Issue : Silent Exception Swallowing in Model Loading

[priyan17singh]

Problem

The model loading code in perceptionmetrics/models/ uses bare except Exception: blocks and assert statements that silently swallow all exceptions, making debugging very difficult.

---

📍 Locations with Issue

• perceptionmetrics/models/torch_segmentation.py (Lines ~217–230, ~548–561)
• perceptionmetrics/models/torch_detection.py (Lines ~265–275)
• perceptionmetrics/models/utils/o3d/randlanet.py (Lines ~5–8)
• perceptionmetrics/models/utils/o3d/kpconv.py (Lines ~5–8)

---

Expected Behavior

• Raise specific exceptions (FileNotFoundError, RuntimeError).
• Preserve original error context.
• Provide clear and actionable error messages.
• Avoid misleading fallback behavior.

---

Current Problematic Code

Segmentation / Detection

if isinstance(model, str):
    assert os.path.isfile(model), "Torch model file not found"  # Uses assert
    model_fname = model
    try:
        model = torch.jit.load(model, map_location=self.device)
        model_type = "compiled"
    except Exception:  # Catches ALL exceptions
        ...
        

Optional Imports

try:
    from open3d._ml3d.datasets.utils import DataProcessing
except Exception:  # Catches ALL exceptions
    print("Open3D-ML3D not available")

---

Why This Is a Problem

1. Catches all exceptions including: FileNotFoundError, PermissionError, MemoryError, Corrupted model files

2. Misleading Behavior like no clear error message and no root cause visibility.

3. Using bare except: is discouraged (PEP 8).

4. Catching Exception hides real bugs.

---

Proposed Solution

Replace assert

if not os.path.isfile(model):
    raise FileNotFoundError(f"Model file not found: {model}")

---

Improve exception handling

try:
    model = torch.jit.load(model, map_location=self.device)
    model_type = "compiled"

except (RuntimeError, EOFError) as jit_err:
    try:
        model = torch.load(model, map_location=self.device)
        model_type = "native"

    except Exception as load_err:
        raise RuntimeError(
        ...

---

Fix optional imports

try:
    from open3d._ml3d.datasets.utils import DataProcessing
except ImportError:
    ...

---

🧪 Testing Plan

• Valid TorchScript model → loads correctly
• Valid PyTorch model → fallback works
• Invalid path → raises FileNotFoundError
• Corrupted file → raises RuntimeError
• Missing dependency → handled via ImportError

---

Impact

• Better debugging
• No silent failures
• Improved robustness
• Aligns with Python best practices

---

[karttikjangid]

Hey @priyan17singh, I noticed the same assert vulnerability in the datasets/ and utils/ files. Since you are already fixing this in the models/ directory for PR #530, do you need me to help patch the ones in datasets and utils, or are you planning to cover those as well? . I also opened a similar issue and maintainer advised me to make a comment here .

[priyan17singh]

Hey @karttikjangid , glad we can coordinate on this!

My PR #530 is scoped only to perceptionmetrics/models/ — so datasets/ and utils/ are completely free for you to take on, no overlap.

@dpascualhe — flagging this so you're aware the datasets/ and utils/ scope will be covered in a follow-up PR from @karttikjangid .