OpenERF

One-line Effective Receptive Field (ERF) extraction for pretrained timm vision models.

import openerf
import timm

model_name = "resnet50.a1_in1k"
model = timm.create_model(model_name, pretrained=True)
openerf.save_erf(model, model_name=model_name)
# -> ./results/OpenERF_resnet50.a1_in1k.png

Visual Gallery

All images below are ERF maps generated from the ImageNet validation subset (./imagenet_val_1000/). The full set is available in ./results/.

ResNet-34	ResNet-50
resnet34.a1_in1k	resnet50.a1_in1k
ResNeXt-101	DenseNet-201
resnext101_32x8d.tv_in1k	densenet201.tv_in1k
ViT-B/16	DeiT-B/16
vit_base_patch16_224.augreg_in1k	deit_base_patch16_224.fb_in1k
CaiT-S24	XCiT-Medium
cait_s24_224.fb_dist_in1k	xcit_medium_24_p16_224.fb_in1k
BEiT-B/16 (224)	SwinV2-Small
beit_base_patch16_224.in22k_ft_in22k_in1k	swinv2_cr_small_224.sw_in1k

Table of Contents

• Features
• Installation
• Quick Start
• CLI Options
• API Reference
• Supported Models
• Dataset
• Repository Layout
• Output Structure
• Distribution Package
• Citing
• License

Features

• One-line API — openerf.save_erf for end-to-end ERF extraction
• Automatic preprocessing — mean/std, interpolation, crop_pct, input size from timm data config
• Unified workflow — CNN and Transformer families through the same interface
• Publication-ready output — ERF heatmaps (plasma by default)
• Optional exports — Gaussian fitting (lmfit), .npy arrays, .json metrics

Installation

pip install openerf

Or install from source:

conda activate OpenERF
pip install -r requirements.txt
pip install -e .
# Optional: Gaussian fitting extras
pip install -e ".[gaussian]"

Quick Start

Python API

import openerf
import timm

model_name = "resnet50.a1_in1k"
model = timm.create_model(model_name, pretrained=True)

result = openerf.save_erf(
    model=model,
    model_name=model_name,
    image_dir="./imagenet_val_1000",
    max_images=1000,
    colormap="plasma",
    save_numpy=True,
    save_metrics=True,
)

print(result["save_path"])
print(result.get("npy_path"))
print(result.get("metrics_path"))

Legacy top-level import is still supported: import OpenERF.

from openerf import compute_erf, save_erf

CLI

openerf --help
openerf --image-dir ./imagenet_val_1000 --max-images 1000

# Show all preset models
openerf --list-models

# Resume long runs
openerf --image-dir ./imagenet_val_1000 --max-images 1000 --skip-existing

# Run selected families
openerf --families vit deit cait xcit beit swin swinv2 \
    --image-dir ./imagenet_val_1000 --max-images 1000

# Module form
python -m openerf.cli --list-models

CLI Options

openerf (or python -m openerf.cli) supports:

Option	Description	Default
--families	Model families to run (when --model-names is not set)	all preset families
--model-names	Explicit timm model names (must include pretrained tag)	None
--image-dir	Input image directory	./imagenet_val_1000
--save-dir	PNG output directory	./results
--npy-dir	NumPy output directory	./results_npy
--metrics-dir	Metrics output directory	./results_metrics
--max-images	Maximum images to process	1000
--target-layer	Override feature source layer	auto (family preset)
--fit-gaussian	Enable 2D Gaussian fit (sigma_x, sigma_y)	disabled
--skip-existing	Skip if output PNG already exists	disabled
--no-save-numpy	Disable .npy export	enabled
--no-save-metrics	Disable .json export	enabled
--stop-on-error	Stop immediately on model failure	disabled

API Reference

openerf.compute_erf(...)

Computes ERF in memory and returns ERFResult:

Field	Description
erf_map	2D ERF array
num_images	Number of images processed
data_config	timm data config used
resolved_target_layer	Actual target layer name
gaussian_fit	(optional) Gaussian fit parameters

openerf.save_erf(...)

Computes ERF, saves PNG, and optionally exports .npy / .json.

Default values: save_numpy=False, save_metrics=False, colormap="plasma"

Return key	Description
model_name	Model identifier
source_model_name	Original timm model name
save_path	Path to saved PNG
num_images	Number of images processed
target_layer	Target layer name
family	Model family
data_config	timm data config used
gaussian_fit	(optional) Gaussian fit result
npy_path	(optional) Path to .npy file
metrics_path	(optional) Path to .json file

Supported Models

11 families × 3 variants = 33 preset models

Family	Models
resnet	resnet18.a1_in1k, resnet34.a1_in1k, resnet50.a1_in1k
resnext	resnext50_32x4d.a1_in1k, resnext50_32x4d.ra_in1k, resnext101_32x8d.tv_in1k
densenet	densenet121.ra_in1k, densenet169.tv_in1k, densenet201.tv_in1k
efficientnet	efficientnet_b0.ra_in1k, efficientnet_b2.ra_in1k, tf_efficientnet_b4.ns_jft_in1k
vit	vit_small_patch16_224.augreg_in1k, vit_base_patch16_224.augreg_in1k, vit_base_patch32_224.augreg_in1k
deit	deit_tiny_patch16_224.fb_in1k, deit_small_patch16_224.fb_in1k, deit_base_patch16_224.fb_in1k
cait	cait_xxs24_224.fb_dist_in1k, cait_xxs36_224.fb_dist_in1k, cait_s24_224.fb_dist_in1k
xcit	xcit_tiny_12_p16_224.fb_in1k, xcit_small_12_p16_224.fb_in1k, xcit_medium_24_p16_224.fb_in1k
beit	beit_base_patch16_224.in22k_ft_in22k, beit_base_patch16_224.in22k_ft_in22k_in1k, beit_base_patch16_384.in22k_ft_in22k_in1k
swin	swin_tiny_patch4_window7_224.ms_in1k, swin_small_patch4_window7_224.ms_in1k, swin_base_patch4_window7_224.ms_in1k
swinv2	swinv2_cr_tiny_ns_224.sw_in1k, swinv2_cr_small_224.sw_in1k, swinv2_cr_small_ns_224.sw_in1k

Dataset

OpenERF uses the ImageNet validation set (or a subset).

• Default subset index: ./imagenet_val_1000.txt
• Image folder: ./imagenet_val_1000/
• Default scope: --image-dir ./imagenet_val_1000 --max-images 1000

Repository Layout

OpenERF/
  src/
    openerf/
      __init__.py
      api.py
      cli.py
      data.py
      erf.py
      feature_ops.py
      fit.py
      metrics.py
      model_zoo.py
      visualization.py
      py.typed
  tests/
    test_smoke.py
  examples/
    example.py
  README.md
  LICENSE
  pyproject.toml
  requirements.txt

Output Structure

./results/         OpenERF_<model_name>.png          # ERF heatmap
./results_npy/     OpenERF_<model_name>.npy          # ERF array
./results_metrics/ OpenERF_<model_name>_metrics.json # ERF metrics
./results_debug/   (smoke/check/intermediate artifacts)

Distribution Package

Source distributions include package source and core metadata while excluding large experiment artifacts.

• Included: src/, README.md, LICENSE, pyproject.toml
• Excluded: results*/, reference/, examples/, tests/, OpenERF.egg-info/

Citing

If OpenERF is useful in your research, please cite:

@inproceedings{KimCJK23,
  author       = {Bum Jun Kim and
                  Hyeyeon Choi and
                  Hyeonah Jang and
                  Sang Woo Kim},
  title        = {Understanding Gaussian Attention Bias of Vision Transformers Using
                  Effective Receptive Fields},
  booktitle    = {{BMVC}},
  pages        = {214},
  publisher    = {{BMVA} Press},
  year         = {2023}
}

@article{KimCJLJK23,
  author       = {Bum Jun Kim and
                  Hyeyeon Choi and
                  Hyeonah Jang and
                  Dong Gu Lee and
                  Wonseok Jeong and
                  Sang Woo Kim},
  title        = {Dead pixel test using effective receptive field},
  journal      = {Pattern Recognit. Lett.},
  volume       = {167},
  pages        = {149--156},
  year         = {2023}
}

License

This project is released under the MIT License.