Upload src/models/unetpp.py with huggingface_hub
Browse files- src/models/unetpp.py +41 -0
src/models/unetpp.py
ADDED
|
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""DiaFoot.AI v2 — U-Net++ via Segmentation Models PyTorch.
|
| 2 |
+
|
| 3 |
+
Phase 2, Commit 9: Baseline single-task segmentation model.
|
| 4 |
+
"""
|
| 5 |
+
|
| 6 |
+
from __future__ import annotations
|
| 7 |
+
|
| 8 |
+
import segmentation_models_pytorch as smp
|
| 9 |
+
import torch.nn as nn # noqa: TC002
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
def build_unetpp(
|
| 13 |
+
encoder_name: str = "efficientnet-b4",
|
| 14 |
+
encoder_weights: str | None = "imagenet",
|
| 15 |
+
in_channels: int = 3,
|
| 16 |
+
classes: int = 1,
|
| 17 |
+
decoder_attention_type: str | None = "scse",
|
| 18 |
+
deep_supervision: bool = True,
|
| 19 |
+
) -> nn.Module:
|
| 20 |
+
"""Build a U-Net++ model via SMP.
|
| 21 |
+
|
| 22 |
+
Args:
|
| 23 |
+
encoder_name: Encoder backbone name (from timm/SMP).
|
| 24 |
+
encoder_weights: Pretrained weights ('imagenet' or None).
|
| 25 |
+
in_channels: Number of input channels.
|
| 26 |
+
classes: Number of output segmentation classes.
|
| 27 |
+
decoder_attention_type: Attention type ('scse' or None).
|
| 28 |
+
deep_supervision: Enable deep supervision for better gradients.
|
| 29 |
+
|
| 30 |
+
Returns:
|
| 31 |
+
SMP UnetPlusPlus model.
|
| 32 |
+
"""
|
| 33 |
+
return smp.UnetPlusPlus(
|
| 34 |
+
encoder_name=encoder_name,
|
| 35 |
+
encoder_weights=encoder_weights,
|
| 36 |
+
in_channels=in_channels,
|
| 37 |
+
classes=classes,
|
| 38 |
+
decoder_attention_type=decoder_attention_type,
|
| 39 |
+
encoder_depth=5,
|
| 40 |
+
decoder_channels=(256, 128, 64, 32, 16),
|
| 41 |
+
)
|