Transformer

An overhaul of structure, adopts pure Transformer now.

No More BatchNormalization

Replace BatchNormalization with constraints on kernels, achieving more stable validation set performance.

Parametric Softplus

Use new activation function, significantly improved accuracy (93% to 98% on training set, 84% to 86% on validation set).

Pretrained Model

Pretrained with the latest code, using TensorFlow 2.2, Cangjie6 with all BMP, Ext A-G CJK Ideographs.

TensorFlow Saved Model

Pre-trained Model, unzip to load