tensorflow 1.13 使用记录

由于生产环境目前还以1.13版本为主, 而google为了推进2.x已经下架了大部分官方教程, 这里简单做一些记录

 

1. name_scope / variable_scope

　　注意 tf.get_variable 不受 name_scope 影响

　　https://blog.csdn.net/u012609509/article/details/80045529

 

2. SavedModel模型的保存与加载 (pb格式,跨语言部署)

　　https://blog.csdn.net/mogoweb/article/details/83021524

 

3. difference between tf.initialize_all_variables() and tf.initialize_local_variables()?

　　sess.run 报错未初始化 acc_op = tf.metrics.accuracy(labels=int_labels, predictions=pred_index, name='acc_op')

　　需要调用 "tf.local_variables_initializer()"

　　原因是:

LOCAL_VARIABLES: the subset of Variable objects that are local to each machine. Usually used for temporarily variables,

like counters. Note: use tf.contrib.framework.local_variable to add to this collection.

https://stackoverflow.com/questions/40220201/what-is-the-difference-between-tf-initialize-all-variables-and-tf-initialize-l

 

3. API

　　tf.einsum(), einsum表达式描述几乎所有矩阵运算

　　　　https://zhuanlan.zhihu.com/p/44954540

　　tf.tile() , 按某些维度复制 tensor

　　　　https://blog.csdn.net/xwd18280820053/article/details/72867818

　　batch_norm / tf.control_dependencies / tf.GraphKeys.UPDATE_OPS

　　　　https://blog.csdn.net/huitailangyz/article/details/85015611

　　L2正则化和collection，tf.GraphKeys.REGULARIZATION_LOSSES

　　　　https://www.jianshu.com/p/5264ff3b3db3

 

4. sparse 2 dense

# split
enum_features_sps = tf.strings.split(tf.reshape(enum_feat_strings, [-1]),
                                     sep=user_params["features_sep"])

# dense shape [batch, feature*seqlen]
enum_features = tf.sparse_to_dense(
    enum_features_sps.indices, 
    [user_params["batch_size"], 
    user_params["n_enumfeatures"] * user_params["n_seq_length"]],
    enum_features_sps.values,
    default_value="yahaha"
)

 

5. mask

mask = tf.cast(tf.not_equal(enum_features, user_params["features_pad"]), tf.int32)

 

6. Unquoted fields cannot have quotes/CRLFs inside

　　use_quote_delim 参数, 设置为True，会把双引号当成引用，在使用tf.decode_csv读取文件的时候如果某一行有双引号会报错，

　　如果设置为False会把双引号当做为一个普通的字符串变量，在这里读取不会报错