def __print_epoch_results(self, ep_no, all_ep, train_tasks, valid_tasks):
result = 'epoch: {}/{}> '.format(ep_no, all_ep)
for cur_task in train_tasks.items():
if cur_task[1].size > 0:
result += '|T: {}, tr-loss: {:.3f}, tr-f1: {:.3f} '.format(
cur_task[0], cur_task[1].loss, cur_task[1].f1)
if valid_tasks is not None:
for cur_task in valid_tasks.items():
if cur_task[1].size > 0:
result += '|T: {}, va-loss: {:.3f}, va-f1: {:.3f} '.format(
cur_task[0], cur_task[1].loss, cur_task[1].f1)
result += '\t' + ELib.get_time()
print(result)
def test(self, test_bundle, return_output_vecs=False, weighted_instance_loss=False,
print_perf=True, title=None, report_number_of_intervals=20, return_output_vecs_get_details=True):
if len(test_bundle.task_list) > 1:
print('only one task is allowed for testing')
return None
if len(test_bundle.tws) == 0:
return list(), list(), list(), list()
if title is None:
title = ''
else:
title += ' '
self.bert_classifier.to(self.config.device)
self.bert_classifier.zero_grad()
self.bert_classifier.eval()
self.setup_objective(weighted_instance_loss)
test_dt = EBertDataset(test_bundle, self.tokenizer, self.config.max_seq)
batches = self.generate_batches([test_dt], self.config, False, False, 0, EInputListMode.sequential)
result_vecs = list()
result_vecs_detail = list()
tasks = {test_bundle.task_list[0] : ETaskState(test_bundle.task_list[0])}
print(title + 'labeling ', end=' ', flush=True)
with torch.no_grad():
for ba_ind, cur_batch in enumerate(batches):
outcome = self.bert_classifier(cur_batch, False)
self.__process_loss(outcome, cur_batch, tasks, False, weighted_instance_loss)
if return_output_vecs:
result_vecs.extend(self.bert_classifier.output_vecs)
if self.bert_classifier.output_vecs_detail is not None and return_output_vecs_get_details:
result_vecs_detail.extend(self.bert_classifier.output_vecs_detail)
if ELib.progress_made(ba_ind, cur_batch['batch_count'], report_number_of_intervals):
print(ELib.progress_percent(ba_ind, cur_batch['batch_count']), end=' ', flush=True)
self.delete_batch_from_gpu(cur_batch, EInputListMode.sequential)
del cur_batch, outcome
print()
task_out = tasks[test_bundle.task_list[0]]
task_out.loss /= task_out.size
perf = ELib.calculate_metrics(task_out.lbl_true, task_out.lbl_pred)
if print_perf:
print('Test Results L1> Loss: {:.3f} F1: {:.3f} Pre: {:.3f} Rec: {:.3f}'.format(
task_out.loss, perf[0], perf[1], perf[2]) + '\t\t' + ELib.get_time())
self.bert_classifier.cpu()
return task_out.lbl_pred, task_out.logits, [result_vecs, result_vecs_detail], perf
def main():
parser = argparse.ArgumentParser()
# general params
parser.add_argument("--cmd",
default=None,
type=str,
required=True,
help='')
parser.add_argument("--TEMP_DIR",
default=temp_dir,
type=str,
required=False,
help='')
# pretrain params
parser.add_argument("--itr", default=1, type=int, required=False, help='')
parser.add_argument("--model_path",
default=None,
type=str,
required=True,
help='')
parser.add_argument("--train_path",
default=None,
type=str,
required=True,
help='')
parser.add_argument("--test_path",
default=None,
type=str,
required=True,
help='')
parser.add_argument("--unlabeled_path",
default=None,
type=str,
required=True,
help='')
parser.add_argument("--output_dir",
default=None,
type=str,
required=True,
help='')
parser.add_argument("--device",
default=None,
type=int,
required=True,
help='')
parser.add_argument("--seed",
default=None,
type=int,
required=True,
help='')
parser.add_argument("--train_sample",
default=None,
type=int,
required=True,
help='')
parser.add_argument("--unlabeled_sample",
default=None,
type=int,
required=True,
help='')
## ignore these!
parser.add_argument("--per_query",
default=False,
type=bool,
required=False,
help='') # ignore it
parser.add_argument("--model_path_2",
default=None,
type=str,
required=False,
help='') # ignore it
parser.add_argument("--lm_model_path",
default=None,
type=str,
required=False,
help='') # ignore it
parser.add_argument("--t_lbl_path_1",
default=None,
type=str,
required=False,
help='') # ignore it
parser.add_argument("--t_lbl_path_2",
default=None,
type=str,
required=False,
help='') # ignore it
parser.add_argument("--valid_path",
default=None,
type=str,
required=False,
help='') # ignore it
parser.add_argument("--device_2",
default=None,
type=int,
required=False,
help='') # ignore it
args, unknown = parser.parse_known_args()
device = 'cpu'
device_name = device
if args.device >= 0:
device = 'cuda:' + str(args.device)
device_name = torch.cuda.get_device_name(args.device)
device_2 = 'cpu'
if 'device_2' in args and (args.device_2 is not None
and args.device_2 >= 0):
device_2 = 'cuda:' + str(args.device_2)
device_name = device_name + ', ' + torch.cuda.get_device_name(
args.device)
print('setup:', '| python>', platform.python_version(), '| numpy>',
np.__version__, '| pytorch>', torch.__version__, '| device>',
device_name)
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
if args.cmd.startswith('bert'):
seed = args.seed
for cur_itr in range(args.itr):
print('------------------------------------')
print('iteration ' + str(cur_itr + 1) + ' began with seed=\'' +
str(seed) + '\' at ' + ELib.get_time())
if cur_itr >= 0:
output_dir = args.output_dir + '_' + str(cur_itr)
EPretrainProj.run(args.cmd, args.per_query, args.train_path,
args.valid_path, args.test_path,
args.unlabeled_path, args.model_path,
args.model_path_2, args.lm_model_path,
args.t_lbl_path_1, args.t_lbl_path_2,
output_dir, device, device_2, seed,
args.train_sample, args.unlabeled_sample)
seed += 1230
ELib.PASS()
ELib.PASS()
device_name)
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
if args.cmd.startswith('bert'):
seed = args.seed
for cur_itr in range(args.itr):
print('------------------------------------')
print('iteration ' + str(cur_itr + 1) + ' began with seed=\'' +
str(seed) + '\' at ' + ELib.get_time())
if cur_itr >= 0:
output_dir = args.output_dir + '_' + str(cur_itr)
EPretrainProj.run(args.cmd, args.per_query, args.train_path,
args.valid_path, args.test_path,
args.unlabeled_path, args.model_path,
args.model_path_2, args.lm_model_path,
args.t_lbl_path_1, args.t_lbl_path_2,
output_dir, device, device_2, seed,
args.train_sample, args.unlabeled_sample)
seed += 1230
ELib.PASS()
ELib.PASS()
if __name__ == "__main__":
print("Started at", ELib.get_time())
main()
print("\nDone at", ELib.get_time())
pass
