mahjong_soul_vision (ONNX)

An ONNX export of pjura/mahjong_soul_vision, a ViT-base image classifier its author fine-tuned to tell the 34 Japanese mahjong tile kinds apart on Mahjong Soul screen crops. Hosted here because the original repository only has PyTorch weights; it is used by an offline mahjong study aid that reads the game window and never touches the game.

Changes from the original

• Exported from revision ea01ca0bcb9be0fc10a5fa09ee8356770c25adc5 with torch.onnx.export (opset 18, dynamic batch axis). The weights are not changed, retrained or quantized.
• On 2989 tile crops the ONNX model and the original PyTorch model give the same answer for every crop; the largest logit difference is 4e-5.

Input and output

• pixel_values: float32 [batch, 3, 224, 224]. Resize the RGB crop to 224x224 with Pillow's bilinear filter, then scale each value as v / 127.5 - 1 (see preprocessor_config.json).
• logits: float32 [batch, 34], in the order of id2label in config.json (n = characters, p = circles, b = bamboo; ew sw ww nw winds; wd gd rd dragons). There is no red-five class.

Files

License and training data

Apache License 2.0, as the original model; see LICENSE. The model was trained by its original author on screenshots of the game Mahjong Soul (pjura/mahjong_souls_tiles). The game art belongs to its owners. This repository contains no images. If you hold rights to the game and want it removed, open a discussion and it will be taken down.