ConfigProto from tensorflow, serialized into byte array. Get with config_proto.SerializeToString()

Dropout coefficient, by default 0.5.

The coefficient of the dropout layer. The value should be between 0.0 and 1.0. Internally, it is used by Tensorflow as: rate = 1.0 - dropout when adding a dropout layer on top of the recurrent layers.

If set, this param specifies the criterion to stop training if performance is not improving.

Default value is 0 which is means that early stopping is not used.

The criterion is set to F1-score if the validationSplit is greater than 0.0 (F1-socre on validation set) or testDataset is defined (F1-score on test set), otherwise it is set to model loss. The priority is as follows: - If testDataset is defined, then the criterion is set to F1-score on test set. - If validationSplit is greater than 0.0, then the criterion is set to F1-score on validation set. - Otherwise, the criterion is set to model loss.

Note that while the F1-score ranges from 0.0 to 1.0, the loss ranges from 0.0 to infinity. So, depending on which case you are in, the value you use for the criterion can be very different. For example, if validationSplit is 0.1, then a criterion of 0.01 means that if the F1-score on the validation set difference from last epoch is greater than 0.01, then the training should stop. However, if there is not validation or test set defined, then a criterion of 2.0 means that if the loss difference between the last epoch and the current one is less than 2.0, then training should stop.

Number of epochs to wait before early stopping if no improvement, by default 5.

Given the earlyStoppingCriterion, if the performance does not improve for the given number of epochs, then the training will stop. If the value is 0, then early stopping will occurs as soon as the criterion is met (no patience).

Whether to optimize for large datasets or not. Enabling this option can slow down training.

In practice, if set to true the training will iterate over the spark Data Frame and retrieve the batches from the Data Frame iterator. This can be slower than the default option as it has to collect the batches on evey bach for every epoch, but it can be useful if the dataset is too large to fit in memory.

It controls if we want the features collected and generated at once and then feed into the network batch by batch (False) or collected and generated by batch and then feed into the network in batches (True) .

If the training data can fit to memory, then it is recommended to set this option to False (default value).

ConfigProto from tensorflow, serialized into byte array. Get with config_proto.SerializeToString()

Whether to optimize for large datasets or not. Enabling this option can slow down training.

Learning rate decay coefficient. Real Learning Rage = lr / (1 + po * epoch)

Whether to use contrib LSTM Cells. Not compatible with Windows. Might slightly improve accuracy.

Path that contains the external graph file.

When specified, the provided file will be used, and no graph search will happen. The path can be a local file path, a distributed file path (HDFS, DBFS), or a cloud storage (S3).

Folder path that contains external graph files.

The path can be a local file path, a distributed file path (HDFS, DBFS), or a cloud storage (S3).

When instantiating the Tensorflow model, uses this folder to search for the adequate Tensorflow graph. The search is done using the name of the .pb file, which should be in this format: blstn_{ntags}_{embedding_dim}_{lstm_size}_{nchars}.pb.

Then, the search follows these rules: - Embedding dimension should be exactly the same as the one used to train the model. - Number of unique tags should be greater than or equal to the number of unique tags in the training data. - Number of unique chars should be greater than or equal to the number of unique chars in the training data.

The returned file will be the first one that satisfies all the conditions.

If the name of the file is ill-formed, errors will occur during training.

Whether to include confidence scores for all tags in annotation metadata or just the score of the predicted tag, by default False.

Needs the includeConfidence parameter to be set to true.

Enabling this may slow down the inference speed.

Whether to include confidence scores in annotation metadata, by default False.

Setting this parameter to True will add the confidence score to the metadata of the NAMED_ENTITY annotation. In addition, if includeAllConfidenceScores is set to true, then the confidence scores of all the tags will be added to the metadata, otherwise only for the predicted tag (the one with maximum score).

Controls whether to override already learned tags when using a pretrained model to initialize the new model. A value of true will override existing tags.

Learning rate decay coefficient (time-based).

This is used to calculate the decayed learning rate at each step as: lr = lr / (1 + po * epoch), meaning that the value of the learning rate is updated on each epoch. By default 0.005.

Path to an already trained MedicalNerModel.

This pretrained model will be used as a starting point for training the new one. The path can be a local file path, a distributed file path (HDFS, DBFS), or a cloud storage (S3).

whether to include the token index for each sentence in annotation metadata, by default false. If the value is true, the process might be slowed down.

ConfigProto from tensorflow, serialized into byte array. Get with config_proto.SerializeToString()

Controls whether to override already learned tags when using a pretrained model to initialize the new model. A value of true will override existing tags.

Learning rate decay coefficient. Real Learning Rage = lr / (1 + po * epoch)

whether to include the token index for each sentence in annotation metadata, by default false. If the value is true, the process might be slowed down.

A map specifying how old tags are mapped to new ones. Maps are specified either using a list of comma separated strings, e.g. ("OLDTAG1,NEWTAG1", "OLDTAG2,NEWTAG2", ...) or by a Map data structure.

A map specifying how old tags are mapped to new ones. Maps are specified either using a list of comma separated strings, e.g. ("OLDTAG1,NEWTAG1", "OLDTAG2,NEWTAG2", ...) or by a Map data structure. It only works if setOverrideExistingTags is false.

Whether to use contrib LSTM Cells. Not compatible with Windows. Might slightly improve accuracy.

A map specifying how old tags are mapped to new ones.

It only works if overrideExistingTags is set to false.

Whether to restore and use the model from the epoch that has achieved the best performance at the end of the training.

By default false (keep the model from the last trained epoch).

The best model depends on the earlyStoppingCriterion, which can be F1-score on test/validation dataset or the value of loss.

whether to use contrib LSTM Cells. Not compatible with Windows. Might slightly improve accuracy. By default true.