Simplify note on checkpoint size
Browse files
README.md
CHANGED
|
@@ -27,6 +27,9 @@ Code and examples: [github.com/HaoranLiu507/LPIFM](https://github.com/HaoranLiu5
|
|
| 27 |
|
| 28 |
EVAFusion fine-tuned weights are **not** hosted here; see the Zenodo companion archive when published.
|
| 29 |
|
|
|
|
|
|
|
|
|
|
| 30 |
## Architecture
|
| 31 |
|
| 32 |
- Backbone: **ConvNeXt-V2** (`convnextv2_base.fcmae_ft_in22k_in1k_384` via `timm`)
|
|
|
|
| 27 |
|
| 28 |
EVAFusion fine-tuned weights are **not** hosted here; see the Zenodo companion archive when published.
|
| 29 |
|
| 30 |
+
**Note on checkpoint size.**
|
| 31 |
+
`lpifm_vifb_baseline_v1.pt` is about **2 GB** because it is a full training checkpoint: three weight copies (`swa` / `ema` / `model`, ~0.4 GB each) plus the optimizer (~0.8 GB). To **use LPIFM** only, keep `swa_state_dict` (+ `config`); ~0.4 GB is enough. Inference already loads SWA by default.
|
| 32 |
+
|
| 33 |
## Architecture
|
| 34 |
|
| 35 |
- Backbone: **ConvNeXt-V2** (`convnextv2_base.fcmae_ft_in22k_in1k_384` via `timm`)
|