Text embedding classifier with a ProtoNet

Abstract class for neural nets with 'keras'/'tensorflow' and 'pytorch'.

This object represents in implementation of a prototypical network for few-shot learning as described by Snell, Swersky, and Zemel (2017). The network uses a multi way contrastive loss described by Zhang et al. (2019). The network learns to scale the metric as described by Oreshkin, Rodriguez, and Lacoste (2018)

Value

Objects of this class are used for assigning texts to classes/categories. For the creation and training of a classifier an object of class EmbeddedText or LargeDataSetForTextEmbeddings and a factor are necessary. The object of class EmbeddedText or LargeDataSetForTextEmbeddings contains the numerical text representations (text embeddings) of the raw texts generated by an object of class TextEmbeddingModel. The factor contains the classes/categories for every text. Missing values (unlabeled cases) are supported. For predictions an object of class EmbeddedText or LargeDataSetForTextEmbeddings has to be used which was created with the same TextEmbeddingModel as for training.

References

Oreshkin, B. N., Rodriguez, P. & Lacoste, A. (2018). TADAM: Task dependent adaptive metric for improved few-shot learning. https://doi.org/10.48550/arXiv.1805.10123

Snell, J., Swersky, K. & Zemel, R. S. (2017). Prototypical Networks for Few-shot Learning. https://doi.org/10.48550/arXiv.1703.05175

Zhang, X., Nie, J., Zong, L., Yu, H. & Liang, W. (2019). One Shot Learning with Margin. In Q. Yang, Z.-H. Zhou, Z. Gong, M.-L. Zhang & S.-J. Huang (Eds.), Lecture Notes in Computer Science. Advances in Knowledge Discovery and Data Mining (Vol. 11440, pp. 305–317). Springer International Publishing. https://doi.org/10.1007/978-3-030-16145-3_24

Super classes

aifeducation::AIFEBaseModel -> aifeducation::TEClassifierRegular -> TEClassifierProtoNet

Methods

Inherited methods

Method `configure()`

Creating a new instance of this class.

Usage

TEClassifierProtoNet$configure(
  ml_framework = "pytorch",
  name = NULL,
  label = NULL,
  text_embeddings = NULL,
  feature_extractor = NULL,
  target_levels = NULL,
  dense_size = 4,
  dense_layers = 0,
  rec_size = 4,
  rec_layers = 2,
  rec_type = "gru",
  rec_bidirectional = FALSE,
  embedding_dim = 2,
  self_attention_heads = 0,
  intermediate_size = NULL,
  attention_type = "fourier",
  add_pos_embedding = TRUE,
  rec_dropout = 0.1,
  repeat_encoder = 1,
  dense_dropout = 0.4,
  recurrent_dropout = 0.4,
  encoder_dropout = 0.1,
  optimizer = "adam"
)

Arguments

ml_framework: string Currently only pytorch is supported (ml_framework="pytorch").
name: string Name of the new classifier. Please refer to common name conventions. Free text can be used with parameter label.
label: string Label for the new classifier. Here you can use free text.
text_embeddings: An object of class TextEmbeddingModel or LargeDataSetForTextEmbeddings.
feature_extractor: Object of class TEFeatureExtractor which should be used in order to reduce the number of dimensions of the text embeddings. If no feature extractor should be applied set NULL.
target_levels: vector containing the levels (categories or classes) within the target data. Please not that order matters. For ordinal data please ensure that the levels are sorted correctly with later levels indicating a higher category/class. For nominal data the order does not matter.
dense_size: int Number of neurons for each dense layer.
dense_layers: int Number of dense layers.
rec_size: int Number of neurons for each recurrent layer.
rec_layers: int Number of recurrent layers.
rec_type: string Type of the recurrent layers.rec_type="gru" for Gated Recurrent Unit and rec_type="lstm" for Long Short-Term Memory.
rec_bidirectional: bool If TRUE a bidirectional version of the recurrent layers is used.
embedding_dim: int determining the number of dimensions for the text embedding.
self_attention_heads: int determining the number of attention heads for a self-attention layer. Only relevant if attention_type="multihead".
intermediate_size: int determining the size of the projection layer within a each transformer encoder.
attention_type: string Choose the relevant attention type. Possible values are "fourier" and "multihead". Please note that you may see different values for a case for different input orders if you choose fourier on linux.
add_pos_embedding: bool TRUE if positional embedding should be used.
rec_dropout: double ranging between 0 and lower 1, determining the dropout between bidirectional recurrent layers.
repeat_encoder: int determining how many times the encoder should be added to the network.
dense_dropout: double ranging between 0 and lower 1, determining the dropout between dense layers.
recurrent_dropout: double ranging between 0 and lower 1, determining the recurrent dropout for each recurrent layer. Only relevant for keras models.
encoder_dropout: double ranging between 0 and lower 1, determining the dropout for the dense projection within the encoder layers.
optimizer: string "adam" or "rmsprop" .

Returns

Returns an object of class TEClassifierProtoNet which is ready for training.

Method `train()`

Method for training a neural net.

Training includes a routine for early stopping. In the case that loss<0.0001 and Accuracy=1.00 and Average Iota=1.00 training stops. The history uses the values of the last trained epoch for the remaining epochs.

After training the model with the best values for Average Iota, Accuracy, and Loss on the validation data set is used as the final model.

Usage

TEClassifierProtoNet$train(
  data_embeddings,
  data_targets,
  data_folds = 5,
  data_val_size = 0.25,
  use_sc = TRUE,
  sc_method = "dbsmote",
  sc_min_k = 1,
  sc_max_k = 10,
  use_pl = TRUE,
  pl_max_steps = 3,
  pl_max = 1,
  pl_anchor = 1,
  pl_min = 0,
  sustain_track = TRUE,
  sustain_iso_code = NULL,
  sustain_region = NULL,
  sustain_interval = 15,
  epochs = 40,
  batch_size = 35,
  Ns = 5,
  Nq = 3,
  loss_alpha = 0.5,
  loss_margin = 0.5,
  sampling_separate = FALSE,
  sampling_shuffle = TRUE,
  dir_checkpoint,
  trace = TRUE,
  ml_trace = 1,
  log_dir = NULL,
  log_write_interval = 10,
  n_cores = auto_n_cores()
)

Arguments

data_embeddings: Object of class EmbeddedText or LargeDataSetForTextEmbeddings.
data_targets: factor containing the labels for cases stored in data_embeddings. Factor must be named and has to use the same names used in data_embeddings.
data_folds: int determining the number of cross-fold samples.
data_val_size: double between 0 and 1, indicating the proportion of cases of each class which should be used for the validation sample during the estimation of the model. The remaining cases are part of the training data.
use_sc: bool TRUE if the estimation should integrate synthetic cases. FALSE if not.
sc_method: vector containing the method for generating synthetic cases. Possible are sc_method="adas", sc_method="smote", and sc_method="dbsmote".
sc_min_k: int determining the minimal number of k which is used for creating synthetic units.
sc_max_k: int determining the maximal number of k which is used for creating synthetic units.
use_pl: bool TRUE if the estimation should integrate pseudo-labeling. FALSE if not.
pl_max_steps: int determining the maximum number of steps during pseudo-labeling.
pl_max: double between 0 and 1, setting the maximal level of confidence for considering a case for pseudo-labeling.
pl_anchor: double between 0 and 1 indicating the reference point for sorting the new cases of every label. See notes for more details.
pl_min: double between 0 and 1, setting the minimal level of confidence for considering a case for pseudo-labeling.
sustain_track: bool If TRUE energy consumption is tracked during training via the python library 'codecarbon'.
sustain_iso_code: string ISO code (Alpha-3-Code) for the country. This variable must be set if sustainability should be tracked. A list can be found on Wikipedia: https://en.wikipedia.org/wiki/List_of_ISO_3166_country_codes.
sustain_region: Region within a country. Only available for USA and Canada See the documentation of codecarbon for more information. https://mlco2.github.io/codecarbon/parameters.html
sustain_interval: int Interval in seconds for measuring power usage.
epochs: int Number of training epochs.
batch_size: int Size of the batches for training.
Ns: int Number of cases for every class in the sample.
Nq: int Number of cases for every class in the query.
loss_alpha: double Value between 0 and 1 indicating how strong the loss should focus on pulling cases to its corresponding prototypes or pushing cases away from other prototypes. The higher the value the more the loss concentrates on pulling cases to its corresponding prototypes.
loss_margin: double Value greater 0 indicating the minimal distance of every case from prototypes of other classes
sampling_separate: bool If TRUE the cases for every class are divided into a data set for sample and for query. These are never mixed. If TRUE sample and query cases are drawn from the same data pool. That is, a case can be part of sample in one epoch and in another epoch it can be part of query. It is ensured that a case is never part of sample and query at the same time. In addition, it is ensured that every cases exists only once during a training step.
sampling_shuffle: bool If TRUE cases a randomly drawn from the data during every step. If FALSE the cases are not shuffled.
dir_checkpoint: string Path to the directory where the checkpoint during training should be saved. If the directory does not exist, it is created.
trace: bool TRUE, if information about the estimation phase should be printed to the console.
ml_trace: int ml_trace=0 does not print any information about the training process from pytorch on the console.
log_dir: string Path to the directory where the log files should be saved. If no logging is desired set this argument to NULL.
log_write_interval: int Time in seconds determining the interval in which the logger should try to update the log files. Only relevant if log_dir is not NULL.
n_cores: int Number of cores which should be used during the calculation of synthetic cases. Only relevant if use_sc=TRUE.
balance_class_weights: bool If TRUE class weights are generated based on the frequencies of the training data with the method Inverse Class Frequency'. If FALSE each class has the weight 1.
balance_sequence_length: bool If TRUE sample weights are generated for the length of sequences based on the frequencies of the training data with the method Inverse Class Frequency'. If FALSE each sequences length has the weight 1.

Details

sc_max_k: All values from sc_min_k up to sc_max_k are successively used. If the number of sc_max_k is too high, the value is reduced to a number that allows the calculating of synthetic units.
pl_anchor: With the help of this value, the new cases are sorted. For this aim, the distance from the anchor is calculated and all cases are arranged into an ascending order.

Returns

Function does not return a value. It changes the object into a trained classifier.

Method `embed()`

Method for embedding documents. Please do not confuse this type of embeddings with the embeddings of texts created by an object of class TextEmbeddingModel. These embeddings embed documents according to their similarity to specific classes.

Usage

TEClassifierProtoNet$embed(embeddings_q = NULL, batch_size = 32)

Arguments

embeddings_q: Object of class EmbeddedText or LargeDataSetForTextEmbeddings containing the text embeddings for all cases which should be embedded into the classification space.
batch_size: int batch size.

Returns

Returns a list containing the following elements

embeddings_q: embeddings for the cases (query sample).
embeddings_prototypes: embeddings of the prototypes which were learned during training. They represents the center for the different classes.

Method `plot_embeddings()`

Method for creating a plot to visualize embeddings and their corresponding centers (prototypes).

Usage

TEClassifierProtoNet$plot_embeddings(
  embeddings_q,
  classes_q = NULL,
  batch_size = 12,
  alpha = 0.5,
  size_points = 3,
  size_points_prototypes = 8,
  inc_unlabeled = TRUE
)

Arguments

embeddings_q: Object of class EmbeddedText or LargeDataSetForTextEmbeddings containing the text embeddings for all cases which should be embedded into the classification space.
classes_q: Named factor containg the true classes for every case. Please note that the names must match the names/ids in embeddings_q.
batch_size: int batch size.
alpha: float Value indicating how transparent the points should be (important if many points overlap). Does not apply to points representing prototypes.
size_points: int Size of the points excluding the points for prototypes.
size_points_prototypes: int Size of points representing prototypes.
inc_unlabeled: bool If TRUE plot includes unlabeled cases as data points.

Returns

Returns a plot of class ggplotvisualizing embeddings.

Method `clone()`

The objects of this class are cloneable with this method.

Usage

TEClassifierProtoNet$clone(deep = FALSE)

Arguments

deep: Whether to make a deep clone.

Value

References

See also

Super classes

Methods

Public methods

Method configure()

Usage

Arguments

Returns

Method train()

Usage

Arguments

Details

Returns

Method embed()

Usage

Arguments

Returns

Method plot_embeddings()

Usage

Arguments

Returns

Method clone()

Usage

Arguments

Method `configure()`

Method `train()`

Method `embed()`

Method `plot_embeddings()`

Method `clone()`