tf.compat.v1.train.Saver.save

Function

Saves a model.

Prototype

def save(self, sess, save_path, global_step=None, latest_filename=None, meta_graph_suffix="meta", write_meta_graph=True, write_state=True, strip_default_attrs=False, save_debug_info=False, is_incremental_checkpoint=False, save_delta=False)

Parameters

Parameter	Type	Mandatory/Optional	Description
sess	Session	Mandatory	Model session to be saved.
save_path	str	Mandatory	Model save path. The local file system and HDFS file system are supported. The length range is [1, 1024]. When the HDFS path is used to save data, the "hdfsWrite: FSDataOutputStream#write error" log is printed. This log does not affect data saving and can be ignored. When multiple nodes use HDFS for training, use the same HDFS path as the save path. When multi-device training is enabled, you do not need to pass the parameter with the device ID to save_path. The multi-device training results are automatically combined and saved in save_path. When the model is loaded in multi-device mode, the parameters of devices are automatically loaded.
global_step	int, np.int64	Optional	Adds the number of training steps to the checkpoint file name. The default value is None, and the value range is [0, 2147483647].
latest_filename	str	Optional	Optional name of the protocol buffer file. The file contains the latest checkpoint list. The default value is None. Value range: [1, 50].
meta_graph_suffix	str	Optional	Suffix of the MetaGraphDef file. The default value is meta, and the suffix length range is [1, 50].
write_meta_graph	bool	Optional	Whether to write data to the MetaGraph file. The default value is True. Value: True: Write data to the MetaGraph file. False: Do not write data to the MetaGraph file.
write_state	bool	Optional	Whether to write data to the CheckpointStateProto file. The default value is True. Value: True: Write data to the CheckpointStateProto file. False: Do not write data to the CheckpointStateProto file.
strip_default_attrs	bool	Optional	Whether to delete the default attribute from NodeDefs when saving a model file. The default value is False. The value True indicates that the default attribute is deleted from NodeDefs during API call. The value False indicates that the default attribute is not deleted.
save_debug_info	bool	Optional	Whether to save the debugging information. The default value is False. If the value is set to True , the graph debugging information is saved to a separate file in the directory corresponding to save_path, and _debug is added before the extension of the generated file. This parameter takes effect only when write_meta_graph is set to True. If this parameter is set to False, debugging information is not saved.
is_incremental_checkpoint	bool	Optional	Whether to save and load the incremental model. The default value is False. True: enabled False: disabled
save_delta	bool	Optional	Whether to save the incremental model. True: saves the incremental model. False: saves the full model.

Return Value

Success: model_checkpoint_path, that is, the path for saving the model.
Failure: An exception is thrown.

Example

The following briefly describes the usage. For details, see the little demo in the Rec SDK code repository.

#1. Import required libraries.
import tensorflow as tf
from mx_rec.util.initialize import init, get_rank_id
# 2. Build a computational graph.
# ...
# 3. Create a saver.
saver = tf.compat.v1.train.Saver() 
# 4. Obtain the rank ID.
rank_id = get_rank_id()
# 5. Set the number of training steps when the model needs to be saved.
global_step = 200
with tf.compat.v1.Session() as sess:
    saver.save(sess, f"./saved-model/model-{rank_id}", global_step=global_step)

tf.compat.v1.train.Saver.save

Function

Prototype

Parameters

Return Value

Example

See Also