unilm_chinese

A Unilm Chinese(pretraind by YunwenTech) Project base tested on couplet Generating

model and code structure

Use the Chinese couplet dataset, more details can be seen in data folder

Examples:

上联：  苦盼郎归，每唤登临，心中只恨东山矮
Bot:   闲思子去，不堪怅望，眼底长留西子愁
原下联：遥思雁去，频添怅慨，石上长留脚印深

上联：  天降甘霖，华夏人民奔富路
Bot:   地生春雨，神州百姓沐和风
原下联：党行善政，崭新时代沐春风

Trained on a single A100(40G) with epoch 3 on about 700k samples, costs about 7 hours. cuda11.2+pytorch1.7

Name		Name	Last commit message	Last commit date
Latest commit History 10 Commits
data/couplet		data/couplet
unilm_chinese		unilm_chinese
README.md		README.md
configuration_unilm.py		configuration_unilm.py
conver_torch_to_tf.py		conver_torch_to_tf.py
convert_unilm_pytorch_checkpoint_to_original_tf.py		convert_unilm_pytorch_checkpoint_to_original_tf.py
decode_seq2seq.py		decode_seq2seq.py
modeling_unilm.py		modeling_unilm.py
predict.ipynb		predict.ipynb
reorganize_couplet_data.py		reorganize_couplet_data.py
run_decode.sh		run_decode.sh
run_seq2seq.py		run_seq2seq.py
run_seq2seq.sh		run_seq2seq.sh
tokenization_unilm.py		tokenization_unilm.py
utils_seq2seq.py		utils_seq2seq.py