Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,57 @@
|
|
| 1 |
---
|
| 2 |
license: mit
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: mit
|
| 3 |
---
|
| 4 |
+
|
| 5 |
+
# MobileNetV3 Backbone
|
| 6 |
+
|
| 7 |
+
A feature extractor based on [MobileNetV3](https://arxiv.org/abs/1905.02244), refactored from https://github.com/xiaolai-sqlai/mobilenetv3: **backbone only — the classification head is removed**, with unified naming.
|
| 8 |
+
|
| 9 |
+
## Models
|
| 10 |
+
|
| 11 |
+
| Model | Variant | Output features | Weight file |
|
| 12 |
+
| --- | --- | --- | --- |
|
| 13 |
+
| `MobileNetV3` | small / large (auto) | 1280 | detected from checkpoint |
|
| 14 |
+
| `MobileNetV3_Small` | Small | 1280 | `mobilenetv3_small.safetensors` |
|
| 15 |
+
| `MobileNetV3_Large` | Large | 1280 | `mobilenetv3_large.safetensors` |
|
| 16 |
+
|
| 17 |
+
Input: `(B, 3, H, W)` images. Output: `(B, 1280)` feature vectors (migrated from ImageNet-pretrained weights).
|
| 18 |
+
|
| 19 |
+
## Weight naming
|
| 20 |
+
|
| 21 |
+
The naming differences between the old and the new code are mapped automatically during migration:
|
| 22 |
+
|
| 23 |
+
| Old name | New name |
|
| 24 |
+
| --- | --- |
|
| 25 |
+
| `bn1/bn2/bn3` | `norm1/norm2/norm3` |
|
| 26 |
+
| `linear3` | `proj` |
|
| 27 |
+
| `Block.se.se.*` | `Block.se.features.*` |
|
| 28 |
+
| `linear4` (classification head) | removed |
|
| 29 |
+
|
| 30 |
+
## Usage
|
| 31 |
+
|
| 32 |
+
```python
|
| 33 |
+
from mobile_net import MobileNetV3_Small, MobileNetV3_Large
|
| 34 |
+
|
| 35 |
+
fe = MobileNetV3_Large().load_pretrained('./mobilenetv3_large.safetensors')
|
| 36 |
+
fe.eval()
|
| 37 |
+
|
| 38 |
+
fe.save_pretrained('./backbone.safetensors')
|
| 39 |
+
fe.save_pretrained('./backbone.pth')
|
| 40 |
+
|
| 41 |
+
import torch
|
| 42 |
+
x = torch.randn(2, 3, 224, 224)
|
| 43 |
+
feat = fe(x) # (2, 1280)
|
| 44 |
+
```
|
| 45 |
+
|
| 46 |
+
### Auto backend detection
|
| 47 |
+
|
| 48 |
+
`MobileNetV3.from_pretrained` inspects the checkpoint and picks Small / Large automatically:
|
| 49 |
+
|
| 50 |
+
```python
|
| 51 |
+
from mobile_net import MobileNetV3
|
| 52 |
+
|
| 53 |
+
fe = MobileNetV3.from_pretrained('./mobilenetv3_large.safetensors') # -> Large, weights loaded
|
| 54 |
+
print(fe.backend) # 'large'
|
| 55 |
+
```
|
| 56 |
+
|
| 57 |
+
Calling it on a pinned class (`MobileNetV3_Large` / `MobileNetV3_Small`) raises if the checkpoint backend disagrees with the class.
|