Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -23,7 +23,7 @@ This repository contains multiple training runs using Cantor fusion architecture
|
|
| 23 |
```
|
| 24 |
vit-beans-v3/
|
| 25 |
├── runs/
|
| 26 |
-
│ ├──
|
| 27 |
│ │ ├── checkpoints/
|
| 28 |
│ │ │ ├── best_model.safetensors
|
| 29 |
│ │ │ ├── best_training_state.pt
|
|
@@ -31,17 +31,17 @@ vit-beans-v3/
|
|
| 31 |
│ │ ├── tensorboard/
|
| 32 |
│ │ ├── config.yaml
|
| 33 |
│ │ └── README.md
|
| 34 |
-
│ ├── cifar100_consciousness_TIMESTAMP/
|
| 35 |
-
│ │ └── ...
|
| 36 |
│ └── ...
|
| 37 |
└── README.md (this file)
|
| 38 |
```
|
| 39 |
|
| 40 |
## Current Run
|
| 41 |
|
| 42 |
-
**Latest**: `
|
| 43 |
- **Dataset**: CIFAR100
|
| 44 |
- **Fusion Mode**: weighted
|
|
|
|
|
|
|
| 45 |
- **Architecture**: 6 blocks, 8 heads
|
| 46 |
- **Simplex**: 4-simplex (5 vertices)
|
| 47 |
|
|
@@ -53,6 +53,14 @@ The Cantor Fusion architecture uses:
|
|
| 53 |
- **Beatrix Consciousness Routing**: Optional consciousness-aware token fusion using the Devil's Staircase
|
| 54 |
- **SafeTensors Format**: All model weights use SafeTensors (not pickle) for security
|
| 55 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 56 |
## Usage
|
| 57 |
|
| 58 |
### Download a Model
|
|
@@ -98,8 +106,11 @@ Each run directory contains:
|
|
| 98 |
|
| 99 |
## Training Details
|
| 100 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 101 |
All models trained with:
|
| 102 |
-
- Optimizer: AdamW
|
| 103 |
- Mixed Precision: Available on A100
|
| 104 |
- Augmentation: AutoAugment (CIFAR10 policy)
|
| 105 |
- Format: SafeTensors (ClamAV safe)
|
|
@@ -110,4 +121,4 @@ Built with geometric consciousness-aware routing using the Devil's Staircase (Be
|
|
| 110 |
|
| 111 |
**Repository maintained by**: [@AbstractPhil](https://huggingface.co/AbstractPhil)
|
| 112 |
|
| 113 |
-
**Latest update**: 2025-11-19 17:
|
|
|
|
| 23 |
```
|
| 24 |
vit-beans-v3/
|
| 25 |
├── runs/
|
| 26 |
+
│ ├── cifar10_weighted_SGD_TIMESTAMP/
|
| 27 |
│ │ ├── checkpoints/
|
| 28 |
│ │ │ ├── best_model.safetensors
|
| 29 |
│ │ │ ├── best_training_state.pt
|
|
|
|
| 31 |
│ │ ├── tensorboard/
|
| 32 |
│ │ ├── config.yaml
|
| 33 |
│ │ └── README.md
|
|
|
|
|
|
|
| 34 |
│ └── ...
|
| 35 |
└── README.md (this file)
|
| 36 |
```
|
| 37 |
|
| 38 |
## Current Run
|
| 39 |
|
| 40 |
+
**Latest**: `cifar100_weighted_SGD_20251119_173038`
|
| 41 |
- **Dataset**: CIFAR100
|
| 42 |
- **Fusion Mode**: weighted
|
| 43 |
+
- **Optimizer**: SGD (momentum=0.9)
|
| 44 |
+
- **Scheduler**: MultiStepLR [40, 60, 80]
|
| 45 |
- **Architecture**: 6 blocks, 8 heads
|
| 46 |
- **Simplex**: 4-simplex (5 vertices)
|
| 47 |
|
|
|
|
| 53 |
- **Beatrix Consciousness Routing**: Optional consciousness-aware token fusion using the Devil's Staircase
|
| 54 |
- **SafeTensors Format**: All model weights use SafeTensors (not pickle) for security
|
| 55 |
|
| 56 |
+
## Training Strategy
|
| 57 |
+
|
| 58 |
+
This model uses the proven **SGD + milestone LR drops** strategy from WideResNet:
|
| 59 |
+
- Initial LR: 0.1
|
| 60 |
+
- Milestones: [40, 60, 80]
|
| 61 |
+
- Decay factor: 0.2 (LR *= 0.2 at each milestone)
|
| 62 |
+
- This causes the dramatic accuracy jumps seen in deep networks!
|
| 63 |
+
|
| 64 |
## Usage
|
| 65 |
|
| 66 |
### Download a Model
|
|
|
|
| 106 |
|
| 107 |
## Training Details
|
| 108 |
|
| 109 |
+
Optimizer options:
|
| 110 |
+
- **SGD**: High momentum (0.9), Nesterov, milestone-based LR drops
|
| 111 |
+
- **AdamW**: Weight decay, cosine annealing with warmup
|
| 112 |
+
|
| 113 |
All models trained with:
|
|
|
|
| 114 |
- Mixed Precision: Available on A100
|
| 115 |
- Augmentation: AutoAugment (CIFAR10 policy)
|
| 116 |
- Format: SafeTensors (ClamAV safe)
|
|
|
|
| 121 |
|
| 122 |
**Repository maintained by**: [@AbstractPhil](https://huggingface.co/AbstractPhil)
|
| 123 |
|
| 124 |
+
**Latest update**: 2025-11-19 17:30:40
|