Parsing Utilities

Overview

Parsing utilities are a set of functions that helps generate parsing spec for tf$parse_example to be used with estimators. If users keep data in tf$Example format, they need to call tf$parse_example with a proper feature spec. There are two main things that these utility functions help:

• Users need to combine parsing spec of features with labels and weights (if any) since they are all parsed from same tf$Example instance. The utility functions combine these specs.

• It is difficult to map expected label by a estimator such as dnn_classifier to corresponding tf$parse_example spec. The utility functions encode it by getting related information from users (key, dtype).

Example output of parsing spec

parsing_spec <- classifier_parse_example_spec(
  feature_columns = column_numeric('a'),
  label_key = 'b',
  weight_column = 'c'
)

For the above example, classifier_parse_example_spec would return the following:

expected_spec <- list(
  a = tf$python$ops$parsing_ops$FixedLenFeature(reticulate::tuple(1L), dtype = tf$float32),
  c = tf$python$ops$parsing_ops$FixedLenFeature(reticulate::tuple(1L), dtype = tf$float32),
  b = tf$python$ops$parsing_ops$FixedLenFeature(reticulate::tuple(1L), dtype = tf$int64)
)

# This should be the same as the one we constructed using `classifier_parse_example_spec`
testthat::expect_equal(parsing_spec, expected_spec)

Example usage with a classifier

Firstly, define features transformations and initiailize your classifier similar to the following:

fcs <- feature_columns(...)

model <- dnn_classifier(
  n_classes = 1000,
  feature_columns = fcs,
  weight_column = 'example-weight',
  label_vocabulary= c('photos', 'keep', ...),
  hidden_units = c(256, 64, 16)
)

Next, create the parsing configuration for tf$parse_example using classifier_parse_example_spec and the feature columns fcs we have just defined:

parsing_spec <- classifier_parse_example_spec(
  feature_columns = fcs,
  label_key = 'my-label',
  label_dtype = tf$string,
  weight_column = 'example-weight'
)

This label configuration tells the classifier the following:

• weights are retrieved with key ‘example-weight’
• label is string and can be one of the following c('photos', 'keep', ...)
• integer id for label ‘photos’ is 0, ‘keep’ is 1, etc

Then define your input function with the help of read_batch_features that reads the batches of features from files in tf$Example format with the parsing configuration parsing_spec we just defined:

input_fn_train <- function() {
  features <- tf$contrib$learn$read_batch_features(
    file_pattern = train_files,
    batch_size = batch_size,
    features = parsing_spec,
    reader = tf$RecordIOReader)
  labels <- features[["my-label"]]
  return(list(features, labels))
}

Finally we can train the model using the training input function parsed by classifier_parse_example_spec:

train(model, input_fn = input_fn_train)