AbstractPhil commited on
Commit
5dc2e68
·
verified ·
1 Parent(s): 9d2bfaa

Update beatrix-cifar100 - Epoch 0 - Acc: 0.0801

Browse files
README.md CHANGED
@@ -11,28 +11,36 @@ license: mit
11
 
12
  # ViT-Beatrix: Fractal PE + Geometric Simplex Vision Transformer
13
 
14
- This model integrates Devil's Staircase positional encoding with geometric simplex features
15
- for vision tasks. Trained on CIFAR-100.
16
 
17
- ## Model Details
 
 
 
 
 
 
 
 
 
18
 
19
  - **Architecture**: Vision Transformer with fractal positional encoding
20
  - **Dataset**: CIFAR-100 (100 classes)
21
- - **Embedding Dimension**: 512
22
- - **Depth**: 4 layers
23
  - **Patch Size**: 4x4
24
  - **PE Levels**: 12
25
  - **Simplex Dimension**: 5-simplex
26
 
27
- ## Training
28
 
29
- - **Dataset**: CIFAR-100
30
- - **Epochs**: 7
31
- - **Best Accuracy**: 0.2841
32
  - **Batch Size**: 512
33
  - **Learning Rate**: 0.001
34
 
35
- ## Loss Components
36
 
37
  - Task Loss Weight: 1.0
38
  - Flow Alignment Weight: 0.5
@@ -46,18 +54,18 @@ from geovocab2.train.model.vit_beatrix import SimplifiedGeometricClassifier
46
  from safetensors.torch import load_file
47
  from huggingface_hub import hf_hub_download
48
 
49
- # Download weights from Hub
50
  weights_path = hf_hub_download(
51
  repo_id="AbstractPhil/vit-beatrix",
52
- filename="weights/beatrix-cifar100/20251007_181852/model.safetensors"
53
  )
54
 
55
  # Load model
56
  model = SimplifiedGeometricClassifier(
57
  num_classes=100,
58
  img_size=32,
59
- embed_dim=512,
60
- depth=4
61
  )
62
 
63
  # Load weights
 
11
 
12
  # ViT-Beatrix: Fractal PE + Geometric Simplex Vision Transformer
13
 
14
+ This repository contains Vision Transformers integrating Devil's Staircase positional encoding
15
+ with geometric simplex features for vision tasks.
16
 
17
+ ## Available Models (Best Checkpoints Only)
18
+
19
+ | Model Name | Accuracy | Epoch | Timestamp | Path |
20
+ |------------|----------|-------|-----------|------|
21
+ | beatrix-cifar100 | 0.0801 | 0 | 20251007_182851 | `weights/beatrix-cifar100/20251007_182851` |
22
+
23
+
24
+ ## Latest Updated Model: beatrix-cifar100
25
+
26
+ ### Model Details
27
 
28
  - **Architecture**: Vision Transformer with fractal positional encoding
29
  - **Dataset**: CIFAR-100 (100 classes)
30
+ - **Embedding Dimension**: 256
31
+ - **Depth**: 12 layers
32
  - **Patch Size**: 4x4
33
  - **PE Levels**: 12
34
  - **Simplex Dimension**: 5-simplex
35
 
36
+ ### Training Details
37
 
38
+ - **Best Accuracy**: 0.0801
39
+ - **Epochs Trained**: 0
 
40
  - **Batch Size**: 512
41
  - **Learning Rate**: 0.001
42
 
43
+ ### Loss Configuration
44
 
45
  - Task Loss Weight: 1.0
46
  - Flow Alignment Weight: 0.5
 
54
  from safetensors.torch import load_file
55
  from huggingface_hub import hf_hub_download
56
 
57
+ # Download weights from Hub (example for beatrix-cifar100)
58
  weights_path = hf_hub_download(
59
  repo_id="AbstractPhil/vit-beatrix",
60
+ filename="weights/beatrix-cifar100/20251007_182851/model.safetensors"
61
  )
62
 
63
  # Load model
64
  model = SimplifiedGeometricClassifier(
65
  num_classes=100,
66
  img_size=32,
67
+ embed_dim=256,
68
+ depth=12
69
  )
70
 
71
  # Load weights
manifest.json ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "beatrix-cifar100": {
3
+ "accuracy": 0.080078125,
4
+ "epoch": 0,
5
+ "timestamp": "20251007_182851",
6
+ "path": "weights/beatrix-cifar100/20251007_182851",
7
+ "embed_dim": 256,
8
+ "depth": 12,
9
+ "batch_size": 512,
10
+ "learning_rate": 0.001
11
+ }
12
+ }
weights/beatrix-cifar100/20251007_182851/config.json ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "num_classes": 100,
3
+ "img_size": 32,
4
+ "patch_size": 4,
5
+ "embed_dim": 256,
6
+ "k_simplex": 5,
7
+ "depth": 12,
8
+ "num_heads": 8,
9
+ "mlp_ratio": 4.0,
10
+ "dropout": 0.0,
11
+ "pe_levels": 12,
12
+ "pe_features_per_level": 2,
13
+ "pe_smooth_tau": 0.25,
14
+ "simplex_feature_weight": 0.2,
15
+ "batch_size": 512,
16
+ "num_epochs": 50,
17
+ "learning_rate": 0.001,
18
+ "weight_decay": 0.005,
19
+ "warmup_epochs": 10,
20
+ "task_loss_weight": 1.0,
21
+ "flow_loss_weight": 0.5,
22
+ "coherence_loss_weight": 0.3,
23
+ "multiscale_loss_weight": 0.2,
24
+ "volume_reg_weight": 0.1,
25
+ "use_augmentation": true,
26
+ "mixup_alpha": 0.2,
27
+ "cutmix_alpha": 1.0,
28
+ "device": "cuda",
29
+ "num_workers": 2,
30
+ "pin_memory": true,
31
+ "save_dir": "./checkpoints",
32
+ "save_every": 10,
33
+ "use_safetensors": true,
34
+ "timestamp_dirs": true,
35
+ "push_to_hub": true,
36
+ "hub_model_id": "AbstractPhil/vit-beatrix",
37
+ "hub_model_name": "beatrix-cifar100",
38
+ "hub_upload_best_only": true,
39
+ "hub_upload_every_n_epochs": 5,
40
+ "use_tensorboard": true,
41
+ "log_dir": "./logs",
42
+ "log_every": 50,
43
+ "train_baseline": false
44
+ }
weights/beatrix-cifar100/20251007_182851/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ab12891e7b1a033c9b4263611a097e576da18e6ab0ac3c4f2c520a1fee6bd705
3
+ size 38117196