Update README.md
Browse files
README.md
CHANGED
|
@@ -34,9 +34,35 @@ model-index:
|
|
| 34 |
|
| 35 |
**Geometric Basin Classification for CIFAR-100**
|
| 36 |
|
| 37 |
-
🚧 **Training in Progress** 🚧
|
| 38 |
|
| 39 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
|
| 41 |
---
|
| 42 |
|
|
|
|
| 34 |
|
| 35 |
**Geometric Basin Classification for CIFAR-100**
|
| 36 |
|
|
|
|
| 37 |
|
| 38 |
+
## Immediate Assessment
|
| 39 |
+
|
| 40 |
+
The geo-beatrix variation is more capable at classification and inferior at geometric capacity than the vit-beatrix.
|
| 41 |
+
|
| 42 |
+
geo-beatrix has a different form of math and a new basin format entirely dependent on teaching traditional structures new behavior.
|
| 43 |
+
|
| 44 |
+
The system is hit or miss, and will be refined over time as the model family evolves.
|
| 45 |
+
|
| 46 |
+
The reality sets in when the classification gets higher, that this is a more capable model than a vit - and yet that vit I built has a far more robust set of tooling and capacity for learning transfer.
|
| 47 |
+
|
| 48 |
+
I'd say this is too big for standard classification tasks, and yet the classifier system does work somewhat - just not as well as SOTA.
|
| 49 |
+
|
| 50 |
+
### Conclusion based on experimentation
|
| 51 |
+
|
| 52 |
+
This requires more experimentation on the subsystem before it can be utilized correctly.
|
| 53 |
+
Optimizations need to happen to components, certain pieces need to be baselined to torch components for faster iterations.
|
| 54 |
+
Even with loops removed this still has some issues with cantor stairs, but the batched stairs will be available on my repo today as well as the full model structure for the family of three here.
|
| 55 |
+
|
| 56 |
+
Alphamix and Fractalmix are hit-or-miss even with Cantor stairs, sometimes improving fidelity, sometimes reducing it.
|
| 57 |
+
|
| 58 |
+
Lacking attention mechanisms I consider this a resounding success as an experiment, and yet it fell short of resnet18 and resnet32 standalones - meaning the head only converted the math into something else, and fell short of the crossentropy goal.
|
| 59 |
+
|
| 60 |
+
That's okay though, I will refine the processes, improve the system, and return with additional trains for this version to further improve classifcation beyond the 69% chance - which may be HIGHER than the vit-beatrix, but it's considerably more shallow in comparison of geometric cohesion than the dual-stream transformer variation.
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
🚧 **Training Concluded** 🚧
|
| 64 |
+
|
| 65 |
+
Current Status: Idle
|
| 66 |
|
| 67 |
---
|
| 68 |
|