def get_params(start_model_path, start_type, info_fn_name, num_gpu):
hp = hyperparams.HPSENLI3()
nli_setting = BertNLI()
set_level_debug()
train_config = NLIExTrainConfig()
train_config.num_gpu = num_gpu
train_config.save_train_payload = True
tokenizer = get_tokenizer()
tf_logging.info("Intializing dataloader")
data_loader = get_modified_data_loader(tokenizer, hp.seq_max,
nli_setting.vocab_filename)
tf_logging.info("loading batches")
data = get_nli_data(hp, nli_setting)
def init_fn(sess):
start_type_generic = {
'nli': 'cls',
'nli_ex': 'cls_ex',
'bert': 'bert',
'cold': 'cold'
}[start_type]
return init_fn_generic(sess, start_type_generic, start_model_path)
informative_fn = get_informative_fn_by_name(info_fn_name)
return data, data_loader, hp, informative_fn, init_fn, train_config
def run(explain_tag, method_name, model_path):
hp = hyperparams.HPSENLI3()
nli_setting = BertNLI()
target_class = enlidef.get_target_class(explain_tag)
data = get_nli_data(hp, nli_setting)
train_batches, dev_batches = data
flat_dev_batches = flatten_filter_batches(dev_batches, target_class)[:2000]
if method_name in ['deletion_seq', "random", 'idf', 'deletion', 'LIME']:
predictor = BaselineExPredictor(hp, nli_setting, model_path,
method_name)
acc_list = eval_fidelity_with_ex_predictor(predictor, flat_dev_batches,
explain_tag)
elif method_name.startswith("nli_ex"):
modeling_option = "co"
predictor = NLIExPredictor(hp, nli_setting, model_path,
modeling_option)
acc_list = eval_fidelity_with_ex_predictor(predictor, flat_dev_batches,
explain_tag)
elif method_name in [
"elrp", "deeplift", "saliency", "grad*input", "intgrad"
]:
acc_list = eval_fidelity_gradient(hp, nli_setting, flat_dev_batches,
explain_tag, method_name, model_path)
else:
raise Exception(
"method_name={} is not in the known method list.".format(
method_name))
print(method_name)
for num_delete in sorted(acc_list.keys()):
print("{}\t{}".format(num_delete, acc_list[num_delete]))
def train_nil_from(model_path, save_dir, resume=False):
print("Load model path : ", model_path)
print("Save dir : ", save_dir)
def load_fn(sess, model_path):
if not resume:
return load_model_w_scope(sess, model_path, "bert")
else:
return load_model(sess, model_path)
steps = 67000
hp = hyperparams.HPSENLI3()
nli_setting = BertNLI()
data = get_nli_data(hp, nli_setting)
set_level_debug()
hp = hyperparams.HPSENLI3()
n_gpu = 2
return train_nli_multi_gpu(hp, nli_setting, save_dir, steps, data,
model_path, load_fn, n_gpu)
def train_nil_from_bert(model_path, save_dir):
def load_fn(sess, model_path):
return load_model_w_scope(sess, model_path, "bert")
max_steps = 61358
max_steps = 36250
hp = hyperparams.HPSENLI3()
set_level_debug()
nli_setting = BertNLI()
data = get_snli_data(hp, nli_setting)
n_gpu = 2
return train_nli_multi_gpu(hp, nli_setting, save_dir, max_steps, data,
model_path, load_fn, n_gpu)
def run(args):
hp = hyperparams.HPSENLI3()
nli_setting = BertNLI()
data_loader = get_modified_data_loader2(hp, nli_setting)
predict_nli_ex(
hp,
nli_setting,
data_loader,
args.tag,
args.data_id,
args.model_path,
args.run_name,
args.modeling_option,
)
def train_nil_from_v2_checkpoint(run_name, model_path):
hp = hyperparams.HPSENLI3()
print(hp.batch_size)
nli_setting = NLI()
nli_setting.vocab_size = 30522
nli_setting.vocab_filename = "bert_voca.txt"
data_loader = nli.DataLoader(hp.seq_max, nli_setting.vocab_filename, True)
tokenizer = get_tokenizer()
CLS_ID = tokenizer.convert_tokens_to_ids(["[CLS]"])[0]
SEP_ID = tokenizer.convert_tokens_to_ids(["[SEP]"])[0]
data_loader.CLS_ID = CLS_ID
data_loader.SEP_ID = SEP_ID
tf_logger.setLevel(logging.INFO)
steps = 12271
data = load_cache("nli_batch16")
if data is None:
tf_logger.info("Encoding data from csv")
data = get_nli_batches_from_data_loader(data_loader, hp.batch_size)
save_to_pickle(data, "nli_batch16")
train_nli(hp, nli_setting, run_name, steps, data, model_path)
def run(args):
hp = hyperparams.HPSENLI3()
nli_setting = BertNLI()
data_loader = get_modified_data_loader2(hp, nli_setting)
if args.method_name in [
'deletion_seq', "random", 'idf', 'deletion', 'LIME'
]:
predictor = nli_baseline_predict
elif args.method_name in [
"elrp",
"deeplift",
"saliency",
"grad*input",
"intgrad",
]:
predictor = nli_attribution_predict
else:
raise Exception(
"method_name={} is not in the known method list.".format(
args.method_name))
predictor(hp, nli_setting, data_loader, args.tag, args.method_name,
args.data_id, args.sub_range, args.model_path)
def train_nil_from(save_dir, model_path, load_fn, max_steps):
hp = hyperparams.HPSENLI3()
set_level_debug()
nli_setting = BertNLI()
data = get_nli_data(hp, nli_setting)
train_nli(hp, nli_setting, save_dir, max_steps, data, model_path, load_fn)
