PaddleSeg¶
This page documents the current PaddleSeg integration status for RankSEG.
Current status¶
The PaddleSeg integration already exists, but it is currently maintained
outside the main rankseg branch.
At this stage, the main repository provides an entry point to that work rather than duplicating or re-implementing the full PaddleSeg integration locally.
How it fits the RankSEG integration pattern¶
The insertion point is the same as the first-party integrations:
PaddleSeg model -> probability map -> convert to PyTorch tensor
-> RankSEG -> prediction mask
The difference is maintenance scope. RankSEG’s public predictor currently
expects a PyTorch tensor, so Paddle outputs need an explicit conversion step
before calling RankSEG.predict.
import torch
from rankseg import RankSEG
# probs_paddle has shape (batch_size, num_classes, height, width)
probs = torch.from_numpy(probs_paddle.numpy())
rankseg = RankSEG(metric="dice", solver="RMA", output_mode="multiclass")
preds = rankseg.predict(probs)
Who should use this path¶
This path is useful if:
you already deploy segmentation pipelines in PaddleSeg;
you want to evaluate RankSEG as an inference-time post-processing module in the Paddle ecosystem;
you are comfortable using an externally maintained integration branch.
Available entry points¶
External integration branch: Leev1s/rankseg (paddleseg branch)
Docker image: ghcr.io/leev1s/rankseg
Notebook: notebooks/rankseg_with_paddleseg.ipynb
Scope and maintenance¶
This integration is currently treated as an external/community-maintained path.
That means:
the main RankSEG repository links to it;
the main RankSEG repository does not yet treat it as a first-party official integration path;
once the integration assets become stable enough, it can be promoted into a first-party maintained path later.
Relationship to official integrations¶
If you are new to RankSEG, the recommended first entry point remains:
That path is maintained directly in this repository and is the current default official integration route.