def read_args(default_config="confs/base.json", **parser_kwargs):
parser = argparse.ArgumentParser(**parser_kwargs)
parser.add_argument("--config", "-c", type=str, default=default_config)
args, _ = parser.parse_known_args()
options = argconf.options_from_json("confs/options.json")
config = argconf.config_from_json(args.config)
return edict(argconf.parse_args(options, config))
def main():
description = "Trains or evaluates a model."
epilog = "Usage:\npython -m deeplm.utils.run -c confs/default.json --action train"
parser = argparse.ArgumentParser(description=description, epilog=epilog)
parser.add_argument("-c", "--config", dest="config", type=str, default="confs/default.json")
parser.add_argument("--action", choices=["train", "eval"], type=str, default="train")
args = parser.parse_args()
conf = argconf.config_from_json(args.config)
if args.action == "train":
train(conf)
def main():
description = "Trains a Magpie model."
epilog = "Usage:\npython -m magpie.utils.train --config confs/cae_model_config.json "\
"--options confs/options.json > cae_train_log"
parser = argparse.ArgumentParser(description=description, epilog=epilog)
parser.add_argument("--config",
type=str,
default="confs/cae_model_config.json")
parser.add_argument("--options", type=str, default="confs/options.json")
args, _ = parser.parse_known_args()
option_dict = argconf.options_from_json(args.options)
config = argconf.config_from_json(args.config)
config = argconf.parse_args(option_dict, config=config)
trainer_cls = model.find_trainer(config["trainer_type"])
trainer = trainer_cls(config)
trainer.train()
def main():
parser = argparse.ArgumentParser(description="Runs the server for tokenization as a service.",
epilog="Usage:\npython -m trident.utils.run_server -c confs/default.json")
parser.add_argument("--config", "-c", type=str, default="confs/default.json")
parser.add_argument("--port", "-p", type=int, default=8080)
args = parser.parse_args()
config = argconf.config_from_json(args.config)
vocab, _ = deeplm.data.Seq2SeqDataset.iters(config)
model = deeplm.model.InterstitialModel(vocab, config)
sd = torch.load(config["resume"])
model.load_state_dict(sd["state"])
model.avg_param = sd["ema"]
model.steps_ema = sd["steps_ema"]
model.cuda()
model.eval()
model.load_ema_params()
cherrypy.config.update({"global": {"engine.autoreload.on": False}})
cherrypy.quickstart(TokenizationServer(vocab, model), "/", {"/": {"request.dispatch": cherrypy.dispatch.MethodDispatcher()}})
def main():
local_rank = -1
parser = argparse.ArgumentParser()
parser.add_argument("--config", "-c", type=str, required=True)
args, _ = parser.parse_known_args()
options = argconf.options_from_json("confs/options.json")
config = argconf.config_from_json(args.config)
args = edict(argconf.parse_args(options, config))
args.local_rank = local_rank
args.on_memory = True
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train:
raise ValueError(
"Training is currently the only implemented execution option. Please set `do_train`."
)
if os.path.exists(args.workspace) and os.listdir(args.workspace):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.workspace))
if not os.path.exists(args.workspace):
os.makedirs(args.workspace)
tokenizer = BertTokenizer.from_pretrained(args.model_file,
do_lower_case=True)
#train_examples = None
num_train_optimization_steps = None
if args.do_train:
print("Loading Train Dataset", args.train_file)
train_dataset = BERTDataset(args.train_file,
tokenizer,
seq_len=args.max_seq_length,
corpus_lines=None,
on_memory=args.on_memory)
num_train_optimization_steps = int(
len(train_dataset) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
# Prepare model
model = BertForPreTraining.from_pretrained(args.model_file)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
if args.do_train:
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset)
else:
#TODO: check if this works with current data generator from disk that relies on next(file)
# (it doesn't return item back by index)
train_sampler = DistributedSampler(train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
loss_fct = nn.CrossEntropyLoss(ignore_index=-1)
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, lm_label_ids = batch
prediction_scores, _ = model(input_ids, segment_ids,
input_mask, lm_label_ids)
loss = loss_fct(
prediction_scores.view(-1, model.module.config.vocab_size),
lm_label_ids.view(-1)).mean()
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
# Save a trained model
logger.info("** ** * Saving fine - tuned model ** ** * ")
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(args.workspace, "pytorch_model.bin")
if args.do_train:
torch.save(model_to_save.state_dict(), output_model_file)
def main():
def evaluate(dataloader, export=None):
eval_loss, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
logits_list = []
iter_idx = 0
corr_x = []
corr_y = []
for input_ids, input_mask, segment_ids, label_ids in tqdm(
dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
tmp_eval_loss = model(input_ids,
segment_ids,
input_mask,
label_ids,
mse=is_float)
logits = model(input_ids,
segment_ids,
input_mask,
mse=is_float)
logits = logits.detach().cpu().numpy()
if export is not None:
logits_list.append(logits)
label_ids = label_ids.to('cpu').numpy()
if is_float:
corr_x.extend(logits.flatten())
corr_y.extend(label_ids.flatten())
tmp_eval_accuracy = accuracy(logits, label_ids)
eval_loss += tmp_eval_loss.mean().item()
eval_accuracy += tmp_eval_accuracy
nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1
# if (iter_idx + 1) % 1000 == 0 and export is not None:
# torch.save((iter_idx, logits_list), export)
iter_idx += 1
if export is not None:
torch.save(logits_list, export)
eval_loss = eval_loss / nb_eval_steps
eval_accuracy = eval_accuracy / nb_eval_examples
loss = tr_loss / nb_tr_steps if args.do_train else None
if is_float:
print(pearsonr(corr_x, corr_y))
print(spearmanr(corr_x, corr_y))
result = {
'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy,
'global_step': global_step,
'loss': loss
}
return result
local_rank = -1
parser = argparse.ArgumentParser()
parser.add_argument("--config", "-c", type=str, required=True)
args, _ = parser.parse_known_args()
options = argconf.options_from_json("confs/options.json")
config = argconf.config_from_json(args.config)
args = edict(argconf.parse_args(options, config))
print(f"Using config: {args}")
bv_utils.set_seed(args.seed)
args.do_train = args.do_train and not args.do_test_only
processors = {
"cola": ColaProcessor,
"mnli": MnliProcessor,
"mrpc": MrpcProcessor,
"sst2": SST2Processor,
'qnli': QnliProcessor,
'rte': RteProcessor,
"imdb": IMDBSentenceProcessor,
"qqp": QuoraProcessor,
"sts": STSProcessor,
"raw_sts_pair": RawSTSPairProcessor,
"raw_pair": RawPairProcessor
}
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
n_gpu = torch.cuda.device_count()
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
if not args.do_train and not args.do_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
# if os.path.exists(args.workspace) and os.listdir(args.workspace) and args.do_train:
# raise ValueError("Output directory ({}) already exists and is not empty.".format(args.workspace))
if not os.path.exists(args.workspace):
os.makedirs(args.workspace)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
num_labels = args.n_labels
label_list = processor.get_labels()
tokenizer = BertTokenizer.from_pretrained(args.model_file,
do_lower_case=args.uncased)
num_train_optimization_steps = None
train_examples = processor.get_train_examples(args.data_dir)
num_train_optimization_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
# Prepare model
cache_dir = os.path.join(PYTORCH_PRETRAINED_BERT_CACHE,
'distributed_{}'.format(local_rank))
model = BertForSequenceClassification.from_pretrained(
args.model_file, cache_dir=cache_dir, num_labels=num_labels)
if args.fp16:
model.half()
model.to(device)
# sd = torch.load('qqp.pt')
# sd = torch.load('sts.pt')
# del sd['classifier.weight']
# del sd['classifier.bias']
# model.load_state_dict(sd, strict=False)
if local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
param_optimizer = list(
filter(
lambda x: x[0] in
("module.classifier.weight", "module.classifier.bias"),
param_optimizer))
print(len(param_optimizer))
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
nb_tr_steps = 0
tr_loss = 0
train_features = convert_examples_to_features(train_examples, label_list,
args.max_seq_length,
tokenizer)
is_float = isinstance(train_features[0].label_id, float)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.float if is_float else torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
all_label_ids)
if local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
# BEGIN SST-2 -> QQP experiments
# END SST-2 -> QQP experiments
if args.do_train:
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
loss = model(input_ids,
segment_ids,
input_mask,
label_ids,
mse=is_float)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
output_model_file = os.path.join(args.workspace, WEIGHTS_NAME)
if args.do_train:
# Save a trained model and the associated configuration
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
torch.save(model_to_save.state_dict(), output_model_file)
output_config_file = os.path.join(args.workspace, CONFIG_NAME)
with open(output_config_file, 'w') as f:
f.write(model_to_save.config.to_json_string())
# Load a trained model and config that you have fine-tuned
config = BertConfig(output_config_file)
model = BertForSequenceClassification(config, num_labels=num_labels)
model.load_state_dict(torch.load(output_model_file))
elif args.do_test_only:
convert = bv_utils.convert_single_to_dp if isinstance(
model, torch.nn.DataParallel) else bv_utils.convert_dp_to_single
model.load_state_dict(convert(torch.load(output_model_file)))
else:
# pass
model = BertForSequenceClassification.from_pretrained(
args.model_file, num_labels=num_labels)
model.to(device)
if args.export:
model.eval()
train_dataloader = DataLoader(train_data,
batch_size=args.eval_batch_size,
shuffle=False)
with torch.no_grad():
evaluate(train_dataloader, export=args.export)
return
if args.visualize:
model.eval()
train_dataloader = DataLoader(train_data,
batch_size=args.eval_batch_size,
shuffle=False)
with open(os.path.join(args.workspace, "viz_results.csv"), "w") as f:
writer = None
dir_a = bv_viz.choose_random_dir(list(model.parameters()))
dir_b = bv_viz.choose_random_dir(list(model.parameters()))
torch.save(dir_a, os.path.join(args.workspace, "viz_dir_a.pt"))
torch.save(dir_b, os.path.join(args.workspace, "viz_dir_b.pt"))
for a, b in bv_viz.contour_2d(model, dir_a, dir_b):
result = evaluate(train_dataloader)
result["a"] = a
result["b"] = b
if writer is None:
writer = csv.DictWriter(f, fieldnames=result.keys())
writer.writeheader()
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.writerow(result)
if args.do_eval and (local_rank == -1
or torch.distributed.get_rank() == 0):
eval_examples = processor.get_test_examples(
args.data_dir
) if args.do_test_only else processor.get_dev_examples(args.data_dir)
eval_features = convert_examples_to_features(eval_examples, label_list,
args.max_seq_length,
tokenizer)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long)
all_label_ids = torch.tensor(
[f.label_id for f in eval_features],
dtype=torch.long
if isinstance(eval_features[0].label_id, int) else torch.float)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
model.eval()
result = evaluate(eval_dataloader)
output_eval_file = os.path.join(args.workspace, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))