苏神https://kexue.fm 学习总结
- crf:crf:crf_keras
- viterbi:crf:word_seg
- attention:attention:attention_keras
- to_mask:attention:attention_keras
- reuse::attention:attention_keras
- beam_search:seq2seq:seq2seq
- OurBidirectional:seq2seq:seq2seq
- Embedding与词向量:https://kexue.fm/archives/4122
- cnn+crf:crf:word_seg
- bilstm:crf:bilstm_fenci
- 分词:样本构建:https://kexue.fm/archives/4245
- 摘要:seq2seq:seq2seq
- Keras中自定义复杂的loss函数 https://kexue.fm/archives/4493
- mask mask = Lambda(lambda x: K.cast(K.greater(K.expand_dims(x, 2), 0), 'float32'))(x) https://kexue.fm/archives/6810
- Keras 训练阶段给输入加入一些噪声，而测试阶段则去掉 https://kexue.fm/archives/5765
- Keras 超参数修改，而测试阶段则去掉 https://kexue.fm/archives/5765
- 支持在各阶段定义回调函数 https://kexue.fm/archives/5765
- 分层的学习率和梯度裁剪（梯度裁剪的作用） https://kexue.fm/archives/6418
- 权重滑动平均 https://kexue.fm/archives/6575
- 层复用 结构复用和权值共享 https://kexue.fm/archives/6985
- 软batch https://kexue.fm/archives/6794