def test_ignore_missing_cues():
weights = xr.DataArray(np.array([[0, 1, 0], [1, 0, 0]]),
coords={
'outcomes': ['o1', 'o2'],
'cues': ['c1', 'c2', 'c3']
},
dims=('outcomes', 'cues'))
events = [(['c1', 'c2', 'c3'], []), (['c1', 'c3'], []), (['c2', 'c4'], []),
(['c1', 'c1'], [])]
reference_activations = np.array([[1, 0, 1, 0], [1, 1, 0, 1]])
with pytest.raises(KeyError):
activations = activation(events,
weights,
number_of_threads=1,
remove_duplicates=True)
activations = activation(events,
weights,
number_of_threads=1,
remove_duplicates=True,
ignore_missing_cues=True)
activations_mp = activation(events,
weights,
number_of_threads=3,
remove_duplicates=True,
ignore_missing_cues=True)
assert np.allclose(reference_activations, activations)
assert np.allclose(reference_activations, activations_mp)
def test_exceptions():
with pytest.raises(ValueError) as e_info:
wm = ndl.dict_ndl(FILE_PATH_SIMPLE,
ALPHA,
BETAS,
remove_duplicates=None)
activation(FILE_PATH_MULTIPLE_CUES, wm)
assert e_info == 'cues or outcomes needs to be unique: cues "a a"; outcomes "A"; use remove_duplicates=True'
with pytest.raises(ValueError) as e_info:
activation(FILE_PATH_MULTIPLE_CUES, "magic")
assert e_info == "Weights other than xarray.DataArray or dicts are not supported."
def test_activation_dict():
weights = defaultdict(lambda: defaultdict(float))
weights['o1']['c1'] = 0
weights['o1']['c2'] = 1
weights['o1']['c3'] = 0
weights['o2']['c1'] = 1
weights['o2']['c2'] = 0
weights['o2']['c3'] = 0
events = [(['c1', 'c2', 'c3'], []), (['c1', 'c3'], []), (['c2'], []),
(['c1', 'c1'], [])]
reference_activations = {'o1': [1, 0, 1, 0], 'o2': [1, 1, 0, 1]}
with pytest.raises(ValueError):
activations = activation(events, weights, number_of_threads=1)
activations = activation(events,
weights,
number_of_threads=1,
remove_duplicates=True)
for outcome, activation_list in activations.items():
assert np.allclose(reference_activations[outcome], activation_list)
def predict_proba_eventfile_NDL(model, data_test, cue_index = None, outcome_index = None, temp_dir = None,
remove_temp_dir = True, T = 1, num_threads = 1, chunksize = None):
""" Generate predicted probabilities for NDL
Parameters
----------
model: class
NDL model outputs (contains weights and activations)
data_test: dataframe or class
dataframe or indexed text file containing test data
cue_index: dict or None
If None, all cues in the event file are used. Otherwise a dictionary that maps cues to indices should
be given. The dictionary should include only the cues to keep in the data. Default: None
outcome_index: dict or None
If None, all outcomes in the event file are used. Otherwise a dictionary that maps outcomes to indices should
be given. The dictionary should include only the outcomes to keep in the data. Default: None
temp_dir: str
directory where to store the converted gz file if a dataframe is passed to data_test
(needed to compute the activation matrix). Default: None (will create a folder
'TEMP_TRAIN_DIRECTORY' in the current working directory
remove_temp_dir: Boolean
whether or not to remove the temporary directory. Default: True
T: float
temperature hyperparameter to adjust the confidence in the predictions from the activations.
Low values increase the confidence in the predictions.
num_threads: int
maximum number of processes to use when computing the activations is the data is unseen. Default: 1
chunksize : int or None
number of lines to use for computing the activation matrix for these lines. If None, all lines will be used.
Default: None
Returns
-------
numpy 2D-array
array of dim (num_events * num_outcomes), which contains, for each event, the predicted probabilities
of the different outcomes
"""
from pyndl.activation import activation
from deep_text_modelling.preprocessing import df_to_gz
# Create a temporary directory if not provided
if not temp_dir:
temp_dir0 = os.path.join(os.getcwd(), 'TEMP_TRAIN_DIRECTORY')
# Add warning if the creation of a temporary directory fails
# (e.g. folder with the same name already existing)
try:
os.mkdir(temp_dir0)
except OSError:
print("Creation of a temporary directory %s failed. This could be because ",
"a folder with the same name already exists or you don't have the ",
"required admin rights on the computer)." % temp_dir0)
else:
temp_dir0 = temp_dir
### Path to the train event file
if isinstance(data_test, str):
events_test_path = data_test
elif isinstance(data_test, pd.DataFrame):
# if temp_dir:
events_test_path = os.path.join(temp_dir, 'data_test_temp.gz')
df_to_gz(data = data_test, gz_outfile = events_test_path)
else:
raise ValueError("data_test should be either a path to an event file or a dataframe")
### Filter the event files by retaining only the cues and outcomes that are in the index system (e.g. most frequent tokens)
### if these index systems are provided by the user. Otherwise, use all cues and/or outcomes
if cue_index or outcome_index: # Filtering only if an index file is provided
# Path to the filtered file
filtered_events_test_path = os.path.join(temp_dir0, 'filtered_events_test.gz')
# Cues
if cue_index:
cues_to_keep = [cue for cue in cue_index.keys()]
else:
cues_to_keep = 'all'
# Outcomes
if outcome_index:
outcomes_to_keep = [outcome for outcome in outcome_index.keys()]
else:
outcomes_to_keep = 'all'
filter_event_file(events_test_path,
filtered_events_test_path,
number_of_processes = num_threads,
keep_cues = cues_to_keep,
keep_outcomes = outcomes_to_keep,
verbose = False)
else:
# Path to the filtered file
filtered_events_test_path = events_test_path
if chunksize:
#N_outcomes = len(outcome_index)
#N_events = len(pd.read_csv(filtered_events_test_path, header = 0, sep='\t', quotechar='"', usecols = ['outcomes']))
#proba_pred = np.empty([N_events, N_outcomes])
proba_pred_arrays = []
events = io.events_from_file(filtered_events_test_path)
for events_chunk in chunk(events, chunksize):
# Generate the activations
activations_test = activation(events = events_chunk,
weights = model.weights,
number_of_threads = 1,
remove_duplicates = True,
ignore_missing_cues = True)
# Predicted probabilities using softmax
proba_pred_arrays.append(activations_to_proba(activations = activations_test, T = T))
proba_pred = np.stack(proba_pred_arrays, axis = 0)
else:
activations_test = activation(events = filtered_events_test_path,
weights = model.weights,
number_of_threads = 1,
remove_duplicates = True,
ignore_missing_cues = True)
# Predicted probabilities using softmax
proba_pred = activations_to_proba(activations = activations_test, T = T)
### Remove temporary directory if it was automatically created and the option was selected by the user
if remove_temp_dir and not temp_dir:
try:
shutil.rmtree(temp_dir0)
except OSError as e:
print("Error: %s : %s" % (temp_dir0, e.strerror))
return proba_pred
def predict_proba_eventfile_NDL(model, data_test, temp_dir = None, remove_temp_dir = True,
T = 1, num_threads = 1, chunksize = 10000):
""" Generate predicted probabilities for NDL
Parameters
----------
model: class
NDL model outputs (contains weights and activations)
data_test: dataframe or class
dataframe or indexed text file containing test data
temp_dir: str
directory where to store the converted gz file if a dataframe is passed to data_test
(needed to compute the activation matrix). Default: None (will create a folder
'TEMP_TRAIN_DIRECTORY' in the current working directory
remove_temp_dir: Boolean
whether or not to remove the temporary directory. Default: True
T: float
temperature hyperparameter to adjust the confidence in the predictions from the activations.
Low values increase the confidence in the predictions.
num_threads: int
maximum number of processes to use when computing the activations is the data is unseen. Default: 1
chunksize : int
number of lines to use for computing the activation matrix for these lines. Default: 10000
Returns
-------
numpy array
array containing the predicted probabilities
"""
from pyndl.activation import activation
from deep_text_modelling.preprocessing import df_to_gz
# Create a temporary directory if not provided
if not temp_dir:
temp_dir0 = os.path.join(os.getcwd(), 'TEMP_TRAIN_DIRECTORY')
# Add warning if the creation of a temporary directory fails
# (e.g. folder with the same name already existing)
try:
os.mkdir(temp_dir0)
except OSError:
print("Creation of a temporary directory %s failed. This could be because ",
"a folder with the same name already exists or you don't have the ",
"required admin rights on the computer)." % temp_dir0)
else:
temp_dir0 = temp_dir
### Path to the train event file
if isinstance(data_test, str):
events_test_path = data_test
elif isinstance(data_test, pd.DataFrame):
# if temp_dir:
events_test_path = os.path.join(temp_dir, 'data_test_temp.gz')
df_to_gz(data = data_test, gz_outfile = events_test_path)
# else:
# raise ValueError("provide a path to a temporary directory for generating a temporary .gz event file")
else:
raise ValueError("data_test should be either a path to an event file or a dataframe")
# Generate the activations
events = io.events_from_file(events_test_path)
for events_chunk in chunk(events, chunksize):
activations_test = activation(events = events_chunk,
weights = model.weights,
number_of_threads = num_threads,
remove_duplicates = True,
ignore_missing_cues = True)
# Predicted probabilities using softmax
proba_pred = activations_to_proba(activations = activations_test, T = T)
### Remove temporary directory if it was automatically created and the option was selected by the user
if remove_temp_dir and temp_dir:
try:
shutil.rmtree(temp_dir0)
except OSError as e:
print("Error: %s : %s" % (temp_dir0, e.strerror))
return proba_pred