AbstractPhil / geo-beatrix

geo-beatrix

Geometric Basin Classification for CIFAR-100

🚧 Training in Progress 🚧

Current Status: Epoch 100/200

Current Performance

Comparison to State-of-the-Art

🎯 Current target : Beat vit-beatrix (66.0%) - Currently -2.12%

Architecture

• Base : ResNet-style with residual blocks
• Channels : 64 → 128 → 256 → 512 → 1024
• Positional Encoding : Devil's Staircase (Cantor function, 1883)
• PE Levels : 20
• PE Features/Level : 10
• Classification : Geometric Basin Compatibility (NO cross-entropy)
• Attention Mechanisms : NONE

Training Configuration

{
  "model_name": "geo-beatrix",
  "model_type": "geometric_basin_classifier",
  "num_classes": 100,
  "batch_size": 256,
  "num_epochs": 200,
  "base_learning_rate": 0.001,
  "weight_decay": 0.05,
  "warmup_epochs": 10,
  "pe_levels": 20,
  "pe_features_per_level": 10,
  "dropout": 0.1,
  "upload_every_n_epochs": 50,
  "alphamix": {
    "enabled": true,
    "range": [
      0.3,
      0.7
    ],
    "spatial_ratio": 0.25,
    "curriculum_start": 0.0,
    "curriculum_end": 0.5
  },
  "architecture": "ResNet-style with Devil's Staircase PE",
  "loss_function": "Geometric Basin Compatibility",
  "cross_entropy": false,
  "attention_mechanisms": false,
  "timestamp": "20251009_234907"
}

Files Structure

├── model.pt                 (BEST overall model - easy access!)
├── model.safetensors        (BEST overall model - easy access!)
├── best_model_info.json     (which epoch/run this came from)
├── README.md
├── weights/geo-beatrix/20251009_234907/
│   ├── model.pt                 (best from this training run)
│   ├── model.safetensors        (best from this training run)
│   ├── config.json
│   ├── training_log.txt
│   └── checkpoints/
│       ├── checkpoint_epoch_10.safetensors
│       ├── checkpoint_epoch_20.safetensors
│       └── checkpoint_epoch_30.safetensors
│       (snapshots every 50 epochs)
└── runs/geo-beatrix/20251009_234907/
    ├── events.out.tfevents.*    (TensorBoard logs)
    └── metrics.csv              (training metrics)

Note : The root model.pt and model.safetensors always contain the best model across all training runs!

Usage

from huggingface_hub import hf_hub_download
import torch

# EASIEST: Download BEST overall model from root (recommended!)
from safetensors.torch import load_file
model_path = hf_hub_download(
    repo_id="AbstractPhil/geo-beatrix",
    filename="model.safetensors"
)
state_dict = load_file(model_path)
# model.load_state_dict(state_dict)

# Check which epoch/run the best model came from
info_path = hf_hub_download(
    repo_id="AbstractPhil/geo-beatrix",
    filename="best_model_info.json"
)
with open(info_path) as f:
    best_info = json.load(f)
    print(f"Best model: epoch {best_info['epoch']}, {best_info['test_accuracy']:.2f}%")

# Or download from specific training run
model_path = hf_hub_download(
    repo_id="AbstractPhil/geo-beatrix",
    filename="weights/geo-beatrix/20251009_234907/model.safetensors"
)

# Download specific epoch checkpoint
epoch_checkpoint = hf_hub_download(
    repo_id="AbstractPhil/geo-beatrix",
    filename="weights/geo-beatrix/20251009_234907/checkpoints/checkpoint_epoch_100.safetensors"
)

Training History

Best Checkpoint

• Epoch: 95
• Train Acc: 76.13%
• Test Acc: 63.88%
• Alpha: 0.4349
• Loss: 0.7637

Latest 5 Epochs

• Epoch 96 : Train 71.44%, Test 0.00%, α=0.4352, Loss=0.7272
• Epoch 97 : Train 73.64%, Test 0.00%, α=0.4369, Loss=0.7389
• Epoch 98 : Train 75.14%, Test 0.00%, α=0.4359, Loss=0.7399
• Epoch 99 : Train 71.31%, Test 0.00%, α=0.4328, Loss=0.7187
• Epoch 100 : Train 69.33%, Test 63.77%, α=0.4342, Loss=0.7044

Training Milestones

• 🎯 50% Accuracy reached at epoch 35
• 🎯 60% Accuracy reached at epoch 55
• 📊 α ≥ 0.40 reached at epoch 9

Innovation

✅ NO attention mechanisms
✅ NO cross-entropy loss
✅ Fractal positional encoding (Cantor function from 1883)
✅ Geometric compatibility classification
✅ Ancient convolutions (ResNet-style from 1990s)

Repository : https://huggingface.co/AbstractPhil/geo-beatrix
Author : AbstractPhil
Framework : PyTorch