Epoch 10 - 44.60% acc

weights/run_003_v2_8redundantscale_4x4patch_d768_4layer_topo_dropout_cutmix_20251202_142542/README.md

@@ -0,0 +1,60 @@
+# Run: run_003_v2_8redundantscale_4x4patch_d768_4layer_topo_dropout_cutmix_20251202_142542
+
+## Results
+- **Best Accuracy**: 44.60%
+- **Dataset**: cifar100
+- **Epochs**: 120
+- **Model Version**: V2
+
+## Model Config
+| Parameter | Value |
+|-----------|-------|
+| Dim | 512 |
+| Layers | 4 |
+| Heads | 16 |
+| Patch Size | 4 |
+
+## Wormhole Routing (V2)
+| Parameter | Value |
+|-----------|-------|
+| Mode | hybrid |
+| Wormholes/Position | 8 |
+| Temperature | 0.1 |
+| Tiles | 16 |
+| Tile Wormholes | 4 |
+
+## Crystal Head
+| Parameter | Value |
+|-----------|-------|
+| Scales | [64, 128, 256, 512] |
+| Scale Copies | [8, 4, 2, 2] |
+| Weighting Mode | geometric |
+| Belly Layers | 2 |
+| Belly Residual | False |
+| Use Spine | True |
+| Use Collective | False |
+
+
+## Training Config
+| Parameter | Value |
+|-----------|-------|
+| Learning Rate | 0.0003 |
+| Weight Decay | 0.05 |
+| Batch Size | 128 |
+| CE Weight | 1.0 |
+| Contrast Weight | 0.5 |
+
+## Augmentation
+| Parameter | Value |
+|-----------|-------|
+| Normalization | standard |
+| Mixup Alpha | 0.0 |
+| CutMix Alpha | 0.0 |
+| AlphaMix | False |
+| Label Smoothing | 0.1 |
+
+
+## Key Findings Applied
+- Routing learns from task pressure (no auxiliary routing losses)
+- Gradients verified to flow through router
+- Cross-contrastive aligns patch↔scale features

weights/run_003_v2_8redundantscale_4x4patch_d768_4layer_topo_dropout_cutmix_20251202_142542/best.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2a2e78a5777dee813459822c3537ce260c49345ae948f6fc4635652db381ee0
+size 98300616

weights/run_003_v2_8redundantscale_4x4patch_d768_4layer_topo_dropout_cutmix_20251202_142542/config.json

@@ -0,0 +1,62 @@
+{
+  "architecture": "DavidBeans_V2",
+  "model_type": "david_beans_v2",
+  "image_size": 32,
+  "patch_size": 4,
+  "in_channels": 3,
+  "dim": 512,
+  "num_layers": 4,
+  "num_heads": 16,
+  "mlp_ratio": 4.0,
+  "num_wormholes": 8,
+  "wormhole_temperature": 0.1,
+  "wormhole_mode": "hybrid",
+  "cantor_weight": 0.3,
+  "num_tiles": 16,
+  "tile_wormholes": 4,
+  "scales": [
+    64,
+    128,
+    256,
+    512
+  ],
+  "num_classes": 100,
+  "use_belly": true,
+  "belly_expand": 2.0,
+  "belly_layers": 2,
+  "belly_residual": false,
+  "weighting_mode": "geometric",
+  "scale_weight_floor": 0.1,
+  "scale_copies": [
+    8,
+    4,
+    2,
+    2
+  ],
+  "copy_theta_step": 0.15,
+  "use_collective": false,
+  "collective_temperature": 0.07,
+  "use_spine": true,
+  "spine_channels": [
+    64,
+    128,
+    256
+  ],
+  "spine_cross_attn": true,
+  "spine_gate_init": 0.0,
+  "contrast_temperature": 0.07,
+  "contrast_weight": 0.5,
+  "routing_weight": 0.0,
+  "dropout": 0.1,
+  "use_topo_dropout": true,
+  "topo_drop_prob": 0.15,
+  "topo_warmup_epochs": 35,
+  "topo_min_routes_keep": 2,
+  "topo_steps_per_epoch": 391,
+  "use_spatial_dropout": true,
+  "spatial_drop_prob": 0.1,
+  "spatial_patch_size": 4,
+  "route_dim_attention": -2,
+  "route_dim_tiles": 2,
+  "pooling": "cls"
+}

weights/run_003_v2_8redundantscale_4x4patch_d768_4layer_topo_dropout_cutmix_20251202_142542/tensorboard/events.out.tfevents.1764685544.7ccc73cc3f03.364.2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06e189f7e84f2e0be759f5128749a6e2210eaec4f2a96971c19426f6442613fa
+size 36912

weights/run_003_v2_8redundantscale_4x4patch_d768_4layer_topo_dropout_cutmix_20251202_142542/training_config.json

@@ -0,0 +1,43 @@
+{
+  "run_name": "8redundantscale_4x4patch_d768_4layer_topo_dropout_cutmix",
+  "run_number": null,
+  "model_version": 2,
+  "dataset": "cifar100",
+  "image_size": 32,
+  "batch_size": 128,
+  "num_workers": 4,
+  "normalization": "standard",
+  "epochs": 120,
+  "warmup_epochs": 10,
+  "learning_rate": 0.0003,
+  "weight_decay": 0.05,
+  "betas": [
+    0.9,
+    0.999
+  ],
+  "scheduler": "cosine",
+  "min_lr": 1e-06,
+  "ce_weight": 1.0,
+  "contrast_weight": 0.5,
+  "gradient_clip": 1.0,
+  "label_smoothing": 0.1,
+  "use_augmentation": true,
+  "mixup_alpha": 0.0,
+  "cutmix_alpha": 0.0,
+  "use_alphamix": false,
+  "alphamix_alpha_range": [
+    0.3,
+    0.7
+  ],
+  "alphamix_spatial_ratio": 0.25,
+  "save_interval": 10,
+  "output_dir": "./checkpoints/cifar100_v2",
+  "resume_from": null,
+  "use_tensorboard": true,
+  "log_interval": 50,
+  "log_routing": true,
+  "push_to_hub": true,
+  "hub_repo_id": "AbstractPhil/geovit-david-beans",
+  "hub_private": false,
+  "device": "cuda"
+}

weights/run_003_v2_8redundantscale_4x4patch_d768_4layer_topo_dropout_cutmix_20251202_142542/training_history.json

@@ -0,0 +1,46 @@
+{
+  "loss": [
+    13.873117962861672,
+    12.275092814518855,
+    11.182369144146259,
+    10.339609992198454,
+    9.659815986339863,
+    9.001137947424864,
+    8.437557945495996,
+    8.010158390876574,
+    7.641950374994522
+  ],
+  "ce": [
+    4.3127842603585655,
+    3.8434408181752913,
+    3.505825737806467,
+    3.2513973236083986,
+    3.0415021242239537,
+    2.835707812431531,
+    2.657775013263409,
+    2.5238647595430033,
+    2.408466780797029
+  ],
+  "contrast": [
+    4.463720106467223,
+    3.955229284824469,
+    3.6059489427468714,
+    3.3343552149259126,
+    3.1169237479185448,
+    2.90509451230367,
+    2.724379973533826,
+    2.5893612372569548,
+    2.471492732488192
+  ],
+  "lr": [
+    2.9999999999999997e-05,
+    5.9999999999999995e-05,
+    8.999999999999999e-05,
+    0.00011999999999999999,
+    0.00015,
+    0.00017999999999999998,
+    0.00020999999999999995,
+    0.00023999999999999998,
+    0.00026999999999999995
+  ]
+}