TransSLR v3 | fix:norm+aug+smallermodel | Top-1:35.33%

config.json

@@ -9,11 +9,11 @@
   "dataset": "luciayen/KSL-W30-Landmarks",
   "num_classes": 30,
   "feat_dim": 225,
-  "d_model": 128,
+  "d_model": 192,
   "nhead": 4,
-  "num_layers": 2,
-  "ffn_dim": 512,
-  "dropout": 0.6,
+  "num_layers": 3,
+  "ffn_dim": 768,
+  "dropout": 0.5,
   "num_frames": 64,
   "label2idx": {
     "Agreement": 0,
@@ -49,21 +49,21 @@
   },
   "normalisation": "train mean/std applied to all splits",
   "augmentation": {
-    "gaussian_noise_std": 0.01,
-    "temporal_jitter_frames": 3,
-    "frame_dropout_fraction": 0.08,
-    "horizontal_flip_prob": 0.5
+    "gaussian_noise_std": 0.008,
+    "temporal_jitter_frames": 2,
+    "frame_dropout_fraction": 0.05,
+    "horizontal_flip_prob": 0.0
   },
   "results": {
-    "top1_acc": 0.2289,
-    "top5_acc": 0.5222,
-    "macro_recall": 0.2289,
-    "recall_at_1": 0.2289,
-    "recall_at_5": 0.5222,
-    "recall_at_10": 0.7022,
-    "median_rank": 5.0,
-    "params_M": 0.43,
-    "flops": "18.764M",
-    "latency_ms": 0.868
+    "top1_acc": 0.3533,
+    "top5_acc": 0.7311,
+    "macro_recall": 0.3533,
+    "recall_at_1": 0.3533,
+    "recall_at_5": 0.7311,
+    "recall_at_10": 0.8533,
+    "median_rank": 2.0,
+    "params_M": 1.38,
+    "flops": "59.759M",
+    "latency_ms": 1.248
   }
 }

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fc9b15a1839c70494af7c6e677269ed59d39f39606ddb59bfb504fbe59be8089
-size 1990622
+oid sha256:9bc4d1de2572c70054443332fc71d22705c005a4f59fd7eaeb35b3c285bf91af
+size 5946312