def make_prediction(Description_A, Description_B, model_path, USE_GPU=True):
# Loading tokenized using a stored Pickle Object, as it is more reliable
# In case you'd like to create the object, you can do so here: bert_tokenizer = BertTokenizer.from_pretrained("bert-large-uncased", do_lower_case=True)
pickle_off = open("tokenizer.pkl", "rb")
tokenizer = pickle.load(pickle_off)
# Tokenizes the words into format required by Bert ex: "I am playing" -> ["I","am","play","##ing"]
hypothesis = tokenizer.tokenize(Description_A)
#If sequence is too long it truncates it to ensure it fits into BERT's max seq len and changes the words into numbers
if len(hypothesis) > 512 - 3:
hypothesis = hypothesis[:512 - 3]
input_ids = tokenizer.convert_tokens_to_ids(
['[CLS]'] + hypothesis + ['[SEP]'])
#Determnines what sentence it's in, doesn't really matter for single sentence, but important for 2 sentence classification
type_ids = [0] * (len(hypothesis) + 2)
#Concatenates all the important labels into a dictionary
#UID : id number (no importance) ; label: "ground truth" (no importance when making a prediction)
# token_id: representation of words ; type_id: position within a sentence
features = {'uid': 0, 'label': 0,
'token_id': input_ids, 'type_id': type_ids}
# Loads data into a BatchGen object which is needed for making a prediction, nothing needed to change here
dev_data = BatchGen([features],
batch_size=8,
gpu=True, is_train=False,
task_id=0,
maxlen=512,
pairwise=False,
data_type=0,
task_type=0)
# function to convert token ids back to words
print(tokenizer.convert_ids_to_tokens([101, 100, 5208, 2024, 17662, 9119, 2096, 3173, 2000, 2175, 14555, 2044, 2074, 5983, 6265, 1012, 102, 100, 2308, 2024, 23581, 2096, 3173, 2000, 2175, 14555, 1012, 102]))
#hyper parameters: whatever is necessary is added as variables at the top
opt = {'init_checkpoint': model_path, 'data_dir': 'data/domain_adaptation', 'data_sort_on': False, 'name': 'farmer', 'train_datasets': ['sst'], 'test_datasets': ['sst'], 'pw_tasks': ['qnnli'], 'update_bert_opt': 0, 'multi_gpu_on': False, 'mem_cum_type': 'simple', 'answer_num_turn': 5, 'answer_mem_drop_p': 0.1, 'answer_att_hidden_size': 128, 'answer_att_type': 'bilinear', 'answer_rnn_type': 'gru', 'answer_sum_att_type': 'bilinear', 'answer_merge_opt': 1, 'answer_mem_type': 1, 'answer_dropout_p': 0.1, 'answer_weight_norm_on': False, 'dump_state_on': False, 'answer_opt': [
0], 'label_size': '2', 'mtl_opt': 0, 'ratio': 0, 'mix_opt': 0, 'max_seq_len': 512, 'init_ratio': 1, 'cuda': USE_GPU, 'log_per_updates': 500, 'epochs': 5, 'batch_size': 32, 'batch_size_eval': 8, 'optimizer': 'adamax', 'grad_clipping': 0.0, 'global_grad_clipping': 1.0, 'weight_decay': 0, 'learning_rate': 5e-05, 'momentum': 0, 'warmup': 0.1, 'warmup_schedule': 'warmup_linear', 'vb_dropout': True, 'dropout_p': 0.1, 'dropout_w': 0.0, 'bert_dropout_p': 0.1, 'ema_opt': 0, 'ema_gamma': 0.995, 'have_lr_scheduler': True, 'multi_step_lr': '10,20,30', 'freeze_layers': -1, 'embedding_opt': 0, 'lr_gamma': 0.5, 'bert_l2norm': 0.0, 'scheduler_type': 'ms', 'output_dir': 'checkpoints/scitail_tl_adamax_answer_opt0_gc0_ggc1_7_2_19', 'seed': 2018, 'task_config_path': 'configs/tasks_config.json', 'tasks_dropout_p': [0.1]}
state_dict = torch.load(model_path)
config = state_dict['config']
config['attention_probs_dropout_prob'] = 0.1
config['hidden_dropout_prob'] = 0.1
opt.update(config)
model = MTDNNModel(opt, state_dict=state_dict, num_train_step=50)
#actual prediction to be made: main outputs are predictions which is a list of size 1, and scores which is confidence in prediction for each class
dev_metrics, dev_predictions, scores, golds, dev_ids = eval_model(
model, dev_data, 0,use_cuda=True, with_label =False)
#model, data, metric_meta, use_cuda=True, with_label=True
return dev_predictions, scores
def main():
logger.info('Launching the MT-DNN training')
opt = vars(args)
# update data dir
opt['data_dir'] = data_dir
batch_size = args.batch_size
train_data_list = []
tasks = {}
tasks_class = {}
nclass_list = []
decoder_opts = []
dropout_list = []
for dataset in args.train_datasets:
prefix = dataset.split('_')[0]
if prefix in tasks: continue
assert prefix in task_defs.n_class_map
assert prefix in task_defs.data_type_map
data_type = task_defs.data_type_map[prefix]
nclass = task_defs.n_class_map[prefix]
task_id = len(tasks)
if args.mtl_opt > 0:
task_id = tasks_class[nclass] if nclass in tasks_class else len(
tasks_class)
task_type = task_defs.task_type_map[prefix]
pw_task = False
if task_type == TaskType.Ranking:
pw_task = True
dopt = generate_decoder_opt(task_defs.enable_san_map[prefix],
opt['answer_opt'])
if task_id < len(decoder_opts):
decoder_opts[task_id] = min(decoder_opts[task_id], dopt)
else:
decoder_opts.append(dopt)
if prefix not in tasks:
tasks[prefix] = len(tasks)
if args.mtl_opt < 1: nclass_list.append(nclass)
if (nclass not in tasks_class):
tasks_class[nclass] = len(tasks_class)
if args.mtl_opt > 0: nclass_list.append(nclass)
dropout_p = task_defs.dropout_p_map.get(prefix, args.dropout_p)
dropout_list.append(dropout_p)
train_path = os.path.join(data_dir, '{}_train.json'.format(dataset))
logger.info('Loading {} as task {}'.format(train_path, task_id))
train_data = BatchGen(BatchGen.load(train_path,
True,
pairwise=pw_task,
maxlen=args.max_seq_len),
batch_size=batch_size,
dropout_w=args.dropout_w,
gpu=args.cuda,
task_id=task_id,
maxlen=args.max_seq_len,
pairwise=pw_task,
data_type=data_type,
task_type=task_type,
encoder_type=encoder_type)
train_data_list.append(train_data)
opt['answer_opt'] = decoder_opts
opt['tasks_dropout_p'] = dropout_list
args.label_size = ','.join([str(l) for l in nclass_list])
logger.info(args.label_size)
dev_data_list = []
test_data_list = []
for dataset in args.test_datasets:
prefix = dataset.split('_')[0]
task_id = tasks_class[
task_defs.
n_class_map[prefix]] if args.mtl_opt > 0 else tasks[prefix]
task_type = task_defs.task_type_map[prefix]
pw_task = False
if task_type == TaskType.Ranking:
pw_task = True
assert prefix in task_defs.data_type_map
data_type = task_defs.data_type_map[prefix]
dev_path = os.path.join(data_dir, '{}_dev.json'.format(dataset))
dev_data = None
if os.path.exists(dev_path):
dev_data = BatchGen(BatchGen.load(dev_path,
False,
pairwise=pw_task,
maxlen=args.max_seq_len),
batch_size=args.batch_size_eval,
gpu=args.cuda,
is_train=False,
task_id=task_id,
maxlen=args.max_seq_len,
pairwise=pw_task,
data_type=data_type,
task_type=task_type,
encoder_type=encoder_type)
dev_data_list.append(dev_data)
test_path = os.path.join(data_dir, '{}_test.json'.format(dataset))
test_data = None
if os.path.exists(test_path):
test_data = BatchGen(BatchGen.load(test_path,
False,
pairwise=pw_task,
maxlen=args.max_seq_len),
batch_size=args.batch_size_eval,
gpu=args.cuda,
is_train=False,
task_id=task_id,
maxlen=args.max_seq_len,
pairwise=pw_task,
data_type=data_type,
task_type=task_type,
encoder_type=encoder_type)
test_data_list.append(test_data)
logger.info('#' * 20)
logger.info(opt)
logger.info('#' * 20)
all_iters = [iter(item) for item in train_data_list]
all_lens = [len(bg) for bg in train_data_list]
# div number of grad accumulation.
num_all_batches = args.epochs * sum(
all_lens) // args.grad_accumulation_step
logger.info('############# Gradient Accumulation Info #############')
logger.info('number of step: {}'.format(args.epochs * sum(all_lens)))
logger.info('number of grad grad_accumulation step: {}'.format(
args.grad_accumulation_step))
logger.info('adjusted number of step: {}'.format(num_all_batches))
logger.info('############# Gradient Accumulation Info #############')
if len(train_data_list) > 1 and args.ratio > 0:
num_all_batches = int(args.epochs * (len(train_data_list[0]) *
(1 + args.ratio)))
bert_model_path = args.init_checkpoint
state_dict = None
if encoder_type == EncoderModelType.BERT:
if os.path.exists(bert_model_path):
state_dict = torch.load(bert_model_path)
config = state_dict['config']
config['attention_probs_dropout_prob'] = args.bert_dropout_p
config['hidden_dropout_prob'] = args.bert_dropout_p
opt.update(config)
else:
logger.error('#' * 20)
logger.error(
'Could not find the init model!\n The parameters will be initialized randomly!'
)
logger.error('#' * 20)
config = BertConfig(vocab_size_or_config_json_file=30522).to_dict()
opt.update(config)
elif encoder_type == EncoderModelType.ROBERTA:
bert_model_path = '{}/model.pt'.format(bert_model_path)
if os.path.exists(bert_model_path):
new_state_dict = {}
state_dict = torch.load(bert_model_path)
for key, val in state_dict['model'].items():
if key.startswith('decoder.sentence_encoder'):
key = 'bert.model.{}'.format(key)
new_state_dict[key] = val
elif key.startswith('classification_heads'):
key = 'bert.model.{}'.format(key)
new_state_dict[key] = val
state_dict = {'state': new_state_dict}
model = MTDNNModel(opt,
state_dict=state_dict,
num_train_step=num_all_batches)
if args.resume and args.model_ckpt:
logger.info('loading model from {}'.format(args.model_ckpt))
model.load(args.model_ckpt)
#### model meta str
headline = '############# Model Arch of MT-DNN #############'
### print network
logger.info('\n{}\n{}\n'.format(headline, model.network))
# dump config
config_file = os.path.join(output_dir, 'config.json')
with open(config_file, 'w', encoding='utf-8') as writer:
writer.write('{}\n'.format(json.dumps(opt)))
writer.write('\n{}\n{}\n'.format(headline, model.network))
logger.info("Total number of params: {}".format(model.total_param))
for epoch in range(0, args.epochs):
logger.warning('At epoch {}'.format(epoch))
for train_data in train_data_list:
train_data.reset()
start = datetime.now()
all_indices = []
if len(train_data_list) > 1 and args.ratio > 0:
main_indices = [0] * len(train_data_list[0])
extra_indices = []
for i in range(1, len(train_data_list)):
extra_indices += [i] * len(train_data_list[i])
random_picks = int(
min(len(train_data_list[0]) * args.ratio, len(extra_indices)))
extra_indices = np.random.choice(extra_indices,
random_picks,
replace=False)
if args.mix_opt > 0:
extra_indices = extra_indices.tolist()
random.shuffle(extra_indices)
all_indices = extra_indices + main_indices
else:
all_indices = main_indices + extra_indices.tolist()
else:
for i in range(1, len(train_data_list)):
all_indices += [i] * len(train_data_list[i])
if args.mix_opt > 0:
random.shuffle(all_indices)
all_indices += [0] * len(train_data_list[0])
if args.mix_opt < 1:
random.shuffle(all_indices)
for i in range(len(all_indices)):
task_id = all_indices[i]
batch_meta, batch_data = next(all_iters[task_id])
model.update(batch_meta, batch_data)
if (model.local_updates) % (args.log_per_updates *
args.grad_accumulation_step
) == 0 or model.local_updates == 1:
ramaining_time = str((datetime.now() - start) / (i + 1) *
(len(all_indices) - i - 1)).split('.')[0]
logger.info(
'Task [{0:2}] updates[{1:6}] train loss[{2:.5f}] remaining[{3}]'
.format(task_id, model.updates, model.train_loss.avg,
ramaining_time))
if args.save_per_updates_on and (
(model.local_updates) %
(args.save_per_updates * args.grad_accumulation_step) == 0):
model_file = os.path.join(
output_dir, 'model_{}_{}.pt'.format(epoch, model.updates))
logger.info('Saving mt-dnn model to {}'.format(model_file))
model.save(model_file)
for idx, dataset in enumerate(args.test_datasets):
prefix = dataset.split('_')[0]
label_dict = task_defs.global_map.get(prefix, None)
dev_data = dev_data_list[idx]
if dev_data is not None:
dev_metrics, dev_predictions, scores, golds, dev_ids = eval_model(
model,
dev_data,
metric_meta=task_defs.metric_meta_map[prefix],
use_cuda=args.cuda)
for key, val in dev_metrics.items():
logger.warning(
'Task {0} -- epoch {1} -- Dev {2}: {3:.3f}'.format(
dataset, epoch, key, val))
score_file = os.path.join(
output_dir, '{}_dev_scores_{}.json'.format(dataset, epoch))
results = {
'metrics': dev_metrics,
'predictions': dev_predictions,
'uids': dev_ids,
'scores': scores
}
dump(score_file, results)
official_score_file = os.path.join(
output_dir, '{}_dev_scores_{}.tsv'.format(dataset, epoch))
submit(official_score_file, results, label_dict)
# test eval
test_data = test_data_list[idx]
if test_data is not None:
test_metrics, test_predictions, scores, golds, test_ids = eval_model(
model,
test_data,
metric_meta=task_defs.metric_meta_map[prefix],
use_cuda=args.cuda,
with_label=False)
score_file = os.path.join(
output_dir,
'{}_test_scores_{}.json'.format(dataset, epoch))
results = {
'metrics': test_metrics,
'predictions': test_predictions,
'uids': test_ids,
'scores': scores
}
dump(score_file, results)
official_score_file = os.path.join(
output_dir, '{}_test_scores_{}.tsv'.format(dataset, epoch))
submit(official_score_file, results, label_dict)
logger.info('[new test scores saved.]')
model_file = os.path.join(output_dir, 'model_{}.pt'.format(epoch))
model.save(model_file)
def get_confidence(words, chars):
dev_data = BatchGen(words, is_train=False, gpu=True,
batch_size=1, maxlen=512)
dev_metrics, dev_predictions, scores, golds, dev_ids = eval_model(
model, dev_data, 0, use_cuda=True, with_label = False)
return max(scores)
# load data
test_data = BatchGen(BatchGen.load(args.prep_input, False, pairwise=pw_task, maxlen=args.max_seq_len),
batch_size=args.batch_size_eval,
gpu=args.cuda, is_train=False,
task_id=args.task_id,
maxlen=args.max_seq_len,
pairwise=pw_task,
data_type=data_type,
task_type=task_type)
# load model
checkpoint_path = args.checkpoint
assert os.path.exists(checkpoint_path)
if args.cuda:
state_dict = torch.load(checkpoint_path)
else:
state_dict = torch.load(checkpoint_path, map_location="cpu")
config = state_dict['config']
config["cuda"] = args.cuda
model = MTDNNModel(config, state_dict=state_dict)
test_metrics, test_predictions, scores, golds, test_ids = eval_model(model, test_data,
metric_meta=metric_meta,
use_cuda=args.cuda, with_label=args.with_label)
results = {'metrics': test_metrics, 'predictions': test_predictions, 'uids': test_ids, 'scores': scores}
dump(args.score, results)
if args.with_label:
print(test_metrics)