RSReasonNet / README.md
sosppxo's picture
Replace highway service area example with storage tank 04077
b036ed3 verified
|
Raw History Blame Contribute Delete
3.44 kB

A newer version of the Gradio SDK is available: 6.32.0

Upgrade
metadata
title: RSReasonNet
emoji: 🛰️
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 4.44.1
python_version: 3.10.13
app_file: app.py
suggested_hardware: zero-a10g
startup_duration_timeout: 1h
models:
  - sosppxo/RSReasonNet-RAO-w
pinned: false

RSReasonNet

RSReasonNet is a private research demo for language-guided reasoning segmentation. It accepts one image and an English text instruction, then returns a text response and a segmentation overlay when the model emits a segmentation token.

Using the demo

The layout places the image and English instruction beside the segmentation overlay and model response. Select Run segmentation (or submit the instruction) to run inference. Clear resets the inputs and outputs.

Three bundled examples are available below the workspace: a ship (03261.jpg), a storage tank (04077.jpg), and a tennis court (02026.jpg). Clicking a thumbnail loads both the image and its original instruction from example/annotations.json. The ship example is preloaded on opening the demo; selecting an example does not run inference or consume GPU quota. Changing an input clears the previous result.

The layout takes inspiration from the clear input/output columns and visual example selectors in Kolors Virtual Try-On and IllusionDiffusion.

Runtime configuration

Configure these values in the Space Settings page before starting the app:

Kind Name Value
Secret HF_TOKEN Fine-grained token with read access only to the private model repository
Variable MODEL_REPO_ID sosppxo/RSReasonNet-RAO-w
Variable MODEL_REVISION Full 40-character commit SHA from the model repository
Variable PRECISION fp16
Variable LOAD_IN_4BIT 0
Variable MAX_NEW_TOKENS 512

The token must never be committed to this repository. The app downloads the private model at the fixed revision during runtime and loads it once in a synchronous startup phase. Requests are queued with a concurrency of one.

Hardware

This Space is designed for ZeroGPU large, backed by half of an NVIDIA RTX Pro 6000 Blackwell with 48 GB GPU memory. The model is placed on emulated CUDA once during startup, while real GPU capacity is allocated only for the decorated inference function. The function has a five-minute maximum GPU duration and uses the account's ZeroGPU quota.

Limitations

  • The checkpoint is intended for research and may produce incorrect text or masks.
  • Prompts are limited to standard English characters and 1,024 characters.
  • FP16 inference is required by this deployment; LOAD_IN_4BIT must remain 0.
  • Runtime storage is ephemeral, so model files may be downloaded again after a rebuild or restart.
  • Model, dataset, and third-party license compatibility has not yet been fully reviewed. The Space and model repository must remain private, and the weights must not be redistributed or made public without a separate review and explicit authorization.

Upstream

The inference architecture includes components derived from LISA, LLaVA, and Segment Anything, and SAM 2.