def main():
parser = argparse.ArgumentParser()
model_config(parser)
set_config(parser)
train_config(parser)
args = parser.parse_args()
layer_indexes = [int(x) for x in args.layers.split(",")]
set_environment(args.seed)
# process data
data, is_single_sentence = process_data(args)
data_type = DataFormat.PremiseOnly if is_single_sentence else DataFormat.PremiseAndOneHypothesis
collater = Collater(gpu=args.cuda, is_train=False, data_type=data_type)
batcher = DataLoader(data, batch_size=args.batch_size, collate_fn=collater.collate_fn, pin_memory=args.cuda)
opt = vars(args)
# load model
if os.path.exists(args.checkpoint):
state_dict = torch.load(args.checkpoint)
config = state_dict['config']
config['dump_feature'] = True
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)
return
num_all_batches = len(batcher)
model = MTDNNModel(
opt,
state_dict=state_dict,
num_train_step=num_all_batches)
if args.cuda:
model.cuda()
features_dict = {}
for batch_meta, batch_data in batcher:
batch_meta, batch_data = Collater.patch_data(args.cuda, batch_meta, batch_data)
all_encoder_layers, _ = model.extract(batch_meta, batch_data)
embeddings = [all_encoder_layers[idx].detach().cpu().numpy()
for idx in layer_indexes]
uids = batch_meta['uids']
masks = batch_data[batch_meta['mask']].detach().cpu().numpy().tolist()
for idx, uid in enumerate(uids):
slen = sum(masks[idx])
features = {}
for yidx, layer in enumerate(layer_indexes):
features[layer] = str(embeddings[yidx][idx][:slen].tolist())
features_dict[uid] = features
# save features
with open(args.foutput, 'w', encoding='utf-8') as writer:
for sample in data:
uid = sample['uid']
tokens = sample['tokens']
feature = features_dict[uid]
feature['tokens'] = tokens
feature['uid'] = uid
writer.write('{}\n'.format(json.dumps(feature)))
def eval_model(
model,
data,
metric_meta,
device,
with_label=True,
label_mapper=None,
task_type=TaskType.Classification,
):
predictions = []
golds = []
scores = []
ids = []
metrics = {}
for (batch_info, batch_data) in tqdm(data, total=len(data)):
batch_info, batch_data = Collater.patch_data(device, batch_info,
batch_data)
score, pred, gold = model.predict(batch_info, batch_data)
scores = merge(score, scores)
golds = merge(gold, golds)
predictions = merge(pred, predictions)
ids = merge(batch_info["uids"], ids)
if task_type == TaskType.Span:
predictions, golds = postprocess_qa_predictions(
golds, scores, version_2_with_negative=False)
elif task_type == TaskType.SpanYN:
predictions, golds = postprocess_qa_predictions(
golds, scores, version_2_with_negative=True)
if with_label:
metrics = calc_metrics(metric_meta, golds, predictions, scores,
label_mapper)
return metrics, predictions, scores, golds, ids
def eval_model(model,
data,
metric_meta,
use_cuda=True,
with_label=True,
label_mapper=None):
if use_cuda:
model.cuda()
predictions = []
golds = []
scores = []
ids = []
metrics = {}
for batch_info, batch_data in data:
batch_info, batch_data = Collater.patch_data(use_cuda, batch_info,
batch_data)
score, pred, gold = model.predict(batch_info, batch_data)
predictions.extend(pred)
golds.extend(gold)
scores.extend(score)
ids.extend(batch_info['uids'])
if with_label:
metrics = calc_metrics(metric_meta, golds, predictions, scores,
label_mapper)
return metrics, predictions, scores, golds, ids
def eval_model(model,
data,
metric_meta,
device,
with_label=True,
label_mapper=None,
task_type=TaskType.Classification):
predictions = []
golds = []
scores = []
ids = []
metrics = {}
for (batch_info, batch_data) in data:
batch_info, batch_data = Collater.patch_data(device, batch_info,
batch_data)
score, pred, gold = model.predict(batch_info, batch_data)
predictions.extend(pred)
golds.extend(gold)
scores.extend(score)
ids.extend(batch_info['uids'])
if task_type == TaskType.Span:
from experiments.squad import squad_utils
golds = squad_utils.merge_answers(ids, golds)
predictions, scores = squad_utils.select_answers(
ids, predictions, scores)
if with_label:
metrics = calc_metrics(metric_meta, golds, predictions, scores,
label_mapper)
return metrics, predictions, scores, golds, ids
def eval_model(model,
data,
metric_meta,
use_cuda=True,
with_label=True,
label_mapper=None,
task_type=TaskType.Classification):
if use_cuda:
model.cuda()
predictions = []
golds = []
scores = []
ids = []
metrics = {}
for idx, (batch_info, batch_data) in enumerate(data):
# if idx % 100 == 0:
# print("predicting {}".format(idx))
batch_info, batch_data = Collater.patch_data(use_cuda, batch_info,
batch_data)
score, pred, gold = model.predict(batch_info, batch_data)
predictions.extend(pred)
golds.extend(gold)
scores.extend(score)
ids.extend(batch_info['uids'])
if task_type == TaskType.Span:
from experiments.squad import squad_utils
golds = squad_utils.merge_answers(ids, golds)
predictions, scores = squad_utils.select_answers(
ids, predictions, scores)
if with_label:
metrics = calc_metrics(metric_meta, golds, predictions, scores,
label_mapper)
return metrics, predictions, scores, golds, ids
def main(args):
# load task info
task_defs = TaskDefs(args.task_def)
assert args.task in task_defs.task_type_map
assert args.task in task_defs.data_type_map
assert args.task in task_defs.metric_meta_map
data_type = task_defs.data_type_map[args.task]
task_type = task_defs.task_type_map[args.task]
metric_meta = task_defs.metric_meta_map[args.task]
# 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)
model.load(checkpoint_path)
encoder_type = config.get('encoder_type', EncoderModelType.BERT)
# load data
test_data_set = SingleTaskDataset(args.prep_input,
False,
task_type=task_type,
maxlen=args.max_seq_len)
collater = Collater(is_train=False, encoder_type=encoder_type)
test_data = DataLoader(test_data_set,
batch_size=args.batch_size_eval,
collate_fn=collater.collate_fn,
pin_memory=args.cuda)
with torch.no_grad():
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)
def extract_encoding(model, data, use_cuda=True):
if use_cuda:
model.cuda()
sequence_outputs = []
max_seq_len = 0
for idx, (batch_info, batch_data) in enumerate(data):
batch_info, batch_data = Collater.patch_data(use_cuda, batch_info, batch_data)
sequence_output = model.encode(batch_info, batch_data)
sequence_outputs.append(sequence_output)
max_seq_len = max(max_seq_len, sequence_output.shape[1])
new_sequence_outputs = []
for sequence_output in sequence_outputs:
new_sequence_output = torch.zeros(sequence_output.shape[0], max_seq_len, sequence_output.shape[2])
new_sequence_output[:, :sequence_output.shape[1], :] = sequence_output
new_sequence_outputs.append(new_sequence_output)
return torch.cat(new_sequence_outputs)
def eval_model(model,
data,
metric_meta,
device,
with_label=True,
label_mapper=None,
task_type=TaskType.Classification):
predictions = []
golds = []
scores = []
ids = []
metrics = {}
print("****device={}".format(device))
for (batch_info, batch_data) in data:
batch_info, batch_data = Collater.patch_data(device, batch_info,
batch_data)
score, pred, gold = model.predict(batch_info, batch_data)
predictions.extend(pred)
golds.extend(gold)
scores.extend(score)
ids.extend(batch_info['uids'])
if task_type == TaskType.Span:
from experiments.squad import squad_utils
golds = squad_utils.merge_answers(ids, golds)
predictions, scores = squad_utils.select_answers(
ids, predictions, scores)
if with_label:
metrics = calc_metrics(metric_meta, golds, predictions, scores,
label_mapper)
for i in range(min(len(ids), 10)):
print("{}\t{}\t{}\t{}\n".format(ids[i], predictions[i], scores[2 * i],
scores[2 * i + 1]))
#print("score heads={}".format(scores[:10]))
return metrics, predictions, scores, golds, ids
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)
model.load(checkpoint_path)
encoder_type = config.get('encoder_type', EncoderModelType.BERT)
# load data
test_data_set = SingleTaskDataset(args.prep_input,
False,
task_type=task_type,
maxlen=args.max_seq_len)
collater = Collater(gpu=args.cuda,
is_train=False,
task_id=args.task_id,
task_type=task_type,
data_type=data_type,
encoder_type=encoder_type)
test_data = DataLoader(test_data_set,
batch_size=args.batch_size_eval,
collate_fn=collater.collate_fn,
pin_memory=args.cuda)
with torch.no_grad():
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)
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
tasks = {}
task_def_list = []
dropout_list = []
train_datasets = []
for dataset in args.train_datasets:
prefix = dataset.split('_')[0]
if prefix in tasks:
continue
task_id = len(tasks)
tasks[prefix] = task_id
task_def = task_defs.get_task_def(prefix)
task_def_list.append(task_def)
train_path = os.path.join(data_dir, '{}_train.json'.format(dataset))
logger.info('Loading {} as task {}'.format(train_path, task_id))
train_data_set = SingleTaskDataset(train_path,
True,
maxlen=args.max_seq_len,
task_id=task_id,
task_def=task_def)
train_datasets.append(train_data_set)
train_collater = Collater(dropout_w=args.dropout_w,
encoder_type=encoder_type,
soft_label=args.mkd_opt > 0)
multi_task_train_dataset = MultiTaskDataset(train_datasets)
multi_task_batch_sampler = MultiTaskBatchSampler(train_datasets,
args.batch_size,
args.mix_opt, args.ratio)
multi_task_train_data = DataLoader(multi_task_train_dataset,
batch_sampler=multi_task_batch_sampler,
collate_fn=train_collater.collate_fn,
pin_memory=args.cuda)
opt['task_def_list'] = task_def_list
dev_data_list = []
test_data_list = []
test_collater = Collater(is_train=False, encoder_type=encoder_type)
for dataset in args.test_datasets:
prefix = dataset.split('_')[0]
task_def = task_defs.get_task_def(prefix)
task_id = tasks[prefix]
task_type = task_def.task_type
data_type = task_def.data_type
dev_path = os.path.join(data_dir, '{}_dev.json'.format(dataset))
dev_data = None
if os.path.exists(dev_path):
dev_data_set = SingleTaskDataset(dev_path,
False,
maxlen=args.max_seq_len,
task_id=task_id,
task_def=task_def)
dev_data = DataLoader(dev_data_set,
batch_size=args.batch_size_eval,
collate_fn=test_collater.collate_fn,
pin_memory=args.cuda)
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_set = SingleTaskDataset(test_path,
False,
maxlen=args.max_seq_len,
task_id=task_id,
task_def=task_def)
test_data = DataLoader(test_data_set,
batch_size=args.batch_size_eval,
collate_fn=test_collater.collate_fn,
pin_memory=args.cuda)
test_data_list.append(test_data)
logger.info('#' * 20)
logger.info(opt)
logger.info('#' * 20)
# div number of grad accumulation.
num_all_batches = args.epochs * len(
multi_task_train_data) // args.grad_accumulation_step
logger.info('############# Gradient Accumulation Info #############')
logger.info('number of step: {}'.format(args.epochs *
len(multi_task_train_data)))
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 #############')
init_model = args.init_checkpoint
state_dict = None
if os.path.exists(init_model):
state_dict = torch.load(init_model)
config = state_dict['config']
else:
if opt['encoder_type'] not in EncoderModelType._value2member_map_:
raise ValueError("encoder_type is out of pre-defined types")
literal_encoder_type = EncoderModelType(
opt['encoder_type']).name.lower()
config_class, model_class, tokenizer_class = MODEL_CLASSES[
literal_encoder_type]
config = config_class.from_pretrained(
init_model, output_hidden_states=True).to_dict(
) # change here to enable multi-layer output
config['output_hidden_states'] = True
config['attention_probs_dropout_prob'] = args.bert_dropout_p
config['hidden_dropout_prob'] = args.bert_dropout_p
config['multi_gpu_on'] = opt["multi_gpu_on"]
if args.num_hidden_layers != -1:
config['num_hidden_layers'] = args.num_hidden_layers
opt.update(config)
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))
# tensorboard
if args.tensorboard:
args.tensorboard_logdir = os.path.join(args.output_dir,
args.tensorboard_logdir)
tensorboard = SummaryWriter(log_dir=args.tensorboard_logdir)
if args.encode_mode:
for idx, dataset in enumerate(args.test_datasets):
prefix = dataset.split('_')[0]
test_data = test_data_list[idx]
with torch.no_grad():
encoding = extract_encoding(model,
test_data,
use_cuda=args.cuda)
torch.save(
encoding,
os.path.join(output_dir, '{}_encoding.pt'.format(dataset)))
return
for epoch in range(0, args.epochs):
logger.warning('At epoch {}'.format(epoch))
start = datetime.now()
for i, (batch_meta, batch_data) in enumerate(multi_task_train_data):
batch_meta, batch_data = Collater.patch_data(
args.cuda, batch_meta, batch_data)
task_id = batch_meta['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(multi_task_train_data) - 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.tensorboard:
tensorboard.add_scalar('train/loss',
model.train_loss.avg,
global_step=model.updates)
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]
task_def = task_defs.get_task_def(prefix)
label_dict = task_def.label_vocab
dev_data = dev_data_list[idx]
if dev_data is not None:
with torch.no_grad():
dev_metrics, dev_predictions, scores, golds, dev_ids = eval_model(
model,
dev_data,
metric_meta=task_def.metric_meta,
use_cuda=args.cuda,
label_mapper=label_dict,
task_type=task_def.task_type)
for key, val in dev_metrics.items():
if args.tensorboard:
tensorboard.add_scalar('dev/{}/{}'.format(
dataset, key),
val,
global_step=epoch)
if isinstance(val, str):
logger.warning(
'Task {0} -- epoch {1} -- Dev {2}:\n {3}'.format(
dataset, epoch, key, val))
else:
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)
if args.glue_format_on:
from experiments.glue.glue_utils import submit
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:
with torch.no_grad():
test_metrics, test_predictions, scores, golds, test_ids = eval_model(
model,
test_data,
metric_meta=task_def.metric_meta,
use_cuda=args.cuda,
with_label=False,
label_mapper=label_dict,
task_type=task_def.task_type)
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)
if args.glue_format_on:
from experiments.glue.glue_utils import submit
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)
if args.tensorboard:
tensorboard.close()
def main():
# set up dist
device = torch.device("cuda")
if args.local_rank > -1:
device = initialize_distributed(args)
elif torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
opt = vars(args)
# update data dir
opt['data_dir'] = data_dir
batch_size = args.batch_size
print_message(logger, 'Launching the MT-DNN training')
#return
tasks = {}
task_def_list = []
dropout_list = []
printable = args.local_rank in [-1, 0]
train_datasets = []
for dataset in args.train_datasets:
prefix = dataset.split('_')[0]
if prefix in tasks:
continue
task_id = len(tasks)
tasks[prefix] = task_id
task_def = task_defs.get_task_def(prefix)
task_def_list.append(task_def)
train_path = os.path.join(data_dir, '{}_train.json'.format(dataset))
print_message(logger,
'Loading {} as task {}'.format(train_path, task_id))
train_data_set = SingleTaskDataset(train_path,
True,
maxlen=args.max_seq_len,
task_id=task_id,
task_def=task_def,
printable=printable)
train_datasets.append(train_data_set)
train_collater = Collater(dropout_w=args.dropout_w,
encoder_type=encoder_type,
soft_label=args.mkd_opt > 0,
max_seq_len=args.max_seq_len,
do_padding=args.do_padding)
multi_task_train_dataset = MultiTaskDataset(train_datasets)
if args.local_rank != -1:
multi_task_batch_sampler = DistMultiTaskBatchSampler(
train_datasets,
args.batch_size,
args.mix_opt,
args.ratio,
rank=args.local_rank,
world_size=args.world_size)
else:
multi_task_batch_sampler = MultiTaskBatchSampler(
train_datasets,
args.batch_size,
args.mix_opt,
args.ratio,
bin_on=args.bin_on,
bin_size=args.bin_size,
bin_grow_ratio=args.bin_grow_ratio)
multi_task_train_data = DataLoader(multi_task_train_dataset,
batch_sampler=multi_task_batch_sampler,
collate_fn=train_collater.collate_fn,
pin_memory=args.cuda)
opt['task_def_list'] = task_def_list
dev_data_list = []
test_data_list = []
test_collater = Collater(is_train=False,
encoder_type=encoder_type,
max_seq_len=args.max_seq_len,
do_padding=args.do_padding)
for dataset in args.test_datasets:
prefix = dataset.split('_')[0]
task_def = task_defs.get_task_def(prefix)
task_id = tasks[prefix]
task_type = task_def.task_type
data_type = task_def.data_type
dev_path = os.path.join(data_dir, '{}_dev.json'.format(dataset))
dev_data = None
if os.path.exists(dev_path):
dev_data_set = SingleTaskDataset(dev_path,
False,
maxlen=args.max_seq_len,
task_id=task_id,
task_def=task_def,
printable=printable)
if args.local_rank != -1:
dev_data_set = DistTaskDataset(dev_data_set, task_id)
single_task_batch_sampler = DistSingleTaskBatchSampler(
dev_data_set,
args.batch_size_eval,
rank=args.local_rank,
world_size=args.world_size)
dev_data = DataLoader(dev_data_set,
batch_sampler=single_task_batch_sampler,
collate_fn=test_collater.collate_fn,
pin_memory=args.cuda)
else:
dev_data = DataLoader(dev_data_set,
batch_size=args.batch_size_eval,
collate_fn=test_collater.collate_fn,
pin_memory=args.cuda)
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_set = SingleTaskDataset(test_path,
False,
maxlen=args.max_seq_len,
task_id=task_id,
task_def=task_def,
printable=printable)
if args.local_rank != -1:
test_data_set = DistTaskDataset(test_data_set, task_id)
single_task_batch_sampler = DistSingleTaskBatchSampler(
test_data_set,
args.batch_size_eval,
rank=args.local_rank,
world_size=args.world_size)
test_data = DataLoader(test_data_set,
batch_sampler=single_task_batch_sampler,
collate_fn=test_collater.collate_fn,
pin_memory=args.cuda)
else:
test_data = DataLoader(test_data_set,
batch_size=args.batch_size_eval,
collate_fn=test_collater.collate_fn,
pin_memory=args.cuda)
test_data_list.append(test_data)
print_message(logger, '#' * 20)
print_message(logger, opt)
print_message(logger, '#' * 20)
# div number of grad accumulation.
num_all_batches = args.epochs * len(
multi_task_train_data) // args.grad_accumulation_step
print_message(logger,
'############# Gradient Accumulation Info #############')
print_message(
logger,
'number of step: {}'.format(args.epochs * len(multi_task_train_data)))
print_message(
logger, 'number of grad grad_accumulation step: {}'.format(
args.grad_accumulation_step))
print_message(logger,
'adjusted number of step: {}'.format(num_all_batches))
print_message(logger,
'############# Gradient Accumulation Info #############')
init_model = args.init_checkpoint
state_dict = None
if os.path.exists(init_model):
if encoder_type == EncoderModelType.BERT or \
encoder_type == EncoderModelType.DEBERTA or \
encoder_type == EncoderModelType.ELECTRA:
state_dict = torch.load(init_model, map_location=device)
config = state_dict['config']
elif encoder_type == EncoderModelType.ROBERTA or encoder_type == EncoderModelType.XLM:
model_path = '{}/model.pt'.format(init_model)
state_dict = torch.load(model_path, map_location=device)
arch = state_dict['args'].arch
arch = arch.replace('_', '-')
if encoder_type == EncoderModelType.XLM:
arch = "xlm-{}".format(arch)
# convert model arch
from data_utils.roberta_utils import update_roberta_keys
from data_utils.roberta_utils import patch_name_dict
state = update_roberta_keys(
state_dict['model'], nlayer=state_dict['args'].encoder_layers)
state = patch_name_dict(state)
literal_encoder_type = EncoderModelType(
opt['encoder_type']).name.lower()
config_class, model_class, tokenizer_class = MODEL_CLASSES[
literal_encoder_type]
config = config_class.from_pretrained(arch).to_dict()
state_dict = {'state': state}
else:
if opt['encoder_type'] not in EncoderModelType._value2member_map_:
raise ValueError("encoder_type is out of pre-defined types")
literal_encoder_type = EncoderModelType(
opt['encoder_type']).name.lower()
config_class, model_class, tokenizer_class = MODEL_CLASSES[
literal_encoder_type]
config = config_class.from_pretrained(init_model).to_dict()
config['attention_probs_dropout_prob'] = args.bert_dropout_p
config['hidden_dropout_prob'] = args.bert_dropout_p
config['multi_gpu_on'] = opt["multi_gpu_on"]
if args.num_hidden_layers > 0:
config['num_hidden_layers'] = args.num_hidden_layers
opt.update(config)
model = MTDNNModel(opt,
device=device,
state_dict=state_dict,
num_train_step=num_all_batches)
if args.resume and args.model_ckpt:
print_message(logger, 'loading model from {}'.format(args.model_ckpt))
model.load(args.model_ckpt)
#### model meta str
headline = '############# Model Arch of MT-DNN #############'
### print network
print_message(logger, '\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))
print_message(logger,
"Total number of params: {}".format(model.total_param))
# tensorboard
tensorboard = None
if args.tensorboard:
args.tensorboard_logdir = os.path.join(args.output_dir,
args.tensorboard_logdir)
tensorboard = SummaryWriter(log_dir=args.tensorboard_logdir)
if args.encode_mode:
for idx, dataset in enumerate(args.test_datasets):
prefix = dataset.split('_')[0]
test_data = test_data_list[idx]
with torch.no_grad():
encoding = extract_encoding(model,
test_data,
use_cuda=args.cuda)
torch.save(
encoding,
os.path.join(output_dir, '{}_encoding.pt'.format(dataset)))
return
for epoch in range(0, args.epochs):
print_message(logger, 'At epoch {}'.format(epoch), level=1)
start = datetime.now()
for i, (batch_meta, batch_data) in enumerate(multi_task_train_data):
batch_meta, batch_data = Collater.patch_data(
device, batch_meta, batch_data)
task_id = batch_meta['task_id']
model.update(batch_meta, batch_data)
if (model.updates) % (
args.log_per_updates) == 0 or model.updates == 1:
ramaining_time = str(
(datetime.now() - start) / (i + 1) *
(len(multi_task_train_data) - i - 1)).split('.')[0]
if args.adv_train and args.debug:
debug_info = ' basic loss[%.5f] adv loss[%.5f] emb val[%.8f] noise val[%.8f] noise grad val[%.8f] no proj noise[%.8f] ' % (
model.basic_loss.avg, model.adv_loss.avg,
model.emb_val.avg, model.noise_val.avg,
model.noise_grad_val.avg, model.no_proj_noise_val.avg)
else:
debug_info = ' '
print_message(
logger,
'Task [{0:2}] updates[{1:6}] train loss[{2:.5f}]{3}remaining[{4}]'
.format(task_id, model.updates, model.train_loss.avg,
debug_info, ramaining_time))
if args.tensorboard:
tensorboard.add_scalar('train/loss',
model.train_loss.avg,
global_step=model.updates)
if args.save_per_updates_on and (
(model.local_updates) %
(args.save_per_updates * args.grad_accumulation_step)
== 0) and args.local_rank in [-1, 0]:
model_file = os.path.join(
output_dir, 'model_{}_{}.pt'.format(epoch, model.updates))
evaluation(model,
args.test_datasets,
dev_data_list,
task_defs,
output_dir,
epoch,
n_updates=args.save_per_updates,
with_label=True,
tensorboard=tensorboard,
glue_format_on=args.glue_format_on,
test_on=False,
device=device,
logger=logger)
evaluation(model,
args.test_datasets,
test_data_list,
task_defs,
output_dir,
epoch,
n_updates=args.save_per_updates,
with_label=False,
tensorboard=tensorboard,
glue_format_on=args.glue_format_on,
test_on=True,
device=device,
logger=logger)
print_message(logger,
'Saving mt-dnn model to {}'.format(model_file))
model.save(model_file)
evaluation(model,
args.test_datasets,
dev_data_list,
task_defs,
output_dir,
epoch,
with_label=True,
tensorboard=tensorboard,
glue_format_on=args.glue_format_on,
test_on=False,
device=device,
logger=logger)
evaluation(model,
args.test_datasets,
test_data_list,
task_defs,
output_dir,
epoch,
with_label=False,
tensorboard=tensorboard,
glue_format_on=args.glue_format_on,
test_on=True,
device=device,
logger=logger)
print_message(logger, '[new test scores at {} saved.]'.format(epoch))
if args.local_rank in [-1, 0]:
model_file = os.path.join(output_dir, 'model_{}.pt'.format(epoch))
model.save(model_file)
if args.tensorboard:
tensorboard.close()
def load_model_for_viz_0(task_def_path,
checkpoint_path,
input_path,
model_type='bert-base-cased',
do_lower_case=False,
use_cuda=True):
# load task info
task = os.path.splitext(os.path.basename(task_def_path))[0]
task_defs = TaskDefs(task_def_path)
assert task in task_defs._task_type_map
assert task in task_defs._data_type_map
assert task in task_defs._metric_meta_map
prefix = task.split('_')[0]
task_def = task_defs.get_task_def(prefix)
data_type = task_defs._data_type_map[task]
task_type = task_defs._task_type_map[task]
metric_meta = task_defs._metric_meta_map[task]
# load model
assert os.path.exists(checkpoint_path)
state_dict = torch.load(checkpoint_path)
config = state_dict['config']
config["cuda"] = use_cuda
task_def = task_defs.get_task_def(prefix)
task_def_list = [task_def]
config['task_def_list'] = task_def_list
####### temp fix #######
config['fp16'] = False
config['answer_opt'] = 0
config['adv_train'] = False
del state_dict['optimizer']
#########################
model = MTDNNModel(config, state_dict=state_dict)
encoder_type = config.get('encoder_type', EncoderModelType.BERT)
root = os.path.basename(task_def_path)
literal_model_type = model_type.split('-')[0].upper()
encoder_model = EncoderModelType[literal_model_type]
literal_model_type = literal_model_type.lower()
mt_dnn_suffix = literal_model_type
if 'base' in model_type:
mt_dnn_suffix += "_base"
elif 'large' in model_type:
mt_dnn_suffix += "_large"
# load tokenizer
config_class, model_class, tokenizer_class = MODEL_CLASSES[
literal_model_type]
tokenizer = tokenizer_class.from_pretrained(model_type,
do_lower_case=do_lower_case)
# load data
prep_input = input_path
test_data_set = SingleTaskDataset(prep_input,
False,
maxlen=512,
task_id=0,
task_def=task_def)
collater = Collater(is_train=False, encoder_type=encoder_type)
test_data = DataLoader(test_data_set,
batch_size=1,
collate_fn=collater.collate_fn,
pin_memory=True)
idx = 0
results = []
return model.mnetwork.module.bert, config, test_data
def load_model_for_viz_1(task_def_path,
checkpoint_path,
input_path,
model_type='bert-base-cased',
do_lower_case=False,
use_cuda=True):
# load task info
task = os.path.splitext(os.path.basename(task_def_path))[0]
task_defs = TaskDefs(task_def_path)
assert task in task_defs._task_type_map
assert task in task_defs._data_type_map
assert task in task_defs._metric_meta_map
prefix = task.split('_')[0]
task_def = task_defs.get_task_def(prefix)
data_type = task_defs._data_type_map[task]
task_type = task_defs._task_type_map[task]
metric_meta = task_defs._metric_meta_map[task]
# load model
assert os.path.exists(checkpoint_path)
state_dict = torch.load(checkpoint_path)
config = state_dict['config']
config["cuda"] = use_cuda
device = torch.device("cuda" if use_cuda else "cpu")
task_def = task_defs.get_task_def(prefix)
task_def_list = [task_def]
config['task_def_list'] = task_def_list
## temp fix
config['fp16'] = False
config['answer_opt'] = 0
config['adv_train'] = False
#del state_dict['optimizer']
config['output_attentions'] = True
config['local_rank'] = -1
model = MTDNNModel(config, device, state_dict=state_dict)
encoder_type = config.get('encoder_type', EncoderModelType.BERT)
root = os.path.basename(task_def_path)
literal_model_type = model_type.split('-')[0].upper()
encoder_model = EncoderModelType[literal_model_type]
literal_model_type = literal_model_type.lower()
mt_dnn_suffix = literal_model_type
if 'base' in model_type:
mt_dnn_suffix += "_base"
elif 'large' in model_type:
mt_dnn_suffix += "_large"
# load tokenizer
config_class, model_class, tokenizer_class = MODEL_CLASSES[
literal_model_type]
tokenizer = tokenizer_class.from_pretrained(model_type,
do_lower_case=do_lower_case)
# load data
prep_input = input_path
test_data_set = SingleTaskDataset(prep_input,
False,
maxlen=512,
task_id=0,
task_def=task_def)
collater = Collater(is_train=False, encoder_type=encoder_type)
test_data = DataLoader(test_data_set,
batch_size=1,
collate_fn=collater.collate_fn,
pin_memory=True)
idx = 0
results = []
for batch_meta, batch_data in tqdm(test_data):
if idx < 360:
idx += 1
continue
batch_meta, batch_data = Collater.patch_data(device, batch_meta,
batch_data)
model.network.eval()
task_id = batch_meta['task_id']
task_def = TaskDef.from_dict(batch_meta['task_def'])
task_type = task_def.task_type
task_obj = tasks.get_task_obj(task_def)
inputs = batch_data[:batch_meta['input_len']]
if len(inputs) == 3:
inputs.append(None)
inputs.append(None)
inputs.append(task_id)
input_ids = inputs[0]
token_type_ids = inputs[1]
attention = model.mnetwork.module.bert(
input_ids, token_type_ids=token_type_ids)[-1]
batch_size = batch_data[0].shape[0]
for i in range(batch_size):
attention = tuple([item[i:i + 1, :, :, :] for item in attention])
input_id_list = input_ids[i].tolist()
tokens = tokenizer.convert_ids_to_tokens(input_id_list)
idx_sep = listRightIndex(tokens, '[SEP]') + 1
tokens = tokens[:idx_sep]
attention = tuple(
[item[:, :, :idx_sep, :idx_sep] for item in attention])
results.append((attention, tokens))
idx += batch_size
return results
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
# tensorboard
tensorboard = None
if args.tensorboard:
args.tensorboard_logdir = os.path.join(args.output_dir,
args.tensorboard_logdir)
tensorboard = SummaryWriter(log_dir=args.tensorboard_logdir)
json_logfile = os.path.join(args.output_dir, "runtime_log.json")
tasks = {}
tasks_class = {}
nclass_list = []
decoder_opts = []
task_types = []
dropout_list = []
loss_types = []
kd_loss_types = []
train_datasets = []
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]
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)
task_types.append(task_type)
loss_types.append(task_defs.loss_map[prefix])
kd_loss_types.append(task_defs.kd_loss_map[prefix])
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_set = SingleTaskDataset(train_path,
True,
maxlen=args.max_seq_len,
task_id=task_id,
task_type=task_type,
data_type=data_type)
train_datasets.append(train_data_set)
train_collater = Collater(dropout_w=args.dropout_w,
encoder_type=encoder_type)
multi_task_train_dataset = MultiTaskDataset(train_datasets)
# MTSampler = SAMPLERS[args.sampler]
n_tasks = len(tasks)
dataset_sizes = [len(dataset) for dataset in train_datasets]
if "random" in args.controller:
controller = CONTROLLERS[args.controller](
n_task=n_tasks,
dataset_names=args.train_datasets,
dataset_sizes=dataset_sizes,
batch_size=args.batch_size,
rebatch_size=args.batch_size_train,
tensorboard=tensorboard,
log_filename=json_logfile)
else:
controller = CONTROLLERS[args.controller](
n_task=n_tasks,
phi=args.phi,
K=args.concurrent_cnt,
dataset_names=args.train_datasets,
dataset_sizes=dataset_sizes,
max_cnt=args.max_queue_cnt,
batch_size=args.batch_size,
rebatch_size=args.batch_size_train,
tensorboard=tensorboard,
log_filename=json_logfile)
multi_task_batch_sampler = ACLSampler(train_datasets,
args.batch_size,
controller=controller)
# controller.max_step = len(multi_task_batch_sampler)
multi_task_train_data = DataLoader(multi_task_train_dataset,
batch_sampler=multi_task_batch_sampler,
collate_fn=train_collater.collate_fn,
pin_memory=args.cuda)
opt['answer_opt'] = decoder_opts
opt['task_types'] = task_types
opt['tasks_dropout_p'] = dropout_list
opt['loss_types'] = loss_types
opt['kd_loss_types'] = kd_loss_types
args.label_size = ','.join([str(l) for l in nclass_list])
logger.info(args.label_size)
dev_data_list = []
test_data_list = []
test_collater = Collater(is_train=False, encoder_type=encoder_type)
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_set = SingleTaskDataset(dev_path,
False,
maxlen=args.max_seq_len,
task_id=task_id,
task_type=task_type,
data_type=data_type)
dev_data = DataLoader(dev_data_set,
batch_size=args.batch_size_eval,
collate_fn=test_collater.collate_fn,
pin_memory=args.cuda)
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_set = SingleTaskDataset(test_path,
False,
maxlen=args.max_seq_len,
task_id=task_id,
task_type=task_type,
data_type=data_type)
test_data = DataLoader(test_data_set,
batch_size=args.batch_size_eval,
collate_fn=test_collater.collate_fn,
pin_memory=args.cuda)
test_data_list.append(test_data)
logger.info('#' * 20)
logger.info(opt)
logger.info('#' * 20)
# div number of grad accumulation.
num_all_batches = args.epochs * len(
multi_task_train_data) // args.grad_accumulation_step
logger.info('############# Gradient Accumulation Info #############')
logger.info('number of step: {}'.format(args.epochs *
len(multi_task_train_data)))
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 #############')
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
config['multi_gpu_on'] = opt["multi_gpu_on"]
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()
config['multi_gpu_on'] = opt["multi_gpu_on"]
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}
# add score history
score_history = [[] for _ in range(len(args.test_datasets))]
total_scores = []
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 {0}/{1}'.format(epoch + 1, args.epochs))
start = datetime.now()
total_len = len(controller)
controller.set_epoch(epoch)
for i, (batch_meta, batch_data) in enumerate(multi_task_train_data):
batch_meta, batch_data = Collater.patch_data(
args.cuda, batch_meta, batch_data)
task_id = batch_meta['task_id']
loss = model.calculate_loss(batch_meta, batch_data)
controller.insert(task_id, (batch_meta, batch_data), loss.item())
if i % args.log_per_updates == 0:
ramaining_time = str(
(datetime.now() - start) / (controller.cur_step + 1) *
(total_len - controller.cur_step - 1)).split('.')[0]
logger.info("Epoch {0} Progress {1} / {2} ({3:.2%})".format(
epoch + 1, controller.cur_step, total_len,
controller.cur_step * 1.0 / total_len))
# logger.info("Progress {0} / {1} ({2:.2f}%)".format(i, total_len, i*100.0/total_len))
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))
summary_str = controller.summary()
for line in summary_str.split("\n"):
logger.info(line)
# avg_loss, out_loss, loss_change, min_loss, min_out_loss = controller.get_loss()
# logger.info('List of loss {}'.format(",".join(avg_loss)))
# logger.info('List of out_loss {}'.format(",".join(out_loss)))
# logger.info('List of loss_change {}'.format(",".join(loss_change)))
# logger.info('List of min_loss {}'.format(",".join(min_loss)))
# logger.info('List of min_out_loss {}'.format(",".join(min_out_loss)))
# chosen = [ "%s:%.3f "%(k,v) for k, v in controller.scaled_dict.items()]
# logger.info('List of Scaled Choosen time {}'.format(",".join(chosen)))
if args.tensorboard:
tensorboard.add_scalar('train/loss',
model.train_loss.avg,
global_step=model.updates)
controller.step(model=model)
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)
total_average_score = 0.0
scoring_cnt = 0
score_dict = dict()
scoring_datasets = "cola,sst,mrpc,stsb,qqp,mnli,qnli,rte,wnli".split(
",")
logger.info('Start Testing')
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:
with torch.no_grad():
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,
label_mapper=label_dict,
task_type=task_defs.task_type_map[prefix])
task_score = 0.0
for key, val in dev_metrics.items():
if args.tensorboard:
tensorboard.add_scalar('dev/{}/{}'.format(
dataset, key),
val,
global_step=epoch)
if isinstance(val, str):
logger.warning(
'Task {0} -- epoch {1} -- Dev {2}:\n {3}'.format(
dataset, epoch + 1, key, val))
else:
logger.warning(
'Task {0} -- epoch {1} -- Dev {2}: {3:.2f}'.format(
dataset, epoch + 1, key, val))
task_score += val
if len(dev_metrics) > 1:
task_score /= len(dev_metrics)
logger.warning(
'Task {0} -- epoch {1} -- Dev {2}: {3:.2f}'.format(
dataset, epoch + 1, "Average", task_score))
if prefix in scoring_datasets:
scoring_cnt += 1
if prefix not in score_dict:
score_dict[prefix] = task_score
else:
score_dict[prefix] = (score_dict[prefix] +
task_score) / 2
total_average_score += task_score
score_history[idx].append("%.2f" % task_score)
logger.warning('Task {0} -- epoch {1} -- Dev {2}: {3}'.format(
dataset, epoch + 1, "History", score_history[idx]))
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)
if args.glue_format_on:
from experiments.glue.glue_utils import submit
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:
with torch.no_grad():
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,
label_mapper=label_dict,
task_type=task_defs.task_type_map[prefix])
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)
if args.glue_format_on:
from experiments.glue.glue_utils import submit
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.]')
scoreing_cnt = len(score_dict)
if scoreing_cnt > 0:
mean_value = np.mean([v for k, v in score_dict.items()])
logger.warning(
'Epoch {0} -- Dev {1} Tasks, Average Score : {2:.3f}'.format(
epoch + 1, scoring_cnt, mean_value))
score_dict['avg'] = mean_value
total_scores.append(score_dict)
model_file = os.path.join(output_dir, 'model_{}.pt'.format(epoch))
model.save(model_file)
for i, total_score in enumerate(total_scores):
logger.info(total_score)
if args.tensorboard:
tensorboard.close()
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
tasks = {}
tasks_class = {}
nclass_list = []
decoder_opts = []
task_types = []
dropout_list = []
loss_types = []
kd_loss_types = []
train_datasets = []
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]
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)
task_types.append(task_type)
loss_types.append(task_defs.loss_map[prefix])
kd_loss_types.append(task_defs.kd_loss_map[prefix])
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_set = SingleTaskDataset(train_path,
True,
maxlen=args.max_seq_len,
task_id=task_id,
task_type=task_type,
data_type=data_type)
train_datasets.append(train_data_set)
train_collater = Collater(dropout_w=args.dropout_w,
encoder_type=encoder_type)
multi_task_train_dataset = MultiTaskDataset(train_datasets)
multi_task_batch_sampler = MultiTaskBatchSampler(train_datasets,
args.batch_size,
args.mix_opt, args.ratio)
multi_task_train_data = DataLoader(multi_task_train_dataset,
batch_sampler=multi_task_batch_sampler,
collate_fn=train_collater.collate_fn,
pin_memory=args.cuda)
opt['answer_opt'] = decoder_opts
opt['task_types'] = task_types
opt['tasks_dropout_p'] = dropout_list
opt['loss_types'] = loss_types
opt['kd_loss_types'] = kd_loss_types
args.label_size = ','.join([str(l) for l in nclass_list])
logger.info(args.label_size)
dev_data_list = []
test_data_list = []
test_collater = Collater(is_train=False, encoder_type=encoder_type)
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_set = SingleTaskDataset(dev_path,
False,
maxlen=args.max_seq_len,
task_id=task_id,
task_type=task_type,
data_type=data_type)
dev_data = DataLoader(dev_data_set,
batch_size=args.batch_size_eval,
collate_fn=test_collater.collate_fn,
pin_memory=args.cuda)
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_set = SingleTaskDataset(test_path,
False,
maxlen=args.max_seq_len,
task_id=task_id,
task_type=task_type,
data_type=data_type)
test_data = DataLoader(test_data_set,
batch_size=args.batch_size_eval,
collate_fn=test_collater.collate_fn,
pin_memory=args.cuda)
test_data_list.append(test_data)
logger.info('#' * 20)
logger.info(opt)
logger.info('#' * 20)
# div number of grad accumulation.
num_all_batches = args.epochs * len(
multi_task_train_data) // args.grad_accumulation_step
logger.info('############# Gradient Accumulation Info #############')
logger.info('number of step: {}'.format(args.epochs *
len(multi_task_train_data)))
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 #############')
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
config['multi_gpu_on'] = opt["multi_gpu_on"]
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()
config['multi_gpu_on'] = opt["multi_gpu_on"]
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))
# tensorboard
if args.tensorboard:
args.tensorboard_logdir = os.path.join(args.output_dir,
args.tensorboard_logdir)
tensorboard = SummaryWriter(log_dir=args.tensorboard_logdir)
for epoch in range(0, args.epochs):
logger.warning('At epoch {}'.format(epoch))
start = datetime.now()
for i, (batch_meta, batch_data) in enumerate(multi_task_train_data):
batch_meta, batch_data = Collater.patch_data(
args.cuda, batch_meta, batch_data)
task_id = batch_meta['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(multi_task_train_data) - 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.tensorboard:
tensorboard.add_scalar('train/loss',
model.train_loss.avg,
global_step=model.updates)
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:
with torch.no_grad():
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,
label_mapper=label_dict,
task_type=task_defs.task_type_map[prefix])
for key, val in dev_metrics.items():
if args.tensorboard:
tensorboard.add_scalar('dev/{}/{}'.format(
dataset, key),
val,
global_step=epoch)
if isinstance(val, str):
logger.warning(
'Task {0} -- epoch {1} -- Dev {2}:\n {3}'.format(
dataset, epoch, key, val))
else:
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)
if args.glue_format_on:
from experiments.glue.glue_utils import submit
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:
with torch.no_grad():
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,
label_mapper=label_dict,
task_type=task_defs.task_type_map[prefix])
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)
if args.glue_format_on:
from experiments.glue.glue_utils import submit
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)
if args.tensorboard:
tensorboard.close()
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
task_def = task_defs.get_task_def(prefix)
task_def_list = [task_def]
config['task_def_list'] = task_def_list
## temp fix
config['answer_opt'] = 0
config['adv_train'] = False
model = MTDNNModel(config, state_dict=state_dict)
model.load(checkpoint_path)
encoder_type = config.get('encoder_type', EncoderModelType.BERT)
# load data
test_data_set = SingleTaskDataset(args.prep_input, False, maxlen=args.max_seq_len, task_id=args.task_id, task_def=task_def)
collater = Collater(is_train=False, encoder_type=encoder_type)
test_data = DataLoader(test_data_set, batch_size=args.batch_size_eval, collate_fn=collater.collate_fn, pin_memory=args.cuda)
with torch.no_grad():
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)
def main():
global tokenizer, test_collater, model
logger.info('Launching the MT-DNN training')
opt = vars(args)
# update data dir
opt['data_dir'] = data_dir
batch_size = args.batch_size
tasks = {}
tasks_class = {}
nclass_list = []
decoder_opts = []
task_types = []
dropout_list = []
loss_types = []
kd_loss_types = []
#train_datasets = []
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]
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)
task_types.append(task_type)
loss_types.append(task_defs.loss_map[prefix])
kd_loss_types.append(task_defs.kd_loss_map[prefix])
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_set = SingleTaskDataset(train_path, True, maxlen=args.max_seq_len, task_id=task_id,
# task_type=task_type, data_type=data_type)
# train_datasets.append(train_data_set)
#train_collater = Collater(dropout_w=args.dropout_w, encoder_type=encoder_type)
# multi_task_train_dataset = MultiTaskDataset(train_datasets)
# multi_task_batch_sampler = MultiTaskBatchSampler(train_datasets, args.batch_size, args.mix_opt, args.ratio)
# multi_task_train_data = DataLoader(multi_task_train_dataset, batch_sampler=multi_task_batch_sampler,
# collate_fn=train_collater.collate_fn, pin_memory=args.cuda)
opt['answer_opt'] = decoder_opts
opt['task_types'] = task_types
opt['tasks_dropout_p'] = dropout_list
opt['loss_types'] = loss_types
opt['kd_loss_types'] = kd_loss_types
args.label_size = ','.join([str(l) for l in nclass_list])
logger.info(args.label_size)
dev_data_list = []
test_data_list = []
test_collater = Collater(is_train=False, encoder_type=encoder_type)
logger.info('#' * 20)
logger.info(opt)
logger.info('#' * 20)
bert_model_path = 'checkpoints/my_mnli/model_0.pt'
state_dict = None
if encoder_type == EncoderModelType.BERT:
if os.path.exists(bert_model_path):
state_dict = torch.load(bert_model_path, map_location=torch.device('cpu'))
config = state_dict['config']
config['attention_probs_dropout_prob'] = args.bert_dropout_p
config['hidden_dropout_prob'] = args.bert_dropout_p
config['multi_gpu_on'] = opt["multi_gpu_on"]
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()
config['multi_gpu_on'] = opt["multi_gpu_on"]
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)
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))
# # tensorboard
# if args.tensorboard:
# args.tensorboard_logdir = os.path.join(args.output_dir, args.tensorboard_logdir)
# tensorboard = SummaryWriter(log_dir=args.tensorboard_logdir)
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased', do_lower_case=True)
def main():
parser = argparse.ArgumentParser()
model_config(parser)
set_config(parser)
train_config(parser)
args = parser.parse_args()
encoder_type = args.encoder_type
layer_indexes = [int(x) for x in args.layers.split(",")]
set_environment(args.seed)
# process data
data, is_single_sentence = process_data(args)
data_type = (DataFormat.PremiseOnly
if is_single_sentence else DataFormat.PremiseAndOneHypothesis)
fout_temp = "{}.tmp".format(args.finput)
dump_data(data, fout_temp)
collater = Collater(is_train=False, encoder_type=encoder_type)
dataset = SingleTaskDataset(
fout_temp,
False,
maxlen=args.max_seq_length,
)
batcher = DataLoader(
dataset,
batch_size=args.batch_size,
collate_fn=collater.collate_fn,
pin_memory=args.cuda,
)
opt = vars(args)
# load model
if os.path.exists(args.checkpoint):
state_dict = torch.load(args.checkpoint)
config = state_dict["config"]
config["dump_feature"] = True
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)
return
num_all_batches = len(batcher)
model = MTDNNModel(opt,
state_dict=state_dict,
num_train_step=num_all_batches)
if args.cuda:
model.cuda()
features_dict = {}
for batch_meta, batch_data in batcher:
batch_meta, batch_data = Collater.patch_data(args.cuda, batch_meta,
batch_data)
all_encoder_layers, _ = model.extract(batch_meta, batch_data)
embeddings = [
all_encoder_layers[idx].detach().cpu().numpy()
for idx in layer_indexes
]
uids = batch_meta["uids"]
masks = batch_data[batch_meta["mask"]].detach().cpu().numpy().tolist()
for idx, uid in enumerate(uids):
slen = sum(masks[idx])
features = {}
for yidx, layer in enumerate(layer_indexes):
features[layer] = str(embeddings[yidx][idx][:slen].tolist())
features_dict[uid] = features
# save features
with open(args.foutput, "w", encoding="utf-8") as writer:
for sample in data:
uid = sample["uid"]
feature = features_dict[uid]
feature["uid"] = uid
writer.write("{}\n".format(json.dumps(feature)))
def compute_heads_importance(args,
model,
eval_dataloader,
compute_entropy=True,
compute_importance=True,
head_mask=None,
actually_pruned=False,
verbose=True):
"""This method shows how to compute:
- head attention entropy
- head importance scores according to http://arxiv.org/abs/1905.10650
"""
# Prepare our tensors
device = torch.device("cuda" if args['cuda'] else "cpu")
n_layers = model.mnetwork.module.bert.config.num_hidden_layers
n_heads = model.mnetwork.module.bert.config.num_attention_heads
head_importance = torch.zeros(n_layers, n_heads).to(device)
attn_entropy = torch.zeros(n_layers, n_heads).to(device)
if head_mask is None:
head_mask = torch.ones(n_layers, n_heads).to(device)
head_mask.requires_grad_(requires_grad=True)
# If actually pruned attention multi-head, set head mask to None to avoid shape mismatch
if actually_pruned:
head_mask = None
preds = None
labels = None
tot_tokens = 0.0
for batch_meta, batch_data in tqdm(eval_dataloader):
for i in range(len(batch_data[1])):
batch_data[1][i] = batch_data[1][i].to(device)
y = batch_data[batch_meta['label']]
y = model._to_cuda(y)
# Do a forward pass (not with torch.no_grad() since we need gradients for importance score - see below)
batch_meta, batch_data = Collater.patch_data(device, batch_meta,
batch_data)
logits, attention = model.update(batch_meta,
batch_data,
head_mask=head_mask)
if compute_entropy:
for layer, attn in enumerate(attention):
masked_entropy = entropy(attn.detach()) * batch_data[
batch_meta['mask']].float().unsqueeze(1)
attn_entropy[layer] += masked_entropy.sum(-1).sum(0).detach()
if compute_importance:
head_importance += head_mask.grad.abs().detach()
# Also store our logits/labels if we want to compute metrics afterwards
if preds is None:
preds = logits.detach().cpu().numpy()
labels = y.detach().cpu().numpy()
else:
preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
labels = np.append(labels, y.detach().cpu().numpy(), axis=0)
tot_tokens += batch_data[
batch_meta['mask']].float().detach().sum().data
# Normalize
attn_entropy /= tot_tokens
head_importance /= tot_tokens
# Layerwise importance normalization
if not args['dont_normalize_importance_by_layer']:
exponent = 2
norm_by_layer = torch.pow(
torch.pow(head_importance, exponent).sum(-1), 1 / exponent)
head_importance /= norm_by_layer.unsqueeze(-1) + 1e-20
if not args['dont_normalize_global_importance']:
head_importance = (head_importance - head_importance.min()) / (
head_importance.max() - head_importance.min())
# Print/save matrices
np.save(os.path.join(args['output_dir'], "attn_entropy.npy"),
attn_entropy.detach().cpu().numpy())
np.save(os.path.join(args['output_dir'], "head_importance.npy"),
head_importance.detach().cpu().numpy())
if verbose:
print("Attention entropies")
print_2d_tensor(attn_entropy)
print("Head importance scores")
print_2d_tensor(head_importance)
print("Head ranked by importance scores")
head_ranks = torch.zeros(head_importance.numel(),
dtype=torch.long,
device=device)
head_ranks[head_importance.view(-1).sort(
descending=True)[1]] = torch.arange(head_importance.numel(),
device=device)
head_ranks = head_ranks.view_as(head_importance)
if verbose:
print_2d_tensor(head_ranks)
return attn_entropy, head_importance, preds, labels
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument("--task_def",
type=str,
required=True,
default="experiments/glue/glue_task_def.yml")
parser.add_argument("--task", type=str, required=True)
parser.add_argument("--task_id",
type=int,
default=0,
help="the id of this task when training")
parser.add_argument("--checkpoint",
default='mt_dnn_models/bert_model_base_uncased.pt',
type=str)
parser.add_argument(
"--output_dir",
default=
'/content/gdrive/My Drive/Colab Notebooks/cs99/mt-dnn/checkpoints/bert-cased_lcp-single_2020-12-23T2029/',
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written.",
)
parser.add_argument(
"--prep_input",
default=
'/content/gdrive/My Drive/Colab Notebooks/cs99/mt-dnn/data_complex/bert_base_cased/lcp_dev.json',
type=str,
required=True,
)
parser.add_argument(
'--bert_model_type',
default='bert-base-cased',
type=str,
help="What type of bert model should we be using",
)
# Other parameters
parser.add_argument(
"--config_name",
default="",
type=str,
help=
"Pretrained config name or path if not the same as model_name_or_path",
)
parser.add_argument(
"--tokenizer_name",
default="",
type=str,
help=
"Pretrained tokenizer name or path if not the same as model_name_or_path",
)
parser.add_argument(
"--cache_dir",
default=None,
type=str,
help=
"Where do you want to store the pre-trained models downloaded from huggingface.co",
)
parser.add_argument(
"--data_subset",
type=int,
default=-1,
help="If > 0: limit the data to a subset of data_subset instances.")
parser.add_argument("--overwrite_output_dir",
action="store_true",
help="Whether to overwrite data in output directory")
parser.add_argument(
"--overwrite_cache",
action="store_true",
help="Overwrite the cached training and evaluation sets")
parser.add_argument("--dont_normalize_importance_by_layer",
action="store_true",
help="Don't normalize importance score by layers")
parser.add_argument(
"--dont_normalize_global_importance",
action="store_true",
help="Don't normalize all importance scores between 0 and 1",
)
parser.add_argument(
"--try_masking",
action="store_true",
help="Whether to try to mask head until a threshold of accuracy.")
parser.add_argument(
"--masking_threshold",
default=0.9,
type=float,
help=
"masking threshold in term of metrics (stop masking when metric < threshold * original metric value).",
)
parser.add_argument(
"--masking_amount",
default=0.1,
type=float,
help="Amount to heads to masking at each masking step.")
parser.add_argument("--metric_name",
default="acc",
type=str,
help="Metric to use for head masking.")
parser.add_argument(
"--max_seq_length",
default=512,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, sequences shorter padded.",
)
# temp fix: technically these parameters should've already bin in checkpoint's config...
parser.add_argument("--world_size",
type=int,
default=1,
help="For distributed training: world size")
parser.add_argument("--batch_size",
default=8,
type=int,
help="Batch size.")
parser.add_argument("--seed", type=int, default=2018)
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--cuda',
type=bool,
default=torch.cuda.is_available(),
help='whether to use GPU acceleration.')
parser.add_argument("--server_ip",
type=str,
default="",
help="Can be used for distant debugging.")
parser.add_argument("--server_port",
type=str,
default="",
help="Can be used for distant debugging.")
parser.add_argument("--do_proper",
type=str,
default=False,
help="Can be used for distant debugging.")
parser.add_argument("--do_improper",
type=str,
default=False,
help="Can be used for distant debugging.")
args = parser.parse_args()
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port),
redirect_output=True)
ptvsd.wait_for_attach()
# Setup devices and distributed training
device = torch.device("cuda")
if args.local_rank > -1:
device = initialize_distributed(args)
elif torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
# load task info
task = args.task
task_defs = TaskDefs(args.task_def)
assert args.task in task_defs._task_type_map
assert args.task in task_defs._data_type_map
assert args.task in task_defs._metric_meta_map
prefix = task.split('_')[0]
task_def = task_defs.get_task_def(prefix)
data_type = task_defs._data_type_map[args.task]
task_type = task_defs._task_type_map[args.task]
metric_meta = task_defs._metric_meta_map[args.task]
# 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")
opt = state_dict['config']
args.bin_on = False
opt.update(vars(args))
model = MTDNNModel(opt, device=device, state_dict=state_dict)
# Load pretrained model and tokenizer
# Load data
data = pd.read_csv('data_complex/lcp_test.tsv',
sep='\t',
header=None,
names=['idx', 'complexity', 'sentence', 'token'])
data['complexity'] = np.load(
'/content/gdrive/My Drive/Colab Notebooks/cs99/from_macbook/single_test_labels.npy'
)
data['class'] = pd.cut(data['complexity'],
labels=[1, 2, 3, 4, 5],
bins=[0, 0.2, 0.4, 0.6, 0.8, 1],
include_lowest=True)
data['sent_len'] = data['sentence'].str.len()
with open(
'/content/gdrive/My Drive/Colab Notebooks/cs99/new-mt-dnn/checkpoints/bert-cased_lcp-single_2021-01-19T0309/lcp_test_scores_epoch_4.json',
'r') as file:
single_dev_bert_scores = json.load(file)
data['finetuned_complexity'] = single_dev_bert_scores['scores']
data['finetuned_error'] = data['finetuned_complexity'] - data[
'complexity']
data['finetuned_abs_error'] = (data['finetuned_complexity'] -
data['complexity']).abs()
with open(
'/content/gdrive/My Drive/Colab Notebooks/cs99/new-mt-dnn/checkpoints/bert-cased_lcp-single_2021-01-19T0309/pretrained.json',
'r') as file:
single_dev_bert_scores = json.load(file)
data['pretrained_complexity'] = single_dev_bert_scores['scores']
data['pretrained_error'] = data['pretrained_complexity'] - data[
'complexity']
data['pretrained_abs_error'] = (data['pretrained_complexity'] -
data['complexity']).abs()
data['improvement'] = data['pretrained_abs_error'] - data[
'finetuned_abs_error']
data['proper'] = data['token'].apply(lambda x: x[0].isupper())
# Distributed training:
# download model & vocab.
printable = opt['local_rank'] in [-1, 0]
encoder_type = opt.get('encoder_type', EncoderModelType.BERT)
collater = Collater(is_train=True,
encoder_type=encoder_type,
max_seq_len=opt['max_seq_len'],
do_padding=opt['do_padding'])
dev_data = SingleTaskDataset(opt['prep_input'],
True,
maxlen=opt['max_seq_len'],
task_id=opt['task_id'],
task_def=task_def,
printable=printable)
if args.do_proper:
dev_data._data = np.array(
dev_data._data)[data[data['proper']]['idx'].to_numpy()].tolist()
if args.do_improper:
dev_data._data = np.array(
dev_data._data)[data[~data['proper']]['idx'].to_numpy()].tolist()
dev_data_loader = DataLoader(dev_data,
batch_size=opt['batch_size_eval'],
collate_fn=collater.collate_fn,
pin_memory=opt['cuda'])
# Compute head entropy and importance score
results = []
for seed in tqdm(range(2010 + 1, 2020 + 1)): # Set seeds
set_seed(seed)
attn_entropy, head_importance, preds, labels = compute_heads_importance(
opt, model, dev_data_loader)
results.append((attn_entropy, head_importance))
pkl.dump(
results,
open('checkpoints/bert-cased_lcp-single_2021-01-19T0309/results.pkl',
'wb'))
# Try head masking (set heads to zero until the score goes under a threshold)
# and head pruning (remove masked heads and see the effect on the network)
if args.try_masking and args.masking_threshold > 0.0 and args.masking_threshold < 1.0:
head_mask = mask_heads(opt, model, dev_data_loader)
def main():
task_def_path = 'data_complex/lcp.yml'
task = os.path.splitext(os.path.basename(task_def_path))[0]
task_defs = TaskDefs(task_def_path)
prefix = task.split('_')[0]
task_def = task_defs.get_task_def(prefix)
parser = argparse.ArgumentParser()
model_config(parser)
set_config(parser)
train_config(parser)
args = parser.parse_args()
encoder_type = args.encoder_type
layer_indexes = [int(x) for x in args.layers.split(",")]
set_environment(args.seed)
# process data
data, is_single_sentence = process_data(args)
data_type = DataFormat.PremiseOnly if is_single_sentence else DataFormat.PremiseAndOneHypothesis
fout_temp = '{}.tmp'.format(args.finput)
dump_data(data, fout_temp)
collater = Collater(is_train=False, encoder_type=encoder_type)
dataset = SingleTaskDataset(fout_temp, False, maxlen=args.max_seq_length, task_def=task_def)#, data_type=data_type)
batcher = DataLoader(dataset, batch_size=args.batch_size, collate_fn=collater.collate_fn, pin_memory=args.cuda)
opt = vars(args)
# load model
if os.path.exists(args.checkpoint):
state_dict = torch.load(args.checkpoint)
config = state_dict['config']
config['dump_feature'] = True
config['local_rank'] = -1
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)
return
num_all_batches = len(batcher)
model = MTDNNModel(
opt,
state_dict=state_dict,
num_train_step=num_all_batches)
if args.cuda:
model.cuda()
features_dict = {}
for batch_meta, batch_data in batcher:
batch_meta, batch_data = Collater.patch_data(args.cuda, batch_meta, batch_data)
all_encoder_layers, _ = model.extract(batch_meta, batch_data)
embeddings = [all_encoder_layers[idx].detach().cpu().numpy()
for idx in layer_indexes]
#import pdb; pdb.set_trace()
uids = batch_meta['uids']
masks = batch_data[batch_meta['mask']].detach().cpu().numpy().tolist()
for idx, uid in enumerate(uids):
slen = sum(masks[idx])
features = {}
for yidx, layer in enumerate(layer_indexes):
features[layer] = str(embeddings[yidx][idx][:slen].tolist())
features_dict[uid] = features
# save features
with open(args.foutput, 'w', encoding='utf-8') as writer:
for sample in data:
uid = sample['uid']
tokens = sample['tokens']
feature = features_dict[uid]
feature['tokens'] = tokens
feature['uid'] = uid
writer.write('{}\n'.format(json.dumps(feature)))
