KSL-TransSLR / config.json
luciayen's picture
TransSLR v3 | fix:norm+aug+smallermodel | Top-1:40.00%
9b3a333 verified
Raw
History Blame Contribute Delete
1.38 kB
{
"model_type": "TransSLR_v3",
"version": "v3",
"changes_from_v2": [
"train-stats normalisation fix",
"data augmentation",
"smaller model (128d 2L)"
],
"dataset": "luciayen/KSL-W30-Landmarks",
"num_classes": 30,
"feat_dim": 225,
"d_model": 192,
"nhead": 4,
"num_layers": 3,
"ffn_dim": 768,
"dropout": 0.5,
"num_frames": 64,
"label2idx": {
"Agreement": 0,
"Apple": 1,
"Colour": 2,
"Friend": 3,
"Gift": 4,
"Market": 5,
"Monday": 6,
"No_100": 7,
"No_125": 8,
"No_17": 9,
"No_22": 10,
"No_268": 11,
"No_35": 12,
"No_388": 13,
"No_444": 14,
"No_48": 15,
"No_54": 16,
"No_66": 17,
"No_73": 18,
"No_89": 19,
"No_9": 20,
"No_91": 21,
"Picture": 22,
"Proud": 23,
"Sweater": 24,
"Teach": 25,
"Tomatoes": 26,
"Tortoise": 27,
"Twin": 28,
"Ugali": 29
},
"normalisation": "train mean/std applied to all splits",
"augmentation": {
"gaussian_noise_std": 0.008,
"temporal_jitter_frames": 2,
"frame_dropout_fraction": 0.05,
"horizontal_flip_prob": 0.0
},
"results": {
"top1_acc": 0.4,
"top5_acc": 0.7289,
"macro_recall": 0.4,
"recall_at_1": 0.4,
"recall_at_5": 0.7289,
"recall_at_10": 0.8356,
"median_rank": 2.0,
"params_M": 1.38,
"flops": "59.759M",
"latency_ms": 1.227
}
}