fix: compute micro_auroc correctly and fix CPU-bound label tensor conversion

[PuneetKumar1790]

Fixes #164

Summary

This PR fixes two bugs in the Classification Transformers sub-project (DeepLense_Classification_Transformers_Archil_Srivastava):

Bug 1: micro_auroc always returns NaN

micro_auroc was initialized as an empty list and never populated. np.mean([]) always returned nan, corrupting W&B experiment logs.

Fix: Compute micro_auroc using auroc_fn(logits, y, num_classes=NUM_CLASSES, average="weighted").

Bug 2: CPU-bound label tensor conversion

labels.type(torch.LongTensor) always creates a CPU tensor, regardless of the current device. In train.py it was immediately moved back with .to(device), wasting an allocation. In eval.py it was never moved to device at all.

Fix: Use labels.to(device, dtype=torch.long) which converts dtype and moves to device in one step.

Files Changed

• eval.py — Fixed micro_auroc computation and label device placement
• train.py — Fixed label device placement
• tests/test_eval_and_train.py — New verification script with 5 test cases

Verification

All 5 tests pass locally (CPU, Python 3.10, PyTorch 2.10.0, torchmetrics 1.8.2):

TEST 1a: Bug confirmed: original code produces nan        [PASS]
TEST 1b: micro_auroc (fixed) = 0.5284                    [PASS]
TEST 2:  labels correctly placed on target device         [PASS]
TEST 3:  train_step loss = 1.3436                         [PASS]
TEST 4:  Full evaluate integration test                   [PASS]
RESULTS: 5 passed, 0 failed

Note on diff size: The +291 additions come mostly from the new verification test script (tests/test_eval_and_train.py) with setup, before/after comparisons, and integration tests. The actual fixes in eval.py and train.py are small (~15–20 lines changed total).