def collect_gcnn(name, most_similar_name, perc):
gcnn_results = []
data = dataTools.Authorship(name, ratioTrain, ratioValid, dataPath)
logging.info('Training GCNN on {0}'.format(name))
h_params = ClusterUtils.load_best_hyperparams(name)
phi, indices = load_phi(name, data, percentage=perc)
if indices.shape[0] < 2:
return [{'acc': 0, 'f1': 0, 'auc': 0, 'prec': 0, 'time': 0}]
for split_n in range(NO_OF_SPLITS):
data.get_split(name, ratioTrain, ratioValid)
data.reduce_dim(indices)
start = time.perf_counter()
gcnn = train_helper.train_net(data, h_params, phi=phi)
end = time.perf_counter()
gcnn_eval = evaluate_gcnn(gcnn, data)
gcnn_eval['time'] = start - end
gcnn_results.append(gcnn_eval)
logging.info(
'SPLIT {0}: GCNN results successfully collected: {1}'.format(
split_n, gcnn_results[split_n]))
return gcnn_results
def collect_gcnn(name, most_similar_name, phi, ind):
gcnn_results = []
data = dataTools.AuthorshipOneVsOne(name, most_similar_name, ratioTrain,
ratioValid, dataPath)
logging.info('Training GCNN on {0} v.s. {1}'.format(
name, most_similar_name))
h_params = ClusterUtils.load_best_hyperparams(name)
for split_n in range(NO_OF_SPLITS):
data.get_split(name, most_similar_name, ratioTrain, ratioValid)
data.reduce_dim(ind)
start = time.perf_counter()
gcnn = train_helper.train_net(data, h_params, phi=phi)
end = time.perf_counter()
gcnn_eval = evaluate_gcnn(gcnn, data)
gcnn_eval['time'] = start - end
gcnn_results.append(gcnn_eval)
logging.info(
'SPLIT {0}: GCNN results successfully collected: {1}'.format(
split_n, gcnn_results[split_n]))
return gcnn_results
def collect_gcnn(name, most_similar_name):
gcnn_results = []
data = dataTools.AuthorshipOneVsOne(name, most_similar_name, ratioTrain, ratioValid, dataPath)
logging.info('Training GCNN on {0} v.s. {1}'.format(name, most_similar_name))
h_params = ClusterUtils.load_best_hyperparams(name)
for split_n in range(NO_OF_SPLITS):
data.get_split(name, most_similar_name, ratioTrain, ratioValid)
gcnn = train_helper.train_net(data, h_params)
gcnn_results.append(evaluate_gcnn(gcnn, data))
logging.info('SPLIT {0}: GCNN results successfully collected: {1}'.format(split_n, gcnn_results[split_n]))
return gcnn_results
def delete_active_author(name, active_file, signal, frame):
print('EXIT F-TION')
ClusterUtils.delete_from_active(active_file, name)
# randomStates[0]['state'] = numpyState
# randomStates.append({})
# randomStates[1]['module'] = 'torch'
# randomStates[1]['state'] = torchState
# randomStates[1]['seed'] = torchSeed
# This list and dictionary follows the format to then be loaded, if needed,
# by calling the loadSeed function in Utils.miscTools
# saveSeed(randomStates, saveDir)
########
# DATA #
########
# find the next author
combinations = [1, 3, 5, 7, 9]
authorName = ClusterUtils.get_author_name(ACTIVE_AUTHORS_FILE, BASE_FILE_NAME,
combinations)
try:
atexit.register(delete_active_author, authorName, ACTIVE_AUTHORS_FILE,
None, None)
for sig in signal.Signals:
try:
# signal.signal(sig, test_fn)
signal.signal(
sig,
partial(delete_active_author, authorName, ACTIVE_AUTHORS_FILE))
except (ValueError, OSError):
print('invalid: ' + str(sig))
# signal.signal(signal.SIGHUP, partial(delete_active_author, authorName, ACTIVE_AUTHORS_FILE))
randomStates[0]['state'] = numpyState
randomStates.append({})
randomStates[1]['module'] = 'torch'
randomStates[1]['state'] = torchState
randomStates[1]['seed'] = torchSeed
# This list and dictionary follows the format to then be loaded, if needed,
# by calling the loadSeed function in Utils.miscTools
saveSeed(randomStates, saveDir)
########
# DATA #
########
# find the next author
authorName = ClusterUtils.get_author_name(ACTIVE_AUTHORS_FILE,
BASE_FILE_NAME, [],
hyperparams_path=HYPER_PARAM_FILE)
if authorName is None:
exit()
# Possible authors: (just use the names in ' ')
# jacob 'abbott', robert louis 'stevenson', louisa may 'alcott',
# horatio 'alger', james 'allen', jane 'austen', emily 'bronte', james 'cooper',
# charles 'dickens', hamlin 'garland', nathaniel 'hawthorne', henry 'james',
# herman 'melville', 'page', herny 'thoreau', mark 'twain',
# arthur conan 'doyle', washington 'irving', edgar allan 'poe',
# sarah orne 'jewett', edith 'wharton'
try:
atexit.register(delete_active_author, authorName, ACTIVE_AUTHORS_FILE,
None, None)
file_name = "{0}{1}.txt".format(BASE_FILE_NAME, authorName)
# create empty files so that other jobs would skip this author
with open(file_name, mode='w+') as f:
pass
# Possible authors: (just use the names in ' ')
# jacob 'abbott', robert louis 'stevenson', louisa may 'alcott',
# horatio 'alger', james 'allen', jane 'austen', emily 'bronte', james 'cooper',
# charles 'dickens', hamlin 'garland', nathaniel 'hawthorne', henry 'james',
# herman 'melville', 'page', herny 'thoreau', mark 'twain',
# arthur conan 'doyle', washington 'irving', edgar allan 'poe',
# sarah orne 'jewett', edith 'wharton'
nFeatures, nShifts = ClusterUtils.load_best_hyperparams(authorName)
if doPrint:
print('Author: {0}, Combination: {1}'.format(authorName, str((nFeatures, nShifts))))
# set training params
nClasses = 1 # Either authorName or not
ratioTrain = 0.6 # Ratio of training samples
ratioValid = 0.2 # Ratio of validation samples (out of the total training
# samples)
# Final split is:
# nValidation = round(ratioValid * ratioTrain * nTotal)
# nTrain = round((1 - ratioValid) * ratioTrain * nTotal)
# nTest = nTotal - nTrain - nValidation
nDataSplits = 7 # Number of data realizations