def test_sequence_tagger_param_selector(results_base_path, tasks_base_path):
corpus = flair.datasets.ColumnCorpus(
data_folder=tasks_base_path / "fashion", column_format={0: "text", 2: "ner"}
)
# define search space
search_space = SearchSpace()
# sequence tagger parameter
search_space.add(
Parameter.EMBEDDINGS,
hp.choice,
options=[StackedEmbeddings([glove_embedding])],
)
search_space.add(Parameter.USE_CRF, hp.choice, options=[True, False])
search_space.add(Parameter.DROPOUT, hp.uniform, low=0.25, high=0.75)
search_space.add(Parameter.WORD_DROPOUT, hp.uniform, low=0.0, high=0.25)
search_space.add(Parameter.LOCKED_DROPOUT, hp.uniform, low=0.0, high=0.5)
search_space.add(Parameter.HIDDEN_SIZE, hp.choice, options=[64, 128])
search_space.add(Parameter.RNN_LAYERS, hp.choice, options=[1, 2])
# model trainer parameter
search_space.add(Parameter.OPTIMIZER, hp.choice, options=[SGD])
# training parameter
search_space.add(Parameter.MINI_BATCH_SIZE, hp.choice, options=[4, 8, 32])
search_space.add(Parameter.LEARNING_RATE, hp.uniform, low=0.01, high=1)
search_space.add(Parameter.ANNEAL_FACTOR, hp.uniform, low=0.3, high=0.75)
search_space.add(Parameter.PATIENCE, hp.choice, options=[3, 5])
search_space.add(Parameter.WEIGHT_DECAY, hp.uniform, low=0.01, high=1)
# find best parameter settings
optimizer = SequenceTaggerParamSelector(
corpus, "ner", results_base_path, max_epochs=2
)
optimizer.optimize(search_space, max_evals=2)
# clean up results directory
shutil.rmtree(results_base_path)
del optimizer, search_space
def test_text_classifier_param_selector(results_base_path, tasks_base_path):
corpus = flair.datasets.ClassificationCorpus(tasks_base_path / "imdb")
search_space = SearchSpace()
# document embeddings parameter
search_space.add(Parameter.EMBEDDINGS, hp.choice, options=[[glove_embedding]])
search_space.add(Parameter.HIDDEN_SIZE, hp.choice, options=[64, 128, 256, 512])
search_space.add(Parameter.RNN_LAYERS, hp.choice, options=[1, 2])
search_space.add(Parameter.REPROJECT_WORDS, hp.choice, options=[True, False])
search_space.add(Parameter.REPROJECT_WORD_DIMENSION, hp.choice, options=[64, 128])
search_space.add(Parameter.BIDIRECTIONAL, hp.choice, options=[True, False])
search_space.add(Parameter.DROPOUT, hp.uniform, low=0.25, high=0.75)
search_space.add(Parameter.WORD_DROPOUT, hp.uniform, low=0.25, high=0.75)
search_space.add(Parameter.LOCKED_DROPOUT, hp.uniform, low=0.25, high=0.75)
# training parameter
search_space.add(Parameter.LEARNING_RATE, hp.uniform, low=0, high=1)
search_space.add(Parameter.MINI_BATCH_SIZE, hp.choice, options=[4, 8, 16, 32])
search_space.add(Parameter.ANNEAL_FACTOR, hp.uniform, low=0, high=0.75)
search_space.add(Parameter.PATIENCE, hp.choice, options=[3, 5])
param_selector = TextClassifierParamSelector(
corpus, False, results_base_path, document_embedding_type="lstm", max_epochs=2
)
param_selector.optimize(search_space, max_evals=2)
# clean up results directory
shutil.rmtree(results_base_path)
del param_selector, search_space
def select_hyperparameters(params, corpus):
search_space = SearchSpace()
embeddings = create_embeddings(params)
search_space.add(Parameter.EMBEDDINGS, hp.choice, options=[embeddings])
search_space.add(Parameter.HIDDEN_SIZE,
hp.choice,
options=[32, 64, 128, 256, 512])
# search_space.add(Parameter.HIDDEN_SIZE, hp.choice, options=[256])
search_space.add(Parameter.RNN_LAYERS, hp.choice, options=[1, 2])
search_space.add(Parameter.DROPOUT,
hp.choice,
options=[0.3, 0.4, 0.5, 0.6, 0.7])
# search_space.add(Parameter.LEARNING_RATE, hp.loguniform, low=-np.log(0.00001), high=np.log(1.0))
# search_space.add(Parameter.OPTIMIZER, hp.choice, options=[Parameter.NESTEROV])
search_space.add(Parameter.MINI_BATCH_SIZE, hp.choice, options=[16])
print("Downsampling the training set to %10 of the original...")
corpus.downsample(percentage=0.1, only_downsample_train=True)
param_selector = SequenceTaggerParamSelector(
corpus=corpus,
tag_type=params['tag_type'],
base_path=os.path.join("hyperparameter_search",
params['model_output_dirpath']),
max_epochs=5,
training_runs=3,
evaluation_metric=EvaluationMetric.MICRO_F1_SCORE,
optimization_value=OptimizationValue.DEV_SCORE)
param_selector.optimize(search_space, max_evals=10)
print(
"Now observe %s to decide on the best hyperparameters" %
(os.path.join("hyperparameter_search", params['model_output_dirpath'],
"param_selection.txt")))
def test_text_classifier_param_selector(results_base_path, tasks_base_path):
corpus = NLPTaskDataFetcher.load_corpus(u'imdb', base_path=tasks_base_path)
glove_embedding = WordEmbeddings(u'en-glove')
search_space = SearchSpace()
search_space.add(Parameter.EMBEDDINGS,
hp.choice,
options=[[glove_embedding]])
search_space.add(Parameter.HIDDEN_SIZE,
hp.choice,
options=[64, 128, 256, 512])
search_space.add(Parameter.RNN_LAYERS, hp.choice, options=[1, 2])
search_space.add(Parameter.REPROJECT_WORDS,
hp.choice,
options=[True, False])
search_space.add(Parameter.REPROJECT_WORD_DIMENSION,
hp.choice,
options=[64, 128])
search_space.add(Parameter.BIDIRECTIONAL, hp.choice, options=[True, False])
search_space.add(Parameter.DROPOUT, hp.uniform, low=0.25, high=0.75)
search_space.add(Parameter.WORD_DROPOUT, hp.uniform, low=0.25, high=0.75)
search_space.add(Parameter.LOCKED_DROPOUT, hp.uniform, low=0.25, high=0.75)
search_space.add(Parameter.LEARNING_RATE, hp.uniform, low=0, high=1)
search_space.add(Parameter.MINI_BATCH_SIZE,
hp.choice,
options=[4, 8, 16, 32])
search_space.add(Parameter.ANNEAL_FACTOR, hp.uniform, low=0, high=0.75)
search_space.add(Parameter.PATIENCE, hp.choice, options=[3, 5])
param_selector = TextClassifierParamSelector(
corpus,
False,
results_base_path,
document_embedding_type=u'lstm',
max_epochs=2)
param_selector.optimize(search_space, max_evals=2)
shutil.rmtree(results_base_path)
def test_sequence_tagger_param_selector(results_base_path, tasks_base_path):
corpus = NLPTaskDataFetcher.load_corpus(NLPTask.FASHION,
base_path=tasks_base_path)
search_space = SearchSpace()
search_space.add(Parameter.EMBEDDINGS,
hp.choice,
options=[
StackedEmbeddings([WordEmbeddings(u'glove')]),
StackedEmbeddings([
WordEmbeddings(u'glove'),
FlairEmbeddings(u'news-forward'),
FlairEmbeddings(u'news-backward')
])
])
search_space.add(Parameter.USE_CRF, hp.choice, options=[True, False])
search_space.add(Parameter.DROPOUT, hp.uniform, low=0.25, high=0.75)
search_space.add(Parameter.WORD_DROPOUT, hp.uniform, low=0.0, high=0.25)
search_space.add(Parameter.LOCKED_DROPOUT, hp.uniform, low=0.0, high=0.5)
search_space.add(Parameter.HIDDEN_SIZE, hp.choice, options=[64, 128])
search_space.add(Parameter.RNN_LAYERS, hp.choice, options=[1, 2])
search_space.add(Parameter.OPTIMIZER, hp.choice, options=[SGD])
search_space.add(Parameter.MINI_BATCH_SIZE, hp.choice, options=[4, 8, 32])
search_space.add(Parameter.LEARNING_RATE, hp.uniform, low=0.01, high=1)
search_space.add(Parameter.ANNEAL_FACTOR, hp.uniform, low=0.3, high=0.75)
search_space.add(Parameter.PATIENCE, hp.choice, options=[3, 5])
search_space.add(Parameter.WEIGHT_DECAY, hp.uniform, low=0.01, high=1)
optimizer = SequenceTaggerParamSelector(corpus,
u'ner',
results_base_path,
max_epochs=2)
optimizer.optimize(search_space, max_evals=2)
shutil.rmtree(results_base_path)
def test_text_classifier_param_selector(results_base_path, tasks_base_path):
corpus = flair.datasets.ClassificationCorpus(tasks_base_path / "imdb")
label_type = "sentiment"
search_space = SearchSpace()
# document embeddings parameter
search_space.add(Parameter.TRANSFORMER_MODEL,
hp.choice,
options=["albert-base-v1"])
search_space.add(Parameter.LAYERS, hp.choice, options=["-1", "-2"])
# training parameter
search_space.add(Parameter.LEARNING_RATE, hp.uniform, low=0, high=1)
search_space.add(Parameter.MINI_BATCH_SIZE,
hp.choice,
options=[4, 8, 16, 32])
search_space.add(Parameter.ANNEAL_FACTOR, hp.uniform, low=0, high=0.75)
search_space.add(Parameter.PATIENCE, hp.choice, options=[3, 5])
param_selector = TextClassifierParamSelector(corpus,
label_type,
False,
results_base_path,
max_epochs=2)
param_selector.optimize(search_space, max_evals=2)
# clean up results directory
shutil.rmtree(results_base_path)
del param_selector, search_space
def test_sequence_tagger_param_selector(results_base_path, tasks_base_path):
corpus = flair.datasets.ColumnCorpus(data_folder=(tasks_base_path /
'fashion'),
column_format={
0: 'text',
2: 'ner',
})
search_space = SearchSpace()
search_space.add(Parameter.EMBEDDINGS,
hp.choice,
options=[
StackedEmbeddings([WordEmbeddings('glove')]),
StackedEmbeddings([
WordEmbeddings('glove'),
FlairEmbeddings('news-forward-fast'),
FlairEmbeddings('news-backward-fast')
])
])
search_space.add(Parameter.USE_CRF, hp.choice, options=[True, False])
search_space.add(Parameter.DROPOUT, hp.uniform, low=0.25, high=0.75)
search_space.add(Parameter.WORD_DROPOUT, hp.uniform, low=0.0, high=0.25)
search_space.add(Parameter.LOCKED_DROPOUT, hp.uniform, low=0.0, high=0.5)
search_space.add(Parameter.HIDDEN_SIZE, hp.choice, options=[64, 128])
search_space.add(Parameter.RNN_LAYERS, hp.choice, options=[1, 2])
search_space.add(Parameter.OPTIMIZER, hp.choice, options=[SGD])
search_space.add(Parameter.MINI_BATCH_SIZE, hp.choice, options=[4, 8, 32])
search_space.add(Parameter.LEARNING_RATE, hp.uniform, low=0.01, high=1)
search_space.add(Parameter.ANNEAL_FACTOR, hp.uniform, low=0.3, high=0.75)
search_space.add(Parameter.PATIENCE, hp.choice, options=[3, 5])
search_space.add(Parameter.WEIGHT_DECAY, hp.uniform, low=0.01, high=1)
optimizer = SequenceTaggerParamSelector(corpus,
'ner',
results_base_path,
max_epochs=2)
optimizer.optimize(search_space, max_evals=2)
shutil.rmtree(results_base_path)
