def objective_function_veracity_branchLSTM_fullPHEME(params):
path = 'saved_data_fullPHEME'
train = [
'ebola-essien', 'ferguson', 'gurlitt', 'ottawashooting',
'prince-toronto', 'putinmissing', 'sydneysiege'
]
test = 'charliehebdo'
max_branch_len = 25
x_train = []
y_train = []
for t in train:
temp_x_train = np.load(os.path.join(path, t, 'train_array.npy'))
temp_y_train = np.load(os.path.join(path, t, 'labels.npy'))
temp_x_train = pad_sequences(temp_x_train,
maxlen=max_branch_len,
dtype='float32',
padding='post',
truncating='post',
value=0.)
x_train.extend(temp_x_train)
y_train.extend(temp_y_train)
x_train = np.asarray(x_train)
y_train = np.asarray(y_train)
x_test = np.load(os.path.join(path, test, 'train_array.npy'))
y_test = np.load(os.path.join(path, test, 'labels.npy'))
ids_test = np.load(os.path.join(path, test, 'ids.npy'))
#%
y_train = to_categorical(y_train, num_classes=None)
model = heteroscedastic_model(x_train, y_train, params, output_classes=3)
mb_size = params['mb_size']
num_epochs = params['num_epochs']
model.fit(x_train, [y_train, y_train],
batch_size=mb_size,
epochs=num_epochs,
shuffle=False,
class_weight=None)
verbose = False
predictions_test = model.predict(x_test,
batch_size=mb_size,
verbose=verbose)
softmax_test = predictions_test[1]
y_pred = np.argmax(softmax_test, axis=1)
trees, tree_prediction, tree_label = branch2treelabels(
ids_test, y_test, y_pred)
mactest_F = f1_score(tree_label, tree_prediction, average='macro')
output = {'loss': 1 - mactest_F, 'Params': params, 'status': STATUS_OK}
#%%
return output
def objective_function_branchLSTM_Twitter15(params):
#%%
path = 'preprocessing/saved_data_15'
# fold 0 is development set
train = '0/train'
test = '0/test'
max_branch_len = 25
x_train = []
y_train = []
temp_x_train = np.load(os.path.join(path, train, 'train_array.npy'))
y_train = np.load(os.path.join(path, train, 'labels.npy'))
# pad sequences to the size of the largest
x_train = pad_sequences(temp_x_train,
maxlen=max_branch_len,
dtype='float32',
padding='post',
truncating='post',
value=0.)
x_test = np.load(os.path.join(path, test, 'train_array.npy'))
y_test = np.load(os.path.join(path, test, 'labels.npy'))
ids_test = np.load(os.path.join(path, test, 'ids.npy'))
#%
y_train = to_categorical(y_train, num_classes=None)
# y_pred, confidence = LSTM_model(x_train, y_train, x_test, params)
model = heteroscedastic_model(x_train, y_train, params, output_classes=4)
mb_size = params['mb_size']
num_epochs = params['num_epochs']
model.fit(x_train, [y_train, y_train],
batch_size=mb_size,
epochs=num_epochs,
shuffle=False,
class_weight=None)
verbose = False
predictions_test = model.predict(x_test,
batch_size=mb_size,
verbose=verbose)
softmax_test = predictions_test[1]
y_pred = np.argmax(softmax_test, axis=1)
trees, tree_prediction, tree_label = branch2treelabels(
ids_test, y_test, y_pred)
mactest_F = f1_score(tree_label, tree_prediction, average='macro')
output = {'loss': 1 - mactest_F, 'Params': params, 'status': STATUS_OK}
#%%
return output
def eval_veracity_LSTM_CV(params,dataset='15'):
path = 'preprocessing/saved_data_'+dataset
folds = ['0','1', '2','3', '4']
num_epochs = params['num_epochs']
mb_size = params['mb_size']
for f in folds:
print(f)
test = f+'/test'
train = f+'/train'
x_test = np.load(os.path.join(path,test, 'train_array.npy'))
y_test = np.load(os.path.join(path,test, 'labels.npy'))
ids_test = np.load(os.path.join(path,test, 'ids.npy'))
predictions_train = []
x_train = np.load(os.path.join(path,train, 'train_array.npy'))
y_train = np.load(os.path.join(path,train, 'labels.npy'))
y_train = to_categorical(y_train, num_classes=4)
ids_train = np.load(os.path.join(path,train, 'ids.npy'))
model = heteroscedastic_model(x_train, y_train, params, output_classes=4)
model.fit(x_train,[y_train, y_train], batch_size=mb_size,
epochs=num_epochs, shuffle=False, class_weight=None)
predictions = models.predict_on_data(model,params, x_train, y_train, x_test, y_test, num_classes=3, verbose=True)
tree_results_train = branch2tree(ids_train, predictions['train'])
predictions['train']['tree_results'] = tree_results_train
predictions_train.append(predictions['train'])
filename = 'output/model'+f+'.h5'
model.save(filename)
json_string = model.to_json()
with open('output/my_model_architecture'+f+'.h5','w') as fout:
json.dump(json_string,fout)
model.save_weights('output/my_model_weights'+f+'.h5')
# I need to improve this
tree_results_test = branch2tree(ids_test, predictions['test'])
predictions['test']['tree_results'] = tree_results_test
predictions['train']['tree_results'] = predictions_train
filename = 'output/predictions'+f+'.pkl'
fout = open(filename, "wb")
pickle.dump(predictions, fout)
fout.close()
eval_info_test = eval_branches(tree_results_test)
# eval_info_train = [eval_branches(i) for i in tree_results_train]
eval_info = {}
eval_info['test'] = eval_info_test
# eval_info['train'] = eval_info_train
filename = 'output/eval_info'+f+'.pkl'
fout = open(filename, "wb")
pickle.dump(eval_info, fout)
fout.close()
def eval_veracity_LSTM_CV(params, branch=True):
#%%
if branch:
path = 'saved_data_fullPHEME'
else:
path = 'saved_data_timelinefullPHEME'
folds = [
'ebola-essien', 'ferguson', 'gurlitt', 'ottawashooting',
'prince-toronto', 'putinmissing', 'sydneysiege', 'charliehebdo',
'germanwings-crash'
]
num_epochs = params['num_epochs']
mb_size = params['mb_size']
for number in range(len(folds)):
x_temp = np.load(os.path.join(path, 'ebola-essien', 'train_array.npy'))
y_temp = np.load(os.path.join(path, 'ebola-essien', 'labels.npy'))
model = heteroscedastic_model(x_temp, y_temp, params, output_classes=3)
# del x_temp
print(number)
test = folds[number]
train = deepcopy(folds)
del train[number]
x_test = np.load(os.path.join(path, test, 'train_array.npy'))
y_test = np.load(os.path.join(path, test, 'labels.npy'))
ids_test = np.load(os.path.join(path, test, 'ids.npy'))
predictions_train = []
for t in train:
x_train = np.load(os.path.join(path, t, 'train_array.npy'))
y_train = np.load(os.path.join(path, t, 'labels.npy'))
y_train = to_categorical(y_train, num_classes=3)
ids_train = np.load(os.path.join(path, t, 'ids.npy'))
model.fit(x_train, [y_train, y_train],
batch_size=mb_size,
epochs=num_epochs,
shuffle=False,
class_weight=None)
predictions = models.predict_on_data(model,
params,
x_train,
y_train,
x_test,
y_test,
num_classes=3,
verbose=True)
tree_results_train = branch2tree(ids_train, predictions['train'])
predictions['train']['tree_results'] = tree_results_train
predictions_train.append(predictions['train'])
filename = 'output/model' + str(test) + '.h5'
model.save(filename)
json_string = model.to_json()
with open('output/my_model_architecture' + str(test) + '.h5',
'w') as f:
json.dump(json_string, f)
model.save_weights('output/my_model_weights' + str(test) + '.h5')
# I need to improve this
tree_results_test = branch2tree(ids_test, predictions['test'])
predictions['test']['tree_results'] = tree_results_test
predictions['train']['tree_results'] = predictions_train
filename = 'output/predictions' + str(test) + '.pkl'
f = open(filename, "wb")
pickle.dump(predictions, f)
f.close()
eval_info_test = eval_branches(tree_results_test)
eval_info = {}
eval_info['test'] = eval_info_test
filename = 'output/eval_info' + str(test) + '.pkl'
f = open(filename, "wb")
pickle.dump(eval_info, f)
f.close()