AbstractPhil commited on
Commit
61808e9
·
verified ·
1 Parent(s): 1880e0e

Update README - Run 20251012_040642

Browse files
Files changed (1) hide show
  1. README.md +14 -16
README.md CHANGED
@@ -12,7 +12,7 @@ datasets:
12
  metrics:
13
  - accuracy
14
  model-index:
15
- - name: David-decoupled-deep_efficiency
16
  results:
17
  - task:
18
  type: image-classification
@@ -21,7 +21,7 @@ model-index:
21
  type: imagenet-1k
22
  metrics:
23
  - type: accuracy
24
- value: 75.53
25
  ---
26
 
27
  # David: Multi-Scale Crystal Classifier
@@ -32,16 +32,16 @@ as class prototypes with role-weighted similarity computation (Rose Loss).
32
  ## Model Details
33
 
34
  ### Architecture
35
- - **Preset**: high_accuracy
36
- - **Sharing Mode**: decoupled
37
  - **Fusion Mode**: deep_efficiency
38
- - **Scales**: [256, 512, 768, 1024, 1280]
39
- - **Feature Dim**: 512
40
  - **Parameters**: ~8.8M
41
 
42
  ### Training Configuration
43
  - **Dataset**: AbstractPhil/imagenet-clip-features-orderly
44
- - **Model Variant**: clip_vit_b16
45
  - **Epochs**: 20
46
  - **Batch Size**: 1024
47
  - **Learning Rate**: 0.01
@@ -51,14 +51,12 @@ as class prototypes with role-weighted similarity computation (Rose Loss).
51
  ## Performance
52
 
53
  ### Best Results
54
- - **Validation Accuracy**: 75.53%
55
- - **Best Epoch**: 8
56
- - **Final Train Accuracy**: 76.21%
57
 
58
  ### Per-Scale Performance
59
- - **Scale 256**: 75.53%
60
- - **Scale 512**: 77.55%
61
- - **Scale 768**: 77.80%
62
 
63
 
64
  ## Usage
@@ -124,7 +122,7 @@ with torch.no_grad():
124
  ## Architecture Overview
125
 
126
  ### Multi-Scale Processing
127
- David processes inputs at multiple scales (256, 512, 768, 1024, 1280),
128
  allowing it to capture both coarse and fine-grained features.
129
 
130
  ### Crystal Geometry
@@ -166,7 +164,7 @@ score = w_anchor * sim(z, anchor) + w_need * sim(z, need) + ...
166
  author = {AbstractPhil},
167
  year = {2025},
168
  url = {https://huggingface.co/AbstractPhil/gated-david},
169
- note = {Run ID: 20251012_032356}
170
  }
171
  ```
172
 
@@ -181,4 +179,4 @@ Special thanks to Claude (Anthropic) for debugging assistance.
181
 
182
  ---
183
 
184
- *Generated on 2025-10-12 03:54:33*
 
12
  metrics:
13
  - accuracy
14
  model-index:
15
+ - name: David-partial_shared-deep_efficiency
16
  results:
17
  - task:
18
  type: image-classification
 
21
  type: imagenet-1k
22
  metrics:
23
  - type: accuracy
24
+ value: 79.49
25
  ---
26
 
27
  # David: Multi-Scale Crystal Classifier
 
32
  ## Model Details
33
 
34
  ### Architecture
35
+ - **Preset**: clip_vit_l14
36
+ - **Sharing Mode**: partial_shared
37
  - **Fusion Mode**: deep_efficiency
38
+ - **Scales**: [384, 768, 1024, 1280]
39
+ - **Feature Dim**: 768
40
  - **Parameters**: ~8.8M
41
 
42
  ### Training Configuration
43
  - **Dataset**: AbstractPhil/imagenet-clip-features-orderly
44
+ - **Model Variant**: clip_vit_l14
45
  - **Epochs**: 20
46
  - **Batch Size**: 1024
47
  - **Learning Rate**: 0.01
 
51
  ## Performance
52
 
53
  ### Best Results
54
+ - **Validation Accuracy**: 79.49%
55
+ - **Best Epoch**: 0
56
+ - **Final Train Accuracy**: 67.10%
57
 
58
  ### Per-Scale Performance
59
+ - **Scale 384**: 79.49%
 
 
60
 
61
 
62
  ## Usage
 
122
  ## Architecture Overview
123
 
124
  ### Multi-Scale Processing
125
+ David processes inputs at multiple scales (384, 768, 1024, 1280),
126
  allowing it to capture both coarse and fine-grained features.
127
 
128
  ### Crystal Geometry
 
164
  author = {AbstractPhil},
165
  year = {2025},
166
  url = {https://huggingface.co/AbstractPhil/gated-david},
167
+ note = {Run ID: 20251012_040642}
168
  }
169
  ```
170
 
 
179
 
180
  ---
181
 
182
+ *Generated on 2025-10-12 04:09:09*