updates to GT and LG

Author: Phil26AT

gluefactory/geometry/gt_generation.py

@@ -16,7 +16,7 @@ def gt_matches_from_pose_depth(
     epi_th=None,
     cc_th=None,
     min_overlap: float | None = None,
-    add_epi_outliers=True,
+    add_epi_outliers: bool = True,
     **kw,
 ):
     if kp0.shape[1] == 0 or kp1.shape[1] == 0:

gluefactory/models/extractors/mixed.py

@@ -39,16 +39,8 @@ class MixedExtractor(BaseModel):
     def _init(self, conf):
         if conf.detector.name:
             self.detector = get_model(conf.detector.name)(to_ctr(conf.detector))
-        else:
-            self.required_data_keys += ["cache"]
-            self.required_cache_keys += ["keypoints"]
-
         if conf.descriptor.name:
             self.descriptor = get_model(conf.descriptor.name)(to_ctr(conf.descriptor))
-        else:
-            self.required_data_keys += ["cache"]
-            self.required_cache_keys += ["descriptors"]
-
         if conf.refiner.name:
             self.refiner = get_model(conf.refiner.name)(to_ctr(conf.refiner))
 

gluefactory/models/matchers/depth_matcher.py

@@ -37,7 +37,6 @@ def _init(self, conf):
             ]
 
     @misc.filter_batch_for_jit
-    @misc.AMP_CUSTOM_FWD_F32
     def _forward(self, data):
         return self.match_with_depth(data)
 

gluefactory/models/matchers/lightglue.py

@@ -119,7 +119,7 @@ def forward(self, q, k, v, mask: Optional[torch.Tensor] = None) -> torch.Tensor:
         if self.enable_flash and q.device.type == "cuda":
             # use torch 2.0 scaled_dot_product_attention with flash
             if FLASH_AVAILABLE:
-                args = [x.half().contiguous() for x in [q, k, v]]
+                args = [x.contiguous() for x in [q, k, v]]
                 v = F.scaled_dot_product_attention(*args, attn_mask=mask).to(q.dtype)
                 return v if mask is None else v.nan_to_num()
         elif FLASH_AVAILABLE:

gluefactory/trainer.py

@@ -710,7 +710,7 @@ def train_step(
             for k, v in loss_metrics.items():
                 val = v.detach()
                 if self.distributed:
-                    torch.distributed.all_reduce(val)
+                    torch.distributed.all_reduce(val.contiguous())
                     val = val / self.num_gpus
                 loss_metrics[k] = val
             self.step_timer.measure("loss_fn")