Adaptive Sensing Repo

This is an adaptive sensing research codebase for selecting the best image capture option from a fixed set of camera-parameter candidates. For each ImageNet-ES-Diverse sample, the dataset provides an auto-exposure image and 27 parameter-controlled candidate images. The main task is to train a lightweight policy network that predicts which candidate should be selected to maximize downstream recognition performance.

Repository Layout

Path	Description
data/ImageNet-ES-Diverse/	Dataset metadata, manifests, ImageNet class index, Lens labels, oracle labels, OpenCLIP labels, and derived label subsets.
lens/	Baseline utilities for manifest construction, image loading, classifier-based candidate selection, and Lens expert-label generation.
policy_network/static_pred/	Main policy-learning code: dataset loader, policy model, oracle-label generation, training entry point, organized experiment scripts, and debug tools.
policy_network/results/	Checkpoints, JSON metrics, cached downstream results, and experiment summaries.
policy_network/vis_results/	Generated plots and visualizations for option/index distributions and multiview probes.
openclip_ds_policy/	OpenCLIP-based downstream experiments, including soft-label generation, baselines, policy evaluation, and xlsx result summaries.
reproduce/src/reproduce_in_ae/print_pdf.py	Converts original ImageNet images into directly printable US Letter PDF pages.

Core Workflow

1. Build or load manifest_all.json, where each sample contains 27 candidate capture options.
2. Generate supervision labels:
   • Lens labels select the candidate with the highest classifier confidence.
   • Oracle labels select candidates based on downstream correctness using the ground-truth class.
   • OpenCLIP labels use image-text similarity over the 200-class subset.
3. Train a policy network with single, dual, or multiview inputs.
4. Evaluate both policy-index accuracy and downstream top-1/top-k accuracy.
5. Analyze results with distribution plots, ablations, heatmaps, and probe experiments.

Main Entry Points

File or Folder	Purpose
policy_network/static_pred/train_policy.py	General policy training entry point.
policy_network/static_pred/generate_oracle_policy_labels.py	Generates oracle hard/soft labels from downstream classifier behavior.
policy_network/static_pred/scripts/	Organized training, evaluation, label-generation, and visualization shell scripts. See policy_network/static_pred/scripts/README.md.
policy_network/static_pred/debug/	Organized analysis and debug tools. See policy_network/static_pred/debug/README.md.
openclip_ds_policy/	OpenCLIP downstream pipeline and documentation.

Policy Inputs

The policy network supports three input modes:

• single: one image, either AE/baseline or a fixed candidate option.
• dual: AE/baseline plus one fixed candidate option.
• multiview: multiple fixed candidate options, optionally including AE.

The training code supports MobileNetV3-Small, ResNet18, tiny convolutional models, and DINOv2 through a shared SensorPolicyNetwork interface.

Notes

• Most scripts are intended to be run from the repository root.
• Shell scripts usually expose overrides such as PYTHON_BIN, DEVICE, BATCH_SIZE, NUM_WORKERS, RESULTS_DIR, and sweep ranges.
• The code assumes a Python environment with PyTorch, torchvision, timm, and related scientific Python packages. OpenCLIP experiments additionally require open_clip_torch.

Letter-size ImageNet Print PDF

The print utility follows the ImageNet-ES-Diverse paper layout: one source image per page, centered without cropping and with its aspect ratio preserved. The paper size is US Letter rather than A4, and the default image width is 60% of the page (5.1 inches). Run it from the repository root after installing the reproduce package dependencies:

PYTHONPATH=reproduce/src python -m reproduce_in_ae.print_pdf \
  --output-pdf imagenet_letter_print.pdf

The default input is the 1,000-image data/ImageNet-ES-Diverse/es-diverse-test/sampled_tin_no_resize2 reference set. Use --input-dir for another image tree, --label none for pages without file names, or --max-pages-per-pdf 500 to split a large job. A CSV manifest is written beside the PDF. Existing outputs are protected unless --overwrite is passed.

Print using US Letter, Portrait, Actual Size / 100%, with Fit-to-page disabled, and select the printer's 600 DPI (or highest quality) mode. The PDF embeds the source raster data without applying ImageNet preprocessing or artificial 600-DPI upsampling.