geo-beatrix-resnet

geo-beatrix-resnet34-step20-feats1000

Geometric Basin Classification for CIFAR-100

🎉 Training Complete 🎉

Final Status: Epoch 200/200

---

Current Performance

All Training Runs

Autogen bug, they all have different test accs.

Comparison to State-of-the-Art

🎯 Current target: Beat geo-beatrix (69.0%) - Currently -12.88%

---

Architecture

• Base: ResNet34 (torchvision)
• Pretrained: From scratch
• Features: 512-dim from ResNet34
• Positional Encoding: Devil's Staircase (Cantor function, 1883)
• PE Levels: 20
• PE Features/Level: 1000
• Classification: Geometric Basin Compatibility (NO cross-entropy)
• Attention Mechanisms: NONE
• Mixing: Standard (single patch)

---

Training Configuration

{
  "model_name": "geo-beatrix-resnet34-step20-feats1000",
  "model_type": "geometric_basin_classifier",
  "num_classes": 100,
  "batch_size": 512,
  "num_epochs": 200,
  "base_learning_rate": 0.001,
  "weight_decay": 0.05,
  "warmup_epochs": 10,
  "pe_levels": 20,
  "pe_features_per_level": 1000,
  "dropout": 0.1,
  "pretrained_resnet": false,
  "frozen_resnet": false,
  "a100_optimizations": {
    "mixed_precision": true,
    "torch_compile": false,
    "channels_last": true,
    "gradient_checkpointing": false
  },
  "alphamix": {
    "enabled": true,
    "fractal_mode": false,
    "range": [
      0.3,
      0.7
    ],
    "spatial_ratio": 0.1,
    "curriculum_start": 0.0,
    "curriculum_end": 0.75,
    "fractal_steps": [
      1,
      3
    ],
    "fractal_scales": [
      0.3333333333333333,
      0.1111111111111111,
      0.037037037037037035
    ]
  },
  "architecture": "ResNet34 + Devil's Staircase PE",
  "loss_function": "Geometric Basin Compatibility",
  "cross_entropy": false,
  "attention_mechanisms": false,
  "timestamp": "20251011_041716"
}

---

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)
├── runs_history.json        (all training runs and their results)
├── README.md
├── weights/geo-beatrix-resnet34-step20-feats1000/20251011_041716/
│   ├── model.pt                 (best from this training run)
│   ├── model.safetensors        (best from this training run)
│   ├── config.json
│   ├── training_log.txt
│   └── checkpoints/
│       ├── checkpoint_epoch_50.safetensors
│       ├── checkpoint_epoch_100.safetensors
│       └── checkpoint_epoch_150.safetensors
│       (snapshots every 10 epochs)
└── runs/geo-beatrix-resnet34-step20-feats1000/20251011_041716/
    ├── 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-resnet",
    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-resnet",
    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-resnet",
    filename="weights/geo-beatrix-resnet34-step20-feats1000/20251011_041716/model.safetensors"
)

# Download specific epoch checkpoint
epoch_checkpoint = hf_hub_download(
    repo_id="AbstractPhil/geo-beatrix-resnet",
    filename="weights/geo-beatrix-resnet34-step20-feats1000/20251011_041716/checkpoints/checkpoint_epoch_100.safetensors"
)

---

Training History

Best Checkpoint

• Epoch: 160
• Train Acc: 59.43%
• Test Acc: 51.64%
• Alpha: 0.4071
• Loss: 0.7570

Latest 5 Epochs

• Epoch 196: Train 62.03%, Test 0.00%, α=0.4032, Loss=0.7300
• Epoch 197: Train 59.02%, Test 0.00%, α=0.4031, Loss=0.6201
• Epoch 198: Train 58.49%, Test 0.00%, α=0.4031, Loss=0.6571
• Epoch 199: Train 59.32%, Test 0.00%, α=0.4031, Loss=0.6543
• Epoch 200: Train 59.29%, Test 51.51%, α=0.4031, Loss=0.6505

Training Milestones

• 🎯 50% Accuracy reached at epoch 120
• 📊 α ≥ 0.40 reached at epoch 17

---

Innovation

✅ NO attention mechanisms
✅ NO cross-entropy loss
✅ Fractal positional encoding (Cantor function from 1883)
✅ Geometric compatibility classification
✅ ResNet34 backbone (proven CNN architecture)

---

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