benjaminjonard/ram-plus-onnx

RAM++ (Recognize Anything Plus) — ONNX export

ONNX export of xinyu1205/recognize-anything-plus-model (ram_plus_swin_large_14m.pth, revision 84d4aee3a0265c4e0df1f714f0572011d1bf2ec3), for CPU inference in Mendako's tagging sidecar.

All credit for the model belongs to its authors. This repository adds no training and no weights of its own — it is the same network, exported.

Files

File	Description
model.onnx	fp32 graph. Input image [1, 3, 384, 384] float32 → output logits [1, 4585] float32
tags.txt	4585 tag names, one per line, in output order
thresholds.txt	RAM++'s own per-tag decision threshold, one float per line, same order

Usage

Preprocess exactly as upstream's get_transform does — resize to 384×384 squashed (no aspect-ratio padding, bilinear), then normalize with the ImageNet statistics:

import numpy as np, onnxruntime as ort
from PIL import Image

MEAN = (0.485, 0.456, 0.406)
STD = (0.229, 0.224, 0.225)

image = Image.open("photo.jpg").convert("RGB").resize((384, 384), Image.BILINEAR)
x = np.asarray(image, dtype=np.float32) / 255.0
x = (x - np.asarray(MEAN, dtype=np.float32)) / np.asarray(STD, dtype=np.float32)
x = np.ascontiguousarray(x.transpose(2, 0, 1)[np.newaxis, ...])

session = ort.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])
logits = np.asarray(session.run(None, {"image": x})[0]).reshape(-1)

tags = [line.strip() for line in open("tags.txt", encoding="utf-8")]
thresholds = [float(line) for line in open("thresholds.txt", encoding="utf-8")]

scores = 1.0 / (1.0 + np.exp(-np.clip(logits.astype(np.float64), -30.0, 30.0)))
fired = [tag for tag, score, t in zip(tags, scores, thresholds) if score > t]

Each tag is scored independently — this is multi-label classification, not a softmax over classes. Compare with strict >: a few tags ship a threshold of 1.0, which is how RAM++ disables them, and a saturated logit rounded in float32 would otherwise revive them.

Verification

The export script refuses to write anything unless the result reproduces upstream's own tagging decision. On upstream's images/demo/demo1.jpg:

• max |eager − onnx| logit drift: 1.5×10⁻⁵
• tags produced: identical set of 19, compared against ram.inference_ram()

Reproducing

See ml/tools/export_ram_plus.py in the Mendako repository.

License

Apache-2.0, inherited from the upstream model (© OPPO). See the Recognize Anything repository for the paper and original code.