Fix fp16 compatibility issues in multiple segmentation models

- IDRNet: Change negative value from -1e16 to -1e4 to prevent fp16 overflow
- ISNet: Fix tensor creation to ensure proper device and dtype handling
- OCRNet: Fix spatial gather module input size mismatch by interpolating seg_logits_aux

These changes ensure models work correctly with fp16 (half precision) training.

🤖 Generated with [Claude Code](https://claude.ai/code)

Co-Authored-By: Claude <noreply@anthropic.com>

Author: sieu-n

ssseg/modules/models/segmentors/idrnet/idrnet.py

@@ -323,7 +323,7 @@ def obtainidcontext(self, context, logits, class_relations, intervention_clsids=
                 cls_contexts = torch.stack(cls_contexts)
                 selected_class_relations = torch.cat(selected_class_relations, dim=1)
                 if remove_negative_cls_relation:
-                    selected_class_relations[selected_class_relations <= 0] = -1e16
+                    selected_class_relations[selected_class_relations <= 0] = -1e4
                 selected_class_relations = F.softmax(selected_class_relations, dim=1)
                 selected_class_relations_tmp = []
                 for cls_id in valid_clsids:

ssseg/modules/models/segmentors/isnet/semanticlevel.py

@@ -31,7 +31,7 @@ def forward(self, x, preds, feats_il):
         inputs = x
         batch_size, num_channels, h, w = x.size()
         num_classes = preds.size(1)
-        feats_sl = torch.zeros(batch_size, h*w, num_channels).type_as(x)
+        feats_sl = torch.zeros(batch_size, h*w, num_channels, device=x.device, dtype=torch.float32)
         for batch_idx in range(batch_size):
             # (C, H, W), (num_classes, H, W) --> (H*W, C), (H*W, num_classes)
             feats_iter, preds_iter = x[batch_idx], preds[batch_idx]

ssseg/modules/models/segmentors/ocrnet/ocrnet.py

@@ -56,7 +56,7 @@ def forward(self, data_meta):
         # feed to bottleneck
         feats = self.bottleneck(backbone_outputs[-1])
         # feed to ocr module
-        context = self.spatial_gather_module(feats, seg_logits_aux)
+        context = self.spatial_gather_module(feats, F.interpolate(seg_logits_aux, size=feats.shape[2:]))
         feats = self.object_context_block(feats, context)
         # feed to decoder
         seg_logits = self.decoder(feats)