LT8/Kanji_ETL8G

Deprecated — superseded by LT8/japanese-handwriting-onnx

Please use LT8/japanese-handwriting-onnx instead. It is a full rebuild from the raw ETL binaries, not a retrain of this model.

	this model	the replacement
classes	956	3,082 — all JIS X 0208 Level 1 kanji, plus hiragana and katakana
preprocessing	yours to reimplement	compiled into the ONNX graph
evaluation	random split	writer-disjoint, plus a second corpus never trained on
accuracy	—	99.79 % unseen writers · 98.98 % unseen corpus
batching	fixed at 1	dynamic batch axis

Why these files are kept but not recommended

Both artifacts here have problems that are awkward to work around:

• kanji_ETL8G.pth does not load on current PyTorch. It was saved as a pickled whole model rather than a state_dict, so torch.load refuses it under the weights_only=True default and it depends on the original class definitions being importable.
• kanji_ETL8G.onnx takes a flattened [1, 4096] vector, so the caller has to resize to 64×64, flatten, and match the original normalisation exactly — and the batch size is fixed at 1.

The replacement addresses both directly: weights are stored as plain tensors, and the exported graph accepts an ordinary (N, 1, 128, 128) grayscale image and does its own contrast normalisation, bounding box, padding and resize. There is no preprocessing left for a caller to get wrong.

The files below are left in place so existing references keep resolving.