fix: make classification transformers module runnable on modern dependencies and fix evaluation pipeline

[DhruvrajSinhZala24]

Problem

The DeepLense_Classification_Transformers_Archil_Srivastava/ module is not runnable on a fresh setup.

Running either train.py or eval.py results in multiple failures due to missing dependencies, outdated APIs, and incorrect evaluation logic. This blocks the full training/evaluation pipeline and makes the module unusable for new contributors.

This affects both:

• Issue [Bug] DeepLense_Classification_Transformers: Missing dependencies, broken APIs, and deprecated calls prevent running #191 (runtime failure)
• Issue [Bug] DeepLense_Classification_Transformers: Evaluation metrics bugs — micro_auroc always NaN, missing softmax dim, hardcoded W&B entity #192 (evaluation pipeline issues)

---

Root Cause

The failures originate from a combination of dependency gaps, API changes, and logic errors:

• No module-level requirements.txt, causing import errors (wandb, timm, einops, torchmetrics)
• torchmetrics API changes (missing required task argument)
• Deprecated PyTorch usage:
  • torch.has_mps no longer valid
  • CosineAnnealingWarmRestarts(verbose=...) no longer supported
• Incorrect evaluation logic:
  • micro_auroc computed as mean([]) → NaN
  • Missing softmax(dim=-1) in evaluation
  • Incorrect predictions used for confusion matrix
• W&B entity hardcoded, preventing external usage

---

Solution

1. Dependency Fix (Reproducibility)

• Added module-level requirements.txt
• Ensures the module can be installed and run independently

---

2. torchmetrics Compatibility

• Updated metric calls to match current torchmetrics API
• Added required task parameter

---

3. Evaluation Fixes

• Fixed AUROC computation to avoid NaN (proper micro AUROC handling)
• Added missing softmax(dim=-1)
• Corrected prediction handling for confusion matrix

---

4. PyTorch Compatibility

• Replaced torch.has_mps with:
  • torch.backends.mps.is_available()
  • torch.backends.mps.is_built()
• Removed unsupported verbose=True from scheduler

---

5. W&B Configuration

• Added --entity argument (defaults to $WANDB_ENTITY)
• Added --run_id support
• Removed hardcoded entity dependency

---

6. Documentation

• Updated README.md with setup instructions:
  • pip install -r requirements.txt
  • Usage of --entity

---

Changes Made

• requirements.txt → added missing dependencies
• eval.py
  • Updated torchmetrics usage
  • Fixed AUROC computation
  • Added softmax(dim=-1)
  • Fixed confusion matrix predictions
  • Added configurable W&B arguments
• train.py
  • Removed deprecated scheduler argument
  • Added --entity
• utils.py
  • Updated MPS device detection
• README.md
  • Added setup + usage instructions

---

Validation

• Verified syntax using:
  python -m compileall -q .
• Ran synthetic evaluate() test:
  • No runtime errors
  • AUROC and accuracy computed correctly
• Confirmed training loop initializes without scheduler errors
• Verified MPS fallback behavior

---

Impact

• Restores full training + evaluation pipeline for this module
• Makes the module runnable on modern environments
• Fixes incorrect evaluation outputs (NaN AUROC)
• Improves usability for new contributors

---

Closes #191
Closes #192