Epoch 10 - 42.76% acc

weights/run_001_v2_4redundantscale_4x4patch_d512_4layer_alphamix_fractalreg_20251201_181252/README.md

@@ -0,0 +1,60 @@
+# Run: run_001_v2_4redundantscale_4x4patch_d512_4layer_alphamix_fractalreg_20251201_181252
+
+## Results
+- **Best Accuracy**: 42.76%
+- **Dataset**: cifar100
+- **Epochs**: 200
+- **Model Version**: V2
+
+## Model Config
+| Parameter | Value |
+|-----------|-------|
+| Dim | 512 |
+| Layers | 4 |
+| Heads | 8 |
+| 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 | [4, 2, 1, 1] |
+| 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 | 0.2 |
+| CutMix Alpha | 1.0 |
+| AlphaMix | True |
+| 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_001_v2_4redundantscale_4x4patch_d512_4layer_alphamix_fractalreg_20251201_181252/best.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:795912e438d503fe07e95efbea1e6d13c00270c7e7f4356c391936858188e1ac
+size 90024320

weights/run_001_v2_4redundantscale_4x4patch_d512_4layer_alphamix_fractalreg_20251201_181252/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": 8,
+  "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": [
+    4,
+    2,
+    1,
+    1
+  ],
+  "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": 2,
+  "route_dim_attention": -2,
+  "route_dim_tiles": 2,
+  "pooling": "cls"
+}

weights/run_001_v2_4redundantscale_4x4patch_d512_4layer_alphamix_fractalreg_20251201_181252/tensorboard/events.out.tfevents.1764612796.c39ac074cc38.1028.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b69eb6e838ac7b3b392ff4d21d11af49c0cbcad8513bb85e2e52653426fb96f
+size 30281

weights/run_001_v2_4redundantscale_4x4patch_d512_4layer_alphamix_fractalreg_20251201_181252/training_config.json

@@ -0,0 +1,43 @@
+{
+  "run_name": "4redundantscale_4x4patch_d512_4layer_alphamix_fractalreg",
+  "run_number": null,
+  "model_version": 2,
+  "dataset": "cifar100",
+  "image_size": 32,
+  "batch_size": 128,
+  "num_workers": 4,
+  "normalization": "standard",
+  "epochs": 200,
+  "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": 0.2,
+  "cutmix_alpha": 1.0,
+  "use_alphamix": true,
+  "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_001_v2_4redundantscale_4x4patch_d512_4layer_alphamix_fractalreg_20251201_181252/training_history.json

@@ -0,0 +1,46 @@
+{
+  "loss": [
+    10.410657728635348,
+    9.608673051687388,
+    9.134598242930878,
+    8.782778321779691,
+    8.438618367757552,
+    8.11135796033419,
+    7.961723312964806,
+    7.672892722105368,
+    7.5494623196430695
+  ],
+  "ce": [
+    4.413112790767963,
+    4.092811764203585,
+    3.914777290515411,
+    3.7460260110023693,
+    3.581492412395966,
+    3.4646466028996,
+    3.398411958034222,
+    3.2972507623525766,
+    3.2184918000147893
+  ],
+  "contrast": [
+    4.547770396257058,
+    4.203475647706252,
+    3.983390191885141,
+    3.84815374032045,
+    3.714197435745826,
+    3.5545716676956567,
+    3.4918246801082904,
+    3.349622757618244,
+    3.3176276726600453
+  ],
+  "lr": [
+    2.9999999999999997e-05,
+    5.9999999999999995e-05,
+    8.999999999999999e-05,
+    0.00011999999999999999,
+    0.00015,
+    0.00017999999999999998,
+    0.00020999999999999995,
+    0.00023999999999999998,
+    0.00026999999999999995
+  ]
+}