[Bug] micro_auroc always NaN and CPU-bound label tensor in Classification Transformers

[PuneetKumar1790]

Description

Two bugs in DeepLense_Classification_Transformers_Archil_Srivastava affect evaluation metric correctness and device placement efficiency.

Bug 1: micro_auroc always returns NaN

In eval.py, micro_auroc is initialized as an empty list on line 47 but is never computed or appended to. When np.mean(micro_auroc) is called on line 69, it returns nan because np.mean([]) is nan.

# Line 47: initialized as empty list
loss, accuracy, class_auroc, micro_auroc, macro_auroc = [], [], [], [], []

# Line 69: np.mean([]) = nan
"micro_auroc": np.mean(micro_auroc),  # Always nan!

Impact: Every evaluation reports nan for micro_auroc, corrupting W&B experiment logs.

Bug 2: labels.type(torch.LongTensor) creates CPU tensor

In both train.py (line 47-49) and eval.py (line 54-55), labels are converted using labels.type(torch.LongTensor). torch.LongTensor always creates a CPU tensor, regardless of the current device.

# train.py line 47-49
images, labels = images.to(device, dtype=torch.float), labels.type(
    torch.LongTensor  # Always CPU!
).to(device)

# eval.py line 54-55
batch_X, batch_y = batch_X.to(device, dtype=torch.float), batch_y.type(
    torch.LongTensor  # Always CPU, never moved to device
)

In train.py, the CPU tensor is immediately moved back to GPU with .to(device), causing an unnecessary CPU allocation. In eval.py, batch_y is never moved to device at all — it stays on CPU while logits are moved to CPU for comparison, silently working but wasting GPU potential.

Impact: Unnecessary CPU tensor allocation on every batch; inefficient for GPU training.

Proposed Fix

Bug 1: Remove micro_auroc from the initialization list and compute it directly:

micro_auroc = auroc_fn(logits, y, num_classes=NUM_CLASSES, average="weighted")

Bug 2: Replace labels.type(torch.LongTensor) with device-aware conversion:

labels.to(device, dtype=torch.long)

Environment

• Python 3.10, PyTorch 2.10.0, torchmetrics 1.8.2