Batch Object Detection

At a glance

Workload: Run DETR object detection on a batch of COCO 2017 images. Runner: gpu=t4, cpu=4, ram=16, tenancy=spot ($0.004/min). Estimated cost: ~$0.05 per run (~10 min on a few hundred images).

This page shows how to run batch object detection on machine.dev GPU runners. The kind of pipeline you’d use to label a dataset, generate annotations, or analyse a video frame by frame.

When to use this

A few reasons you might run GPU-accelerated batch object detection in CI:

• Process large image collections without hand-rolling a server
• Pull objects, people, vehicles, or other entities out of a corpus
• Generate metadata and annotations for a CV training set
• Build analytics from visual content with no manual review

How it works

The pipeline uses the DETR model (facebook/detr-resnet-50) via Hugging Face Transformers. It runs as a GitHub Actions workflow you trigger on demand with input parameters.

The job:

1. Sets up Python and CUDA on a GPU runner
2. Downloads images from the COCO2017 dataset
3. Runs DETR over each image
4. Writes annotations and a per-image CSV
5. Uploads the CSV as a GitHub Actions artifact

Workflow

name: Batch Object Detection

on:
  workflow_dispatch:
    inputs:
      tenancy:
        type: choice
        required: false
        description: 'The tenancy of the machine'
        default: 'spot'
        options:
          - 'spot'
          - 'on_demand'

jobs:
  detect_objects:
    name: Detect Objects
    runs-on: machine/gpu=t4/cpu=4/ram=16/architecture=x64/tenancy=${{ inputs.tenancy }}
    steps:

      - uses: actions/checkout@v4

      - name: Set up Python 3.10
        uses: actions/setup-python@v5
        with:
          python-version: '3.10'

      - name: Install Dependencies
        run: |
          pip install -r requirements.txt

      - name: Run Object Detection
        run: python3 object_detection.py

      - name: Upload Detection Results CSV
        uses: actions/upload-artifact@v4
        with:
          name: detection-results-csv
          path: detection_output/detections.csv

Runner config

The default runner used here:

• T4 GPU: 16 GB VRAM, plenty for DETR-resnet-50
• Spot tenancy: ~70-90% cheaper than on-demand
• Configurable CPU, RAM, and architecture

The T4 is fast enough for batch CV pipelines. For heavier models or larger images, step up to L4 or L40S.

Why CI

A few things you get from running this in GitHub Actions rather than a long-lived box:

• GPU acceleration without a dedicated server
• Datasets fetched on demand, no persistent disk
• Pipeline runs end to end with no manual review needed
• Output is a CSV artifact, easy to attach to a PR or download
• The whole repo is a template you can fork

Getting started

1. Use MachineDotDev/gpu-batch-object-detection as a template
2. Open the Actions tab in your repository
3. Pick the “Batch Object Detection” workflow
4. Click “Run workflow” and pick spot or on-demand tenancy
5. Run it and wait
6. Download the detection-results-csv artifact to inspect the detections

Tips

• Spot tenancy is fine for batch CV jobs. Worst case you re-run.
• Tune batch size to fit GPU memory comfortably
• Pre-filter inputs so you don’t waste minutes on images you don’t care about
• For very large datasets, chunk the input and parallelise across multiple workflow runs
• Print progress every N images so a stalled run is easy to spot

How to adapt this

• Larger images or heavier model: swap gpu=t4 for gpu=l4 (24 GB VRAM, $0.006/min spot)
• Different model: change the model ID in the Python script. Any Hugging Face object-detection model works.
• Run on a schedule: add on: schedule: cron: '0 2 * * *' to process new images nightly
• Trigger from a webhook: add on: repository_dispatch and POST to GitHub when new data lands

Next steps

• Working repo: fork or use as a template
• CPU vs GPU: picking GPU size for CV workloads
• Parallel Hyperparameter Tuning: fan the same workflow out across multiple model variants