def _yaml_config_run(self, predict_yaml_config_path,
tmp_output_dir, model_meta_file):
predict_config_rewritten = rewrite_config(predict_yaml_config_path,
tmp_output_dir,
replace_dict={'model_meta_file':
('replace with model_meta_file',
model_meta_file),
'output_dir':
('replace with tmp_output_dir',
str(tmp_output_dir))})
predict_outputs_before = glob(os.path.join(str(tmp_output_dir),
'predict_output*',
'features_created*'))
hvc.predict(predict_config_rewritten)
# helper function with assertions shared by all
# tests for hvc.select run with config.yml files
predict_outputs_after = glob(os.path.join(str(tmp_output_dir),
'predict_output*',
'features_created*'))
predict_output = [after for after in predict_outputs_after
if after not in predict_outputs_before]
# should only be one summary output file
if len(predict_output) != 1:
raise ValueError('found wrong number of predict outputs after '
'running .yaml config {}.\n'
'This was the output found: {}'
.format(predict_config_rewritten, predict_output))
else:
predict_output = predict_output[0]
predict = joblib.load(predict_output)
self._generic_predict_asserts(predict)
def test_predict_flatwindow_data_dirs(self, tmp_output_dir, test_data_dir):
# tests predict with svm model, using data dirs
data_dirs = [
'cbins/gy6or6/032312',
'cbins/gy6or6/032412']
data_dirs = [
os.path.join(test_data_dir,
os.path.normpath(data_dir))
for data_dir in data_dirs
]
file_format = 'cbin'
model_meta_file = os.path.join(test_data_dir,
'model_files',
'flatwindow.meta')
output_dir = tmp_output_dir
# explicitly set segment to None because we want to test
# that default behavior works that happens when
# we supply argument for data_dirs parameter, **and**
# segment is set to None (as it should be by default)
segment = None
predict_proba = False
return_predictions = True
predict = hvc.predict(data_dirs=data_dirs,
file_format=file_format,
model_meta_file=model_meta_file,
segment=segment,
output_dir=str(tmp_output_dir),
predict_proba=predict_proba,
return_predictions=return_predictions)
self._generic_predict_asserts(predict)
def test_data_dirs_cbins(self, tmp_output_dir, test_data_dir):
"""test that calling predict doesn't fail when we
pass a data_dirs list that contain cbin audio files"""
data_dirs = [
'cbins/gy6or6/032312',
'cbins/gy6or6/032412']
data_dirs = [
os.path.join(test_data_dir,
os.path.normpath(data_dir))
for data_dir in data_dirs
]
file_format = 'cbin'
model_meta_file = ''
output_dir = tmp_output_dir
# explicitly set segment to None because we want to test
# that default behavior works that happens when
# we supply argument for data_dirs parameter, **and**
# segment is set to None (as it should be by default)
segment = None
predict_proba = False
convert_to = 'cbin' # to check that this works
return_predictions = True
predict = hvc.predict(data_dirs=data_dirs,
file_format=file_format,
model_meta_file=model_meta_file,
segment=segment,
predict_proba=predict_proba,
convert_to=convert_to,
return_predictions=return_predictions)
assert type(predict) == dict
for key in ['labels', 'pred_labels', 'songfile_IDs', 'onsets_Hz', 'offsets_Hz',
'features',]:
assert key in predict
def test_predict_knn_data_dirs_notmat(self, tmp_output_dir, test_data_dir):
# tests predict with knn model, using data dirs, and
# converting output to notmat files
data_dirs = [
'cbins/gy6or6/032312',
'cbins/gy6or6/032412']
data_dirs = [
os.path.join(test_data_dir,
os.path.normpath(data_dir))
for data_dir in data_dirs
]
file_format = 'cbin'
model_meta_file = os.path.join(test_data_dir,
'model_files',
'knn.meta')
output_dir = tmp_output_dir
# explicitly set segment to None because we want to test
# that default behavior works that happens when
# we supply argument for data_dirs parameter, **and**
# segment is set to None (as it should be by default)
segment = None
predict_proba = False
convert_to = 'notmat' # to check that this works
return_predictions = True
predict = hvc.predict(data_dirs=data_dirs,
file_format=file_format,
model_meta_file=model_meta_file,
segment=segment,
output_dir=str(tmp_output_dir),
predict_proba=predict_proba,
convert_to=convert_to,
return_predictions=return_predictions)
assert type(predict) == dict
for key in ['labels', 'pred_labels', 'songfile_IDs', 'onsets_Hz', 'offsets_Hz',
'features',]:
assert key in predict
def run_main_workflow(tmp_output_dir, script_tuple_dict, configs_path):
"""tests main workflow for hybrid-vocal-classifier
by iterating through test_main_workflow_dict,
running the scripts named in each tuple in the dict
"""
extract_config_filename = os.path.join(configs_path,
script_tuple_dict['extract'])
replace_dict = {
'output_dir': ('replace with tmp_output_dir', str(tmp_output_dir))
}
# have to put tmp_output_dir into yaml file
extract_config_rewritten = rewrite_config(extract_config_filename,
tmp_output_dir, replace_dict)
hvc.extract(extract_config_rewritten)
extract_outputs = list(
filter(os.path.isdir,
glob(os.path.join(str(tmp_output_dir), '*extract*'))))
extract_outputs.sort(key=os.path.getmtime)
extract_output_dir = (extract_outputs[-1]
) # [-1] is newest dir, after sort
assert check_extract_output(extract_output_dir)
feature_file = glob(os.path.join(extract_output_dir, 'features_created*'))
feature_file = feature_file[0] # because glob returns list
os.remove(extract_config_rewritten)
select_and_predict_tuples = script_tuple_dict['select and predict']
for select_and_predict_tuple in select_and_predict_tuples:
(select_config_filename,
predict_config_filename) = select_and_predict_tuple
select_config_filename = os.path.join(configs_path,
select_config_filename)
select_config_rewritten = rewrite_config(
select_config_filename,
tmp_output_dir,
replace_dict={
'feature_file': ('replace with feature_file', feature_file),
'output_dir':
('replace with tmp_output_dir', str(tmp_output_dir))
})
hvc.select(select_config_rewritten)
select_outputs = list(
filter(os.path.isdir,
glob(os.path.join(str(tmp_output_dir), '*select*'))))
select_outputs.sort(key=os.path.getmtime)
select_output_dir = (select_outputs[-1]
) # [-1] is newest dir, after sort
assert check_select_output(select_config_rewritten, select_output_dir)
os.remove(select_config_rewritten)
select_outputs.sort(key=os.path.getmtime)
select_output_dir = (select_outputs[-1]
) # [-1] is newest dir, after sort
model_meta_files = glob(os.path.join(select_output_dir, '*', '*meta*'))
replace_dict = {
'model_meta_file':
('replace with model_file', model_meta_files[-1]),
'output_dir': ('replace with tmp_output_dir', str(tmp_output_dir))
}
predict_config_filename_with_path = os.path.join(
configs_path, predict_config_filename)
predict_config_rewritten = rewrite_config(
predict_config_filename_with_path, tmp_output_dir, replace_dict)
hvc.predict(predict_config_rewritten)
os.remove(predict_config_rewritten)
predict_outputs = list(
filter(os.path.isdir,
glob(os.path.join(str(tmp_output_dir), '*predict*'))))
predict_outputs.sort(key=os.path.getmtime)
predict_output_dir = (predict_outputs[-1]
) # [-1] is newest dir, after sort
feature_files = glob(os.path.join(predict_output_dir, 'feature*'))
for ftr_filename in feature_files:
ftr_file = joblib.load(ftr_filename)
assert 'pred_labels' in ftr_file
if 'predict_proba_True' in extract_config_filename:
assert 'pred_probs' in ftr_file
assert ftr_file['pred_labels'].shape[0] == ftr_file[
'pred_probs'].shape[0]
from glob import glob
import hvc
# 0. create training data
# In this case, we download already labeled data from an open repository.
# String in quotes matches with the name of one of the folders in the repository.
hvc.utils.fetch('gy6or6.032612')
# 1. pick a model and 2. extract features for that model
# Model and features are defined in extract.config.yml file.
hvc.extract('gy6or6_autolabel_example.knn.extract.config.yml')
# 3. pick hyperparameters for model
# Load summary feature file to use with helper functions for
# finding best hyperparameters.
summary_file = glob('./extract_output*/summary*')
summary_data = hvc.load_feature_file(summary_file)
# In this case, we picked a k-nearest neighbors model
# and we want to find what value of k will give us the highest accuracy
cv_scores, best_k = hvc.utils.find_best_k(summary_data['features'],
summary_data['labels'],
k_range=range(1, 11))
# 4. Fit the **model** to the data and 5. Select the **best** model
hvc.select('gy6or6_autolabel.example.select.knn.config.yml')
# 6. **Predict** labels for unlabeled data using the fit model.
hvc.predict('gy6or6_autolabel.example.predict.knn.config.yml')