Commit 04b847a2 authored by Vít Starý Novotný's avatar Vít Starý Novotný
Browse files

Evaluate more often with fewer samples

parent b3810c32

ahisto_named_entity_search/default.ini

+3 −2
Original line number Diff line number Diff line
@@ -94,8 +94,9 @@ base_model = xlm-roberta-base 
batch_size = 4

gradient_accumulation_steps = 4

log_every_n_steps = 100

evaluate_every_n_steps = 10000

save_every_n_steps = 10000

evaluate_every_n_steps = 1000

save_every_n_steps = 1000

number_of_validation_samples = 1000

number_of_training_epochs = 10

schedule = fair-sequential-schedule

ahisto_named_entity_search/recognition/model.py

+4 −0
Original line number Diff line number Diff line
@@ -33,6 +33,7 @@ class NerModel: 
    LOGGING_STEPS = CONFIG.getint('log_every_n_steps')

    NUM_TRAIN_EPOCHS = CONFIG.getint('number_of_training_epochs')

    SCHEDULE_NAME = CONFIG['schedule']

    NUM_VALIDATION_SAMPLES = CONFIG.getint('number_of_validation_samples')



    def __init__(self, model: AutoModelForTokenClassification):

        self.model = model
@@ -52,6 +53,7 @@ class NerModel: 
        # Set up masked language modeling (MLM) training

        mlm_training_texts = list(Document.load_sentences(training_sentence_basename))

        mlm_validation_texts = list(Document.load_sentences(validation_sentence_basename))

        mlm_validation_texts = mlm_validation_texts[:cls.NUM_VALIDATION_SAMPLES]



        mlm_objective = MaskedLanguageModeling(lang_module,

                                               batch_size=cls.BATCH_SIZE,
@@ -68,6 +70,8 @@ class NerModel: 


        ner_training_texts, ner_training_labels = load_ner_dataset(training_tagged_sentence_basename)

        ner_validation_texts, ner_validation_labels = load_ner_dataset(validation_tagged_sentence_basename)

        ner_validation_texts = ner_validation_texts[:cls.NUM_VALIDATION_SAMPLES]

        ner_validation_labels = ner_validation_labels[:cls.NUM_VALIDATION_SAMPLES]



        ner_evaluators = [MeanFScore(decides_convergence=True)]

        ner_objective = TokenClassification(lang_module,