Pin dependencies to a verified stack; align container to Python 3.12

[Alvin-Nahabwe]

Why

Every dependency in the DL service was unpinned, so a fresh pip install resolved whatever was latest on PyPI. That is the direct root cause of the runtime breakage fixed in #1: the code had drifted behind transformers 5.x, albumentations 2.x, and segmentation-models-pytorch 0.5. A clean install produced a service that crashed on two of its three segmentation model families and on object detection.

What

Pins requirements.txt to the exact versions verified end-to-end — full train → evaluate → save → inference — for all three tasks and both model families per task, on Python 3.12 / CPU torch:

Task	Models verified
Image classification	ViT
Object detection	DETR (HF) + YOLO (Ultralytics)
Semantic segmentation	SegFormer (HF) + U-Net (SMP)

Key pins: torch==2.13.0, transformers==5.13.1, datasets==5.0.0, albumentations==2.0.8, segmentation-models-pytorch==0.5.0, ultralytics==8.4.92, numpy==2.4.4, pandas==3.0.3.

• torch is pinned by version only (no +cpu); on the CUDA base image PyPI serves the matching CUDA build (confirmed 2.13.0 is on the default index).
• gradio and wandb removed — neither is imported anywhere in the codebase (training sets WANDB_DISABLED=true), and gradio pulls a large unused dependency tree. Re-add with a pin if a demo UI / W&B logging is built.
• The pinned stack needs Python ≥ 3.11 (numpy 2.4 / pandas 3.0), so the Dockerfile base moves ubuntu22.04/py3.10 → ubuntu24.04/py3.12 and installs into a venv (24.04's system Python is externally-managed, PEP 668). The ubuntu24.04 CUDA base tag was confirmed to exist.

Verification

• ✅ Pins run all five model paths on CPU / Python 3.12 (train + inference).
• ⚠️ Not verified here: the CUDA container build itself — needs a GPU host. torch==2.13.0 and the ubuntu24.04 CUDA tag were confirmed to exist, but the image has not been built. Please docker build + one smoke run on the VM before deploying.

Merge order

Independent of #1 (touches only requirements.txt, Dockerfile, .gitignore), but they're two halves of one story: #1 makes the code run on the current libraries; this makes "the current libraries" a fixed, reproducible set. Recommend merging together.

🤖 Generated with Claude Code