parallel Routes at a Glance
+
+| Route | Name | Paradigm | Matching | Temporal | When to Use |
+|-------|------|----------|----------|----------|------------|
+| A | `local_spatial_slot` | DETR-style slots | Per-frame Hungarian | Weak | Frequent entry/exit |
+| B | `local_spatial_track` | EINV2-style queries | Clip-level Hungarian | Strong | Continuous trajectories |
+| C | `local_spatial_accdoa` | DCASE ACCDOA | None | Implicit | Simple, no overlap |
+
+## Version Series Quick Compare
+
+| Version | Route | Key Innovation | Loss Weights | Status |
+|---------|-------|-----------------|--------------|--------|
+| v7 | B | Clip-level single-source | — | Legacy |
+| v8a | B | Per-frame frame-track | 1/1/4/1 | Legacy |
+| v9 | B | Class spectral demixer | 1/1/4/1 | **Production** |
+| v10 | B | Phase-1 class focus | 0.5/1/0/0 → 1/1/4/1 | Staging |
+| v10b | B | Activity rebalancing | 1/0.8 adjustments | Staging |
+| v11a | B | Spatial demixer | 1/1/4/1 + DOA demixer | Testing |
+| v11b | B | Local spatial KV | 1/1/4/1 + local KV | Testing |
+| v11c | C | ACCDOA paradigm | 4/0/0/1 | Testing |
+| v11d | — | Activity calibration | Post-hoc decode only | Tool |
+
+## Key Code Locations
+
+### Architecture Classes
+- **Spatial-AST**: `PreTrunkASTPredictionHeads` (line 1177)
+- **EINV2 queries**: `SourceQueryDecoder` (line 1569)
+- **EINV2 heads**: `FrameTrackPredictionHeads` (line 1685)
+- **DCASE ACCDOA**: `ACCDOAHeads` (line 2132)
+- **DETR slots**: `FrameSlotHead` (line 1484)
+
+### Loss Functions
+- **Route A losses**: `compute_frame_slot_losses()` (line 2573)
+- **Route B losses**: `compute_frame_track_losses()` (line 2682)
+- **Route C losses**: `compute_frame_accdoa_losses()` (line 2857)
+- **DCASE metrics**: `OfficialDCASESELDMetrics` (line 3300)
+
+### Configuration Factories
+- **Spatial-AST**: `make_ov1_ast_config()` (line 570)
+- **v9 Route B**: `make_ov1_local_spatial_v9_ov123_top4_config()` (line 2228)
+- **v11a demixer**: `make_ov1_local_spatial_v11a_ov123_top4_config()` (line 2286)
+- **v11c ACCDOA**: `make_ov123_local_spatial_accdoa_config()` (line 2987)
+
+## Common Loss Weight Patterns
+
+```python
+# Route A (Slot) / Route B (Track)
+lambda_frame_activity = 1.0
+lambda_frame_class = 1.0
+lambda_frame_direction = 4.0  # 4x weighted
+lambda_frame_distance = 1.0
+lambda_clip_aux = 0.1
+
+# Route C (ACCDOA)
+lambda_frame_activity = 4.0   # MSE dominates
+lambda_frame_class = 0.0      # (implicit in per-class slot)
+lambda_frame_direction = 0.0  # (implicit in ACCDOA)
+lambda_frame_distance = 1.0
+lambda_clip_aux = 0.1
+```
+
+## Hot-Start Checkpoint Pattern
+
+All routes (A/B/C) typically start from:
+```
+checkpoints/spatial_beats_ov1_local_spatial_run1/best.pt
+```
+
+Load with `strict=False`:
+- ✓ Reuse: BEATs trunk, LocalSpatialEncoder, fusion stack
+- ✗ Skip: Incompatible prediction heads
+- New params: zero-initialized (demixer gate=1e-2)
+- Guarantee: epoch-0 identical to baseline
+
+## Shell Scripts for Running Experiments
+
+```bash
+# Route A
+./run_ov123_local_spatial_slot.sh
+
+# Route B (v9 current production)
+./run_ov1_v9_ov123_top4.sh
+
+# Route B (v11a with spatial demixer)
+./run_ov1_v11a_ov123_top4.sh
+
+# Route B (v11b with local spatial KV)
+./run_ov1_v11b_ov123_top4.sh
+
+# Route C (ACCDOA)
+./run_ov1_v11c_ov123_accdoa.sh
+
+# v11d (activity calibration, post-hoc)
+python3 scripts/calibrate_activity.py --dump-dir <csv_dir>
+```
+
+## Failure Mode → Experiment Mapping (v11 Series)
+
+| Split | Symptom | Root Cause | Experiment |
+|-------|---------|-----------|------------|
+| real_ov1 | 37% loss post-threshold | Ranking/activity calibration | v11d |
+| real_ov2 | 73.9% class_right_angle_wrong | DOA head bottleneck | v11a / v11b |
+| real_ov3 | 24.5% no_same_class_candidate | Query-to-source binding | v11c |
+
+## Key Constraints & Design Decisions
+
+�� **All coexist** through conditional compilation (readout_scheme selection)
+✓ **Zero-initialized demixers** ensure epoch-0 bit-equivalence to baselines
+✓ **Shared clip aux head** provides stable classification signal for all routes
+✓ **No matching** for Route C (ACCDOA) — eliminates binding complexity
+✓ **Frozen trunk** during stage 1 training (warm-start stability)
+✓ **official DCASE evaluator** for reproducible cross-framework comparison
+
+## References in Code
+
+- **Spatial-AST inspiration**: `.gitignore:9` (protected directory)
+- **DCASE SELD baseline**: `https://github.com/sharathadavanne/seld-dcase2023/`
+- **BEATs foundational model**: `https://arxiv.org/abs/2212.09058`
+- **EINV2 paradigm**: Docs lines mention "EINV2 track-wise" in route comparisons
+
+## For Comparative Studies
+
+Use this table to rapidly compare framework properties:
+
+```python
+FRAMEWORKS = {
+    "Spatial-AST": {
+        "class": "PreTrunkASTPredictionHeads",
+        "paradigm": "pre-trunk task tokens",
+        "sources": 1,  # single-source only
+        "matching": None,
+        "temporal": "none",
+        "file": "spatial_modules.py:1177",
+    },
+    "DCASE-ACCDOA": {
+        "class": "ACCDOAHeads",
+        "paradigm": "per-class vector field",
+        "sources": "multi (ov2/ov3)",
+        "matching": "none",
+        "temporal": "implicit per-class",
+        "file": "spatial_modules.py:2132",
+    },
+    "EINV2": {
+        "class": "SourceQueryDecoder + FrameTrackPredictionHeads",
+        "paradigm": "learnable track queries",
+        "sources": "multi (ov2/ov3)",
+        "matching": "clip-level Hungarian",
+        "temporal": "strong (query persistence)",
+        "file": "spatial_modules.py:1569,1685",
+    },
+    "DETR-inspired": {
+        "class": "FrameSlotHead",
+        "paradigm": "per-frame slot allocation",
+        "sources": "multi (ov2/ov3)",
+        "matching": "per-frame Hungarian",
+        "temporal": "weak (per-step independent)",
+        "file": "spatial_modules.py:1484",
+    },
+}
+```
+
+---
+
+**Last Updated**: 2026-04-27
+**For full analysis**: See `docs/SPATIAL_AUDIO_FRAMEWORKS_ANALYSIS.md`

docs/spatial_beats_design_guide.md

@@ -0,0 +1,603 @@
+# Spatial-BEATs 设计与训练指南
+
+## 1. 文档目标
+
+本文档用于整理本项目中 `Spatial-BEATs` 的任务定义、模型改造方向、保留与替换的模块、训练方法，以及后续接入 LLM 的接口约定。
+
+目标是基于公开的 BEATs 框架，构建一个独立的 `Spatial Encoder`：
+
+- 输入为 `FOA 音频` 及其派生的 `FOA 空间特征`
+- 主干尽可能复用 `BEATs backbone` 和其预训练权重
+- 输出为一组 `Spatial Tokens`
+- 这些 `Spatial Tokens` 作为独立模态输入给 LLM
+- 原有的语义 audio encoder 保持不动，避免直接与空间 encoder 混合后产生职责不清或语义冲突
+
+本设计不是在原有 LLM audio encoder 上强行加入空间分支，而是单独训练一个 `Spatial-BEATs`，让其专注于空间感知和空间结构建模。
+
+## 2. 任务定义
+
+### 2.1 核心任务
+
+`Spatial-BEATs` 的核心任务不是通用音频语义分类，而是：
+
+1. 从 `FOA` 输入中提取与空间位置相关的结构化表示
+2. 在多源场景中输出 `source-level spatial tokens`
+3. 每个 token 尽量对应一个潜在声源，编码其空间信息
+4. 后续由 LLM 使用这些 token 完成空间关系理解和推理
+
+### 2.2 输入与输出
+
+输入：
+
+- 原始 `FOA waveform`
+- 或由 `FOA waveform` 计算得到的 `FOA 特征图`
+
+推荐特征：
+
+- `W, X, Y, Z` 的 log-mel
+- `IV (Intensity Vector)`，例如 `IVx, IVy, IVz`
+- 可选的 `diffuseness / coherence / phase-related` 特征
+
+输出：
+
+- `K` 个 `Spatial Tokens`
+- 每个 token 对应一个潜在声源或一个空间实体
+- 每个 token 供下游预测：
+  - `objectness`
+  - `azimuth`
+  - `elevation`
+  - `distance`
+  - 可选的 `class embedding / source type embedding`
+
+### 2.3 为什么不是只用 W
+
+只让 `W` 通道经过 backbone，本质上更像单通道语义编码，空间线索主要被放到外挂 adapter 中。  
+这不符合本项目目标，因为这里希望：
+
+- 整个 `FOA` 特征都经过主干
+- 主干本身学习空间结构
+- `Spatial-BEATs` 成为一个真正的空间 encoder，而不是一个“语义 encoder + 小空间补丁”
+
+因此，本项目的推荐路线是：
+
+- `整套 FOA 特征 -> patch embedding -> BEATs backbone -> source-level spatial tokens`
+
+而不是：
+
+- `W-only -> BEATs`
+- `W-only BEATs + 外挂小 adapter`
+
+## 3. 与原始 BEATs 的关系
+
+### 3.1 BEATs 中值得最大化复用的部分
+
+当前仓库中的 BEATs 主干主要包括：
+
+- `post_extract_proj`
+- `TransformerEncoder`
+- `Transformer layers`
+- `conv_pos`
+- `LayerNorm / FFN / attention`
+
+这些模块位于：
+
+- `BEATs.py`
+- `backbone.py`
+
+这些部分是最应该保留并加载预训练权重的。
+
+### 3.2 BEATs 中不适合直接保留的部分
+
+原始 BEATs 代码是单通道设计，关键假设包括：
+
+- `preprocess()` 只生成单通道 `fbank`
+- `patch_embedding` 是 `Conv2d(1, embed_dim, ...)`
+- 下游输出默认是整段时间序列平均后的分类预测
+
+因此，下列部分不应直接照搬：
+
+1. 单通道 `preprocess`
+2. 单通道 `patch_embedding`
+3. 最终的 clip-level 平均池化分类输出方式
+4. 原始 `predictor` 作为最终目标头
+
+### 3.3 对 BEATs 的总体改造原则
+
+原则是：
+
+- **尽量保留 trunk**
+- **必要时重做 stem**
+- **完全重做 spatial head**
+
+也就是：
+
+- `输入端` 改
+- `输出端` 改
+- `中间主干` 尽量不改
+
+## 4. Spatial-AST 相比 AudioMAE 的改造经验
+
+Spatial-AST 对本项目最有借鉴价值的不是其 binaural 细节，而是其改造模式。
+
+### 4.1 Spatial-AST 做了什么
+
+相对于原始 AudioMAE/ViT，Spatial-AST 主要改了四类模块：
+
+1. **输入前端**
+   - 从原始单通道 spectrogram 输入，改成 `双耳 log-mel + IPD`
+   - 在输入前端加入 `STFT / LogMel / IPD / conv_downsample`
+
+2. **token 设计**
+   - 把原来的单个 `cls token` 改为 `3 个任务专用 token`
+   - 分别对应：
+     - 分类
+     - 距离
+     - 方向
+
+3. **输出头**
+   - 除原分类 head 外，新增：
+     - `distance_head`
+     - `azimuth_head`
+     - `elevation_head`
+
+4. **训练目标**
+   - 从单任务分类，改成多任务训练
+   - 同时训练：
+     - sound event detection
+     - distance prediction
+     - direction prediction
+
+### 4.2 Spatial-AST 没有怎么改
+
+Spatial-AST 没有重写 Transformer block 本体。  
+它保留了 AudioMAE/ViT 的核心 encoder 结构，而把改动集中在：
+
+- front-end
+- tokens
+- heads
+- objectives
+
+### 4.3 对本项目的可迁移结论
+
+可直接借鉴的思想：
+
+1. 用预训练音频主干初始化空间 encoder
+2. 重新设计输入前端，让空间特征真正进入 backbone
+3. 使用专门的空间 token，而不是只做全局池化
+4. 使用多任务监督训练空间能力
+
+不能直接照搬的部分：
+
+1. Spatial-AST 的 `binaural + IPD` 前端
+2. 只面向单/双耳的空间 cue 设计
+3. 只输出全局 token 的思路
+
+本项目是 `FOA`，因此应该把输入前端换成 `FOA 专属空间特征`。
+
+## 5. Spatial-BEATs 的推荐任务形式
+
+### 5.1 单场景 token 不够
+
+如果只输出一个全局 spatial token，它只能表示整个 scene 的压缩摘要，不适合做：
+
+- 多声源关系理解
+- “谁在谁左边”
+- “某个类对应的声源在什么位置”
+- source-wise grounding
+
+既然数据中已经有多源标注，推荐直接把任务定义为：
+
+- `multi-source set prediction`
+
+### 5.2 推荐输出形式
+
+令模型输出固定数量的 `K` 个 spatial queries/tokens。
+
+每个 token 预测：
+
+- `p(obj)`
+- `azimuth`
+- `elevation`
+- `distance`
+- 可选 `class logits` 或 `class embedding`
+
+训练时使用：
+
+- `Hungarian matching`
+- 或者其他 set prediction matching
+
+把 `K` 个预测 token 与当前样本中的 `N` 个 GT 声源做一一匹配。
+
+这会比单一 scene token 更适合后续接入 LLM 做空间关系推理。
+
+## 6. Spatial-BEATs 的推荐模型结构
+
+推荐结构如下：
+
+```text
+FOA waveform
+  -> FOA front-end
+  -> FOA feature map
+  -> FOA patch embedding
+  -> BEATs Transformer trunk
+  -> source queries / spatial decoder
+  -> K spatial tokens
+  -> spatial heads
+```
+
+### 6.1 FOA front-end
+
+输入可以是：
+
+- `WXYZ log-mel`
+- `WXYZ log-mel + IVx/IVy/IVz`
+
+推荐第一版就至少使用：
+
+- `WXYZ`
+- `IV`
+
+原因：
+
+- 只用 `WXYZ` 仍然需要模型自己从通道关系中恢复空间线索
+- `IV` 直接提供有物理意义的方向信息
+- 对空间收敛会更稳
+
+### 6.2 FOA patch embedding
+
+这一层应替换原始 BEATs 的单通道 patch stem。
+
+原始：
+
+- `Conv2d(1, embed_dim, kernel_size=patch, stride=patch)`
+
+新的思路：
+
+- `Conv2d(C_foa, embed_dim, kernel_size=patch, stride=patch)`
+
+其中 `C_foa` 可以是：
+
+- `4`，如果只用 `WXYZ`
+- `7`，如果用 `WXYZ + IVxyz`
+- 更大，如果加入更多派生空间特征
+
+### 6.3 BEATs trunk
+
+尽量保留以下模块：
+
+- `post_extract_proj`
+- `TransformerEncoder`
+- `attention`
+- `FFN`
+- `conv_pos`
+- `LayerNorm`
+
+这是整个“最大化复用预训练权重”的核心。
+
+### 6.4 Spatial token 模块
+
+不要继续使用原始 BEATs 的：
+
+- mean pooling
+- clip-level predictor
+
+推荐改为：
+
+- `K` 个 learnable source queries
+- queries 对 trunk 输出做 attention
+- 得到 `K` 个 source-level spatial tokens
+
+如果实现上希望更简单，第一版也可以：
+
+- 先直接在 trunk 输出后接一个轻量 decoder
+- 再输出 `K` 个 tokens
+
+### 6.5 Spatial heads
+
+每个 token 对应：
+
+- `objectness head`
+- `azimuth head`
+- `elevation head`
+- `distance head`
+- 可选 `class head`
+
+如果担心与原 LLM audio encoder 产生语义冲突，则建议：
+
+- 把 `class head` 只作为辅助监督
+- 不把其输出作为最终送入 LLM 的主要表示
+
+## 7. 保留的部分
+
+下面这些建议尽量保留：
+
+### 7.1 保留原有 LLM audio encoder
+
+原始语义 audio encoder 不动，继续负责：
+
+- 音频内容语义
+- 事件类别理解
+- 与现有 LLM 接口保持兼容
+
+### 7.2 保留 Spatial-BEATs 作为独立 encoder
+
+`Spatial-BEATs` 单独负责：
+
+- 方向
+- 距离
+- 多源空间结构
+- 可选 source-wise 辅助类别信息
+
+### 7.3 保留 BEATs 主干参数初始化
+
+建议保留：
+
+- trunk 的预训练参数
+- 尽量避免从零训练整个 spatial encoder
+
+## 8. 需要替换或新增的部分
+
+### 8.1 必改模块
+
+必须修改：
+
+1. `preprocess`
+2. `patch_embedding`
+3. `forward / extract_features` 的输出方式
+4. 下游 `predictor`
+
+### 8.2 必增模块
+
+必须新增：
+
+1. `FOA spatial front-end`
+2. `spatial query / token module`
+3. `multi-head spatial prediction heads`
+4. `set matching / multi-source loss`
+
+### 8.3 建议新增模块
+
+建议新增：
+
+1. `source confidence / objectness`
+2. `auxiliary class supervision`
+3. `LLM projection head`
+
+## 9. 训练方法
+
+## 9.1 第一阶段是否需要 SSL
+
+当前结论是：
+
+- **第一版不需要重新做 BEATs 式 SSL**
+
+原因：
+
+1. 已经有多源 `GT relative positions`
+2. 目标不是再学通用音频语义，而是让模型学空间结构
+3. 已有 BEATs 预训练权重可作为稳定初始化
+4. 先做监督式空间学习，工程收益最高
+
+因此，推荐第一阶段直接做 `supervised multi-task training`。
+
+### 9.2 第一阶段训练目标
+
+基础目标：
+
+- `L_obj`
+- `L_azimuth`
+- `L_elevation`
+- `L_distance`
+
+可选目标：
+
+- `L_class_aux`
+
+总损失可写为：
+
+```text
+L = lambda_obj * L_obj
+  + lambda_azi * L_azimuth
+  + lambda_ele * L_elevation
+  + lambda_dist * L_distance
+  + lambda_cls * L_class_aux
+```
+
+这里建议：
+
+- 空间任务作为主目标
+- 类别只做辅助目标
+
+因为本项目中语义主责已经由原始 audio encoder 承担。
+
+### 9.3 多源匹配训练
+
+如果每个样本有多个声源标注，推荐：
+
+1. 模型输出固定数量 `K` 个 tokens
+2. 用 Hungarian matching 在 token 与 GT 源之间做匹配
+3. 对 matched token 计算位置损失
+4. 对 unmatched token 计算 no-object loss
+
+这是比“对所有 token 平均做 scene 监督”更合适的做法。
+
+### 9.4 训练阶段建议
+
+推荐三步走：
+
+#### Stage A: Stem + Head Warmup
+
+- 冻结大部分 BEATs trunk
+- 只训练：
+  - FOA front-end
+  - FOA patch embedding
+  - spatial token/query module
+  - spatial heads
+
+目的：
+
+- 让新输入 stem 和新 heads 先适配预训练 trunk
+
+#### Stage B: Upper Trunk Finetune
+
+- 解冻 BEATs 上层若干层
+- 使用较小学习率微调
+- 使用 layer-wise lr decay
+
+目的：
+
+- 让 trunk 逐步适配 FOA 分布和空间任务
+
+#### Stage C: Full or Near-Full Finetune
+
+- 在稳定后解冻更多层
+- 继续以空间目标微调
+
+目的：
+
+- 提升空间 token 的表达能力
+
+### 9.5 训练数据组织
+
+每个样本应包含：
+
+- `foa waveform`
+- `num_sources`
+- `per-source azimuth`
+- `per-source elevation`
+- `per-source distance`
+- 可选 `per-source class`
+
+推荐统一成：
+
+```text
+sample = {
+  "audio": ...,
+  "sources": [
+    {"azimuth": ..., "elevation": ..., "distance": ..., "label": ...},
+    ...
+  ]
+}
+```
+
+## 10. 与 LLM 的接口
+
+### 10.1 推荐输入形式
+
+最终不要把 Spatial-BEATs 的全部 dense patch tokens 都喂给 LLM。  
+推荐只输出：
+
+- `K` 个 `Spatial Tokens`
+
+每个 token 代表一个潜在空间实体。
+
+### 10.2 避免语义冲突的策略
+
+本项目中“避免语义冲突”的关键不是完全不学类别，而是：
+
+1. 原始 audio encoder 继续承担主要语义理解
+2. Spatial-BEATs 主要承担空间结构建模
+3. Spatial-BEATs 输出给 LLM 的是 `source-level spatial tokens`
+4. 辅助类别监督只用于训练，不一定直接暴露为最终模态表示
+
+这样两套 encoder 的职责边界更清晰：
+
+- 原语义 encoder：`what`
+- Spatial-BEATs：`where / relation / spatial structure`
+
+## 11. 当前推荐方案总结
+
+当前最推荐的路线不是：
+
+- `W-only BEATs`
+- `W-only + adapter`
+
+而是：
+
+- `FOA full-feature Spatial-BEATs`
+- `独立空间 encoder`
+- `最大化复用 BEATs trunk`
+- `重做输入 stem`
+- `重做多源 spatial token heads`
+
+用一句话总结就是：
+
+> 用 BEATs 的主干做 FOA 空间建模，而不是只拿 BEATs 当单通道语义骨干，再在旁边打一个空间补丁。
+
+## 12. 推荐实施顺序
+
+建议按以下顺序推进：
+
+1. 明确 `FOA feature schema`
+   - 是否使用 `WXYZ`
+   - 是否加入 `IV`
+   - 是否加入其他物理特征
+
+2. 设计 `Spatial-BEATs` 的新输入 stem
+   - 替换单通道 preprocess
+   - 替换 patch embedding
+
+3. 设计 `K-source spatial tokens`
+   - 确定 token 数量
+   - 确定 query 机制
+
+4. 实现多头空间预测
+   - objectness
+   - distance
+   - azimuth
+   - elevation
+   - optional class
+
+5. 实现多源匹配训练
+   - Hungarian matching
+   - no-object loss
+
+6. 先做监督训练
+   - 不急于加入 SSL
+
+7. 训练稳定后，再评估是否需要第二阶段自监督
+   - masked spatial cue prediction
+   - view consistency
+   - teacher distillation
+
+## 13. 第二阶段可选方向
+
+在第一版监督训练稳定后，可考虑加入：
+
+1. `Spatial SSL`
+   - masked FOA cue prediction
+   - spatial consistency loss
+   - teacher-student distillation
+
+2. `source-conditioned tokens`
+   - 类别条件的 source token
+
+3. `LLM-side alignment`
+   - 将 spatial token 投影到 LLM hidden space
+   - 与文本空间词汇进行弱对齐
+
+4. `更复杂的多源推理训练`
+   - source relation supervision
+   - pairwise spatial relation labels
+
+## 14. 结论
+
+本项目的推荐方向已经比较明确：
+
+- 使用 `FOA 全特征`
+- 让 `FOA 特征` 真正经过 `BEATs backbone`
+- 将 `BEATs` 改造成独立的 `Spatial-BEATs`
+- 保留原语义 audio encoder 不动
+- 以 `多源 spatial token prediction` 为核心任务
+- 第一阶段采用 `监督式空间训练`
+- 最大化复用 `BEATs trunk` 的预训练权重
+
+如果后续开始实现，建议优先落地以下最小可行版本：
+
+1. `WXYZ + IV` 输入
+2. 多通道 patch embedding
+3. BEATs trunk 复用
+4. `K` 个 spatial queries
+5. `objectness + azimuth + elevation + distance` 多头监督
+
+这会比任何 `W-only` 或“仅外挂 adapter”的方案更符合项目目标，也更适合最终作为 LLM 的空间模态输入。

docs/spatial_beats_token_interface_note.md

@@ -0,0 +1,176 @@
+# Spatial-BEATs Token Interface Clarification
+
+## 1. Why `25 * 4 = 100` Is Not the Right Final Token Count
+
+The previous discussion mixed two different concepts:
+
+- internal multi-source slot capacity
+- final LLM-visible token rate
+
+These should be separated.
+
+For the current design:
+
+- `2.5 Hz` means the **final spatial token rate visible to the LLM**
+- for a `10 s` clip, this means:
+  - `T_s = 10 * 2.5 = 25`
+
+So the correct final token count is:
+
+- `25 spatial tokens`
+
+not:
+
+- `25 * 4 = 100`
+
+The `4` only refers to:
+
+- internal source slots per time step
+
+It is an internal modeling capacity, not an external token-rate multiplier.
+
+## 2. Corrected Design
+
+The corrected interface is:
+
+```text
+FOA waveform
+  -> FOA features
+  -> BEATs trunk
+  -> temporal memory at 2.5 Hz              [B, T_s, D]
+  -> per-step source slots (K=4)            [B, T_s, K, D]
+  -> objectness-weighted slot pooling       [B, T_s, D]
+  -> MLP projector
+  -> final LLM spatial tokens               [B, T_s, d_llm]
+```
+
+With the default setup:
+
+- `T_s = 25`
+- `K = 4`
+
+So:
+
+- internal representation: `[B, 25, 4, D]`
+- final LLM tokens: `[B, 25, d_llm]`
+
+## 3. What `objectness-weighted pooling + MLP projector` Means
+
+At each time step `t`, the model first predicts `K=4` source slots:
+
+- `z_{t,1}, z_{t,2}, z_{t,3}, z_{t,4}`
+
+Each slot also has an objectness score:
+
+- `o_{t,1}, o_{t,2}, o_{t,3}, o_{t,4}`
+
+These objectness scores are normalized across the `K` slots:
+
+```text
+alpha_{t,k} = softmax(o_{t,:})_k
+```
+
+Then the slot latents are pooled:
+
+```text
+h_t = sum_{k=1..K} alpha_{t,k} * z_{t,k}
+```
+
+This produces one pooled latent for this time step:
+
+- `h_t`
+
+The same idea is used to pool the structured slot-level predictions:
+
+```text
+c_t = sum_k alpha_{t,k} * c_{t,k}
+u_t = sum_k alpha_{t,k} * u_{t,k}
+d_t = sum_k alpha_{t,k} * d_{t,k}
+o_t = sum_k alpha_{t,k} * e_{obj,t,k}
+```
+
+where:
+
+- `c_{t,k}` is the slot-level class-context embedding
+- `u_{t,k}` is the slot-level direction embedding/vector
+- `d_{t,k}` is the slot-level distance embedding
+- `e_{obj,t,k}` is the slot-level confidence embedding
+
+Then the final per-step spatial token is formed as:
+
+```text
+s_t = Proj([h_t ; c_t ; u_t ; d_t ; o_t])
+```
+
+where:
+
+- `Proj` is an MLP projector into the LLM hidden space
+
+So the final sequence is:
+
+```text
+S = [s_1, s_2, ..., s_{T_s}]
+```
+
+For a `10 s` clip:
+
+- `S` has `25` tokens
+
+## 4. Why This Is Better
+
+This corrected design keeps both goals:
+
+1. multi-source capacity inside the model
+2. fixed low-rate spatial tokens outside the model
+
+Advantages:
+
+- the model can still represent up to `4` sources at each time step
+- the final LLM token count stays fixed at `2.5 Hz`
+- the external token interface is simpler and easier to scale
+- it avoids unnecessarily inflating the LLM token count by `K`
+
+## 5. Corrected Tensor Shapes
+
+Recommended tensor shapes:
+
+- `temporal_memory`: `[B, T_s, D]`
+- `slot_tokens`: `[B, T_s, K, D]`
+- `pred_obj`: `[B, T_s, K]`
+- `pred_azi_logits`: `[B, T_s, K, 360]`
+- `pred_ele_logits`: `[B, T_s, K, 180]`
+- `pred_dist`: `[B, T_s, K, 1]`
+- `pred_class_logits`: `[B, T_s, K, C_cls]`
+- `pooled_spatial_latents`: `[B, T_s, D]`
+- `llm_spatial_tokens`: `[B, T_s, d_llm]`
+
+For the default setup:
+
+- `T_s = 25`
+- `K = 4`
+
+Therefore:
+
+- internal slots: `[B, 25, 4, D]`
+- final LLM tokens: `[B, 25, d_llm]`
+
+## 6. What Should Be Updated in the Main Design
+
+The main design should be interpreted as:
+
+- internal `4` slots
+- external fixed `25` spatial tokens for a `10 s` clip
+
+So any previous statement implying:
+
+- `2.5 Hz * 4 = 10 tokens / second`
+
+should be considered obsolete for the final LLM interface.
+
+The correct statement is:
+
+- final LLM-visible spatial tokens are `2.5 tokens / second`
+
+and:
+
+- `K=4` is only internal source-slot capacity.

docs/v13d_full_hyperparameters.md

@@ -0,0 +1,230 @@
+# Spatial-BEATs v13d 完整超参数与实现细节附录
+
+> 适用于 NeurIPS 论文附录。本附录详尽地列出 v13d 模型的全部架构超参数、训练超参数、损失函数权重、数据预处理参数以及优化器配置。所有数值均与 `train_spatial_beats.py::make_ov1_unified_v13d_config()` 以及 `run_ov1_unified_v13d.sh` 中的代码一致。
+
+---
+
+## A. 输入与特征提取
+
+| 参数 | 取值 | 说明 |
+|---|---|---|
+| 采样率 | 16 kHz | FOA 4 通道，顺序 [W, X, Y, Z] |
+| 单 clip 时长 | 10 s | 输入波形形状 [B, 4, 160000] |
+| STFT n_fft | 400 | Qwen-2.5-Omni 对齐 |
+| STFT hop_length | 160 | 时间步长 10 ms |
+| STFT win_length | 400 | 窗长 25 ms |
+| 窗函数 | Hann | — |
+| Mel 滤波器组数 | 128 | f_min=0, f_max=8000 |
+| 时间帧数 T_f | 1000 | 10 s × 100 帧/s |
+| 输入特征通道数 | 7 | 4 个 mel (W/X/Y/Z) + 3 个 IV (x/y/z) |
+| IV 公式 | `IV_d = Re[W · conj(X_d)] / (\|W\|² + ε)` | ε=1e-8，IV 经 mel 投影后 clamp 到 ±10 |
+| W 通道归一化 | mean=15.41663, std=6.55582 | BEATs 预训练统计量 |
+| SpecAugment（仅 W 通道） | 2 个时间 mask × 100 帧, 2 个频率 mask × 27 bin | 训练时启用 |
+
+---
+
+## B. 模型架构超参数
+
+### B.1 SpatialDeltaPatchAdapter (v1)
+| 参数 | 取值 |
+|---|---|
+| 输入通道数 | 7 |
+| 隐藏通道数 | 32 |
+| 输出维度 | 512（patch embedding 维度） |
+| Patch size | (16, 16), stride=16 |
+| 残差缩放 α 初始化 | 0.1（可学习） |
+| 结构 | Conv2d(7→32, 1×1) → GELU → DWConv(32, 3×3) → GELU → Conv2d(32→512, 16×16, s=16) |
+
+### B.2 SpatialPatchEmbedding（继承 BEATs）
+- 单通道（W）patch embedding，预训练权重不修改
+- 输出 token 数 = 496（10 s clip）
+- Hidden = 512，再投影至 768
+
+### B.3 BEATs Transformer Trunk
+| 参数 | 取值 |
+|---|---|
+| Layer 数 | 12 |
+| Hidden 维度 | 768 |
+| 注意力头数 | 12 |
+| FFN 维度 | 3072 |
+| 相对位置偏置 | sinusoidal + GRU gating |
+| Trunk adapter | 1 层 spectral demixer，零门控初始化（继承 v11a 的 `use_spatial_head_demixer=True`） |
+
+### B.4 LocalSpatialEncoder（并行空间分支）
+| 参数 | 取值 |
+|---|---|
+| 输入 | 7 通道 FOA 特征 [B, 7, T_f, 128] |
+| CNN block 1 | Conv2d(7→64, 3×3) + GroupNorm(8) + GELU |
+| CNN block 2 | Conv2d(64→128, 3×3, stride=(1,2)) + GroupNorm(8) + GELU |
+| CNN block 3 | Conv2d(128→256, 3×3, stride=(1,2)) + GroupNorm(16) + GELU |
+| 频率维度处理 | 在最终 GN 后对频率轴做 mean → [B, T_f, 256] |
+| Transformer 层数 | 2 |
+| Transformer hidden | 256 |
+| Transformer heads | 4 |
+| Norm 顺序 | norm_first = True (pre-LN) |
+| Dropout | 0.1 |
+| 输出投影 | Linear(256 → 768) |
+
+### B.5 FrequencyPool + TemporalResampler
+- FrequencyPool：reshape [B, 496, 768] → [B, 62, 8, 768]，频率轴均值 → [B, 62, 768]
+- TemporalResampler：线性插值到 10 Hz 网格 → [B, T_s=100, 768]
+- **Token 频率 = 10 Hz**（继承自 v9_real_balanced_10hz）
+
+### B.6 LocalSpatialCrossFuser（语义-空间融合）
+| 参数 | 取值 |
+|---|---|
+| 模式 | `cross_attn_gated` |
+| 层数 | 2 |
+| Embed 维度 | 768 |
+| 注意力头数 | 8 |
+| Gate bias | -2.0（即 sigmoid(-2.0)≈0.119 初始化） |
+| Direct gate bias | -1.5（sigmoid≈0.182） |
+| ShallowTemporalReadout | 1 层 Transformer + LayerNorm |
+| 输出 | fused_tokens [B, T_s=100, 768] |
+
+### B.7 SourceQueryDecoder（多源解耦）
+| 参数 | 取值 |
+|---|---|
+| Track query 数 K | 4 |
+| Stage-1 层数 | 2（TransformerDecoder） |
+| Stage-2 层数 | 1（per-frame refinement + LN） |
+| 注意力头数 | 8 |
+| FFN 维度 | 3072 |
+| 时间位置编码 | 可学习 [T_s, 768] |
+| 输出 | [B, K=4, T_s=100, 768] |
+
+### B.8 FrameTrackPredictionHeads（每个 (track, frame) 4 个预测头）
+| Head | 结构 | 输出 |
+|---|---|---|
+| Activity | LayerNorm + Linear(768→1) | logit ℓ ∈ ℝ |
+| Class | MLP + 残差 + spectral demixer | 63 类 logits |
+| Direction | MLP(768→768→3) + L2 normalize | 单位向量 ∈ ℝ³ |
+| Distance | MLP(768→768→1) + softplus | 距离（米） |
+
+---
+
+## C. 损失函数与权重
+
+### C.1 损失项与权重
+| 损失项 | 权重 | 备注 |
+|---|---|---|
+| `lambda_frame_class` | 1.0 | 63 类 cross-entropy |
+| `lambda_frame_activity` | 1.0 | **Top-K rank loss**（v13d 核心改动） |
+| `lambda_frame_direction` | 1.0 | 1 - cos(pred, gt) |
+| `lambda_frame_distance` | 1.0 | smooth-L1 |
+| `lambda_frame_hemisphere` | 1.0 | 半球 BCE（继承 v11a） |
+
+### C.2 Top-K Rank Activity Loss（D-2）
+$$\mathcal{L}_{\text{rank}} = \frac{1}{|P|}\sum_{(i,j)\in P}\max(0, m + \ell_j - \ell_i),\quad \mathcal{L}_{\text{act}} = \mathcal{L}_{\text{rank}} + 0.1 \cdot \mathcal{L}_{\text{BCE}}$$
+
+| 超参数 | 取值 |
+|---|---|
+| `frame_activity_loss_type` | `topk_rank` |
+| margin m | 2.0 |
+| BCE anchor 权重 | 0.1 |
+
+### C.3 Spatial loss warmup / ramp（D-1）
+| 阶段 | Epoch 范围 | 空间 loss 权重 |
+|---|---|---|
+| cls-only warmup | 0 – 7（共 8 ep） | 0 |
+| linear ramp | 8 – 9（共 2 ep） | 0 → 1 |
+| full joint training | 10 – 24 | 1 |
+
+对应 cfg：`frame_spatial_loss_warmup_epochs=8`, `frame_spatial_loss_ramp_epochs=2`.
+
+---
+
+## D. 训练超参数
+
+### D.1 优化器
+| 参数 | 取值 |
+|---|---|
+| Optimizer | AdamW |
+| β₁, β₂ | 0.9, 0.999 |
+| ε | 1e-8 |
+| Weight decay | 0.01 |
+| Gradient clipping | 1.0（global L2 norm） |
+| Resume optimizer state | True（D-5：从 v12 best.pt 继承 Adam momentum） |
+
+### D.2 学习率（Cosine schedule，D-1）
+| 参数 | 取值 |
+|---|---|
+| Peak LR | 1.5e-5 |
+| Linear warmup epochs | 3（LR 从 0 → peak） |
+| Cosine decay epochs | 22（peak → peak × min_ratio） |
+| Min LR ratio | 0.05（最低 LR = 7.5e-7） |
+| `use_cosine_lr` | True |
+
+### D.3 训练规模
+| 参数 | 取值 |
+|---|---|
+| 总 epoch 数 | 25 |
+| GPUs | 8 × A100 |
+| 单 GPU batch size | 8 |
+| 等效 batch size | 64 |
+| 数据并行 | torchrun + DDP |
+| 精度 | fp32 |
+| Num workers | 8 / GPU |
+| Hot-start checkpoint | v12 best.pt（strict=False，missing=0/unexpected=0） |
+
+### D.4 EMA shadow weights（D-6）
+| 参数 | 取值 |
+|---|---|
+| `use_ema` | True |
+| EMA decay | 0.9995 |
+| 启动 epoch | 3（避开 LR warmup 噪声） |
+| 应用范围 | 验证、保存 best.pt 时使用 EMA 权重；训练 forward/backward 仍用原权重 |
+| 实现方式 | swap → evaluate → restore（不污染训练梯度） |
+
+---
+
+## E. 数据集与采样
+
+| 参数 | 取值 |
+|---|---|
+| 训练 manifest | `unified_spatial_foa_fsd63_all/train.jsonl` |
+| 训练样本总数 | 约 329 K |
+| - sim_static | 304 K |
+| - dcase_real | 20 K |
+| - qa_sim | 74 K |
+| Manifest replication | (1,)（v13d 不做真实数据加权） |
+| 词表 | FSD50K 衍生 63 类（`final_vocabulary.csv`） |
+| 验证集 | ov1/ov2/ov3 sim + ov1/ov2/ov3 real + dcase_starss_valid + unified_valid（约 35 K） |
+| 数据增广 | 仅 W 通道 SpecAugment；不开启 v13b 的 random gain / channel dropout / lowpass |
+
+---
+
+## F. Hungarian 匹配与推理
+
+| 参数 | 取值 |
+|---|---|
+| 匹配粒度 | 段级（segment-level，相同 active set 窗口内稳定分配） |
+| 匹配代价 | activity + class CE + direction cosine + distance L1 加权和 |
+| 推理时活跃 track 选择 | top-K̂（DCASE SELD evaluator 统一标准），与训练 Top-K rank loss 对齐 |
+
+---
+
+## G. 实际训练曲线（参考）
+
+| Epoch | F20 | oracle_cls | azi MAE |
+|---|---|---|---|
+| 0 | 0.311 | 0.650 | 28.6° |
+| 7（cls warmup 末） | 0.193 | 0.786 | 31.0° |
+| 8（spatial 启动） | 0.397 | 0.876 | 18.5° |
+| 10（当前最佳） | 0.402 | 0.864 | 17.2° |
+| 25（预期） | 0.43 ~ 0.46 | ~0.88 | 17~19° |
+
+ep1→ep7 期间 F20 下降是 **预期行为**：cls warmup 中 trunk 逐步适配类别学习，但空间梯度被 mask 为 0，方向头无监督信号导致 azi 漂移。ep8 空间 loss 解锁后 F20 单 epoch 跃升 +107%（0.193 → 0.397），证明 D-1 ~ D-6 的训练机制改造工作正常。
+
+---
+
+## H. 复现命令
+
+```bash
+# 默认 8 GPU、bs=8/GPU、peak LR=1.5e-5、25 epochs
+GPUS=8 BATCH_SIZE=8 SPATIAL_EPOCHS=25 SPATIAL_LR=1.5e-5 \
+  RESUME_CKPT=checkpoints/spatial_beats_ov1_unified_v12_exp/03_ov123_top4/best.pt \
+  ./run_ov1_unified_v13d.sh
+```
+
+所有改动通过 cfg flag 控制，默认 False，因此 v12 / v13b / v13c 实验不受影响。

eval_voxaudio_ood_results/dacvae/per_sample.json

@@ -0,0 +1,2096 @@
+{
+  "s1_rank0_opensource_emilia_zh_rawdata_English_003050": {
+    "gt_top_class": "speech",
+    "rc_top_class": "speech",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.573,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.45,
+          "azi_deg": -42.47,
+          "ele_deg": 43.03,
+          "dist_m": 1.291
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.779,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.503,
+          "azi_deg": -42.38,
+          "ele_deg": 42.74,
+          "dist_m": 1.258
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.784,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.523,
+          "azi_deg": -42.73,
+          "ele_deg": 42.45,
+          "dist_m": 1.252
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.771,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.549,
+          "azi_deg": -43.1,
+          "ele_deg": 42.33,
+          "dist_m": 1.256
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.756,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.568,
+          "azi_deg": -43.15,
+          "ele_deg": 42.18,
+          "dist_m": 1.258
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.623,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.521,
+          "azi_deg": 34.03,
+          "ele_deg": 37.12,
+          "dist_m": 1.241
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.817,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.592,
+          "azi_deg": 32.3,
+          "ele_deg": 37.82,
+          "dist_m": 1.175
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.826,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.616,
+          "azi_deg": 33.55,
+          "ele_deg": 37.57,
+          "dist_m": 1.164
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.816,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.64,
+          "azi_deg": 34.84,
+          "ele_deg": 37.71,
+          "dist_m": 1.173
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.805,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.66,
+          "azi_deg": 35.57,
+          "ele_deg": 37.47,
+          "dist_m": 1.179
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 56,
+      "activity_gt_frac": 0.14732142857142858,
+      "activity_rc_frac": 0.10267857142857142,
+      "activity_jaccard": 0.696969696969697,
+      "activity_precision_rc_vs_gt": 1.0,
+      "activity_recall_rc_vs_gt": 0.696969696969697,
+      "activity_f1_rc_vs_gt": 0.8214285714285715,
+      "class_match_rate": 1.0,
+      "doa_angular_error_deg_mean": 59.86446613244225,
+      "doa_angular_error_deg_median": 60.656924189839735,
+      "distance_mae_m": 0.056358031928539276,
+      "top_class_agreement": 1,
+      "gt_top_class": "speech",
+      "rc_top_class": "speech"
+    },
+    "sample": "s1_rank0_opensource_emilia_zh_rawdata_English_003050"
+  },
+  "s1_rank10_fsd50k_000123": {
+    "gt_top_class": "insect",
+    "rc_top_class": "insect",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.919,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.876,
+          "azi_deg": 175.31,
+          "ele_deg": -4.01,
+          "dist_m": 3.163
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.942,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.87,
+          "azi_deg": 175.91,
+          "ele_deg": -3.56,
+          "dist_m": 3.156
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.948,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.869,
+          "azi_deg": 176.06,
+          "ele_deg": -3.29,
+          "dist_m": 3.154
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.949,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.871,
+          "azi_deg": 176.09,
+          "ele_deg": -2.86,
+          "dist_m": 3.15
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.95,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.87,
+          "azi_deg": 176.3,
+          "ele_deg": -2.76,
+          "dist_m": 3.163
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.774,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.819,
+          "azi_deg": -9.79,
+          "ele_deg": -0.5,
+          "dist_m": 3.845
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.814,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.809,
+          "azi_deg": -15.55,
+          "ele_deg": 1.56,
+          "dist_m": 3.876
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.831,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.806,
+          "azi_deg": -18.77,
+          "ele_deg": 0.67,
+          "dist_m": 3.847
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.841,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.81,
+          "azi_deg": -18.84,
+          "ele_deg": 0.81,
+          "dist_m": 3.83
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.849,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.807,
+          "azi_deg": -19.82,
+          "ele_deg": 0.35,
+          "dist_m": 3.857
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 59,
+      "activity_gt_frac": 0.25,
+      "activity_rc_frac": 0.25,
+      "activity_jaccard": 1.0,
+      "activity_precision_rc_vs_gt": 1.0,
+      "activity_recall_rc_vs_gt": 1.0,
+      "activity_f1_rc_vs_gt": 1.0,
+      "class_match_rate": 1.0,
+      "doa_angular_error_deg_mean": 156.0560686538037,
+      "doa_angular_error_deg_median": 155.71115908405488,
+      "distance_mae_m": 0.6611307263374329,
+      "top_class_agreement": 1,
+      "gt_top_class": "insect",
+      "rc_top_class": "insect"
+    },
+    "sample": "s1_rank10_fsd50k_000123"
+  },
+  "s1_rank10_insectset_000529": {
+    "gt_top_class": "metal_clink",
+    "rc_top_class": "insect",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.598,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.279,
+          "azi_deg": -80.4,
+          "ele_deg": 23.47,
+          "dist_m": 1.418
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.598,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.28,
+          "azi_deg": -80.17,
+          "ele_deg": 23.68,
+          "dist_m": 1.402
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.6,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.28,
+          "azi_deg": -80.23,
+          "ele_deg": 23.71,
+          "dist_m": 1.402
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.618,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.285,
+          "azi_deg": -80.66,
+          "ele_deg": 23.91,
+          "dist_m": 1.4
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.639,
+          "class_idx": 12,
+          "class_name": "metal_clink",
+          "class_conf": 0.283,
+          "azi_deg": -81.02,
+          "ele_deg": 24.07,
+          "dist_m": 1.401
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.507,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.46,
+          "azi_deg": 111.58,
+          "ele_deg": 20.34,
+          "dist_m": 1.732
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.507,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.46,
+          "azi_deg": 111.26,
+          "ele_deg": 20.38,
+          "dist_m": 1.732
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.508,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.46,
+          "azi_deg": 111.3,
+          "ele_deg": 20.37,
+          "dist_m": 1.731
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.516,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.465,
+          "azi_deg": 111.77,
+          "ele_deg": 20.48,
+          "dist_m": 1.726
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.53,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.464,
+          "azi_deg": 112.0,
+          "ele_deg": 20.43,
+          "dist_m": 1.717
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 303,
+      "activity_gt_frac": 0.18976897689768976,
+      "activity_rc_frac": 0.08745874587458746,
+      "activity_jaccard": 0.4608695652173913,
+      "activity_precision_rc_vs_gt": 1.0,
+      "activity_recall_rc_vs_gt": 0.4608695652173913,
+      "activity_f1_rc_vs_gt": 0.6309523809523809,
+      "class_match_rate": 0.05660377358490566,
+      "doa_angular_error_deg_mean": 136.71160597941594,
+      "doa_angular_error_deg_median": 137.4092032601551,
+      "distance_mae_m": 0.27179428935050964,
+      "top_class_agreement": 0,
+      "gt_top_class": "metal_clink",
+      "rc_top_class": "insect"
+    },
+    "sample": "s1_rank10_insectset_000529"
+  },
+  "s1_rank10_laion_630k_002480": {
+    "gt_top_class": "keyboard_instrument",
+    "rc_top_class": "keyboard_instrument",
+    "gt_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "rc_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "comparison": {
+      "T_s": 300,
+      "activity_gt_frac": 0.23416666666666666,
+      "activity_rc_frac": 0.0725,
+      "activity_jaccard": 0.3096085409252669,
+      "activity_precision_rc_vs_gt": 1.0,
+      "activity_recall_rc_vs_gt": 0.3096085409252669,
+      "activity_f1_rc_vs_gt": 0.47282608695652173,
+      "class_match_rate": 1.0,
+      "doa_angular_error_deg_mean": 138.44985927817157,
+      "doa_angular_error_deg_median": 140.29241128549629,
+      "distance_mae_m": 0.7546095848083496,
+      "top_class_agreement": 1,
+      "gt_top_class": "keyboard_instrument",
+      "rc_top_class": "keyboard_instrument"
+    },
+    "sample": "s1_rank10_laion_630k_002480"
+  },
+  "s1_rank10_temporal_order_001229": {
+    "gt_top_class": "human_vocalization",
+    "rc_top_class": "human_vocalization",
+    "gt_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      [
+        {
+          "track": 0,
+          "activity": 0.529,
+          "class_idx": 6,
+          "class_name": "human_vocalization",
+          "class_conf": 0.807,
+          "azi_deg": 90.28,
+          "ele_deg": 20.51,
+          "dist_m": 1.839
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.689,
+          "class_idx": 6,
+          "class_name": "human_vocalization",
+          "class_conf": 0.348,
+          "azi_deg": 89.27,
+          "ele_deg": 13.42,
+          "dist_m": 2.255
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.843,
+          "class_idx": 6,
+          "class_name": "human_vocalization",
+          "class_conf": 0.539,
+          "azi_deg": 89.49,
+          "ele_deg": 13.74,
+          "dist_m": 2.25
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.865,
+          "class_idx": 6,
+          "class_name": "human_vocalization",
+          "class_conf": 0.548,
+          "azi_deg": 89.13,
+          "ele_deg": 14.13,
+          "dist_m": 2.258
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.879,
+          "class_idx": 6,
+          "class_name": "human_vocalization",
+          "class_conf": 0.569,
+          "azi_deg": 89.23,
+          "ele_deg": 14.32,
+          "dist_m": 2.269
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.895,
+          "class_idx": 6,
+          "class_name": "human_vocalization",
+          "class_conf": 0.573,
+          "azi_deg": 89.42,
+          "ele_deg": 14.41,
+          "dist_m": 2.277
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 29,
+      "activity_gt_frac": 0.15517241379310345,
+      "activity_rc_frac": 0.20689655172413793,
+      "activity_jaccard": 0.75,
+      "activity_precision_rc_vs_gt": 0.75,
+      "activity_recall_rc_vs_gt": 1.0,
+      "activity_f1_rc_vs_gt": 0.8571428571428571,
+      "class_match_rate": 1.0,
+      "doa_angular_error_deg_mean": 6.153626667733724,
+      "doa_angular_error_deg_median": 6.147921418235693,
+      "distance_mae_m": 0.45835474133491516,
+      "top_class_agreement": 1,
+      "gt_top_class": "human_vocalization",
+      "rc_top_class": "human_vocalization"
+    },
+    "sample": "s1_rank10_temporal_order_001229"
+  },
+  "s1_rank14_audioset_keyword_000653": {
+    "gt_top_class": "human_vocalization",
+    "rc_top_class": "singing",
+    "gt_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 2,
+          "activity": 0.697,
+          "class_idx": 41,
+          "class_name": "singing",
+          "class_conf": 0.405,
+          "azi_deg": 123.85,
+          "ele_deg": -8.07,
+          "dist_m": 2.299
+        }
+      ],
+      [
+        {
+          "track": 2,
+          "activity": 0.722,
+          "class_idx": 41,
+          "class_name": "singing",
+          "class_conf": 0.546,
+          "azi_deg": 126.55,
+          "ele_deg": -7.12,
+          "dist_m": 2.28
+        }
+      ],
+      [
+        {
+          "track": 2,
+          "activity": 0.768,
+          "class_idx": 41,
+          "class_name": "singing",
+          "class_conf": 0.54,
+          "azi_deg": 131.08,
+          "ele_deg": -6.14,
+          "dist_m": 2.281
+        }
+      ],
+      [
+        {
+          "track": 2,
+          "activity": 0.797,
+          "class_idx": 41,
+          "class_name": "singing",
+          "class_conf": 0.549,
+          "azi_deg": 132.51,
+          "ele_deg": -5.47,
+          "dist_m": 2.275
+        }
+      ],
+      [
+        {
+          "track": 2,
+          "activity": 0.824,
+          "class_idx": 41,
+          "class_name": "singing",
+          "class_conf": 0.572,
+          "azi_deg": 133.35,
+          "ele_deg": -5.2,
+          "dist_m": 2.276
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 103,
+      "activity_gt_frac": 0.2087378640776699,
+      "activity_rc_frac": 0.24029126213592233,
+      "activity_jaccard": 0.8686868686868687,
+      "activity_precision_rc_vs_gt": 0.8686868686868687,
+      "activity_recall_rc_vs_gt": 1.0,
+      "activity_f1_rc_vs_gt": 0.9297297297297298,
+      "class_match_rate": 0.0,
+      "doa_angular_error_deg_mean": 6.5160168715399385,
+      "doa_angular_error_deg_median": 6.497438991141886,
+      "distance_mae_m": 0.2764527499675751,
+      "top_class_agreement": 0,
+      "gt_top_class": "human_vocalization",
+      "rc_top_class": "singing"
+    },
+    "sample": "s1_rank14_audioset_keyword_000653"
+  },
+  "s1_rank14_fsd50k_000024": {
+    "gt_top_class": "keyboard_instrument",
+    "rc_top_class": "musical_instrument",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.741,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.85,
+          "azi_deg": 106.65,
+          "ele_deg": 15.69,
+          "dist_m": 1.944
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.833,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.853,
+          "azi_deg": 108.18,
+          "ele_deg": 16.01,
+          "dist_m": 1.926
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.83,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.849,
+          "azi_deg": 108.66,
+          "ele_deg": 15.48,
+          "dist_m": 1.927
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.837,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.851,
+          "azi_deg": 108.29,
+          "ele_deg": 15.24,
+          "dist_m": 1.927
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.844,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.849,
+          "azi_deg": 108.11,
+          "ele_deg": 14.97,
+          "dist_m": 1.931
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.894,
+          "class_idx": 28,
+          "class_name": "musical_instrument",
+          "class_conf": 0.458,
+          "azi_deg": 58.25,
+          "ele_deg": -1.86,
+          "dist_m": 1.559
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.909,
+          "class_idx": 28,
+          "class_name": "musical_instrument",
+          "class_conf": 0.479,
+          "azi_deg": 58.86,
+          "ele_deg": 4.65,
+          "dist_m": 1.595
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.886,
+          "class_idx": 28,
+          "class_name": "musical_instrument",
+          "class_conf": 0.491,
+          "azi_deg": 62.04,
+          "ele_deg": 2.91,
+          "dist_m": 1.601
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.883,
+          "class_idx": 28,
+          "class_name": "musical_instrument",
+          "class_conf": 0.491,
+          "azi_deg": 62.44,
+          "ele_deg": -0.01,
+          "dist_m": 1.588
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.877,
+          "class_idx": 28,
+          "class_name": "musical_instrument",
+          "class_conf": 0.5,
+          "azi_deg": 62.67,
+          "ele_deg": -1.44,
+          "dist_m": 1.594
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 24,
+      "activity_gt_frac": 0.25,
+      "activity_rc_frac": 0.25,
+      "activity_jaccard": 1.0,
+      "activity_precision_rc_vs_gt": 1.0,
+      "activity_recall_rc_vs_gt": 1.0,
+      "activity_f1_rc_vs_gt": 1.0,
+      "class_match_rate": 0.0,
+      "doa_angular_error_deg_mean": 48.053440319441144,
+      "doa_angular_error_deg_median": 47.67819369056817,
+      "distance_mae_m": 0.35104572772979736,
+      "top_class_agreement": 0,
+      "gt_top_class": "keyboard_instrument",
+      "rc_top_class": "musical_instrument"
+    },
+    "sample": "s1_rank14_fsd50k_000024"
+  },
+  "s1_rank14_temporal_order_003256": {
+    "gt_top_class": "keyboard_instrument",
+    "rc_top_class": "keyboard_instrument",
+    "gt_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "rc_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "comparison": {
+      "T_s": 13,
+      "activity_gt_frac": 0.0,
+      "activity_rc_frac": 0.0,
+      "activity_jaccard": 0.0,
+      "activity_precision_rc_vs_gt": 0.0,
+      "activity_recall_rc_vs_gt": 0.0,
+      "activity_f1_rc_vs_gt": 0.0,
+      "class_match_rate": NaN,
+      "doa_angular_error_deg_mean": NaN,
+      "doa_angular_error_deg_median": NaN,
+      "distance_mae_m": NaN,
+      "top_class_agreement": 1,
+      "gt_top_class": "keyboard_instrument",
+      "rc_top_class": "keyboard_instrument"
+    },
+    "sample": "s1_rank14_temporal_order_003256"
+  },
+  "s1_rank15_fsd50k_000489": {
+    "gt_top_class": "telephone_alarm",
+    "rc_top_class": "telephone_alarm",
+    "gt_frames_preview": [
+      [],
+      [
+        {
+          "track": 0,
+          "activity": 0.512,
+          "class_idx": 23,
+          "class_name": "telephone_alarm",
+          "class_conf": 0.862,
+          "azi_deg": 48.36,
+          "ele_deg": 28.03,
+          "dist_m": 1.952
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.702,
+          "class_idx": 23,
+          "class_name": "telephone_alarm",
+          "class_conf": 0.868,
+          "azi_deg": 48.63,
+          "ele_deg": 28.02,
+          "dist_m": 1.937
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.846,
+          "class_idx": 23,
+          "class_name": "telephone_alarm",
+          "class_conf": 0.871,
+          "azi_deg": 49.05,
+          "ele_deg": 28.02,
+          "dist_m": 1.925
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.875,
+          "class_idx": 23,
+          "class_name": "telephone_alarm",
+          "class_conf": 0.874,
+          "azi_deg": 49.06,
+          "ele_deg": 27.97,
+          "dist_m": 1.918
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "comparison": {
+      "T_s": 155,
+      "activity_gt_frac": 0.22258064516129034,
+      "activity_rc_frac": 0.21935483870967742,
+      "activity_jaccard": 0.9571428571428572,
+      "activity_precision_rc_vs_gt": 0.9852941176470589,
+      "activity_recall_rc_vs_gt": 0.9710144927536232,
+      "activity_f1_rc_vs_gt": 0.9781021897810219,
+      "class_match_rate": 1.0,
+      "doa_angular_error_deg_mean": 10.624633151373759,
+      "doa_angular_error_deg_median": 10.627009159697835,
+      "distance_mae_m": 0.258139431476593,
+      "top_class_agreement": 1,
+      "gt_top_class": "telephone_alarm",
+      "rc_top_class": "telephone_alarm"
+    },
+    "sample": "s1_rank15_fsd50k_000489"
+  },
+  "s1_rank17_audioset_keyword_001725": {
+    "gt_top_class": "guitar",
+    "rc_top_class": "guitar",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.978,
+          "class_idx": 2,
+          "class_name": "guitar",
+          "class_conf": 0.871,
+          "azi_deg": -6.63,
+          "ele_deg": 1.21,
+          "dist_m": 2.677
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.979,
+          "class_idx": 2,
+          "class_name": "guitar",
+          "class_conf": 0.871,
+          "azi_deg": -6.29,
+          "ele_deg": 0.95,
+          "dist_m": 2.66
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.982,
+          "class_idx": 2,
+          "class_name": "guitar",
+          "class_conf": 0.876,
+          "azi_deg": -6.28,
+          "ele_deg": 0.53,
+          "dist_m": 2.64
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.982,
+          "class_idx": 2,
+          "class_name": "guitar",
+          "class_conf": 0.874,
+          "azi_deg": -7.43,
+          "ele_deg": 0.61,
+          "dist_m": 2.621
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.983,
+          "class_idx": 2,
+          "class_name": "guitar",
+          "class_conf": 0.879,
+          "azi_deg": -8.48,
+          "ele_deg": 0.63,
+          "dist_m": 2.606
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.965,
+          "class_idx": 2,
+          "class_name": "guitar",
+          "class_conf": 0.785,
+          "azi_deg": 7.39,
+          "ele_deg": 5.51,
+          "dist_m": 2.516
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.963,
+          "class_idx": 2,
+          "class_name": "guitar",
+          "class_conf": 0.793,
+          "azi_deg": 8.29,
+          "ele_deg": 5.46,
+          "dist_m": 2.503
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.964,
+          "class_idx": 2,
+          "class_name": "guitar",
+          "class_conf": 0.809,
+          "azi_deg": 9.0,
+          "ele_deg": 5.33,
+          "dist_m": 2.483
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.966,
+          "class_idx": 2,
+          "class_name": "guitar",
+          "class_conf": 0.813,
+          "azi_deg": 8.13,
+          "ele_deg": 5.62,
+          "dist_m": 2.465
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.969,
+          "class_idx": 2,
+          "class_name": "guitar",
+          "class_conf": 0.826,
+          "azi_deg": 7.11,
+          "ele_deg": 5.77,
+          "dist_m": 2.453
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 104,
+      "activity_gt_frac": 0.24759615384615385,
+      "activity_rc_frac": 0.25,
+      "activity_jaccard": 0.9903846153846154,
+      "activity_precision_rc_vs_gt": 0.9903846153846154,
+      "activity_recall_rc_vs_gt": 1.0,
+      "activity_f1_rc_vs_gt": 0.9951690821256038,
+      "class_match_rate": 1.0,
+      "doa_angular_error_deg_mean": 13.613160991148467,
+      "doa_angular_error_deg_median": 13.479843693995544,
+      "distance_mae_m": 0.14917439222335815,
+      "top_class_agreement": 1,
+      "gt_top_class": "guitar",
+      "rc_top_class": "guitar"
+    },
+    "sample": "s1_rank17_audioset_keyword_001725"
+  },
+  "s1_rank18_opensource_emilia_zh_rawdata_English_000545": {
+    "gt_top_class": "speech",
+    "rc_top_class": "speech",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.616,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.859,
+          "azi_deg": 54.28,
+          "ele_deg": 32.99,
+          "dist_m": 1.793
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.644,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.864,
+          "azi_deg": 54.3,
+          "ele_deg": 32.91,
+          "dist_m": 1.788
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.729,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.87,
+          "azi_deg": 54.21,
+          "ele_deg": 32.44,
+          "dist_m": 1.792
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.809,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.876,
+          "azi_deg": 54.15,
+          "ele_deg": 32.04,
+          "dist_m": 1.808
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.824,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.877,
+          "azi_deg": 53.89,
+          "ele_deg": 31.84,
+          "dist_m": 1.814
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.577,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.868,
+          "azi_deg": 53.63,
+          "ele_deg": 32.63,
+          "dist_m": 1.75
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.606,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.874,
+          "azi_deg": 53.62,
+          "ele_deg": 32.51,
+          "dist_m": 1.746
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.696,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.878,
+          "azi_deg": 53.58,
+          "ele_deg": 32.12,
+          "dist_m": 1.745
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.784,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.881,
+          "azi_deg": 53.63,
+          "ele_deg": 31.87,
+          "dist_m": 1.748
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.801,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.882,
+          "azi_deg": 53.5,
+          "ele_deg": 31.77,
+          "dist_m": 1.749
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 147,
+      "activity_gt_frac": 0.12585034013605442,
+      "activity_rc_frac": 0.11394557823129252,
+      "activity_jaccard": 0.9054054054054054,
+      "activity_precision_rc_vs_gt": 1.0,
+      "activity_recall_rc_vs_gt": 0.9054054054054054,
+      "activity_f1_rc_vs_gt": 0.950354609929078,
+      "class_match_rate": 1.0,
+      "doa_angular_error_deg_mean": 0.7895429014465694,
+      "doa_angular_error_deg_median": 0.843487291291565,
+      "distance_mae_m": 0.07034817337989807,
+      "top_class_agreement": 1,
+      "gt_top_class": "speech",
+      "rc_top_class": "speech"
+    },
+    "sample": "s1_rank18_opensource_emilia_zh_rawdata_English_000545"
+  },
+  "s1_rank20_laion_630k_002181": {
+    "gt_top_class": "drum",
+    "rc_top_class": "drum",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.895,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.562,
+          "azi_deg": -117.81,
+          "ele_deg": 27.03,
+          "dist_m": 1.466
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.896,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.568,
+          "azi_deg": -118.0,
+          "ele_deg": 26.75,
+          "dist_m": 1.479
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.908,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.578,
+          "azi_deg": -118.28,
+          "ele_deg": 26.29,
+          "dist_m": 1.475
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.922,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.576,
+          "azi_deg": -118.63,
+          "ele_deg": 25.93,
+          "dist_m": 1.462
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.933,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.572,
+          "azi_deg": -118.84,
+          "ele_deg": 25.37,
+          "dist_m": 1.459
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.664,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.71,
+          "azi_deg": -115.78,
+          "ele_deg": 5.0,
+          "dist_m": 1.49
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.675,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.711,
+          "azi_deg": -115.98,
+          "ele_deg": 4.51,
+          "dist_m": 1.502
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.736,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.716,
+          "azi_deg": -116.66,
+          "ele_deg": 4.57,
+          "dist_m": 1.487
+        },
+        {
+          "track": 1,
+          "activity": 0.571,
+          "class_idx": 49,
+          "class_name": "paper",
+          "class_conf": 0.264,
+          "azi_deg": 43.47,
+          "ele_deg": 33.66,
+          "dist_m": 1.097
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.78,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.709,
+          "azi_deg": -117.42,
+          "ele_deg": 5.16,
+          "dist_m": 1.461
+        },
+        {
+          "track": 1,
+          "activity": 0.701,
+          "class_idx": 49,
+          "class_name": "paper",
+          "class_conf": 0.277,
+          "azi_deg": 43.15,
+          "ele_deg": 35.27,
+          "dist_m": 1.045
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.8,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.705,
+          "azi_deg": -117.79,
+          "ele_deg": 5.13,
+          "dist_m": 1.454
+        },
+        {
+          "track": 1,
+          "activity": 0.76,
+          "class_idx": 49,
+          "class_name": "paper",
+          "class_conf": 0.27,
+          "azi_deg": 42.93,
+          "ele_deg": 35.7,
+          "dist_m": 1.039
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 174,
+      "activity_gt_frac": 0.27873563218390807,
+      "activity_rc_frac": 0.3620689655172414,
+      "activity_jaccard": 0.6766917293233082,
+      "activity_precision_rc_vs_gt": 0.7142857142857143,
+      "activity_recall_rc_vs_gt": 0.9278350515463918,
+      "activity_f1_rc_vs_gt": 0.8071748878923768,
+      "class_match_rate": 0.5555555555555556,
+      "doa_angular_error_deg_mean": 20.903164025298974,
+      "doa_angular_error_deg_median": 21.441705235349964,
+      "distance_mae_m": 0.06737590581178665,
+      "top_class_agreement": 1,
+      "gt_top_class": "drum",
+      "rc_top_class": "drum"
+    },
+    "sample": "s1_rank20_laion_630k_002181"
+  },
+  "s1_rank20_vocalsound_000310": {
+    "gt_top_class": "laughter",
+    "rc_top_class": "laughter",
+    "gt_frames_preview": [
+      [],
+      [
+        {
+          "track": 0,
+          "activity": 0.514,
+          "class_idx": 14,
+          "class_name": "laughter",
+          "class_conf": 0.735,
+          "azi_deg": -54.88,
+          "ele_deg": 6.41,
+          "dist_m": 2.726
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.505,
+          "class_idx": 14,
+          "class_name": "laughter",
+          "class_conf": 0.779,
+          "azi_deg": -59.56,
+          "ele_deg": 7.71,
+          "dist_m": 2.714
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.532,
+          "class_idx": 14,
+          "class_name": "laughter",
+          "class_conf": 0.788,
+          "azi_deg": -63.85,
+          "ele_deg": 9.31,
+          "dist_m": 2.698
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.535,
+          "class_idx": 14,
+          "class_name": "laughter",
+          "class_conf": 0.81,
+          "azi_deg": -66.25,
+          "ele_deg": 10.24,
+          "dist_m": 2.685
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 1,
+          "activity": 0.702,
+          "class_idx": 9,
+          "class_name": "tool",
+          "class_conf": 0.658,
+          "azi_deg": -39.89,
+          "ele_deg": 8.84,
+          "dist_m": 1.331
+        }
+      ],
+      [
+        {
+          "track": 1,
+          "activity": 0.559,
+          "class_idx": 9,
+          "class_name": "tool",
+          "class_conf": 0.707,
+          "azi_deg": -43.23,
+          "ele_deg": 9.31,
+          "dist_m": 1.336
+        }
+      ],
+      [],
+      [],
+      []
+    ],
+    "comparison": {
+      "T_s": 53,
+      "activity_gt_frac": 0.37735849056603776,
+      "activity_rc_frac": 0.4009433962264151,
+      "activity_jaccard": 0.27906976744186046,
+      "activity_precision_rc_vs_gt": 0.4235294117647059,
+      "activity_recall_rc_vs_gt": 0.45,
+      "activity_f1_rc_vs_gt": 0.43636363636363634,
+      "class_match_rate": 0.9722222222222222,
+      "doa_angular_error_deg_mean": 11.065543586032767,
+      "doa_angular_error_deg_median": 6.664019736793584,
+      "distance_mae_m": 0.4840873181819916,
+      "top_class_agreement": 1,
+      "gt_top_class": "laughter",
+      "rc_top_class": "laughter"
+    },
+    "sample": "s1_rank20_vocalsound_000310"
+  },
+  "s1_rank21_insectset_000328": {
+    "gt_top_class": "water",
+    "rc_top_class": "water",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.578,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.565,
+          "azi_deg": 143.21,
+          "ele_deg": 5.5,
+          "dist_m": 2.413
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.578,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.542,
+          "azi_deg": 142.53,
+          "ele_deg": 5.38,
+          "dist_m": 2.417
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.581,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.542,
+          "azi_deg": 142.57,
+          "ele_deg": 5.38,
+          "dist_m": 2.416
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.604,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.534,
+          "azi_deg": 143.0,
+          "ele_deg": 5.6,
+          "dist_m": 2.409
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.629,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.533,
+          "azi_deg": 143.29,
+          "ele_deg": 5.62,
+          "dist_m": 2.402
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.7,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.529,
+          "azi_deg": 11.86,
+          "ele_deg": 9.44,
+          "dist_m": 1.891
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.7,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.535,
+          "azi_deg": 18.39,
+          "ele_deg": 5.93,
+          "dist_m": 1.854
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.702,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.535,
+          "azi_deg": 18.02,
+          "ele_deg": 6.05,
+          "dist_m": 1.852
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.719,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.529,
+          "azi_deg": 15.45,
+          "ele_deg": 9.66,
+          "dist_m": 1.843
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.739,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.526,
+          "azi_deg": 13.27,
+          "ele_deg": 10.83,
+          "dist_m": 1.831
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 303,
+      "activity_gt_frac": 0.2433993399339934,
+      "activity_rc_frac": 0.25,
+      "activity_jaccard": 0.9735973597359736,
+      "activity_precision_rc_vs_gt": 0.9735973597359736,
+      "activity_recall_rc_vs_gt": 1.0,
+      "activity_f1_rc_vs_gt": 0.9866220735785952,
+      "class_match_rate": 0.8033898305084746,
+      "doa_angular_error_deg_mean": 115.9973936791736,
+      "doa_angular_error_deg_median": 119.49698684207382,
+      "distance_mae_m": 0.6184173822402954,
+      "top_class_agreement": 1,
+      "gt_top_class": "water",
+      "rc_top_class": "water"
+    },
+    "sample": "s1_rank21_insectset_000328"
+  },
+  "s1_rank21_opensource_emilia_zh_rawdata_English_002504": {
+    "gt_top_class": "speech",
+    "rc_top_class": "speech",
+    "gt_frames_preview": [
+      [],
+      [],
+      [],
+      [
+        {
+          "track": 0,
+          "activity": 0.754,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.898,
+          "azi_deg": -177.3,
+          "ele_deg": 2.6,
+          "dist_m": 3.281
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.846,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.899,
+          "azi_deg": -176.82,
+          "ele_deg": 2.84,
+          "dist_m": 3.314
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [],
+      [],
+      [
+        {
+          "track": 2,
+          "activity": 0.51,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.884,
+          "azi_deg": 8.96,
+          "ele_deg": 13.06,
+          "dist_m": 5.013
+        }
+      ],
+      [
+        {
+          "track": 2,
+          "activity": 0.594,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.885,
+          "azi_deg": 7.9,
+          "ele_deg": 13.47,
+          "dist_m": 5.131
+        }
+      ],
+      [
+        {
+          "track": 2,
+          "activity": 0.621,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.886,
+          "azi_deg": 6.98,
+          "ele_deg": 13.78,
+          "dist_m": 5.132
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 166,
+      "activity_gt_frac": 0.23795180722891565,
+      "activity_rc_frac": 0.03765060240963856,
+      "activity_jaccard": 0.0,
+      "activity_precision_rc_vs_gt": 0.0,
+      "activity_recall_rc_vs_gt": 0.0,
+      "activity_f1_rc_vs_gt": 0.0,
+      "class_match_rate": NaN,
+      "doa_angular_error_deg_mean": NaN,
+      "doa_angular_error_deg_median": NaN,
+      "distance_mae_m": NaN,
+      "top_class_agreement": 1,
+      "gt_top_class": "speech",
+      "rc_top_class": "speech"
+    },
+    "sample": "s1_rank21_opensource_emilia_zh_rawdata_English_002504"
+  },
+  "s1_rank22_audiosl_000699": {
+    "gt_top_class": "water",
+    "rc_top_class": "metal_clink",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.552,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.635,
+          "azi_deg": -136.67,
+          "ele_deg": 6.51,
+          "dist_m": 1.162
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.667,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.481,
+          "azi_deg": -137.17,
+          "ele_deg": 6.84,
+          "dist_m": 1.152
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.676,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.503,
+          "azi_deg": -137.0,
+          "ele_deg": 7.0,
+          "dist_m": 1.152
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.686,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.485,
+          "azi_deg": -136.81,
+          "ele_deg": 7.29,
+          "dist_m": 1.165
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.701,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.477,
+          "azi_deg": -136.63,
+          "ele_deg": 7.32,
+          "dist_m": 1.168
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.522,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.351,
+          "azi_deg": -165.62,
+          "ele_deg": 2.62,
+          "dist_m": 1.905
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.532,
+          "class_idx": 12,
+          "class_name": "metal_clink",
+          "class_conf": 0.33,
+          "azi_deg": -172.48,
+          "ele_deg": 4.14,
+          "dist_m": 1.919
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.545,
+          "class_idx": 12,
+          "class_name": "metal_clink",
+          "class_conf": 0.335,
+          "azi_deg": -174.27,
+          "ele_deg": 4.09,
+          "dist_m": 1.946
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.567,
+          "class_idx": 12,
+          "class_name": "metal_clink",
+          "class_conf": 0.325,
+          "azi_deg": -176.62,
+          "ele_deg": 2.58,
+          "dist_m": 1.986
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.601,
+          "class_idx": 12,
+          "class_name": "metal_clink",
+          "class_conf": 0.318,
+          "azi_deg": -177.65,
+          "ele_deg": 3.23,
+          "dist_m": 1.989
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 18,
+      "activity_gt_frac": 0.25,
+      "activity_rc_frac": 0.25,
+      "activity_jaccard": 1.0,
+      "activity_precision_rc_vs_gt": 1.0,
+      "activity_recall_rc_vs_gt": 1.0,
+      "activity_f1_rc_vs_gt": 1.0,
+      "class_match_rate": 0.05555555555555555,
+      "doa_angular_error_deg_mean": 41.03007833276686,
+      "doa_angular_error_deg_median": 42.283789633071905,
+      "distance_mae_m": 0.8085865378379822,
+      "top_class_agreement": 0,
+      "gt_top_class": "water",
+      "rc_top_class": "metal_clink"
+    },
+    "sample": "s1_rank22_audiosl_000699"
+  },
+  "s1_rank23_vocalsound_000569": {
+    "gt_top_class": "breathing",
+    "rc_top_class": "breathing",
+    "gt_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "rc_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "comparison": {
+      "T_s": 43,
+      "activity_gt_frac": 0.18023255813953487,
+      "activity_rc_frac": 0.1744186046511628,
+      "activity_jaccard": 0.967741935483871,
+      "activity_precision_rc_vs_gt": 1.0,
+      "activity_recall_rc_vs_gt": 0.967741935483871,
+      "activity_f1_rc_vs_gt": 0.9836065573770492,
+      "class_match_rate": 1.0,
+      "doa_angular_error_deg_mean": 23.05399986579469,
+      "doa_angular_error_deg_median": 23.140233324503665,
+      "distance_mae_m": 0.6998323202133179,
+      "top_class_agreement": 1,
+      "gt_top_class": "breathing",
+      "rc_top_class": "breathing"
+    },
+    "sample": "s1_rank23_vocalsound_000569"
+  },
+  "s1_rank4_opensource_emilia_zh_rawdata_Chinese_003160": {
+    "gt_top_class": "speech",
+    "rc_top_class": "speech",
+    "gt_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "rc_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "comparison": {
+      "T_s": 134,
+      "activity_gt_frac": 0.055970149253731345,
+      "activity_rc_frac": 0.0037313432835820895,
+      "activity_jaccard": 0.0,
+      "activity_precision_rc_vs_gt": 0.0,
+      "activity_recall_rc_vs_gt": 0.0,
+      "activity_f1_rc_vs_gt": 0.0,
+      "class_match_rate": NaN,
+      "doa_angular_error_deg_mean": NaN,
+      "doa_angular_error_deg_median": NaN,
+      "distance_mae_m": NaN,
+      "top_class_agreement": 1,
+      "gt_top_class": "speech",
+      "rc_top_class": "speech"
+    },
+    "sample": "s1_rank4_opensource_emilia_zh_rawdata_Chinese_003160"
+  },
+  "s1_rank5_opensource_emilia_zh_rawdata_English_001914": {
+    "gt_top_class": "speech",
+    "rc_top_class": "speech",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.556,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.867,
+          "azi_deg": 14.18,
+          "ele_deg": 9.65,
+          "dist_m": 1.336
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.748,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.879,
+          "azi_deg": 14.36,
+          "ele_deg": 9.37,
+          "dist_m": 1.305
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.792,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.881,
+          "azi_deg": 15.14,
+          "ele_deg": 9.25,
+          "dist_m": 1.301
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.816,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.882,
+          "azi_deg": 15.62,
+          "ele_deg": 9.07,
+          "dist_m": 1.298
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.831,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.883,
+          "azi_deg": 15.93,
+          "ele_deg": 8.96,
+          "dist_m": 1.303
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "comparison": {
+      "T_s": 46,
+      "activity_gt_frac": 0.25,
+      "activity_rc_frac": 0.0,
+      "activity_jaccard": 0.0,
+      "activity_precision_rc_vs_gt": 0.0,
+      "activity_recall_rc_vs_gt": 0.0,
+      "activity_f1_rc_vs_gt": 0.0,
+      "class_match_rate": NaN,
+      "doa_angular_error_deg_mean": NaN,
+      "doa_angular_error_deg_median": NaN,
+      "distance_mae_m": NaN,
+      "top_class_agreement": 1,
+      "gt_top_class": "speech",
+      "rc_top_class": "speech"
+    },
+    "sample": "s1_rank5_opensource_emilia_zh_rawdata_English_001914"
+  },
+  "s1_rank9_opensource_emilia_zh_rawdata_English_001143": {
+    "gt_top_class": "speech",
+    "rc_top_class": "speech",
+    "gt_frames_preview": [
+      [],
+      [],
+      [
+        {
+          "track": 0,
+          "activity": 0.638,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.889,
+          "azi_deg": 28.91,
+          "ele_deg": -18.71,
+          "dist_m": 1.646
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.754,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.894,
+          "azi_deg": 29.12,
+          "ele_deg": -18.26,
+          "dist_m": 1.635
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.8,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.895,
+          "azi_deg": 29.48,
+          "ele_deg": -18.06,
+          "dist_m": 1.616
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "comparison": {
+      "T_s": 121,
+      "activity_gt_frac": 0.2190082644628099,
+      "activity_rc_frac": 0.04338842975206612,
+      "activity_jaccard": 0.0,
+      "activity_precision_rc_vs_gt": 0.0,
+      "activity_recall_rc_vs_gt": 0.0,
+      "activity_f1_rc_vs_gt": 0.0,
+      "class_match_rate": NaN,
+      "doa_angular_error_deg_mean": NaN,
+      "doa_angular_error_deg_median": NaN,
+      "distance_mae_m": NaN,
+      "top_class_agreement": 1,
+      "gt_top_class": "speech",
+      "rc_top_class": "speech"
+    },
+    "sample": "s1_rank9_opensource_emilia_zh_rawdata_English_001143"
+  }
+}

eval_voxaudio_ood_results/dacvae/summary.json

@@ -0,0 +1,25 @@
+{
+  "mean_activity_jaccard": 0.5918084170858557,
+  "n_valid_activity_jaccard": 20,
+  "mean_activity_precision_rc_vs_gt": 0.6852889043752468,
+  "n_valid_activity_precision_rc_vs_gt": 20,
+  "mean_activity_recall_rc_vs_gt": 0.6344722344150824,
+  "n_valid_activity_recall_rc_vs_gt": 20,
+  "mean_activity_f1_rc_vs_gt": 0.6424736331628711,
+  "n_valid_activity_f1_rc_vs_gt": 20,
+  "mean_class_match_rate": 0.6962217958284476,
+  "n_valid_class_match_rate": 15,
+  "mean_doa_angular_error_deg_mean": 52.59217336237226,
+  "n_valid_doa_angular_error_deg_mean": 15,
+  "mean_doa_angular_error_deg_median": 52.82468845575132,
+  "n_valid_doa_angular_error_deg_median": 15,
+  "mean_distance_mae_m": 0.3990471541881561,
+  "n_valid_distance_mae_m": 15,
+  "mean_top_class_agreement": 0.8,
+  "n_valid_top_class_agreement": 20,
+  "mean_activity_gt_frac": 0.20619253654594938,
+  "n_valid_activity_gt_frac": 20,
+  "mean_activity_rc_frac": 0.16576634449721478,
+  "n_valid_activity_rc_frac": 20,
+  "n_samples": 20
+}

eval_voxaudio_ood_results/foa_vae/summary.json

@@ -0,0 +1,25 @@
+{
+  "mean_activity_jaccard": 0.2869139591647025,
+  "n_valid_activity_jaccard": 70,
+  "mean_activity_precision_rc_vs_gt": 0.3842437781867924,
+  "n_valid_activity_precision_rc_vs_gt": 70,
+  "mean_activity_recall_rc_vs_gt": 0.3355311172735717,
+  "n_valid_activity_recall_rc_vs_gt": 70,
+  "mean_activity_f1_rc_vs_gt": 0.33179939006604536,
+  "n_valid_activity_f1_rc_vs_gt": 70,
+  "mean_class_match_rate": 0.7050047775153738,
+  "n_valid_class_match_rate": 39,
+  "mean_doa_angular_error_deg_mean": 47.81001474877972,
+  "n_valid_doa_angular_error_deg_mean": 39,
+  "mean_doa_angular_error_deg_median": 48.616234655597715,
+  "n_valid_doa_angular_error_deg_median": 39,
+  "mean_distance_mae_m": 0.4801142644137144,
+  "n_valid_distance_mae_m": 39,
+  "mean_top_class_agreement": 0.6857142857142857,
+  "n_valid_top_class_agreement": 70,
+  "mean_activity_gt_frac": 0.21816713600161056,
+  "n_valid_activity_gt_frac": 70,
+  "mean_activity_rc_frac": 0.15041837313585474,
+  "n_valid_activity_rc_frac": 70,
+  "n_samples": 70
+}

eval_voxaudio_ood_results/mono_vae/per_sample.json

@@ -0,0 +1,2216 @@
+{
+  "s1_rank0_opensource_emilia_zh_rawdata_English_003050": {
+    "gt_top_class": "speech",
+    "rc_top_class": "speech",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.564,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.458,
+          "azi_deg": -41.81,
+          "ele_deg": 40.52,
+          "dist_m": 1.332
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.776,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.512,
+          "azi_deg": -41.79,
+          "ele_deg": 40.34,
+          "dist_m": 1.3
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.783,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.532,
+          "azi_deg": -42.09,
+          "ele_deg": 40.04,
+          "dist_m": 1.293
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.769,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.558,
+          "azi_deg": -42.43,
+          "ele_deg": 39.88,
+          "dist_m": 1.298
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.754,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.576,
+          "azi_deg": -42.44,
+          "ele_deg": 39.7,
+          "dist_m": 1.299
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [],
+      [
+        {
+          "track": 0,
+          "activity": 0.623,
+          "class_idx": 32,
+          "class_name": "female_speech",
+          "class_conf": 0.588,
+          "azi_deg": 47.27,
+          "ele_deg": 20.26,
+          "dist_m": 1.751
+        },
+        {
+          "track": 1,
+          "activity": 0.602,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.756,
+          "azi_deg": 33.75,
+          "ele_deg": 34.18,
+          "dist_m": 1.073
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.622,
+          "class_idx": 32,
+          "class_name": "female_speech",
+          "class_conf": 0.578,
+          "azi_deg": 47.44,
+          "ele_deg": 22.41,
+          "dist_m": 1.726
+        },
+        {
+          "track": 1,
+          "activity": 0.629,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.768,
+          "azi_deg": 35.32,
+          "ele_deg": 36.23,
+          "dist_m": 1.076
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.61,
+          "class_idx": 32,
+          "class_name": "female_speech",
+          "class_conf": 0.557,
+          "azi_deg": 47.51,
+          "ele_deg": 23.64,
+          "dist_m": 1.757
+        },
+        {
+          "track": 1,
+          "activity": 0.62,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.779,
+          "azi_deg": 36.76,
+          "ele_deg": 37.92,
+          "dist_m": 1.1
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.594,
+          "class_idx": 32,
+          "class_name": "female_speech",
+          "class_conf": 0.534,
+          "azi_deg": 46.67,
+          "ele_deg": 24.12,
+          "dist_m": 1.771
+        },
+        {
+          "track": 1,
+          "activity": 0.615,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.789,
+          "azi_deg": 37.67,
+          "ele_deg": 38.75,
+          "dist_m": 1.105
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 56,
+      "activity_gt_frac": 0.14732142857142858,
+      "activity_rc_frac": 0.0625,
+      "activity_jaccard": 0.175,
+      "activity_precision_rc_vs_gt": 0.5,
+      "activity_recall_rc_vs_gt": 0.21212121212121213,
+      "activity_f1_rc_vs_gt": 0.2978723404255319,
+      "class_match_rate": 0.0,
+      "doa_angular_error_deg_mean": 74.55371361509434,
+      "doa_angular_error_deg_median": 74.22053995719097,
+      "distance_mae_m": 0.47506183385849,
+      "top_class_agreement": 1,
+      "gt_top_class": "speech",
+      "rc_top_class": "speech"
+    },
+    "sample": "s1_rank0_opensource_emilia_zh_rawdata_English_003050"
+  },
+  "s1_rank10_fsd50k_000123": {
+    "gt_top_class": "insect",
+    "rc_top_class": "insect",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.961,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.878,
+          "azi_deg": 67.07,
+          "ele_deg": -8.42,
+          "dist_m": 2.868
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.975,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.874,
+          "azi_deg": 101.76,
+          "ele_deg": -23.16,
+          "dist_m": 2.837
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.979,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.873,
+          "azi_deg": 128.2,
+          "ele_deg": -28.73,
+          "dist_m": 2.826
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.981,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.872,
+          "azi_deg": 136.34,
+          "ele_deg": -12.68,
+          "dist_m": 2.818
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.982,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.871,
+          "azi_deg": 146.37,
+          "ele_deg": -10.89,
+          "dist_m": 2.817
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "comparison": {
+      "T_s": 59,
+      "activity_gt_frac": 0.25,
+      "activity_rc_frac": 0.0,
+      "activity_jaccard": 0.0,
+      "activity_precision_rc_vs_gt": 0.0,
+      "activity_recall_rc_vs_gt": 0.0,
+      "activity_f1_rc_vs_gt": 0.0,
+      "class_match_rate": NaN,
+      "doa_angular_error_deg_mean": NaN,
+      "doa_angular_error_deg_median": NaN,
+      "distance_mae_m": NaN,
+      "top_class_agreement": 1,
+      "gt_top_class": "insect",
+      "rc_top_class": "insect"
+    },
+    "sample": "s1_rank10_fsd50k_000123"
+  },
+  "s1_rank10_insectset_000529": {
+    "gt_top_class": "metal_clink",
+    "rc_top_class": "tool",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.6,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.279,
+          "azi_deg": -78.82,
+          "ele_deg": 23.04,
+          "dist_m": 1.427
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.6,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.28,
+          "azi_deg": -78.61,
+          "ele_deg": 23.24,
+          "dist_m": 1.41
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.603,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.28,
+          "azi_deg": -78.66,
+          "ele_deg": 23.27,
+          "dist_m": 1.41
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.621,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.285,
+          "azi_deg": -79.05,
+          "ele_deg": 23.46,
+          "dist_m": 1.408
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.641,
+          "class_idx": 35,
+          "class_name": "insect",
+          "class_conf": 0.282,
+          "azi_deg": -79.36,
+          "ele_deg": 23.62,
+          "dist_m": 1.408
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "comparison": {
+      "T_s": 303,
+      "activity_gt_frac": 0.1905940594059406,
+      "activity_rc_frac": 0.0,
+      "activity_jaccard": 0.0,
+      "activity_precision_rc_vs_gt": 0.0,
+      "activity_recall_rc_vs_gt": 0.0,
+      "activity_f1_rc_vs_gt": 0.0,
+      "class_match_rate": NaN,
+      "doa_angular_error_deg_mean": NaN,
+      "doa_angular_error_deg_median": NaN,
+      "distance_mae_m": NaN,
+      "top_class_agreement": 0,
+      "gt_top_class": "metal_clink",
+      "rc_top_class": "tool"
+    },
+    "sample": "s1_rank10_insectset_000529"
+  },
+  "s1_rank10_laion_630k_002480": {
+    "gt_top_class": "keyboard_instrument",
+    "rc_top_class": "keyboard_instrument",
+    "gt_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "rc_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "comparison": {
+      "T_s": 300,
+      "activity_gt_frac": 0.23416666666666666,
+      "activity_rc_frac": 0.1825,
+      "activity_jaccard": 0.7793594306049823,
+      "activity_precision_rc_vs_gt": 1.0,
+      "activity_recall_rc_vs_gt": 0.7793594306049823,
+      "activity_f1_rc_vs_gt": 0.876,
+      "class_match_rate": 1.0,
+      "doa_angular_error_deg_mean": 131.0347148858392,
+      "doa_angular_error_deg_median": 126.96155147876038,
+      "distance_mae_m": 1.121482014656067,
+      "top_class_agreement": 1,
+      "gt_top_class": "keyboard_instrument",
+      "rc_top_class": "keyboard_instrument"
+    },
+    "sample": "s1_rank10_laion_630k_002480"
+  },
+  "s1_rank10_temporal_order_001229": {
+    "gt_top_class": "human_vocalization",
+    "rc_top_class": "human_vocalization",
+    "gt_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      [
+        {
+          "track": 0,
+          "activity": 0.532,
+          "class_idx": 6,
+          "class_name": "human_vocalization",
+          "class_conf": 0.808,
+          "azi_deg": 90.28,
+          "ele_deg": 20.53,
+          "dist_m": 1.838
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.694,
+          "class_idx": 6,
+          "class_name": "human_vocalization",
+          "class_conf": 0.778,
+          "azi_deg": 88.41,
+          "ele_deg": 8.83,
+          "dist_m": 1.051
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.873,
+          "class_idx": 6,
+          "class_name": "human_vocalization",
+          "class_conf": 0.745,
+          "azi_deg": 89.6,
+          "ele_deg": 8.43,
+          "dist_m": 1.028
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.898,
+          "class_idx": 6,
+          "class_name": "human_vocalization",
+          "class_conf": 0.742,
+          "azi_deg": 89.71,
+          "ele_deg": 8.38,
+          "dist_m": 1.011
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.913,
+          "class_idx": 6,
+          "class_name": "human_vocalization",
+          "class_conf": 0.734,
+          "azi_deg": 89.98,
+          "ele_deg": 8.42,
+          "dist_m": 0.994
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.93,
+          "class_idx": 6,
+          "class_name": "human_vocalization",
+          "class_conf": 0.725,
+          "azi_deg": 90.21,
+          "ele_deg": 8.42,
+          "dist_m": 0.997
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 29,
+      "activity_gt_frac": 0.15517241379310345,
+      "activity_rc_frac": 0.1896551724137931,
+      "activity_jaccard": 0.8181818181818182,
+      "activity_precision_rc_vs_gt": 0.8181818181818182,
+      "activity_recall_rc_vs_gt": 1.0,
+      "activity_f1_rc_vs_gt": 0.9,
+      "class_match_rate": 1.0,
+      "doa_angular_error_deg_mean": 11.89599920735741,
+      "doa_angular_error_deg_median": 11.875431948942609,
+      "distance_mae_m": 0.8219536542892456,
+      "top_class_agreement": 1,
+      "gt_top_class": "human_vocalization",
+      "rc_top_class": "human_vocalization"
+    },
+    "sample": "s1_rank10_temporal_order_001229"
+  },
+  "s1_rank14_audioset_keyword_000653": {
+    "gt_top_class": "human_vocalization",
+    "rc_top_class": "singing",
+    "gt_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "rc_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "comparison": {
+      "T_s": 103,
+      "activity_gt_frac": 0.2087378640776699,
+      "activity_rc_frac": 0.0825242718446602,
+      "activity_jaccard": 0.3953488372093023,
+      "activity_precision_rc_vs_gt": 1.0,
+      "activity_recall_rc_vs_gt": 0.3953488372093023,
+      "activity_f1_rc_vs_gt": 0.5666666666666667,
+      "class_match_rate": 0.0,
+      "doa_angular_error_deg_mean": 158.1389025850092,
+      "doa_angular_error_deg_median": 158.05901087896916,
+      "distance_mae_m": 0.2706669867038727,
+      "top_class_agreement": 0,
+      "gt_top_class": "human_vocalization",
+      "rc_top_class": "singing"
+    },
+    "sample": "s1_rank14_audioset_keyword_000653"
+  },
+  "s1_rank14_fsd50k_000024": {
+    "gt_top_class": "keyboard_instrument",
+    "rc_top_class": "water",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.742,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.85,
+          "azi_deg": 106.65,
+          "ele_deg": 15.68,
+          "dist_m": 1.945
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.833,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.853,
+          "azi_deg": 108.18,
+          "ele_deg": 15.99,
+          "dist_m": 1.927
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.83,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.849,
+          "azi_deg": 108.66,
+          "ele_deg": 15.47,
+          "dist_m": 1.928
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.837,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.851,
+          "azi_deg": 108.29,
+          "ele_deg": 15.23,
+          "dist_m": 1.928
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.844,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.849,
+          "azi_deg": 108.11,
+          "ele_deg": 14.96,
+          "dist_m": 1.932
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.939,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.87,
+          "azi_deg": 38.55,
+          "ele_deg": 1.47,
+          "dist_m": 1.649
+        },
+        {
+          "track": 1,
+          "activity": 0.685,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.291,
+          "azi_deg": 27.75,
+          "ele_deg": -31.03,
+          "dist_m": 1.963
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.916,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.893,
+          "azi_deg": 38.07,
+          "ele_deg": 2.53,
+          "dist_m": 1.633
+        },
+        {
+          "track": 1,
+          "activity": 0.705,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.362,
+          "azi_deg": -134.58,
+          "ele_deg": -44.36,
+          "dist_m": 1.892
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.851,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.905,
+          "azi_deg": 38.57,
+          "ele_deg": 3.16,
+          "dist_m": 1.633
+        },
+        {
+          "track": 1,
+          "activity": 0.674,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.341,
+          "azi_deg": -154.14,
+          "ele_deg": -17.61,
+          "dist_m": 1.9
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.814,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.912,
+          "azi_deg": 38.6,
+          "ele_deg": 3.49,
+          "dist_m": 1.649
+        },
+        {
+          "track": 1,
+          "activity": 0.686,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.343,
+          "azi_deg": -148.69,
+          "ele_deg": -17.26,
+          "dist_m": 1.946
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.773,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.913,
+          "azi_deg": 38.91,
+          "ele_deg": 4.05,
+          "dist_m": 1.654
+        },
+        {
+          "track": 1,
+          "activity": 0.686,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.333,
+          "azi_deg": -149.23,
+          "ele_deg": -19.44,
+          "dist_m": 1.959
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 24,
+      "activity_gt_frac": 0.25,
+      "activity_rc_frac": 0.375,
+      "activity_jaccard": 0.25,
+      "activity_precision_rc_vs_gt": 0.3333333333333333,
+      "activity_recall_rc_vs_gt": 0.5,
+      "activity_f1_rc_vs_gt": 0.4,
+      "class_match_rate": 1.0,
+      "doa_angular_error_deg_mean": 67.23609243670141,
+      "doa_angular_error_deg_median": 67.39671260216033,
+      "distance_mae_m": 0.2839207947254181,
+      "top_class_agreement": 0,
+      "gt_top_class": "keyboard_instrument",
+      "rc_top_class": "water"
+    },
+    "sample": "s1_rank14_fsd50k_000024"
+  },
+  "s1_rank14_temporal_order_003256": {
+    "gt_top_class": "musical_instrument",
+    "rc_top_class": "keyboard_instrument",
+    "gt_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.599,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.395,
+          "azi_deg": 75.42,
+          "ele_deg": 6.0,
+          "dist_m": 1.999
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.614,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.402,
+          "azi_deg": 77.8,
+          "ele_deg": 8.12,
+          "dist_m": 1.987
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.575,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.401,
+          "azi_deg": 78.54,
+          "ele_deg": 7.52,
+          "dist_m": 1.973
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.569,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.41,
+          "azi_deg": 78.66,
+          "ele_deg": 7.48,
+          "dist_m": 1.967
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.572,
+          "class_idx": 7,
+          "class_name": "keyboard_instrument",
+          "class_conf": 0.413,
+          "azi_deg": 78.57,
+          "ele_deg": 7.05,
+          "dist_m": 1.966
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 13,
+      "activity_gt_frac": 0.0,
+      "activity_rc_frac": 0.23076923076923078,
+      "activity_jaccard": 0.0,
+      "activity_precision_rc_vs_gt": 0.0,
+      "activity_recall_rc_vs_gt": 0.0,
+      "activity_f1_rc_vs_gt": 0.0,
+      "class_match_rate": NaN,
+      "doa_angular_error_deg_mean": NaN,
+      "doa_angular_error_deg_median": NaN,
+      "distance_mae_m": NaN,
+      "top_class_agreement": 0,
+      "gt_top_class": "musical_instrument",
+      "rc_top_class": "keyboard_instrument"
+    },
+    "sample": "s1_rank14_temporal_order_003256"
+  },
+  "s1_rank15_fsd50k_000489": {
+    "gt_top_class": "telephone_alarm",
+    "rc_top_class": "insect",
+    "gt_frames_preview": [
+      [],
+      [
+        {
+          "track": 0,
+          "activity": 0.514,
+          "class_idx": 23,
+          "class_name": "telephone_alarm",
+          "class_conf": 0.862,
+          "azi_deg": 48.29,
+          "ele_deg": 27.98,
+          "dist_m": 1.946
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.703,
+          "class_idx": 23,
+          "class_name": "telephone_alarm",
+          "class_conf": 0.868,
+          "azi_deg": 48.56,
+          "ele_deg": 27.97,
+          "dist_m": 1.932
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.847,
+          "class_idx": 23,
+          "class_name": "telephone_alarm",
+          "class_conf": 0.871,
+          "azi_deg": 48.98,
+          "ele_deg": 27.98,
+          "dist_m": 1.919
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.875,
+          "class_idx": 23,
+          "class_name": "telephone_alarm",
+          "class_conf": 0.874,
+          "azi_deg": 48.99,
+          "ele_deg": 27.92,
+          "dist_m": 1.913
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 2,
+          "activity": 0.539,
+          "class_idx": 52,
+          "class_name": "knock",
+          "class_conf": 0.502,
+          "azi_deg": -172.43,
+          "ele_deg": -25.56,
+          "dist_m": 2.656
+        }
+      ],
+      [
+        {
+          "track": 2,
+          "activity": 0.543,
+          "class_idx": 52,
+          "class_name": "knock",
+          "class_conf": 0.486,
+          "azi_deg": -168.97,
+          "ele_deg": -25.49,
+          "dist_m": 2.671
+        }
+      ],
+      [
+        {
+          "track": 2,
+          "activity": 0.547,
+          "class_idx": 52,
+          "class_name": "knock",
+          "class_conf": 0.486,
+          "azi_deg": -161.56,
+          "ele_deg": -22.67,
+          "dist_m": 2.711
+        }
+      ],
+      [
+        {
+          "track": 2,
+          "activity": 0.555,
+          "class_idx": 52,
+          "class_name": "knock",
+          "class_conf": 0.489,
+          "azi_deg": -157.76,
+          "ele_deg": -20.51,
+          "dist_m": 2.731
+        }
+      ],
+      [
+        {
+          "track": 2,
+          "activity": 0.526,
+          "class_idx": 52,
+          "class_name": "knock",
+          "class_conf": 0.488,
+          "azi_deg": -157.31,
+          "ele_deg": -19.91,
+          "dist_m": 2.718
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 155,
+      "activity_gt_frac": 0.22258064516129034,
+      "activity_rc_frac": 0.3790322580645161,
+      "activity_jaccard": 0.3814814814814815,
+      "activity_precision_rc_vs_gt": 0.43829787234042555,
+      "activity_recall_rc_vs_gt": 0.7463768115942029,
+      "activity_f1_rc_vs_gt": 0.5522788203753352,
+      "class_match_rate": 1.0,
+      "doa_angular_error_deg_mean": 90.65873059085607,
+      "doa_angular_error_deg_median": 91.4724381595999,
+      "distance_mae_m": 0.1063886359333992,
+      "top_class_agreement": 0,
+      "gt_top_class": "telephone_alarm",
+      "rc_top_class": "insect"
+    },
+    "sample": "s1_rank15_fsd50k_000489"
+  },
+  "s1_rank17_audioset_keyword_001725": {
+    "gt_top_class": "guitar",
+    "rc_top_class": "guitar",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.978,
+          "class_idx": 2,
+          "class_name": "guitar",
+          "class_conf": 0.869,
+          "azi_deg": -7.08,
+          "ele_deg": 1.27,
+          "dist_m": 2.662
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.979,
+          "class_idx": 2,
+          "class_name": "guitar",
+          "class_conf": 0.869,
+          "azi_deg": -6.71,
+          "ele_deg": 1.01,
+          "dist_m": 2.645
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.982,
+          "class_idx": 2,
+          "class_name": "guitar",
+          "class_conf": 0.874,
+          "azi_deg": -6.67,
+          "ele_deg": 0.57,
+          "dist_m": 2.623
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.982,
+          "class_idx": 2,
+          "class_name": "guitar",
+          "class_conf": 0.872,
+          "azi_deg": -7.83,
+          "ele_deg": 0.64,
+          "dist_m": 2.604
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.983,
+          "class_idx": 2,
+          "class_name": "guitar",
+          "class_conf": 0.878,
+          "azi_deg": -8.91,
+          "ele_deg": 0.66,
+          "dist_m": 2.589
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.774,
+          "class_idx": 41,
+          "class_name": "singing",
+          "class_conf": 0.348,
+          "azi_deg": 22.64,
+          "ele_deg": 7.04,
+          "dist_m": 2.347
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.73,
+          "class_idx": 41,
+          "class_name": "singing",
+          "class_conf": 0.333,
+          "azi_deg": 23.98,
+          "ele_deg": 7.07,
+          "dist_m": 2.347
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.634,
+          "class_idx": 2,
+          "class_name": "guitar",
+          "class_conf": 0.382,
+          "azi_deg": 25.59,
+          "ele_deg": 7.1,
+          "dist_m": 2.335
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.584,
+          "class_idx": 2,
+          "class_name": "guitar",
+          "class_conf": 0.437,
+          "azi_deg": 25.17,
+          "ele_deg": 7.23,
+          "dist_m": 2.317
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.567,
+          "class_idx": 2,
+          "class_name": "guitar",
+          "class_conf": 0.491,
+          "azi_deg": 24.76,
+          "ele_deg": 7.35,
+          "dist_m": 2.308
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 104,
+      "activity_gt_frac": 0.24759615384615385,
+      "activity_rc_frac": 0.03125,
+      "activity_jaccard": 0.1262135922330097,
+      "activity_precision_rc_vs_gt": 1.0,
+      "activity_recall_rc_vs_gt": 0.1262135922330097,
+      "activity_f1_rc_vs_gt": 0.22413793103448276,
+      "class_match_rate": 0.8461538461538461,
+      "doa_angular_error_deg_mean": 34.422865995490916,
+      "doa_angular_error_deg_median": 35.16672869426235,
+      "distance_mae_m": 0.25046706199645996,
+      "top_class_agreement": 1,
+      "gt_top_class": "guitar",
+      "rc_top_class": "guitar"
+    },
+    "sample": "s1_rank17_audioset_keyword_001725"
+  },
+  "s1_rank18_opensource_emilia_zh_rawdata_English_000545": {
+    "gt_top_class": "speech",
+    "rc_top_class": "speech",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.615,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.858,
+          "azi_deg": 54.36,
+          "ele_deg": 32.93,
+          "dist_m": 1.773
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.644,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.863,
+          "azi_deg": 54.38,
+          "ele_deg": 32.83,
+          "dist_m": 1.769
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.729,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.87,
+          "azi_deg": 54.29,
+          "ele_deg": 32.37,
+          "dist_m": 1.772
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.809,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.875,
+          "azi_deg": 54.23,
+          "ele_deg": 31.96,
+          "dist_m": 1.787
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.824,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.876,
+          "azi_deg": 53.96,
+          "ele_deg": 31.76,
+          "dist_m": 1.793
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [],
+      [],
+      [],
+      [
+        {
+          "track": 1,
+          "activity": 0.532,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.814,
+          "azi_deg": 69.6,
+          "ele_deg": 18.12,
+          "dist_m": 1.181
+        }
+      ],
+      [
+        {
+          "track": 1,
+          "activity": 0.543,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.822,
+          "azi_deg": 70.36,
+          "ele_deg": 17.33,
+          "dist_m": 1.174
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 147,
+      "activity_gt_frac": 0.12585034013605442,
+      "activity_rc_frac": 0.07993197278911565,
+      "activity_jaccard": 0.0,
+      "activity_precision_rc_vs_gt": 0.0,
+      "activity_recall_rc_vs_gt": 0.0,
+      "activity_f1_rc_vs_gt": 0.0,
+      "class_match_rate": NaN,
+      "doa_angular_error_deg_mean": NaN,
+      "doa_angular_error_deg_median": NaN,
+      "distance_mae_m": NaN,
+      "top_class_agreement": 1,
+      "gt_top_class": "speech",
+      "rc_top_class": "speech"
+    },
+    "sample": "s1_rank18_opensource_emilia_zh_rawdata_English_000545"
+  },
+  "s1_rank20_laion_630k_002181": {
+    "gt_top_class": "drum",
+    "rc_top_class": "drum",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.895,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.562,
+          "azi_deg": -117.83,
+          "ele_deg": 27.07,
+          "dist_m": 1.464
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.896,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.568,
+          "azi_deg": -118.02,
+          "ele_deg": 26.8,
+          "dist_m": 1.478
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.908,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.578,
+          "azi_deg": -118.29,
+          "ele_deg": 26.34,
+          "dist_m": 1.473
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.922,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.576,
+          "azi_deg": -118.65,
+          "ele_deg": 25.98,
+          "dist_m": 1.461
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.933,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.573,
+          "azi_deg": -118.86,
+          "ele_deg": 25.42,
+          "dist_m": 1.458
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.741,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.673,
+          "azi_deg": -95.62,
+          "ele_deg": -12.13,
+          "dist_m": 2.057
+        },
+        {
+          "track": 1,
+          "activity": 0.66,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.587,
+          "azi_deg": 17.81,
+          "ele_deg": 42.29,
+          "dist_m": 1.222
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.743,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.681,
+          "azi_deg": -95.68,
+          "ele_deg": -12.49,
+          "dist_m": 2.07
+        },
+        {
+          "track": 1,
+          "activity": 0.677,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.6,
+          "azi_deg": 20.36,
+          "ele_deg": 42.29,
+          "dist_m": 1.248
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.754,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.688,
+          "azi_deg": -97.51,
+          "ele_deg": -11.47,
+          "dist_m": 2.08
+        },
+        {
+          "track": 1,
+          "activity": 0.768,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.612,
+          "azi_deg": 31.56,
+          "ele_deg": 51.95,
+          "dist_m": 1.221
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.761,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.685,
+          "azi_deg": -99.17,
+          "ele_deg": -10.93,
+          "dist_m": 2.064
+        },
+        {
+          "track": 1,
+          "activity": 0.823,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.617,
+          "azi_deg": 41.36,
+          "ele_deg": 57.8,
+          "dist_m": 1.17
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.763,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.681,
+          "azi_deg": -99.82,
+          "ele_deg": -10.75,
+          "dist_m": 2.075
+        },
+        {
+          "track": 1,
+          "activity": 0.845,
+          "class_idx": 4,
+          "class_name": "drum",
+          "class_conf": 0.609,
+          "azi_deg": 47.64,
+          "ele_deg": 58.23,
+          "dist_m": 1.159
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 174,
+      "activity_gt_frac": 0.27873563218390807,
+      "activity_rc_frac": 0.257183908045977,
+      "activity_jaccard": 0.3970037453183521,
+      "activity_precision_rc_vs_gt": 0.5921787709497207,
+      "activity_recall_rc_vs_gt": 0.5463917525773195,
+      "activity_f1_rc_vs_gt": 0.5683646112600536,
+      "class_match_rate": 0.2358490566037736,
+      "doa_angular_error_deg_mean": 27.102211614582803,
+      "doa_angular_error_deg_median": 22.093031445824906,
+      "distance_mae_m": 0.2159106582403183,
+      "top_class_agreement": 1,
+      "gt_top_class": "drum",
+      "rc_top_class": "drum"
+    },
+    "sample": "s1_rank20_laion_630k_002181"
+  },
+  "s1_rank20_vocalsound_000310": {
+    "gt_top_class": "laughter",
+    "rc_top_class": "laughter",
+    "gt_frames_preview": [
+      [],
+      [
+        {
+          "track": 0,
+          "activity": 0.51,
+          "class_idx": 14,
+          "class_name": "laughter",
+          "class_conf": 0.73,
+          "azi_deg": -54.9,
+          "ele_deg": 6.32,
+          "dist_m": 2.723
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.5,
+          "class_idx": 14,
+          "class_name": "laughter",
+          "class_conf": 0.776,
+          "azi_deg": -59.68,
+          "ele_deg": 7.66,
+          "dist_m": 2.71
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.527,
+          "class_idx": 14,
+          "class_name": "laughter",
+          "class_conf": 0.786,
+          "azi_deg": -64.05,
+          "ele_deg": 9.3,
+          "dist_m": 2.694
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.531,
+          "class_idx": 14,
+          "class_name": "laughter",
+          "class_conf": 0.809,
+          "azi_deg": -66.48,
+          "ele_deg": 10.24,
+          "dist_m": 2.681
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 1,
+          "activity": 0.736,
+          "class_idx": 9,
+          "class_name": "tool",
+          "class_conf": 0.418,
+          "azi_deg": -20.49,
+          "ele_deg": 67.02,
+          "dist_m": 1.491
+        },
+        {
+          "track": 2,
+          "activity": 0.577,
+          "class_idx": 36,
+          "class_name": "typing",
+          "class_conf": 0.371,
+          "azi_deg": -30.13,
+          "ele_deg": 40.13,
+          "dist_m": 1.734
+        }
+      ],
+      [
+        {
+          "track": 3,
+          "activity": 0.719,
+          "class_idx": 14,
+          "class_name": "laughter",
+          "class_conf": 0.827,
+          "azi_deg": -54.38,
+          "ele_deg": 10.57,
+          "dist_m": 2.292
+        }
+      ],
+      [
+        {
+          "track": 3,
+          "activity": 0.831,
+          "class_idx": 14,
+          "class_name": "laughter",
+          "class_conf": 0.834,
+          "azi_deg": -47.86,
+          "ele_deg": 8.67,
+          "dist_m": 2.276
+        }
+      ],
+      [
+        {
+          "track": 3,
+          "activity": 0.895,
+          "class_idx": 14,
+          "class_name": "laughter",
+          "class_conf": 0.82,
+          "azi_deg": -41.0,
+          "ele_deg": 9.06,
+          "dist_m": 2.272
+        }
+      ],
+      [
+        {
+          "track": 3,
+          "activity": 0.919,
+          "class_idx": 14,
+          "class_name": "laughter",
+          "class_conf": 0.826,
+          "azi_deg": -29.83,
+          "ele_deg": 9.62,
+          "dist_m": 2.253
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 53,
+      "activity_gt_frac": 0.3867924528301887,
+      "activity_rc_frac": 0.42924528301886794,
+      "activity_jaccard": 0.2814814814814815,
+      "activity_precision_rc_vs_gt": 0.4175824175824176,
+      "activity_recall_rc_vs_gt": 0.4634146341463415,
+      "activity_f1_rc_vs_gt": 0.4393063583815029,
+      "class_match_rate": 1.0,
+      "doa_angular_error_deg_mean": 148.37263579773634,
+      "doa_angular_error_deg_median": 148.62451307406513,
+      "distance_mae_m": 0.994830310344696,
+      "top_class_agreement": 1,
+      "gt_top_class": "laughter",
+      "rc_top_class": "laughter"
+    },
+    "sample": "s1_rank20_vocalsound_000310"
+  },
+  "s1_rank21_insectset_000328": {
+    "gt_top_class": "water",
+    "rc_top_class": "tool",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.574,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.569,
+          "azi_deg": 138.37,
+          "ele_deg": 5.76,
+          "dist_m": 2.279
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.574,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.547,
+          "azi_deg": 137.61,
+          "ele_deg": 5.59,
+          "dist_m": 2.278
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.577,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.547,
+          "azi_deg": 137.66,
+          "ele_deg": 5.6,
+          "dist_m": 2.277
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.599,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.538,
+          "azi_deg": 138.21,
+          "ele_deg": 5.83,
+          "dist_m": 2.271
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.624,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.538,
+          "azi_deg": 138.61,
+          "ele_deg": 5.86,
+          "dist_m": 2.265
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "comparison": {
+      "T_s": 303,
+      "activity_gt_frac": 0.2433993399339934,
+      "activity_rc_frac": 0.0,
+      "activity_jaccard": 0.0,
+      "activity_precision_rc_vs_gt": 0.0,
+      "activity_recall_rc_vs_gt": 0.0,
+      "activity_f1_rc_vs_gt": 0.0,
+      "class_match_rate": NaN,
+      "doa_angular_error_deg_mean": NaN,
+      "doa_angular_error_deg_median": NaN,
+      "distance_mae_m": NaN,
+      "top_class_agreement": 0,
+      "gt_top_class": "water",
+      "rc_top_class": "tool"
+    },
+    "sample": "s1_rank21_insectset_000328"
+  },
+  "s1_rank21_opensource_emilia_zh_rawdata_English_002504": {
+    "gt_top_class": "speech",
+    "rc_top_class": "speech",
+    "gt_frames_preview": [
+      [],
+      [],
+      [
+        {
+          "track": 0,
+          "activity": 0.598,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.9,
+          "azi_deg": -176.91,
+          "ele_deg": 0.86,
+          "dist_m": 3.275
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.823,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.901,
+          "azi_deg": -175.78,
+          "ele_deg": 1.28,
+          "dist_m": 3.333
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.889,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.901,
+          "azi_deg": -175.28,
+          "ele_deg": 1.55,
+          "dist_m": 3.354
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [],
+      [],
+      [],
+      [
+        {
+          "track": 0,
+          "activity": 0.586,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.854,
+          "azi_deg": -78.27,
+          "ele_deg": 21.7,
+          "dist_m": 5.139
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.668,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.856,
+          "azi_deg": -76.07,
+          "ele_deg": 21.79,
+          "dist_m": 5.066
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 166,
+      "activity_gt_frac": 0.24096385542168675,
+      "activity_rc_frac": 0.1927710843373494,
+      "activity_jaccard": 0.8,
+      "activity_precision_rc_vs_gt": 1.0,
+      "activity_recall_rc_vs_gt": 0.8,
+      "activity_f1_rc_vs_gt": 0.888888888888889,
+      "class_match_rate": 1.0,
+      "doa_angular_error_deg_mean": 103.12276361635858,
+      "doa_angular_error_deg_median": 103.52123227313189,
+      "distance_mae_m": 1.1880497932434082,
+      "top_class_agreement": 1,
+      "gt_top_class": "speech",
+      "rc_top_class": "speech"
+    },
+    "sample": "s1_rank21_opensource_emilia_zh_rawdata_English_002504"
+  },
+  "s1_rank22_audiosl_000699": {
+    "gt_top_class": "water",
+    "rc_top_class": "war_sound",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.553,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.649,
+          "azi_deg": -136.51,
+          "ele_deg": 6.92,
+          "dist_m": 1.167
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.668,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.503,
+          "azi_deg": -137.02,
+          "ele_deg": 7.26,
+          "dist_m": 1.156
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.678,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.525,
+          "azi_deg": -136.84,
+          "ele_deg": 7.42,
+          "dist_m": 1.156
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.688,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.508,
+          "azi_deg": -136.64,
+          "ele_deg": 7.71,
+          "dist_m": 1.168
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.703,
+          "class_idx": 5,
+          "class_name": "water",
+          "class_conf": 0.5,
+          "azi_deg": -136.46,
+          "ele_deg": 7.75,
+          "dist_m": 1.171
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.569,
+          "class_idx": 11,
+          "class_name": "war_sound",
+          "class_conf": 0.517,
+          "azi_deg": 51.47,
+          "ele_deg": 10.91,
+          "dist_m": 2.405
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.514,
+          "class_idx": 11,
+          "class_name": "war_sound",
+          "class_conf": 0.491,
+          "azi_deg": 53.41,
+          "ele_deg": 11.56,
+          "dist_m": 2.371
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.514,
+          "class_idx": 11,
+          "class_name": "war_sound",
+          "class_conf": 0.505,
+          "azi_deg": 54.52,
+          "ele_deg": 11.57,
+          "dist_m": 2.384
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.534,
+          "class_idx": 11,
+          "class_name": "war_sound",
+          "class_conf": 0.462,
+          "azi_deg": 55.12,
+          "ele_deg": 11.62,
+          "dist_m": 2.408
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.563,
+          "class_idx": 11,
+          "class_name": "war_sound",
+          "class_conf": 0.488,
+          "azi_deg": 55.43,
+          "ele_deg": 11.45,
+          "dist_m": 2.38
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 18,
+      "activity_gt_frac": 0.25,
+      "activity_rc_frac": 0.25,
+      "activity_jaccard": 1.0,
+      "activity_precision_rc_vs_gt": 1.0,
+      "activity_recall_rc_vs_gt": 1.0,
+      "activity_f1_rc_vs_gt": 1.0,
+      "class_match_rate": 0.0,
+      "doa_angular_error_deg_mean": 158.0126151635061,
+      "doa_angular_error_deg_median": 157.80501566802064,
+      "distance_mae_m": 1.176613688468933,
+      "top_class_agreement": 0,
+      "gt_top_class": "water",
+      "rc_top_class": "war_sound"
+    },
+    "sample": "s1_rank22_audiosl_000699"
+  },
+  "s1_rank23_vocalsound_000569": {
+    "gt_top_class": "breathing",
+    "rc_top_class": "breathing",
+    "gt_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "rc_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "comparison": {
+      "T_s": 43,
+      "activity_gt_frac": 0.18023255813953487,
+      "activity_rc_frac": 0.18604651162790697,
+      "activity_jaccard": 0.6578947368421053,
+      "activity_precision_rc_vs_gt": 0.78125,
+      "activity_recall_rc_vs_gt": 0.8064516129032258,
+      "activity_f1_rc_vs_gt": 0.7936507936507936,
+      "class_match_rate": 1.0,
+      "doa_angular_error_deg_mean": 25.577542631147775,
+      "doa_angular_error_deg_median": 25.48267970568861,
+      "distance_mae_m": 0.9324265122413635,
+      "top_class_agreement": 1,
+      "gt_top_class": "breathing",
+      "rc_top_class": "breathing"
+    },
+    "sample": "s1_rank23_vocalsound_000569"
+  },
+  "s1_rank4_opensource_emilia_zh_rawdata_Chinese_003160": {
+    "gt_top_class": "speech",
+    "rc_top_class": "speech",
+    "gt_frames_preview": [
+      [],
+      [],
+      [],
+      [],
+      []
+    ],
+    "rc_frames_preview": [
+      [
+        {
+          "track": 2,
+          "activity": 0.504,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.85,
+          "azi_deg": 7.45,
+          "ele_deg": 3.25,
+          "dist_m": 5.727
+        }
+      ],
+      [
+        {
+          "track": 2,
+          "activity": 0.527,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.858,
+          "azi_deg": 7.82,
+          "ele_deg": 3.39,
+          "dist_m": 5.743
+        }
+      ],
+      [
+        {
+          "track": 2,
+          "activity": 0.55,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.866,
+          "azi_deg": 7.71,
+          "ele_deg": 3.69,
+          "dist_m": 5.816
+        }
+      ],
+      [
+        {
+          "track": 2,
+          "activity": 0.555,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.869,
+          "azi_deg": 7.39,
+          "ele_deg": 3.96,
+          "dist_m": 5.895
+        }
+      ],
+      [
+        {
+          "track": 2,
+          "activity": 0.548,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.867,
+          "azi_deg": 7.35,
+          "ele_deg": 3.94,
+          "dist_m": 5.874
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 134,
+      "activity_gt_frac": 0.15111940298507462,
+      "activity_rc_frac": 0.014925373134328358,
+      "activity_jaccard": 0.0,
+      "activity_precision_rc_vs_gt": 0.0,
+      "activity_recall_rc_vs_gt": 0.0,
+      "activity_f1_rc_vs_gt": 0.0,
+      "class_match_rate": NaN,
+      "doa_angular_error_deg_mean": NaN,
+      "doa_angular_error_deg_median": NaN,
+      "distance_mae_m": NaN,
+      "top_class_agreement": 1,
+      "gt_top_class": "speech",
+      "rc_top_class": "speech"
+    },
+    "sample": "s1_rank4_opensource_emilia_zh_rawdata_Chinese_003160"
+  },
+  "s1_rank5_opensource_emilia_zh_rawdata_English_001914": {
+    "gt_top_class": "speech",
+    "rc_top_class": "speech",
+    "gt_frames_preview": [
+      [
+        {
+          "track": 0,
+          "activity": 0.553,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.869,
+          "azi_deg": 13.72,
+          "ele_deg": 12.41,
+          "dist_m": 1.348
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.743,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.88,
+          "azi_deg": 13.84,
+          "ele_deg": 11.95,
+          "dist_m": 1.318
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.787,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.882,
+          "azi_deg": 14.6,
+          "ele_deg": 11.81,
+          "dist_m": 1.314
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.81,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.883,
+          "azi_deg": 15.07,
+          "ele_deg": 11.63,
+          "dist_m": 1.311
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.825,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.885,
+          "azi_deg": 15.36,
+          "ele_deg": 11.5,
+          "dist_m": 1.317
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [],
+      [],
+      [
+        {
+          "track": 0,
+          "activity": 0.517,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.844,
+          "azi_deg": 9.65,
+          "ele_deg": 14.2,
+          "dist_m": 1.282
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.547,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.853,
+          "azi_deg": 10.72,
+          "ele_deg": 14.96,
+          "dist_m": 1.259
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.57,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.86,
+          "azi_deg": 11.48,
+          "ele_deg": 15.47,
+          "dist_m": 1.253
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 46,
+      "activity_gt_frac": 0.25,
+      "activity_rc_frac": 0.07065217391304347,
+      "activity_jaccard": 0.2826086956521739,
+      "activity_precision_rc_vs_gt": 1.0,
+      "activity_recall_rc_vs_gt": 0.2826086956521739,
+      "activity_f1_rc_vs_gt": 0.4406779661016949,
+      "class_match_rate": 1.0,
+      "doa_angular_error_deg_mean": 6.089056468647589,
+      "doa_angular_error_deg_median": 6.163415867663726,
+      "distance_mae_m": 0.0694989487528801,
+      "top_class_agreement": 1,
+      "gt_top_class": "speech",
+      "rc_top_class": "speech"
+    },
+    "sample": "s1_rank5_opensource_emilia_zh_rawdata_English_001914"
+  },
+  "s1_rank9_opensource_emilia_zh_rawdata_English_001143": {
+    "gt_top_class": "speech",
+    "rc_top_class": "speech",
+    "gt_frames_preview": [
+      [],
+      [],
+      [
+        {
+          "track": 0,
+          "activity": 0.636,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.889,
+          "azi_deg": 29.79,
+          "ele_deg": -18.96,
+          "dist_m": 1.704
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.752,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.894,
+          "azi_deg": 29.98,
+          "ele_deg": -18.52,
+          "dist_m": 1.694
+        }
+      ],
+      [
+        {
+          "track": 0,
+          "activity": 0.798,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.895,
+          "azi_deg": 30.34,
+          "ele_deg": -18.31,
+          "dist_m": 1.677
+        }
+      ]
+    ],
+    "rc_frames_preview": [
+      [],
+      [],
+      [
+        {
+          "track": 1,
+          "activity": 0.647,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.851,
+          "azi_deg": 35.77,
+          "ele_deg": 26.09,
+          "dist_m": 1.346
+        }
+      ],
+      [
+        {
+          "track": 1,
+          "activity": 0.783,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.862,
+          "azi_deg": 36.15,
+          "ele_deg": 25.67,
+          "dist_m": 1.35
+        }
+      ],
+      [
+        {
+          "track": 1,
+          "activity": 0.829,
+          "class_idx": 16,
+          "class_name": "speech",
+          "class_conf": 0.865,
+          "azi_deg": 36.62,
+          "ele_deg": 25.73,
+          "dist_m": 1.342
+        }
+      ]
+    ],
+    "comparison": {
+      "T_s": 121,
+      "activity_gt_frac": 0.2190082644628099,
+      "activity_rc_frac": 0.24586776859504134,
+      "activity_jaccard": 0.0,
+      "activity_precision_rc_vs_gt": 0.0,
+      "activity_recall_rc_vs_gt": 0.0,
+      "activity_f1_rc_vs_gt": 0.0,
+      "class_match_rate": NaN,
+      "doa_angular_error_deg_mean": NaN,
+      "doa_angular_error_deg_median": NaN,
+      "distance_mae_m": NaN,
+      "top_class_agreement": 1,
+      "gt_top_class": "speech",
+      "rc_top_class": "speech"
+    },
+    "sample": "s1_rank9_opensource_emilia_zh_rawdata_English_001143"
+  }
+}

eval_voxaudio_ood_results/mono_vae/summary.json

@@ -0,0 +1,25 @@
+{
+  "mean_activity_jaccard": 0.31722869095023537,
+  "n_valid_activity_jaccard": 20,
+  "mean_activity_precision_rc_vs_gt": 0.4940412106193858,
+  "n_valid_activity_precision_rc_vs_gt": 20,
+  "mean_activity_recall_rc_vs_gt": 0.3829143289520885,
+  "n_valid_activity_recall_rc_vs_gt": 20,
+  "mean_activity_f1_rc_vs_gt": 0.39739221883924747,
+  "n_valid_activity_f1_rc_vs_gt": 20,
+  "mean_class_match_rate": 0.6986156079044322,
+  "n_valid_class_match_rate": 13,
+  "mean_doa_angular_error_deg_mean": 79.70906496987138,
+  "n_valid_doa_angular_error_deg_mean": 13,
+  "mean_doa_angular_error_deg_median": 79.14171551956005,
+  "n_valid_doa_angular_error_deg_median": 13,
+  "mean_distance_mae_m": 0.6082516071888117,
+  "n_valid_distance_mae_m": 13,
+  "mean_top_class_agreement": 0.65,
+  "n_valid_top_class_agreement": 20,
+  "mean_activity_gt_frac": 0.2116135538807752,
+  "n_valid_activity_gt_frac": 20,
+  "mean_activity_rc_frac": 0.1629927504276915,
+  "n_valid_activity_rc_frac": 20,
+  "n_samples": 20
+}

results/v13d_test_dcase_starss.json

@@ -0,0 +1,30 @@
+{
+  "checkpoint": "checkpoints/spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt",
+  "preset": "ov1_unified_v13d",
+  "split": "test",
+  "activity_threshold": 0.5,
+  "per_subset": [
+    {
+      "oracle_class_acc": 0.7304948906094069,
+      "oracle_azi_mae_deg": 34.44487453252077,
+      "oracle_ele_mae_deg": 11.19228895008564,
+      "oracle_dist_mae": 1.0144560723565519,
+      "class_acc": 0.884446004871279,
+      "azi_mae_deg": 28.73322555422783,
+      "ele_mae_deg": 9.473692879080772,
+      "dist_mae": 1.011748286895454,
+      "activity_precision": 0.524139404296875,
+      "activity_recall": 0.08432960510253906,
+      "activity_acc": 0.439788818359375,
+      "matched_count": 706.25,
+      "ER20": 0.8175,
+      "F20": 0.075,
+      "LE_CD": 133.91,
+      "LR_CD": 0.1355,
+      "SELD_score": 0.8377,
+      "subset": "dcase_starss",
+      "manifest": "/apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/dcase_starss_foa.test.jsonl",
+      "size": 505
+    }
+  ]
+}

results/v13d_test_dcase_starss.log

@@ -0,0 +1,26 @@
+/opt/conda/lib/python3.11/site-packages/torch/nn/utils/weight_norm.py:143: FutureWarning: `torch.nn.utils.weight_norm` is deprecated in favor of `torch.nn.utils.parametrizations.weight_norm`.
+  WeightNorm.apply(module, name, dim)
+/opt/conda/lib/python3.11/site-packages/torch/nn/modules/transformer.py:379: UserWarning: enable_nested_tensor is True, but self.use_nested_tensor is False because encoder_layer.norm_first was True
+  warnings.warn(
+[Eval] Device: cuda
+[Eval] Checkpoint: checkpoints/spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt
+[Eval] Preset: ov1_unified_v13d
+[Eval] Split: test
+[SpatialDataset] Initialize from /apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/dcase_starss_foa.test.jsonl
+
                                                                            
                                                                                             
                                                                                  
                                                                                      
+
+[Eval][dcase_starss] manifest=/apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/dcase_starss_foa.test.jsonl
+[Eval][dcase_starss] split=('test',)  size=505
+
                                                                  
+
+================================================================================================
+  v12 per-subset (test split)
+================================================================================================
+subset                N    F20   ER20   LE_CD  LR_CD   SELD   o_cls  o_azi  o_ele  o_dst    a_P   a_R
+-----------------------------------------------------------------------------------------------------
+dcase_starss        505 0.0750 0.8175  133.91 0.1355 0.8377  0.7305  34.44  11.19 1.0145  0.524 0.084
+================================================================================================
+  legend:  F20↑ ER20↓ LE_CD↓ LR_CD↑ SELD↓  o_cls=oracle_class_acc  o_azi/ele=oracle doa MAE (deg)  a_P/a_R=activity precision/recall
+================================================================================================
+
+[Eval] Summary saved to results/v13d_test_dcase_starss.json

results/v13d_test_unified.json

@@ -0,0 +1,30 @@
+{
+  "checkpoint": "checkpoints/spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt",
+  "preset": "ov1_unified_v13d",
+  "split": "test",
+  "activity_threshold": 0.5,
+  "per_subset": [
+    {
+      "oracle_class_acc": 0.7837179984777259,
+      "oracle_azi_mae_deg": 12.592206098633284,
+      "oracle_ele_mae_deg": 5.67435007226348,
+      "oracle_dist_mae": 0.4022363679437855,
+      "class_acc": 0.860446607692338,
+      "azi_mae_deg": 9.436391388904388,
+      "ele_mae_deg": 4.500656277421764,
+      "dist_mae": 0.38138905987075844,
+      "activity_precision": 0.8045251830722071,
+      "activity_recall": 0.05630378134135871,
+      "activity_acc": 0.7482400144045186,
+      "matched_count": 785.8321980018166,
+      "ER20": 0.5265,
+      "F20": 0.4871,
+      "LE_CD": 13.35,
+      "LR_CD": 0.5696,
+      "SELD_score": 0.386,
+      "subset": "unified",
+      "manifest": "/apdcephfs_cq12/share_302080740/user/schmittzhu/data/unified_spatial_foa_fsd63_all/test.jsonl",
+      "size": 35231
+    }
+  ]
+}

results/v13d_valid_dcase_starss.json

@@ -0,0 +1,30 @@
+{
+  "checkpoint": "checkpoints/spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt",
+  "preset": "ov1_unified_v13d",
+  "split": "valid",
+  "activity_threshold": 0.5,
+  "per_subset": [
+    {
+      "oracle_class_acc": 0.7087607669715541,
+      "oracle_azi_mae_deg": 33.16232564825761,
+      "oracle_ele_mae_deg": 11.2290494040439,
+      "oracle_dist_mae": 1.0284199475980642,
+      "class_acc": 0.8441169418950092,
+      "azi_mae_deg": 26.882637571870234,
+      "ele_mae_deg": 9.890066773117635,
+      "dist_mae": 1.028246155276633,
+      "activity_precision": 0.506396484375,
+      "activity_recall": 0.07933697616844847,
+      "activity_acc": 0.42705163788377193,
+      "matched_count": 690.7456140350877,
+      "ER20": 0.8213,
+      "F20": 0.0799,
+      "LE_CD": 129.39,
+      "LR_CD": 0.1395,
+      "SELD_score": 0.8302,
+      "subset": "dcase_starss",
+      "manifest": "/apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/dcase_starss_foa.valid.jsonl",
+      "size": 4560
+    }
+  ]
+}

results/v13d_valid_dcase_starss.log

@@ -0,0 +1,26 @@
+/opt/conda/lib/python3.11/site-packages/torch/nn/utils/weight_norm.py:143: FutureWarning: `torch.nn.utils.weight_norm` is deprecated in favor of `torch.nn.utils.parametrizations.weight_norm`.
+  WeightNorm.apply(module, name, dim)
+/opt/conda/lib/python3.11/site-packages/torch/nn/modules/transformer.py:379: UserWarning: enable_nested_tensor is True, but self.use_nested_tensor is False because encoder_layer.norm_first was True
+  warnings.warn(
+[Eval] Device: cuda
+[Eval] Checkpoint: checkpoints/spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt
+[Eval] Preset: ov1_unified_v13d
+[Eval] Split: valid
+[SpatialDataset] Initialize from /apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/dcase_starss_foa.valid.jsonl
+
                                                                            
                                                                                                 
                                                                                    
                                                                                        
+
+[Eval][dcase_starss] manifest=/apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/dcase_starss_foa.valid.jsonl
+[Eval][dcase_starss] split=('valid',)  size=4560
+
                                                                    
+
+================================================================================================
+  v12 per-subset (valid split)
+================================================================================================
+subset                N    F20   ER20   LE_CD  LR_CD   SELD   o_cls  o_azi  o_ele  o_dst    a_P   a_R
+-----------------------------------------------------------------------------------------------------
+dcase_starss       4560 0.0799 0.8213  129.39 0.1395 0.8302  0.7088  33.16  11.23 1.0284  0.506 0.079
+================================================================================================
+  legend:  F20↑ ER20↓ LE_CD↓ LR_CD↑ SELD↓  o_cls=oracle_class_acc  o_azi/ele=oracle doa MAE (deg)  a_P/a_R=activity precision/recall
+================================================================================================
+
+[Eval] Summary saved to results/v13d_valid_dcase_starss.json

results/v13d_valid_ov1_real.log

@@ -0,0 +1,92 @@
+/opt/conda/lib/python3.11/site-packages/torch/nn/utils/weight_norm.py:143: FutureWarning: `torch.nn.utils.weight_norm` is deprecated in favor of `torch.nn.utils.parametrizations.weight_norm`.
+  WeightNorm.apply(module, name, dim)
+/opt/conda/lib/python3.11/site-packages/torch/nn/modules/transformer.py:379: UserWarning: enable_nested_tensor is True, but self.use_nested_tensor is False because encoder_layer.norm_first was True
+  warnings.warn(
+[Eval] Device: cuda
+[Eval] Checkpoint: checkpoints/spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt
+[Eval] Preset: ov1_unified_v13d
+[Eval] Split: valid
+[SpatialDataset] Initialize from /apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/ov1_real_static_foa_mapped.jsonl
+
                                                                            
                                                                                                        
                                                                                         
                                                                                            
+
+[Eval][ov1_real] manifest=/apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/ov1_real_static_foa_mapped.jsonl
+[Eval][ov1_real] split=('valid',)  size=3374
+
                                                      
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/eval_v12_per_subset.py", line 470, in <module>
+    main()
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/eval_v12_per_subset.py", line 420, in main
+    m = eval_one_subset(
+        ^^^^^^^^^^^^^^^^
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/eval_v12_per_subset.py", line 312, in eval_one_subset
+    for batch in tqdm(loader, desc=f"Eval {subset_name}", leave=False):
+  File "/opt/conda/lib/python3.11/site-packages/tqdm/std.py", line 1181, in __iter__
+    for obj in iterable:
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/dataloader.py", line 701, in __next__
+    data = self._next_data()
+           ^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/dataloader.py", line 1465, in _next_data
+    return self._process_data(data)
+           ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/dataloader.py", line 1491, in _process_data
+    data.reraise()
+  File "/opt/conda/lib/python3.11/site-packages/torch/_utils.py", line 715, in reraise
+    raise exception
+RuntimeError: Caught RuntimeError in DataLoader worker process 0.
+Original Traceback (most recent call last):
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 378, in _load_audio_file
+    waveform, sample_rate = sf.read(path, always_2d=True)
+                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/soundfile.py", line 305, in read
+    with SoundFile(file, 'r', samplerate, channels,
+         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/soundfile.py", line 690, in __init__
+    self._file = self._open(file, mode_int, closefd)
+                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/soundfile.py", line 1265, in _open
+    raise LibsndfileError(err, prefix="Error opening {0!r}: ".format(self.name))
+soundfile.LibsndfileError: Error opening '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov1_real_static_foa/valid/ov1_real_static_foa_006985.wav': System error.
+
+During handling of the above exception, another exception occurred:
+
+Traceback (most recent call last):
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 384, in _load_audio_file
+    sample_rate, waveform_np = wavfile.read(path)
+                               ^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/scipy/io/wavfile.py", line 674, in read
+    fid = open(filename, 'rb')
+          ^^^^^^^^^^^^^^^^^^^^
+FileNotFoundError: [Errno 2] No such file or directory: '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov1_real_static_foa/valid/ov1_real_static_foa_006985.wav'
+
+During handling of the above exception, another exception occurred:
+
+Traceback (most recent call last):
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 397, in _load_audio_file
+    with wave.open(path, "rb") as handle:
+         ^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/wave.py", line 631, in open
+    return Wave_read(f)
+           ^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/wave.py", line 279, in __init__
+    f = builtins.open(f, 'rb')
+        ^^^^^^^^^^^^^^^^^^^^^^
+FileNotFoundError: [Errno 2] No such file or directory: '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov1_real_static_foa/valid/ov1_real_static_foa_006985.wav'
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/_utils/worker.py", line 351, in _worker_loop
+    data = fetcher.fetch(index)  # type: ignore[possibly-undefined]
+           ^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/_utils/fetch.py", line 52, in fetch
+    data = [self.dataset[idx] for idx in possibly_batched_index]
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/_utils/fetch.py", line 52, in <listcomp>
+    data = [self.dataset[idx] for idx in possibly_batched_index]
+            ~~~~~~~~~~~~^^^^^
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 1310, in __getitem__
+    waveform = _load_audio_file(str(waveform_path), self.config.mel_config.sample_rate)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 416, in _load_audio_file
+    raise RuntimeError(
+RuntimeError: Failed to load audio file '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov1_real_static_foa/valid/ov1_real_static_foa_006985.wav'. Install soundfile/scipy or provide PCM wav.
+

results/v13d_valid_ov1_sim.json

@@ -0,0 +1,30 @@
+{
+  "checkpoint": "checkpoints/spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt",
+  "preset": "ov1_unified_v13d",
+  "split": "valid",
+  "activity_threshold": 0.5,
+  "per_subset": [
+    {
+      "oracle_class_acc": 0.7937788535902898,
+      "oracle_azi_mae_deg": 25.076975057125093,
+      "oracle_ele_mae_deg": 9.14060350239277,
+      "oracle_dist_mae": 0.6071755435566107,
+      "class_acc": 0.879228694178164,
+      "azi_mae_deg": 21.48522803256909,
+      "ele_mae_deg": 7.427904010117054,
+      "dist_mae": 0.5744369750159483,
+      "activity_precision": 0.8104524739583333,
+      "activity_recall": 0.04147186279296875,
+      "activity_acc": 0.7690266927083333,
+      "matched_count": 391.64,
+      "ER20": 0.6422,
+      "F20": 0.387,
+      "LE_CD": 26.11,
+      "LR_CD": 0.6079,
+      "SELD_score": 0.4481,
+      "subset": "ov1_sim",
+      "manifest": "/apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/ov1_foa.jsonl",
+      "size": 4800
+    }
+  ]
+}

results/v13d_valid_ov1_sim.log

@@ -0,0 +1,26 @@
+/opt/conda/lib/python3.11/site-packages/torch/nn/utils/weight_norm.py:143: FutureWarning: `torch.nn.utils.weight_norm` is deprecated in favor of `torch.nn.utils.parametrizations.weight_norm`.
+  WeightNorm.apply(module, name, dim)
+/opt/conda/lib/python3.11/site-packages/torch/nn/modules/transformer.py:379: UserWarning: enable_nested_tensor is True, but self.use_nested_tensor is False because encoder_layer.norm_first was True
+  warnings.warn(
+[Eval] Device: cuda
+[Eval] Checkpoint: checkpoints/spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt
+[Eval] Preset: ov1_unified_v13d
+[Eval] Split: valid
+[SpatialDataset] Initialize from /apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/ov1_foa.jsonl
+
                                                                            
                                                                                     
                                                                      
                                                                         
+
+[Eval][ov1_sim] manifest=/apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/ov1_foa.jsonl
+[Eval][ov1_sim] split=('valid',)  size=4800
+
                                                               
+
+================================================================================================
+  v12 per-subset (valid split)
+================================================================================================
+subset                N    F20   ER20   LE_CD  LR_CD   SELD   o_cls  o_azi  o_ele  o_dst    a_P   a_R
+-----------------------------------------------------------------------------------------------------
+ov1_sim            4800 0.3870 0.6422   26.11 0.6079 0.4481  0.7938  25.08   9.14 0.6072  0.810 0.041
+================================================================================================
+  legend:  F20↑ ER20↓ LE_CD↓ LR_CD↑ SELD↓  o_cls=oracle_class_acc  o_azi/ele=oracle doa MAE (deg)  a_P/a_R=activity precision/recall
+================================================================================================
+
+[Eval] Summary saved to results/v13d_valid_ov1_sim.json

results/v13d_valid_ov2_real.log

@@ -0,0 +1,92 @@
+/opt/conda/lib/python3.11/site-packages/torch/nn/utils/weight_norm.py:143: FutureWarning: `torch.nn.utils.weight_norm` is deprecated in favor of `torch.nn.utils.parametrizations.weight_norm`.
+  WeightNorm.apply(module, name, dim)
+/opt/conda/lib/python3.11/site-packages/torch/nn/modules/transformer.py:379: UserWarning: enable_nested_tensor is True, but self.use_nested_tensor is False because encoder_layer.norm_first was True
+  warnings.warn(
+[Eval] Device: cuda
+[Eval] Checkpoint: checkpoints/spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt
+[Eval] Preset: ov1_unified_v13d
+[Eval] Split: valid
+[SpatialDataset] Initialize from /apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/ov2_real_static_foa_mapped.jsonl
+
                                                                            
                                                                                                      
                                                                                        
                                                                                            
+
+[Eval][ov2_real] manifest=/apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/ov2_real_static_foa_mapped.jsonl
+[Eval][ov2_real] split=('valid',)  size=2230
+
                                                      
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/eval_v12_per_subset.py", line 470, in <module>
+    main()
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/eval_v12_per_subset.py", line 420, in main
+    m = eval_one_subset(
+        ^^^^^^^^^^^^^^^^
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/eval_v12_per_subset.py", line 312, in eval_one_subset
+    for batch in tqdm(loader, desc=f"Eval {subset_name}", leave=False):
+  File "/opt/conda/lib/python3.11/site-packages/tqdm/std.py", line 1181, in __iter__
+    for obj in iterable:
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/dataloader.py", line 701, in __next__
+    data = self._next_data()
+           ^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/dataloader.py", line 1465, in _next_data
+    return self._process_data(data)
+           ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/dataloader.py", line 1491, in _process_data
+    data.reraise()
+  File "/opt/conda/lib/python3.11/site-packages/torch/_utils.py", line 715, in reraise
+    raise exception
+RuntimeError: Caught RuntimeError in DataLoader worker process 0.
+Original Traceback (most recent call last):
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 378, in _load_audio_file
+    waveform, sample_rate = sf.read(path, always_2d=True)
+                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/soundfile.py", line 305, in read
+    with SoundFile(file, 'r', samplerate, channels,
+         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/soundfile.py", line 690, in __init__
+    self._file = self._open(file, mode_int, closefd)
+                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/soundfile.py", line 1265, in _open
+    raise LibsndfileError(err, prefix="Error opening {0!r}: ".format(self.name))
+soundfile.LibsndfileError: Error opening '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov2_real_static_foa/valid/ov2_real_static_foa_004035.wav': System error.
+
+During handling of the above exception, another exception occurred:
+
+Traceback (most recent call last):
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 384, in _load_audio_file
+    sample_rate, waveform_np = wavfile.read(path)
+                               ^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/scipy/io/wavfile.py", line 674, in read
+    fid = open(filename, 'rb')
+          ^^^^^^^^^^^^^^^^^^^^
+FileNotFoundError: [Errno 2] No such file or directory: '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov2_real_static_foa/valid/ov2_real_static_foa_004035.wav'
+
+During handling of the above exception, another exception occurred:
+
+Traceback (most recent call last):
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 397, in _load_audio_file
+    with wave.open(path, "rb") as handle:
+         ^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/wave.py", line 631, in open
+    return Wave_read(f)
+           ^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/wave.py", line 279, in __init__
+    f = builtins.open(f, 'rb')
+        ^^^^^^^^^^^^^^^^^^^^^^
+FileNotFoundError: [Errno 2] No such file or directory: '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov2_real_static_foa/valid/ov2_real_static_foa_004035.wav'
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/_utils/worker.py", line 351, in _worker_loop
+    data = fetcher.fetch(index)  # type: ignore[possibly-undefined]
+           ^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/_utils/fetch.py", line 52, in fetch
+    data = [self.dataset[idx] for idx in possibly_batched_index]
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/_utils/fetch.py", line 52, in <listcomp>
+    data = [self.dataset[idx] for idx in possibly_batched_index]
+            ~~~~~~~~~~~~^^^^^
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 1310, in __getitem__
+    waveform = _load_audio_file(str(waveform_path), self.config.mel_config.sample_rate)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 416, in _load_audio_file
+    raise RuntimeError(
+RuntimeError: Failed to load audio file '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov2_real_static_foa/valid/ov2_real_static_foa_004035.wav'. Install soundfile/scipy or provide PCM wav.
+

results/v13d_valid_ov2_sim.log

@@ -0,0 +1,92 @@
+/opt/conda/lib/python3.11/site-packages/torch/nn/utils/weight_norm.py:143: FutureWarning: `torch.nn.utils.weight_norm` is deprecated in favor of `torch.nn.utils.parametrizations.weight_norm`.
+  WeightNorm.apply(module, name, dim)
+/opt/conda/lib/python3.11/site-packages/torch/nn/modules/transformer.py:379: UserWarning: enable_nested_tensor is True, but self.use_nested_tensor is False because encoder_layer.norm_first was True
+  warnings.warn(
+[Eval] Device: cuda
+[Eval] Checkpoint: checkpoints/spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt
+[Eval] Preset: ov1_unified_v13d
+[Eval] Split: valid
+[SpatialDataset] Initialize from /apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/ov2_foa.jsonl
+
                                                                            
                                                                                     
                                                                      
                                                                         
+
+[Eval][ov2_sim] manifest=/apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/ov2_foa.jsonl
+[Eval][ov2_sim] split=('valid',)  size=1718
+
                                                     
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/eval_v12_per_subset.py", line 470, in <module>
+    main()
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/eval_v12_per_subset.py", line 420, in main
+    m = eval_one_subset(
+        ^^^^^^^^^^^^^^^^
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/eval_v12_per_subset.py", line 312, in eval_one_subset
+    for batch in tqdm(loader, desc=f"Eval {subset_name}", leave=False):
+  File "/opt/conda/lib/python3.11/site-packages/tqdm/std.py", line 1181, in __iter__
+    for obj in iterable:
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/dataloader.py", line 701, in __next__
+    data = self._next_data()
+           ^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/dataloader.py", line 1465, in _next_data
+    return self._process_data(data)
+           ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/dataloader.py", line 1491, in _process_data
+    data.reraise()
+  File "/opt/conda/lib/python3.11/site-packages/torch/_utils.py", line 715, in reraise
+    raise exception
+RuntimeError: Caught RuntimeError in DataLoader worker process 0.
+Original Traceback (most recent call last):
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 378, in _load_audio_file
+    waveform, sample_rate = sf.read(path, always_2d=True)
+                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/soundfile.py", line 305, in read
+    with SoundFile(file, 'r', samplerate, channels,
+         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/soundfile.py", line 690, in __init__
+    self._file = self._open(file, mode_int, closefd)
+                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/soundfile.py", line 1265, in _open
+    raise LibsndfileError(err, prefix="Error opening {0!r}: ".format(self.name))
+soundfile.LibsndfileError: Error opening '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov2_foa/valid/ov2_000000.wav': System error.
+
+During handling of the above exception, another exception occurred:
+
+Traceback (most recent call last):
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 384, in _load_audio_file
+    sample_rate, waveform_np = wavfile.read(path)
+                               ^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/scipy/io/wavfile.py", line 674, in read
+    fid = open(filename, 'rb')
+          ^^^^^^^^^^^^^^^^^^^^
+FileNotFoundError: [Errno 2] No such file or directory: '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov2_foa/valid/ov2_000000.wav'
+
+During handling of the above exception, another exception occurred:
+
+Traceback (most recent call last):
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 397, in _load_audio_file
+    with wave.open(path, "rb") as handle:
+         ^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/wave.py", line 631, in open
+    return Wave_read(f)
+           ^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/wave.py", line 279, in __init__
+    f = builtins.open(f, 'rb')
+        ^^^^^^^^^^^^^^^^^^^^^^
+FileNotFoundError: [Errno 2] No such file or directory: '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov2_foa/valid/ov2_000000.wav'
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/_utils/worker.py", line 351, in _worker_loop
+    data = fetcher.fetch(index)  # type: ignore[possibly-undefined]
+           ^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/_utils/fetch.py", line 52, in fetch
+    data = [self.dataset[idx] for idx in possibly_batched_index]
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/_utils/fetch.py", line 52, in <listcomp>
+    data = [self.dataset[idx] for idx in possibly_batched_index]
+            ~~~~~~~~~~~~^^^^^
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 1310, in __getitem__
+    waveform = _load_audio_file(str(waveform_path), self.config.mel_config.sample_rate)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 416, in _load_audio_file
+    raise RuntimeError(
+RuntimeError: Failed to load audio file '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov2_foa/valid/ov2_000000.wav'. Install soundfile/scipy or provide PCM wav.
+

results/v13d_valid_ov3_real.log

@@ -0,0 +1,92 @@
+/opt/conda/lib/python3.11/site-packages/torch/nn/utils/weight_norm.py:143: FutureWarning: `torch.nn.utils.weight_norm` is deprecated in favor of `torch.nn.utils.parametrizations.weight_norm`.
+  WeightNorm.apply(module, name, dim)
+/opt/conda/lib/python3.11/site-packages/torch/nn/modules/transformer.py:379: UserWarning: enable_nested_tensor is True, but self.use_nested_tensor is False because encoder_layer.norm_first was True
+  warnings.warn(
+[Eval] Device: cuda
+[Eval] Checkpoint: checkpoints/spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt
+[Eval] Preset: ov1_unified_v13d
+[Eval] Split: valid
+[SpatialDataset] Initialize from /apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/ov3_real_static_foa_mapped.jsonl
+
                                                                            
                                                                                                    
                                                                                        
                                                                                           
+
+[Eval][ov3_real] manifest=/apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/ov3_real_static_foa_mapped.jsonl
+[Eval][ov3_real] split=('valid',)  size=740
+
                                                     
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/eval_v12_per_subset.py", line 470, in <module>
+    main()
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/eval_v12_per_subset.py", line 420, in main
+    m = eval_one_subset(
+        ^^^^^^^^^^^^^^^^
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/eval_v12_per_subset.py", line 312, in eval_one_subset
+    for batch in tqdm(loader, desc=f"Eval {subset_name}", leave=False):
+  File "/opt/conda/lib/python3.11/site-packages/tqdm/std.py", line 1181, in __iter__
+    for obj in iterable:
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/dataloader.py", line 701, in __next__
+    data = self._next_data()
+           ^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/dataloader.py", line 1465, in _next_data
+    return self._process_data(data)
+           ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/dataloader.py", line 1491, in _process_data
+    data.reraise()
+  File "/opt/conda/lib/python3.11/site-packages/torch/_utils.py", line 715, in reraise
+    raise exception
+RuntimeError: Caught RuntimeError in DataLoader worker process 0.
+Original Traceback (most recent call last):
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 378, in _load_audio_file
+    waveform, sample_rate = sf.read(path, always_2d=True)
+                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/soundfile.py", line 305, in read
+    with SoundFile(file, 'r', samplerate, channels,
+         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/soundfile.py", line 690, in __init__
+    self._file = self._open(file, mode_int, closefd)
+                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/soundfile.py", line 1265, in _open
+    raise LibsndfileError(err, prefix="Error opening {0!r}: ".format(self.name))
+soundfile.LibsndfileError: Error opening '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov3_real_static_foa/valid/ov3_real_static_foa_000845.wav': System error.
+
+During handling of the above exception, another exception occurred:
+
+Traceback (most recent call last):
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 384, in _load_audio_file
+    sample_rate, waveform_np = wavfile.read(path)
+                               ^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/scipy/io/wavfile.py", line 674, in read
+    fid = open(filename, 'rb')
+          ^^^^^^^^^^^^^^^^^^^^
+FileNotFoundError: [Errno 2] No such file or directory: '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov3_real_static_foa/valid/ov3_real_static_foa_000845.wav'
+
+During handling of the above exception, another exception occurred:
+
+Traceback (most recent call last):
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 397, in _load_audio_file
+    with wave.open(path, "rb") as handle:
+         ^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/wave.py", line 631, in open
+    return Wave_read(f)
+           ^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/wave.py", line 279, in __init__
+    f = builtins.open(f, 'rb')
+        ^^^^^^^^^^^^^^^^^^^^^^
+FileNotFoundError: [Errno 2] No such file or directory: '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov3_real_static_foa/valid/ov3_real_static_foa_000845.wav'
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/_utils/worker.py", line 351, in _worker_loop
+    data = fetcher.fetch(index)  # type: ignore[possibly-undefined]
+           ^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/_utils/fetch.py", line 52, in fetch
+    data = [self.dataset[idx] for idx in possibly_batched_index]
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/_utils/fetch.py", line 52, in <listcomp>
+    data = [self.dataset[idx] for idx in possibly_batched_index]
+            ~~~~~~~~~~~~^^^^^
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 1310, in __getitem__
+    waveform = _load_audio_file(str(waveform_path), self.config.mel_config.sample_rate)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 416, in _load_audio_file
+    raise RuntimeError(
+RuntimeError: Failed to load audio file '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov3_real_static_foa/valid/ov3_real_static_foa_000845.wav'. Install soundfile/scipy or provide PCM wav.
+

results/v13d_valid_ov3_sim.log

@@ -0,0 +1,92 @@
+/opt/conda/lib/python3.11/site-packages/torch/nn/utils/weight_norm.py:143: FutureWarning: `torch.nn.utils.weight_norm` is deprecated in favor of `torch.nn.utils.parametrizations.weight_norm`.
+  WeightNorm.apply(module, name, dim)
+/opt/conda/lib/python3.11/site-packages/torch/nn/modules/transformer.py:379: UserWarning: enable_nested_tensor is True, but self.use_nested_tensor is False because encoder_layer.norm_first was True
+  warnings.warn(
+[Eval] Device: cuda
+[Eval] Checkpoint: checkpoints/spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt
+[Eval] Preset: ov1_unified_v13d
+[Eval] Split: valid
+[SpatialDataset] Initialize from /apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/ov3_foa.jsonl
+
                                                                            
                                                                                     
                                                                      
                                                                         
+
+[Eval][ov3_sim] manifest=/apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/ov3_foa.jsonl
+[Eval][ov3_sim] split=('valid',)  size=1612
+
                                                     
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/eval_v12_per_subset.py", line 470, in <module>
+    main()
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/eval_v12_per_subset.py", line 420, in main
+    m = eval_one_subset(
+        ^^^^^^^^^^^^^^^^
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/eval_v12_per_subset.py", line 312, in eval_one_subset
+    for batch in tqdm(loader, desc=f"Eval {subset_name}", leave=False):
+  File "/opt/conda/lib/python3.11/site-packages/tqdm/std.py", line 1181, in __iter__
+    for obj in iterable:
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/dataloader.py", line 701, in __next__
+    data = self._next_data()
+           ^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/dataloader.py", line 1465, in _next_data
+    return self._process_data(data)
+           ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/dataloader.py", line 1491, in _process_data
+    data.reraise()
+  File "/opt/conda/lib/python3.11/site-packages/torch/_utils.py", line 715, in reraise
+    raise exception
+RuntimeError: Caught RuntimeError in DataLoader worker process 0.
+Original Traceback (most recent call last):
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 378, in _load_audio_file
+    waveform, sample_rate = sf.read(path, always_2d=True)
+                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/soundfile.py", line 305, in read
+    with SoundFile(file, 'r', samplerate, channels,
+         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/soundfile.py", line 690, in __init__
+    self._file = self._open(file, mode_int, closefd)
+                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/soundfile.py", line 1265, in _open
+    raise LibsndfileError(err, prefix="Error opening {0!r}: ".format(self.name))
+soundfile.LibsndfileError: Error opening '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov3_foa/valid/ov3_000001.wav': System error.
+
+During handling of the above exception, another exception occurred:
+
+Traceback (most recent call last):
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 384, in _load_audio_file
+    sample_rate, waveform_np = wavfile.read(path)
+                               ^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/scipy/io/wavfile.py", line 674, in read
+    fid = open(filename, 'rb')
+          ^^^^^^^^^^^^^^^^^^^^
+FileNotFoundError: [Errno 2] No such file or directory: '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov3_foa/valid/ov3_000001.wav'
+
+During handling of the above exception, another exception occurred:
+
+Traceback (most recent call last):
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 397, in _load_audio_file
+    with wave.open(path, "rb") as handle:
+         ^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/wave.py", line 631, in open
+    return Wave_read(f)
+           ^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/wave.py", line 279, in __init__
+    f = builtins.open(f, 'rb')
+        ^^^^^^^^^^^^^^^^^^^^^^
+FileNotFoundError: [Errno 2] No such file or directory: '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov3_foa/valid/ov3_000001.wav'
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/_utils/worker.py", line 351, in _worker_loop
+    data = fetcher.fetch(index)  # type: ignore[possibly-undefined]
+           ^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/_utils/fetch.py", line 52, in fetch
+    data = [self.dataset[idx] for idx in possibly_batched_index]
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/opt/conda/lib/python3.11/site-packages/torch/utils/data/_utils/fetch.py", line 52, in <listcomp>
+    data = [self.dataset[idx] for idx in possibly_batched_index]
+            ~~~~~~~~~~~~^^^^^
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 1310, in __getitem__
+    waveform = _load_audio_file(str(waveform_path), self.config.mel_config.sample_rate)
+               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_dataset.py", line 416, in _load_audio_file
+    raise RuntimeError(
+RuntimeError: Failed to load audio file '/apdcephfs_cq10/share_1603164/user/schmittzhu/data/ov3_foa/valid/ov3_000001.wav'. Install soundfile/scipy or provide PCM wav.
+

results/v13d_valid_unified.json

@@ -0,0 +1,30 @@
+{
+  "checkpoint": "checkpoints/spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt",
+  "preset": "ov1_unified_v13d",
+  "split": "valid",
+  "activity_threshold": 0.5,
+  "per_subset": [
+    {
+      "oracle_class_acc": 0.7986697095998104,
+      "oracle_azi_mae_deg": 15.619765282153757,
+      "oracle_ele_mae_deg": 6.781703096456646,
+      "oracle_dist_mae": 0.47289481620372786,
+      "class_acc": 0.8735275138492873,
+      "azi_mae_deg": 12.024803064106642,
+      "ele_mae_deg": 5.486542534098445,
+      "dist_mae": 0.4547080461109417,
+      "activity_precision": 0.7795697905364892,
+      "activity_recall": 0.06645036734855028,
+      "activity_acc": 0.7131156938813386,
+      "matched_count": 876.207857469164,
+      "ER20": 0.5535,
+      "F20": 0.4425,
+      "LE_CD": 18.81,
+      "LR_CD": 0.5468,
+      "SELD_score": 0.4172,
+      "subset": "unified",
+      "manifest": "/apdcephfs_cq12/share_302080740/user/schmittzhu/data/unified_spatial_foa_fsd63_all/valid.jsonl",
+      "size": 35021
+    }
+  ]
+}