Upload weights and configs - Run 20251012_040642

weights/best_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:395f5ebc0ca3bce746bd7fce81f21d10e71341304c6e8236478064f779529919
-size 59515088
+oid sha256:294bbd186f6036cf8599e3d257dae1b99f2b797e69dd10829c778672fbcc83e0
+size 64195804

weights/david_config.json

@@ -1,35 +1,33 @@
 {
-  "name": "david_high_accuracy",
-  "uid": "c.david.high_accuracy",
-  "feature_dim": 512,
+  "name": "david_clip_vit_l14",
+  "uid": "c.david.clip_vit_l14",
+  "feature_dim": 768,
   "num_classes": 1000,
   "scales": [
-    256,
-    512,
+    384,
     768,
     1024,
     1280
   ],
-  "sharing_mode": "decoupled",
+  "sharing_mode": "partial_shared",
   "fusion_mode": "deep_efficiency",
   "use_belly": true,
-  "belly_expand": 2.5,
-  "shared_feature_dim": 768,
+  "belly_expand": 2.0,
+  "shared_feature_dim": 1024,
   "shared_layers": 2,
   "shared_dropout": 0.1,
   "fusion_temperature": 1.0,
   "fusion_dropout": 0.1,
   "tree_depth": 3,
-  "num_experts": 5,
-  "compression_ratio": 2,
+  "num_experts": 4,
+  "compression_ratio": 4,
   "expert_dropout": 0.1,
   "attention_dropout": 0.1,
   "progressive_training": true,
   "scale_warmup_epochs": {
-    "256": 0,
-    "512": 3,
-    "768": 6,
-    "1024": 9,
-    "1280": 12
+    "384": 0,
+    "768": 1,
+    "1024": 2,
+    "1280": 3
   }
 }

weights/train_config.json

@@ -1,15 +1,21 @@
 {
   "name": "david_training",
-  "run_id": "20251012_032356",
+  "run_id": "20251012_040642",
   "dataset_name": "AbstractPhil/imagenet-clip-features-orderly",
-  "model_variant": "clip_vit_b16",
+  "model_variant": "clip_vit_l14",
   "num_classes": 1000,
-  "preset": "high_accuracy",
+  "preset": "clip_vit_l14",
   "custom_config_path": null,
   "num_classes_override": null,
   "use_belly_override": null,
   "belly_expand_override": null,
   "progressive_training_override": true,
+  "scale_warmup_epochs_override": {
+    "384": 0,
+    "768": 1,
+    "1024": 2,
+    "1280": 3
+  },
   "num_epochs": 20,
   "batch_size": 1024,
   "learning_rate": 0.01,
@@ -26,8 +32,8 @@
   "gradient_clip": 5.0,
   "scheduler_type": "cosine_restarts",
   "min_lr": 1e-06,
-  "freeze_strategy": "performance",
-  "freeze_threshold": 70.0,
+  "freeze_strategy": "never",
+  "freeze_threshold": 90.0,
   "unfreeze_on_plateau": true,
   "patience": 10,
   "track_gradients": true,