Upload folder using huggingface_hub

app.py

@@ -29,7 +29,7 @@ print(f"Loaded {len(MODELS)} EdgeCrafter checkpoints", flush=True)
 TABLE_HEADERS = ["#", "class", "score", "box (x1, y1, x2, y2)", "mask px / keypoints"]
 
 
-@spaces.GPU(duration=25)
+@spaces.GPU(duration=10)
 def predict(
     image: Image.Image,
     task: str = "Object Detection",

ecmodels.py

@@ -336,15 +336,18 @@ def infer(model, task: str, image: Image.Image, threshold: float, device: str =
     results = []
     if masks is not None:
         img_w, img_h = image.size
-        msks = torch.nn.functional.interpolate(
-            masks[0].unsqueeze(0).float(), size=(img_h, img_w),
-            mode="bilinear", align_corners=False,
-        )[0]
-        msks = (msks > 0.0)[keep]
+        # same as the reference (bilinear upsample of the mask logits to the original
+        # resolution, threshold at 0), but one kept instance at a time so memory stays
+        # bounded on large inputs instead of upsampling all 300 queries at once.
+        kept_masks = masks[0][keep].float()
         for j in range(len(lbls)):
+            m = torch.nn.functional.interpolate(
+                kept_masks[j][None, None], size=(img_h, img_w),
+                mode="bilinear", align_corners=False,
+            )[0, 0]
             results.append(Result(label=int(lbls[j].item()), score=float(scs[j].item()),
                                   box=bxs[j].float().cpu().numpy(),
-                                  mask=msks[j].cpu().numpy()))
+                                  mask=(m > 0.0).cpu().numpy()))
     else:
         for j in range(len(lbls)):
             results.append(Result(label=int(lbls[j].item()), score=float(scs[j].item()),