[Bug] DeepLense_Classification_Transformers: Missing dependencies, broken APIs, and deprecated calls prevent running

[Ganglet]

Description

Attempting to run train.py and eval.py in DeepLense_Classification_Transformers_Archil_Srivastava/
fails due to multiple missing dependencies, breaking API changes, and deprecated function calls.

Tested on: macOS, Python 3.11, PyTorch 2.x

Issues Found

1. No requirements.txt

Running python train.py crashes repeatedly with missing modules:

• ModuleNotFoundError: No module named 'wandb'
• ModuleNotFoundError: No module named 'einops'
• ModuleNotFoundError: No module named 'timm'
• ModuleNotFoundError: No module named 'torchmetrics'

There is no requirements.txt or setup.py to install dependencies.

2. Broken torchmetrics API (eval.py:59-61)

torchmetrics >= 0.11 requires a task= argument for all functional calls. Current code:

accuracy_fn(logits, y, num_classes=NUM_CLASSES)
auroc_fn(logits, y, num_classes=NUM_CLASSES, average=None)

Both throw: TypeError: auroc() missing 1 required positional argument: 'task'

Fix: Add task="multiclass" to both calls.

3. Deprecated torch.has_mps (utils.py:32)

UserWarning: 'has_mps' is deprecated, please use 'torch.backends.mps.is_built()'

4. Removed verbose= param in CosineAnnealingWarmRestarts (train.py:322)

CosineAnnealingWarmRestarts(optimizer, T_0=15, T_mult=1, eta_min=1e-6, verbose=True)

verbose was removed in newer PyTorch versions. Running with --decay_lr crashes.

Proposed Fix

• Add requirements.txt with pinned versions
• Fix torchmetrics calls with task="multiclass"
• Replace torch.has_mps with torch.backends.mps.is_built()
• Remove verbose=True from scheduler

I'll submit a PR addressing all of these.

[Akarshak51]

Great catch, @Ganglet! This is a critical issue affecting the reproducibility of the entire DeepLense_Classification_Transformers pipeline.

Analysis of the Root Causes

The issues you've identified reveal a systematic compatibility problem:

1. Missing requirements.txt - This is the most fundamental issue affecting all users. Without dependency pinning, fresh installations will fail immediately.

2. torchmetrics API Breaking Change - The shift to requiring task= argument is a common pattern in ML libraries for clarity. Your fix suggestion is correct and aligns with torchmetrics best practices.

3. PyTorch Deprecations - These indicate the codebase was developed on older PyTorch versions. The migration pattern you've identified is standard.

Additional Verification Needed (Suggested for PR #194)

While reviewing the PR, I recommend verifying:

1. Python version compatibility - Test across Python 3.8, 3.9, 3.10, 3.11 to ensure maximum compatibility
2. PyTorch versions - Recommend testing with PyTorch 2.0 through 2.3 to establish the supported range
3. torchmetrics versions - Confirm the minimum version that supports the task= argument
4. Cross-platform testing - Verify fixes work on Linux, macOS, and Windows

Testing Strategy Recommendation

I suggest adding:

• A requirements.txt with pinned versions for reproducibility
• A requirements-dev.txt for development/testing dependencies
• GitHub Actions CI workflow to test against multiple Python/PyTorch combinations

Implementation Quality Assessment

Your proposed fixes demonstrate:
✅ Understanding of modern ML library practices
✅ Awareness of API migration patterns
✅ Focus on reproducibility (critical for scientific ML)

I'm happy to help test the PR or contribute additional compatibility improvements. This kind of foundational fix strengthens the entire DeepLense ecosystem.

cc @ML4SCI/DeepLense - This is exactly the type of issue that benefits scientific credibility and contributor onboarding!