def run_hyperparam_search(questions_to_run, directory, q_docpartitions, bertmodel, sents_embed_dir, question_gpu_map,
attention):
hyperparam_search_dir = create_directory('hyperparam_search', directory)
hyperparam_model_search_parallel(questions_to_run, q_docpartitions, bertmodel, sents_embed_dir,
hyperparam_search_dir,
question_gpu_map,
fdtype=torch.float32,
num_epochs=15,
prob_interval_truemax=0.05,
prob_estim=0.95, random_seed=42, attention=attention)
def write_sents_embeddings(directory, bertmodel, sents_embed_dir_name, docs_data_tensor):
# === Generate sents embedding ===
# load BertModel
# define BertEmbedder
bert_config = {'bert_train_flag': False,
'bert_all_output': False}
bertembeder = BertEmbedder(bertmodel, bert_config)
sents_embed_dir = create_directory(sents_embed_dir_name, directory)
fdtype = torch.float32
# generate and dump bert embedding for the tokens inside the specificed embedding directory
bert_proc_docs = generate_sents_embeds_from_docs(docs_data_tensor, bertembeder, sents_embed_dir, fdtype)
ReaderWriter.dump_data(bert_proc_docs, os.path.join(sents_embed_dir, 'bert_proc_docs.pkl'))
def run_predict(q_docpartitions,
q_fold_config_map,
bertmodel,
q_state_dict_path_map,
results_dir,
sents_embed_dir,
question_fold_map,
to_gpu,
gpu_index,
num_epochs=1) -> Dict:
q_predictions = {}
for question in q_fold_config_map:
mconfig, options, __ = q_fold_config_map[question]
options[
'num_epochs'] = num_epochs # override number of epochs using user specified value
# update options fold num to the current fold
options['fold_num'] = question_fold_map[question]
data_partition = q_docpartitions[question][options['fold_num']]
results_path = os.path.join(results_dir,
'question_{}'.format(question),
'fold_{}'.format(options['fold_num']))
results_wrk_dir = create_directory(results_path)
q_predictions[question] = predict_neural_discern(
data_partition,
bertmodel,
mconfig,
options,
results_wrk_dir,
sents_embed_dir,
state_dict_dir=q_state_dict_path_map[question],
to_gpu=to_gpu,
gpu_index=gpu_index)
return q_predictions
def evaluate_on_test_set(directory, q_docpartitions, q_config_map, bertmodel, train_val_dir, sents_embed_dir,
gpu_index):
test_dir = create_directory('test', directory)
test_run(q_docpartitions, q_config_map, bertmodel, train_val_dir, test_dir, sents_embed_dir, gpu_index,
num_epochs=1)
return test_dir
def run_training(directory, q_docpartitions, q_config_map, bertmodel, sents_embed_dir, question_gpu_map, num_epochs,
max_folds):
train_val_dir = create_directory('train_validation', directory)
train_val_run(q_docpartitions, q_config_map, bertmodel, train_val_dir, sents_embed_dir, question_gpu_map,
num_epochs, max_folds)
return train_val_dir
"The pre-built experiment_to_rerun will be used. Hyperparam search will not be run.")
# under test mode (for faster debugging), run a smaller set of partitions and epochs, and no hyper param search
if config['test_mode']:
config['max_folds'] = 1 # max number of data partition folds to run (for faster testing)
config['num_epochs'] = 1
config['run_hyper_param_search'] = False
time_stamp = datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S')
if config['experiment_to_rerun']:
if config['copy_exp_dir']:
from distutils.dir_util import copy_tree
rerun_dir_name = '{}_{}_{}'.format(config['experiment_to_rerun'], 'rerun', time_stamp)
orig_exp_dir = os.path.join(config['base_dir'], 'experiments', config['experiment_to_rerun'])
exp_dir = os.path.join(config['base_dir'], 'experiments', rerun_dir_name)
create_directory(exp_dir)
# copy contents of original exp dir so experiment re-run has everything it needs
print("copying experiment for re-run in {}...".format(exp_dir))
copy_tree(orig_exp_dir, exp_dir)
print("... complete")
else:
exp_dir = exp_dir = os.path.join(config['base_dir'], 'experiments', config['experiment_to_rerun'])
else:
if config['test_mode']:
exp_dir = os.path.join(config['base_dir'], 'experiments', 'tests', time_stamp)
else:
exp_dir = os.path.join(config['base_dir'], 'experiments', time_stamp)
create_directory(exp_dir)
config['exp_dir'] = exp_dir
if config['biobert']: