def run_main(args, datasets, model, exp_logger):
if args.task == 'cds':
utt_data = load_candidates(args)
model.add_candidate_data(*utt_data)
kb_labels = {}
if args.use_kb:
kb_labels['intent'] = list(model.mappings['intent'].keys())
kb_labels['action'] = list(model.mappings['action'].keys())
exp_logger.init_tb_writers()
run_train(args, datasets, model, exp_logger, kb_labels)
if args.do_eval:
result = run_eval(args,
datasets,
model,
exp_logger,
kb_labels,
split='test')
results = dict((k + f'_{args.filename}', v) for k, v in result.items())
print('Test Results -', results)
# Gathers the input arguments
args = get_arguments()
# Gathering variables from arguments
dataset = args.dataset
descriptor = args.descriptor
step = 'val'
fold = args.fold
type = args.type
meta = args.mh
# Random seed for experimental consistency
np.random.seed(fold-1)
# Loads the predictions and labels
preds, y = l.load_candidates(dataset, step, fold)
# If descriptor is global-based
if descriptor == 'global':
# Gets the global predictors
preds = preds[:, :35]
# If descriptor is cnn-based
elif descriptor == 'cnn':
# Gets the CNN predictors
preds = preds[:, 35:]
# Checks if the type of used ensemble was weight-based
if type == 'weight':
# Defining function to be optimized
opt_fn = e.weighted_classifiers(preds, y)
# Gathering variables from arguments
dataset = args.dataset
# Creating empty lists to save outputs
val_accs, test_accs = [], []
val_time, test_time = [], []
print(f'\nPerforming majority voting over {dataset} ...\n')
# For each possible fold
for k in range(c.N_FOLDS):
print(f'Fold {k+1}/{c.N_FOLDS}:')
# Loads the validation step predictions and labels
val_pred, val_y = l.load_candidates(dataset, 'val', k + 1)
# Loads the testing step predictions and labels
test_pred, test_y = l.load_candidates(dataset, 'test', k + 1)
# Defining the starting time of validation majority voting
start_v = time.time()
# Gather the majority votes between validation predictions
val_votes = e.majority_voting(val_pred)
# Defining the ending time of validation majority voting
end_v = time.time()
# Defining the starting time of testing majority voting
start_t = time.time()
descriptor = args.descriptor
fold = args.fold
type = args.type
meta = args.mh
# Defining an input file
input_file = f'output/{meta}_{type}_{dataset}_val_{fold}.pkl'
# Creating a History object
h = History()
# Loading the input file
h.load(input_file)
# Loading the predictions and labels
preds, y = l.load_candidates(dataset, 'test', fold)
# If descriptor is global-based
if descriptor == 'global':
# Gets the global predictors
preds = preds[:, :35]
# If descriptor is cnn-based
elif descriptor == 'cnn':
# Gets the CNN predictors
preds = preds[:, 35:]
# Gathering the best weights
best_weights = np.asarray(h.best_agent[-1][0])
# Checks if the type of used ensemble was weight-based