Reliability-Calibrated Routing (RCR)

Inference checkpoints for When Neural Collapse Fails to Route: Routing Drift in Continual Learning, BMVC 2026.

Code | Project page | Evaluation artifacts

RCR is a classifier for class-incremental image recognition. It keeps the trained ProNC ResNet-18 and ETF classifier fixed, and combines ETF alignment with a radial score based on transported class prototypes. It predicts among all classes seen during training without a task label at inference.

Checkpoints

This repository mirrors the 18 final inference checkpoints from GitHub release v1.0.0. There are three seeds (0, 1, 2) for each dataset and replay budget (200, 500). Each checkpoint contains the backbone weights, ETF vertices, class prototypes, and class radii. It does not contain replay images or optimizer state.

Dataset Checkpoint path pattern
CIFAR-10 checkpoints/seq_cifar10_buffer{200,500}_seed{0,1,2}/final.pt
CIFAR-100 checkpoints/seq_cifar100_buffer{200,500}_seed{0,1,2}/final.pt
TinyImageNet checkpoints/seq_tinyimagenet_buffer{200,500}_seed{0,1,2}/final.pt

Inference

From a checkout of the RCR code, install the dependencies with python -m pip install -r requirements-eval.txt huggingface_hub. Download a checkpoint and load it with the RCR inference loader:

import torch
from huggingface_hub import hf_hub_download
from rcr.rcr import load_inference_checkpoint, rcr_scores

path = hf_hub_download(
    repo_id="towzeur/RCR",
    filename="checkpoints/seq_cifar100_buffer200_seed0/final.pt",
)
net, state = load_inference_checkpoint(path)

# images: a batch of RGB images with the dataset's test normalization.
with torch.inference_mode():
    features = net(images, returnt="features")
    scores = rcr_scores(
        features, state["etf"].T, state["ldc_proto"],
        state["ldcrad_class_radius"], state["classes_per_task"],
    )
    labels = scores.argmax(dim=1)

Use the repository's dataset preprocessing and label order. The checkpoints are trained for the classes in their respective benchmarks, not for open-set recognition or resuming training.

Evaluation

CIFAR-100 with a replay budget of 200, averaged over three seeds: RCR raises Class-IL final average accuracy from 41.71% to 49.74% and lowers final forgetting from 38.55% to 16.27%, using the same trained backbones. See the results table for all six settings. Recomputing forgetting and task-by-task metrics requires the per-task features in GitHub Releases, not just the final checkpoints here.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Paper for towzeur/RCR