From c064457d4485dca6bbab56884e07dd9fc0fc1138 Mon Sep 17 00:00:00 2001 From: Tink_Y <31891223+tink2123@users.noreply.github.com> Date: Wed, 20 Feb 2019 11:45:45 +0800 Subject: [PATCH] fix_error_for_lsr (#683) --- 07.label_semantic_roles/README.cn.md | 85 ++------------------------- 07.label_semantic_roles/index.cn.html | 85 ++------------------------- 2 files changed, 8 insertions(+), 162 deletions(-) diff --git a/07.label_semantic_roles/README.cn.md b/07.label_semantic_roles/README.cn.md index 81307a5..88aa96b 100644 --- a/07.label_semantic_roles/README.cn.md +++ b/07.label_semantic_roles/README.cn.md @@ -232,87 +232,6 @@ def load_parameter(file_name, h, w): return np.fromfile(f, dtype=np.float32).reshape(h, w) ``` -- 8个LSTM单元以“正向/反向”的顺序对所有输入序列进行学习，主要的执行逻辑如下： - 1）为不同的输入特征分别定义embedding层 - -```python -def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark, - **ignored): - # 8 features - predicate_embedding = fluid.layers.embedding( - input=predicate, - size=[pred_dict_len, word_dim], - dtype='float32', - is_sparse=IS_SPARSE, - param_attr='vemb') - - mark_embedding = fluid.layers.embedding( - input=mark, - size=[mark_dict_len, mark_dim], - dtype='float32', - is_sparse=IS_SPARSE) - - word_input = [word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2] - # 因词向量是预训练好的，这里不再训练embedding表， - # 参数属性trainable设置成False阻止了embedding表在训练过程中被更新 - emb_layers = [ - fluid.layers.embedding( - size=[word_dict_len, word_dim], - input=x, - param_attr=fluid.ParamAttr( - name=embedding_name, trainable=False)) for x in word_input - ] - emb_layers.append(predicate_embedding) - emb_layers.append(mark_embedding) -``` -2) 定义深度双向LSTM结构 - -```python - # 共有8个LSTM单元被训练，每个单元的方向为从左到右或从右到左， - # 由参数`is_reverse`确定 - hidden_0_layers = [ - fluid.layers.fc(input=emb, size=hidden_dim, act='tanh') - for emb in emb_layers - ] - - hidden_0 = fluid.layers.sums(input=hidden_0_layers) - - lstm_0 = fluid.layers.dynamic_lstm( - input=hidden_0, - size=hidden_dim, - candidate_activation='relu', - gate_activation='sigmoid', - cell_activation='sigmoid') - - # 用直连的边来堆叠L-LSTM、R-LSTM - input_tmp = [hidden_0, lstm_0] - - for i in range(1, depth): - mix_hidden = fluid.layers.sums(input=[ - fluid.layers.fc(input=input_tmp[0], size=hidden_dim, act='tanh'), - fluid.layers.fc(input=input_tmp[1], size=hidden_dim, act='tanh') - ]) - - lstm = fluid.layers.dynamic_lstm( - input=mix_hidden, - size=hidden_dim, - candidate_activation='relu', - gate_activation='sigmoid', - cell_activation='sigmoid', - is_reverse=((i % 2) == 1)) - - input_tmp = [mix_hidden, lstm] - - # 取最后一个栈式LSTM的输出和这个LSTM单元的输入到隐层映射， - # 经过一个全连接层映射到标记字典的维度，来学习 CRF 的状态特征 - feature_out = fluid.layers.sums(input=[ - fluid.layers.fc(input=input_tmp[0], size=label_dict_len, act='tanh'), - fluid.layers.fc(input=input_tmp[1], size=label_dict_len, act='tanh') - ]) - - return feature_out -``` - ## 训练模型 - 我们根据网络拓扑结构和模型参数来进行训练，在构造时还需指定优化方法，这里使用最基本的SGD方法(momentum设置为0)，同时设定了学习率、正则等。 @@ -432,6 +351,10 @@ feature_out = fluid.layers.sums(input=[ fluid.layers.fc(input=input_tmp[1], size=label_dict_len, act='tanh') ]) +# 标注序列 +target = fluid.layers.data( + name='target', shape=[1], dtype='int64', lod_level=1) + # 学习 CRF 的转移特征 crf_cost = fluid.layers.linear_chain_crf( input=feature_out, diff --git a/07.label_semantic_roles/index.cn.html b/07.label_semantic_roles/index.cn.html index d01cf9f..27e0d3b 100644 --- a/07.label_semantic_roles/index.cn.html +++ b/07.label_semantic_roles/index.cn.html @@ -274,87 +274,6 @@ def load_parameter(file_name, h, w): return np.fromfile(f, dtype=np.float32).reshape(h, w) ``` -- 8个LSTM单元以“正向/反向”的顺序对所有输入序列进行学习，主要的执行逻辑如下： - 1）为不同的输入特征分别定义embedding层 - -```python -def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark, - **ignored): - # 8 features - predicate_embedding = fluid.layers.embedding( - input=predicate, - size=[pred_dict_len, word_dim], - dtype='float32', - is_sparse=IS_SPARSE, - param_attr='vemb') - - mark_embedding = fluid.layers.embedding( - input=mark, - size=[mark_dict_len, mark_dim], - dtype='float32', - is_sparse=IS_SPARSE) - - word_input = [word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2] - # 因词向量是预训练好的，这里不再训练embedding表， - # 参数属性trainable设置成False阻止了embedding表在训练过程中被更新 - emb_layers = [ - fluid.layers.embedding( - size=[word_dict_len, word_dim], - input=x, - param_attr=fluid.ParamAttr( - name=embedding_name, trainable=False)) for x in word_input - ] - emb_layers.append(predicate_embedding) - emb_layers.append(mark_embedding) -``` -2) 定义深度双向LSTM结构 - -```python - # 共有8个LSTM单元被训练，每个单元的方向为从左到右或从右到左， - # 由参数`is_reverse`确定 - hidden_0_layers = [ - fluid.layers.fc(input=emb, size=hidden_dim, act='tanh') - for emb in emb_layers - ] - - hidden_0 = fluid.layers.sums(input=hidden_0_layers) - - lstm_0 = fluid.layers.dynamic_lstm( - input=hidden_0, - size=hidden_dim, - candidate_activation='relu', - gate_activation='sigmoid', - cell_activation='sigmoid') - - # 用直连的边来堆叠L-LSTM、R-LSTM - input_tmp = [hidden_0, lstm_0] - - for i in range(1, depth): - mix_hidden = fluid.layers.sums(input=[ - fluid.layers.fc(input=input_tmp[0], size=hidden_dim, act='tanh'), - fluid.layers.fc(input=input_tmp[1], size=hidden_dim, act='tanh') - ]) - - lstm = fluid.layers.dynamic_lstm( - input=mix_hidden, - size=hidden_dim, - candidate_activation='relu', - gate_activation='sigmoid', - cell_activation='sigmoid', - is_reverse=((i % 2) == 1)) - - input_tmp = [mix_hidden, lstm] - - # 取最后一个栈式LSTM的输出和这个LSTM单元的输入到隐层映射， - # 经过一个全连接层映射到标记字典的维度，来学习 CRF 的状态特征 - feature_out = fluid.layers.sums(input=[ - fluid.layers.fc(input=input_tmp[0], size=label_dict_len, act='tanh'), - fluid.layers.fc(input=input_tmp[1], size=label_dict_len, act='tanh') - ]) - - return feature_out -``` - ## 训练模型 - 我们根据网络拓扑结构和模型参数来进行训练，在构造时还需指定优化方法，这里使用最基本的SGD方法(momentum设置为0)，同时设定了学习率、正则等。 @@ -474,6 +393,10 @@ feature_out = fluid.layers.sums(input=[ fluid.layers.fc(input=input_tmp[1], size=label_dict_len, act='tanh') ]) +# 标注序列 +target = fluid.layers.data( + name='target', shape=[1], dtype='int64', lod_level=1) + # 学习 CRF 的转移特征 crf_cost = fluid.layers.linear_chain_crf( input=feature_out, -- GitLab