Create sam2_model_stub.py
Browse files- sam2_model_stub.py +45 -0
sam2_model_stub.py
ADDED
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# sam2_model_stub.py
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
import torch.nn.functional as F
|
6 |
+
|
7 |
+
class SAM2Hierarchical(nn.Module):
|
8 |
+
def __init__(self, num_classes=6, in_channels=3, backbone="vit_b", freeze_backbone=True, use_cls_head=True):
|
9 |
+
super().__init__()
|
10 |
+
self.use_cls_head = use_cls_head
|
11 |
+
|
12 |
+
# Minimal vision backbone stub (fake transformer or CNN)
|
13 |
+
self.backbone = nn.Sequential(
|
14 |
+
nn.Conv2d(in_channels, 64, kernel_size=3, stride=2, padding=1),
|
15 |
+
nn.BatchNorm2d(64),
|
16 |
+
nn.ReLU(inplace=True),
|
17 |
+
nn.Conv2d(64, 128, kernel_size=3, stride=2, padding=1),
|
18 |
+
nn.BatchNorm2d(128),
|
19 |
+
nn.ReLU(inplace=True)
|
20 |
+
)
|
21 |
+
|
22 |
+
# Segmentation head stub
|
23 |
+
self.segmentation_head = nn.Sequential(
|
24 |
+
nn.Conv2d(128, 64, kernel_size=3, padding=1),
|
25 |
+
nn.ReLU(inplace=True),
|
26 |
+
nn.Conv2d(64, num_classes, kernel_size=1)
|
27 |
+
)
|
28 |
+
|
29 |
+
# Optional classification head
|
30 |
+
if self.use_cls_head:
|
31 |
+
self.cls_head = nn.Linear(128, num_classes)
|
32 |
+
|
33 |
+
if freeze_backbone:
|
34 |
+
for param in self.backbone.parameters():
|
35 |
+
param.requires_grad = False
|
36 |
+
|
37 |
+
def forward(self, x):
|
38 |
+
features = self.backbone(x)
|
39 |
+
logits = self.segmentation_head(features)
|
40 |
+
|
41 |
+
if self.use_cls_head:
|
42 |
+
# Just return segmentation output; inference only cares about logits
|
43 |
+
return logits
|
44 |
+
|
45 |
+
return logits
|