Skip to content

README.md

Latest commit

 

History

History
62 lines (42 loc) · 2 KB

README.md

File metadata and controls

62 lines (42 loc) · 2 KB
Raw

Python side — ResNet image classifier

This folder is the "real" Python workload the .NET wrapper drives: a PyTorch/torchvision ResNet image classifier that runs on CUDA when a compatible GPU is available.

It's a plain command-line program. The rule the whole POC is built around:

If classifier.py runs when you type it at a terminal, it runs identically when launched by the .NET PyBridge wrapper — with stdin, stdout, stderr, and the exit code all captured and handed back to the C# caller.

Contract

  • stdout = machine-readable JSON only (the result).
  • stderr = human-readable progress/log lines.
  • exit code: 0 success · 2 bad usage · 3 missing dependency (e.g. no PyTorch) · 4 runtime failure.

Commands

# Environment report — standard library only, so it ALWAYS runs (even with no PyTorch):
python classifier.py check

# Classify an image with a pretrained ResNet (needs torch + torchvision):
python classifier.py classify --image path/to/photo.jpg --topk 5

# No image? A synthetic test image is generated so the pipeline still runs end to end:
python classifier.py classify

First classify run downloads the pretrained ImageNet weights (~tens of MB) into the torch hub cache; subsequent runs are offline.

Installing PyTorch

Option A — conda (recommended, GPU)

conda env create -f environment.yml
conda activate pybridge
python classifier.py check     # look for "cuda_available": true

Option B — pip + CUDA index (for newer GPUs like the RTX 5060 Ti / Blackwell)

Blackwell cards (sm_120) need CUDA 12.8+ wheels:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128
pip install pillow

Option C — pip, CPU only (quickest, no GPU)

pip install -r requirements.txt   # CPU wheels from PyPI

Driving it from .NET

The PyBridge.Console app in this repo calls every one of the commands above through the PyBridge library and prints what it captured. See the repo root README.md.