def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, config_file,
pytorch_dump_path,
discriminator_or_generator):
# Initialise PyTorch model
config = ElectraConfig.from_json_file(config_file)
print("Building PyTorch model from configuration: {}".format(str(config)))
if discriminator_or_generator == "discriminator":
model = ElectraForPreTraining(config)
elif discriminator_or_generator == "generator":
model = ElectraForMaskedLM(config)
else:
raise ValueError(
"The discriminator_or_generator argument should be either 'discriminator' or 'generator'"
)
# Load weights from tf checkpoint
load_tf_weights_in_electra(
model,
config,
tf_checkpoint_path,
discriminator_or_generator=discriminator_or_generator)
# Save pytorch-model
print("Save PyTorch model to {}".format(pytorch_dump_path))
torch.save(model.state_dict(), pytorch_dump_path)
class ElectraForLanguageModelingModel(PreTrainedModel):
def __init__(self, config, **kwargs):
super(ElectraForLanguageModelingModel, self).__init__(config, **kwargs)
if "generator_config" in kwargs:
generator_config = kwargs["generator_config"]
else:
generator_config = config
self.generator_model = ElectraForMaskedLM(generator_config)
if "discriminator_config" in kwargs:
discriminator_config = kwargs["discriminator_config"]
else:
discriminator_config = config
self.discriminator_model = ElectraForPreTraining(discriminator_config)
self.vocab_size = generator_config.vocab_size
if kwargs.get("tie_generator_and_discriminator_embeddings", True):
self.tie_generator_and_discriminator_embeddings()
def tie_generator_and_discriminator_embeddings(self):
self.discriminator_model.set_input_embeddings(self.generator_model.get_input_embeddings())
def forward(self, inputs, masked_lm_labels, attention_mask=None, token_type_ids=None):
d_inputs = inputs.clone()
# run masked LM.
g_out = self.generator_model(
inputs, masked_lm_labels=masked_lm_labels, attention_mask=attention_mask, token_type_ids=token_type_ids
)
# get samples from masked LM.
sample_probs = torch.softmax(g_out[1], dim=-1, dtype=torch.float32)
sample_probs = sample_probs.view(-1, self.vocab_size)
sampled_tokens = torch.multinomial(sample_probs, 1).view(-1)
sampled_tokens = sampled_tokens.view(d_inputs.shape[0], -1)
# labels have a -100 value to mask out loss from unchanged tokens.
mask = masked_lm_labels.ne(-100)
# replace the masked out tokens of the input with the generator predictions.
d_inputs[mask] = sampled_tokens[mask]
# turn mask into new target labels. 1 (True) for corrupted, 0 otherwise.
# if the prediction was correct, mark it as uncorrupted.
correct_preds = sampled_tokens == masked_lm_labels
d_labels = mask.long()
d_labels[correct_preds] = 0
# run token classification, predict whether each token was corrupted.
d_out = self.discriminator_model(
d_inputs, labels=d_labels, attention_mask=attention_mask, token_type_ids=token_type_ids
)
g_loss = g_out[0]
d_loss = d_out[0]
g_scores = g_out[1]
d_scores = d_out[1]
return g_loss, d_loss, g_scores, d_scores, d_labels
def __init__(self, config, **kwargs):
super(ElectraForLanguageModelingModel, self).__init__(config, **kwargs)
if "generator_config" in kwargs:
generator_config = kwargs["generator_config"]
else:
generator_config = config
self.generator_model = ElectraForMaskedLM(generator_config)
if "discriminator_config" in kwargs:
discriminator_config = kwargs["discriminator_config"]
else:
discriminator_config = config
self.discriminator_model = ElectraForPreTraining(discriminator_config)
self.vocab_size = generator_config.vocab_size
if kwargs.get("tie_generator_and_discriminator_embeddings", True):
self.tie_generator_and_discriminator_embeddings()
def __init__(self, config: ElectraConfig, embeddings):
super().__init__()
self.embed_layer = nn.Embedding(num_embeddings=config.vocab_size,
embedding_dim=config.embedding_size,
padding_idx=config.vocab_size - 1)
self.embed_layer.weight = nn.Parameter(embeddings)
self.discriminator = ElectraForPreTraining(config)
self.sigmoid = nn.Sigmoid()
def main(train_cfg='config/electra_pretrain.json',
model_cfg='config/electra_small.json',
data_file='../tbc/books_large_all.txt',
model_file=None,
data_parallel=True,
vocab='../uncased_L-12_H-768_A-12/vocab.txt',
log_dir='../exp/electra/pretrain/runs',
save_dir='../exp/electra/pretrain',
max_len=128,
max_pred=20,
mask_prob=0.15,
quantize=False):
check_dirs_exist([log_dir, save_dir])
train_cfg = ElectraConfig().from_json_file(train_cfg)
model_cfg = ElectraConfig().from_json_file(model_cfg)
set_seeds(train_cfg.seed)
tokenizer = tokenization.FullTokenizer(vocab_file=vocab,
do_lower_case=True)
tokenize = lambda x: tokenizer.tokenize(tokenizer.convert_to_unicode(x))
pipeline = [
Preprocess4Pretrain(max_pred, mask_prob, list(tokenizer.vocab.keys()),
tokenizer.convert_tokens_to_ids, max_len)
]
data_iter = SentPairDataLoader(data_file,
train_cfg.batch_size,
tokenize,
max_len,
pipeline=pipeline)
# Get distilled-electra and quantized-distilled-electra
generator = ElectraForMaskedLM.from_pretrained(
'google/electra-small-generator')
t_discriminator = ElectraForPreTraining.from_pretrained(
'google/electra-base-discriminator')
s_discriminator = QuantizedElectraForPreTraining(
model_cfg) if quantize else ElectraForPreTraining
s_discriminator = s_discriminator.from_pretrained(
'google/electra-small-discriminator', config=model_cfg) # model
# config is used for model "QuantizedElectraForPreTraining"
model = DistillElectraForPreTraining(generator, t_discriminator,
s_discriminator, model_cfg)
optimizer = optim.optim4GPU(train_cfg, model)
writer = SummaryWriter(log_dir=log_dir) # for tensorboardX
base_trainer_args = (train_cfg, model_cfg, model, data_iter, None,
optimizer, save_dir, get_device())
trainer = QuantizedDistillElectraTrainer(writer, *base_trainer_args)
trainer.train(model_file, None, data_parallel)
trainer._eval()
def create_and_check_electra_for_pretraining(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
fake_token_labels,
):
config.num_labels = self.num_labels
model = ElectraForPreTraining(config=config)
model.to(torch_device)
model.eval()
loss, logits = model(input_ids,
attention_mask=input_mask,
token_type_ids=token_type_ids,
labels=fake_token_labels)
result = {
"loss": loss,
"logits": logits,
}
self.parent.assertListEqual(list(result["logits"].size()),
[self.batch_size, self.seq_length])
self.check_loss_output(result)
def __init__(self, config, **kwargs):
super(ElectraForLanguageModelingModel, self).__init__(config, **kwargs)
if "generator_config" in kwargs:
generator_config = kwargs["generator_config"]
else:
generator_config = config
self.generator_model = ElectraForMaskedLM(generator_config)
if "discriminator_config" in kwargs:
discriminator_config = kwargs["discriminator_config"]
else:
discriminator_config = config
self.discriminator_model = ElectraForPreTraining(discriminator_config)
self.vocab_size = config.vocab_size
def test_inference_no_head_absolute_embedding(self):
model = ElectraForPreTraining.from_pretrained(
"google/electra-small-discriminator")
input_ids = torch.tensor(
[[0, 345, 232, 328, 740, 140, 1695, 69, 6078, 1588, 2]])
output = model(input_ids)[0]
expected_shape = torch.Size((1, 11))
self.assertEqual(output.shape, expected_shape)
expected_slice = torch.tensor([[
-8.9253, -4.0305, -3.9306, -3.8774, -4.1873, -4.1280, 0.9429,
-4.1672, 0.9281, 0.0410, -3.4823
]])
self.assertTrue(torch.allclose(output, expected_slice, atol=1e-4))
def load_huggingface_weights_in_electra(model, config, pytorch_model_path):
try:
from transformers import ElectraForPreTraining
except ImportError:
raise ImportError(
"cannot import transformers, please install transformers first")
hf_model = ElectraForPreTraining.from_pretrained(pytorch_model_path)
model.electra.embeddings.token_embeddings.weight = (
hf_model.electra.embeddings.word_embeddings.weight)
model.electra.embeddings.position_embeddings.weight = (
hf_model.electra.embeddings.position_embeddings.weight)
model.electra.embeddings.token_type_embeddings.weight = (
hf_model.electra.embeddings.token_type_embeddings.weight)
model.electra.embeddings.layer_norm.weight = hf_model.electra.embeddings.LayerNorm.weight
model.electra.embeddings.layer_norm.bias = hf_model.electra.embeddings.LayerNorm.bias
if config.embedding_size != config.hidden_size:
model.electra.embeddings_project.weight = hf_model.electra.embeddings_project.weight
for layer_idx in range(config.num_hidden_layers):
layer = model.electra.encoder.layers[layer_idx]
hf_layer = hf_model.electra.encoder.layer[layer_idx]
layer.self.query.weight = hf_layer.attention.self.query.weight
layer.self.query.bias = hf_layer.attention.self.query.bias
layer.self.key.weight = hf_layer.attention.self.key.weight
layer.self.key.bias = hf_layer.attention.self.key.bias
layer.self.value.weight = hf_layer.attention.self.value.weight
layer.self.value.bias = hf_layer.attention.self.value.bias
layer.self.dense.weight = hf_layer.attention.output.dense.weight
layer.self.dense.bias = hf_layer.attention.output.dense.bias
layer.feed_forward.intermediate.weight = hf_layer.intermediate.dense.weight
layer.feed_forward.intermediate.bias = hf_layer.intermediate.dense.bias
layer.feed_forward.output.weight = hf_layer.output.dense.weight
layer.feed_forward.output.bias = hf_layer.output.dense.bias
layer.add_norm[
0].layer_norm.weight = hf_layer.attention.output.LayerNorm.weight
layer.add_norm[
0].layer_norm.bias = hf_layer.attention.output.LayerNorm.bias
layer.add_norm[1].layer_norm.weight = hf_layer.output.LayerNorm.weight
layer.add_norm[1].layer_norm.bias = hf_layer.output.LayerNorm.bias
def detect_error_demo():
tokenizer = ElectraTokenizer.from_pretrained(D_model_dir)
discriminator = ElectraForPreTraining.from_pretrained(D_model_dir)
sentence = '今天新情很好'
fake_tokens = tokenizer.tokenize(sentence)
print(fake_tokens)
fake_inputs = tokenizer.encode(sentence, return_tensors="pt")
discriminator_outputs = discriminator(fake_inputs)
predictions = torch.round((torch.sign(discriminator_outputs[0]) + 1) / 2)
print(list(zip(fake_tokens, predictions.tolist())))
print("fixed " + '*' * 42)
print(list(zip(fake_tokens, predictions.tolist()[1:-1])))
def __init__(self, d_mdel_dir=os.path.join(pwd_path,
"../data/electra_models/chinese_electra_base_discriminator_pytorch/"),
g_model_dir=os.path.join(pwd_path,
"../data/electra_models/chinese_electra_base_generator_pytorch/"),
):
super(ElectraCorrector, self).__init__()
self.name = 'electra_corrector'
t1 = time.time()
self.g_model = pipeline("fill-mask",
model=g_model_dir,
tokenizer=g_model_dir
)
self.d_model = ElectraForPreTraining.from_pretrained(d_mdel_dir)
if self.g_model:
self.mask = self.g_model.tokenizer.mask_token
logger.debug('Loaded electra model: %s, spend: %.3f s.' % (g_model_dir, time.time() - t1))
def __init__(self,
d_model_dir=config.electra_D_model_dir,
g_model_dir=config.electra_G_model_dir,
device=device_id):
super(ElectraCorrector, self).__init__()
self.name = 'electra_corrector'
t1 = time.time()
self.g_model = pipeline(
"fill-mask",
model=g_model_dir,
tokenizer=g_model_dir,
device=device, # gpu device id
)
self.d_model = ElectraForPreTraining.from_pretrained(d_model_dir)
if self.g_model:
self.mask = self.g_model.tokenizer.mask_token
logger.debug('Loaded electra model: %s, spend: %.3f s.' %
(g_model_dir, time.time() - t1))
def create_and_check_electra_for_pretraining(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
fake_token_labels,
):
config.num_labels = self.num_labels
model = ElectraForPreTraining(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=fake_token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length))
disc_pred = disc_logits > 0
return gen_pred, generated, disc_pred, is_replaced
# %% [markdown]
# # 5. Train
# %%
# Generator and Discriminator
if c.my_model:
gen_hparam['tie_in_out_embedding'] = c.tie_gen_in_out_embedding
generator = ModelForGenerator(gen_hparam)
discriminator = ModelForDiscriminator(disc_hparam)
discriminator.electra.embedding = generator.electra.embedding
else:
generator = ElectraForMaskedLM(gen_config)
discriminator = ElectraForPreTraining(disc_config)
discriminator.electra.embeddings = generator.electra.embeddings
if c.tie_gen_in_out_embedding:
generator.generator_predictions.dense.weight = generator.electra.embeddings.word_embeddings.weight
# ELECTRA training loop
electra_model = ELECTRAModel(generator, discriminator, hf_tokenizer)
electra_loss_func = ELECTRALoss(gen_label_smooth=c.gen_smooth_label, disc_label_smooth=c.disc_smooth_label)
# jit (Haven't fiqured out how to make it work)
# input_ids, sentA_lenths = dls.one_batch()
# masked_inputs, labels, is_mlm_applied = mlm_cb.mask_tokens(input_ids)
# electra_jit_model = torch.jit.trace(electra_model, (masked_inputs, sentA_lenths, is_mlm_applied, labels))
# Optimizer
if c.adam_bias_correction: opt_func = partial(Adam, eps=1e-6, mom=0.9, sqr_mom=0.999, wd=0.01)
def __init__(
self,
model_type,
model_name,
generator_name=None,
discriminator_name=None,
train_files=None,
args=None,
use_cuda=True,
cuda_device=-1,
**kwargs,
):
"""
Initializes a LanguageModelingModel.
Args:
model_type: The type of model (gpt2, openai-gpt, bert, roberta, distilbert, camembert)
model_name: Default Transformer model name or path to a directory containing Transformer model file (pytorch_nodel.bin).
generator_name (optional): A pretrained model name or path to a directory containing an ELECTRA generator model.
discriminator_name (optional): A pretrained model name or path to a directory containing an ELECTRA discriminator model.
args (optional): Default args will be used if this parameter is not provided. If provided, it should be a dict containing the args that should be changed in the default args.
train_files (optional): List of files to be used when training the tokenizer.
use_cuda (optional): Use GPU if available. Setting to False will force model to use CPU only.
cuda_device (optional): Specific GPU that should be used. Will use the first available GPU by default.
**kwargs (optional): For providing proxies, force_download, resume_download, cache_dir and other options specific to the 'from_pretrained' implementation where this will be supplied.
""" # noqa: ignore flake8"
if args and "manual_seed" in args:
random.seed(args["manual_seed"])
np.random.seed(args["manual_seed"])
torch.manual_seed(args["manual_seed"])
if "n_gpu" in args and args["n_gpu"] > 0:
torch.cuda.manual_seed_all(args["manual_seed"])
if use_cuda:
if torch.cuda.is_available():
if cuda_device == -1:
self.device = torch.device("cuda")
else:
self.device = torch.device(f"cuda:{cuda_device}")
else:
raise ValueError(
"'use_cuda' set to True when cuda is unavailable."
" Make sure CUDA is available or set use_cuda=False."
)
else:
self.device = "cpu"
self.results = {}
self.args = {
"dataset_type": "None",
"dataset_class": None,
"custom_tokenizer": None,
"block_size": -1,
"mlm": True,
"mlm_probability": 0.15,
"max_steps": -1,
"config_name": None,
"tokenizer_name": None,
"min_frequency": 2,
"special_tokens": ["<s>", "<pad>", "</s>", "<unk>", "<mask>"],
"sliding_window": False,
"stride": 0.8,
"generator_config": {},
"discriminator_config": {},
"vocab_size": None,
}
self.args.update(global_args)
if not use_cuda:
self.args["fp16"] = False
if args:
self.args.update(args)
self.args["model_name"] = model_name
self.args["model_type"] = model_type
config_class, model_class, tokenizer_class = MODEL_CLASSES[model_type]
self.tokenizer_class = tokenizer_class
new_tokenizer = False
if self.args["tokenizer_name"]:
self.tokenizer = tokenizer_class.from_pretrained(
self.args["tokenizer_name"], cache_dir=self.args["cache_dir"]
)
elif self.args["model_name"]:
self.tokenizer = tokenizer_class.from_pretrained(model_name, cache_dir=self.args["cache_dir"], **kwargs)
self.args["tokenizer_name"] = self.args["model_name"]
else:
if not train_files:
raise ValueError(
"model_name and tokenizer_name are not specified."
"You must specify train_files to train a Tokenizer."
)
else:
self.train_tokenizer(train_files)
new_tokenizer = True
if self.args["config_name"]:
self.config = config_class.from_pretrained(self.args["config_name"], cache_dir=self.args["cache_dir"])
elif self.args["model_name"]:
self.config = config_class.from_pretrained(model_name, cache_dir=self.args["cache_dir"], **kwargs)
else:
self.config = config_class(**self.args["config"], **kwargs)
if self.args["vocab_size"]:
self.config.vocab_size = self.args["vocab_size"]
if new_tokenizer:
self.config.vocab_size = len(self.tokenizer)
if self.args["model_type"] == "electra":
if generator_name:
self.generator_config = ElectraConfig.from_pretrained(generator_name)
elif self.args["model_name"]:
self.generator_config = ElectraConfig.from_pretrained(
os.path.join(self.args["model_name"], "generator_config"), **kwargs,
)
else:
self.generator_config = ElectraConfig(**self.args["generator_config"], **kwargs)
if new_tokenizer:
self.generator_config.vocab_size = len(self.tokenizer)
if discriminator_name:
self.discriminator_config = ElectraConfig.from_pretrained(discriminator_name)
elif self.args["model_name"]:
self.discriminator_config = ElectraConfig.from_pretrained(
os.path.join(self.args["model_name"], "discriminator_config"), **kwargs,
)
else:
self.discriminator_config = ElectraConfig(**self.args["discriminator_config"], **kwargs)
if new_tokenizer:
self.discriminator_config.vocab_size = len(self.tokenizer)
if self.args["block_size"] <= 0:
self.args["block_size"] = min(self.args["max_seq_length"], self.tokenizer.max_len)
else:
self.args["block_size"] = min(self.args["block_size"], self.tokenizer.max_len, self.args["max_seq_length"])
if self.args["model_name"]:
if self.args["model_type"] == "electra":
self.model = model_class.from_pretrained(
model_name,
config=self.config,
cache_dir=self.args["cache_dir"],
generator_config=self.generator_config,
discriminator_config=self.discriminator_config,
**kwargs,
)
self.model.load_state_dict(torch.load(os.path.join(self.args["model_name"], "pytorch_model.bin")))
else:
self.model = model_class.from_pretrained(
model_name, config=self.config, cache_dir=self.args["cache_dir"], **kwargs,
)
else:
logger.info(" Training language model from scratch")
if self.args["model_type"] == "electra":
generator_model = ElectraForMaskedLM(config=self.generator_config)
discriminator_model = ElectraForPreTraining(config=self.discriminator_config)
self.model = ElectraForLanguageModelingModel(
config=self.config,
generator_model=generator_model,
discriminator_model=discriminator_model,
generator_config=self.generator_config,
discriminator_config=self.discriminator_config,
)
model_to_resize = (
self.model.generator_model.module
if hasattr(self.model.generator_model, "module")
else self.model.generator_model
)
model_to_resize.resize_token_embeddings(len(self.tokenizer))
model_to_resize = (
self.model.discriminator_model.module
if hasattr(self.model.discriminator_model, "module")
else self.model.discriminator_model
)
model_to_resize.resize_token_embeddings(len(self.tokenizer))
else:
self.model = model_class(config=self.config)
model_to_resize = self.model.module if hasattr(self.model, "module") else self.model
model_to_resize.resize_token_embeddings(len(self.tokenizer))
if model_type in ["camembert", "xlmroberta"]:
warnings.warn(
f"use_multiprocessing automatically disabled as {model_type}"
" fails when using multiprocessing for feature conversion."
)
self.args["use_multiprocessing"] = False
if self.args["wandb_project"] and not wandb_available:
warnings.warn("wandb_project specified but wandb is not available. Wandb disabled.")
self.args["wandb_project"] = None
import torch
from fio import read
from transformers import ElectraTokenizer, ElectraForPreTraining
"""
python ppl.py [hypothesis file]
"""
hyp_file = sys.argv[1]
assert os.path.exists(hyp_file)
device = torch.device("cuda:0")
tokenizer = ElectraTokenizer.from_pretrained(
'google/electra-small-discriminator')
model = ElectraForPreTraining.from_pretrained(
'google/electra-small-discriminator')
model.to(device)
model.eval()
sigmoid = torch.nn.Sigmoid()
pad_id = tokenizer.pad_token_id
src_tsf_paris = [line.split("\t")[:2] for line in read(hyp_file)]
src_tsf_paris = [(tokenizer.encode(src), tokenizer.encode(tsf))
for src, tsf in src_tsf_paris]
def format_samples(pairs):
src_bat, tsf_bat = zip(*pairs)
max_l_src, max_l_tsf = max([len(src) for src in src_bat
]), max([len(tsf) for tsf in tsf_bat])
def __init__(
self,
model_type,
model_name,
generator_name=None,
discriminator_name=None,
train_files=None,
args=None,
use_cuda=True,
cuda_device=-1,
**kwargs,
):
"""
Initializes a LanguageModelingModel.
Args:
model_type: The type of model (gpt2, openai-gpt, bert, roberta, distilbert, camembert)
model_name: Default Transformer model name or path to a directory containing Transformer model file (pytorch_nodel.bin).
generator_name (optional): A pretrained model name or path to a directory containing an ELECTRA generator model.
discriminator_name (optional): A pretrained model name or path to a directory containing an ELECTRA discriminator model.
args (optional): Default args will be used if this parameter is not provided. If provided, it should be a dict containing the args that should be changed in the default args.
train_files (optional): List of files to be used when training the tokenizer.
use_cuda (optional): Use GPU if available. Setting to False will force model to use CPU only.
cuda_device (optional): Specific GPU that should be used. Will use the first available GPU by default.
**kwargs (optional): For providing proxies, force_download, resume_download, cache_dir and other options specific to the 'from_pretrained' implementation where this will be supplied.
""" # noqa: ignore flake8"
self.args = self._load_model_args(model_name)
if isinstance(args, dict):
self.args.update_from_dict(args)
elif isinstance(args, LanguageModelingArgs):
self.args = args
if "sweep_config" in kwargs:
sweep_config = kwargs.pop("sweep_config")
sweep_values = {key: value["value"] for key, value in sweep_config.as_dict().items() if key != "_wandb"}
self.args.update_from_dict(sweep_values)
if self.args.manual_seed:
random.seed(self.args.manual_seed)
np.random.seed(self.args.manual_seed)
torch.manual_seed(self.args.manual_seed)
if self.args.n_gpu > 0:
torch.cuda.manual_seed_all(self.args.manual_seed)
if self.args.local_rank != -1:
logger.info(f"local_rank: {self.args.local_rank}")
torch.distributed.init_process_group(backend="nccl")
cuda_device = self.args.local_rank
if use_cuda:
if torch.cuda.is_available():
if cuda_device == -1:
self.device = torch.device("cuda")
else:
self.device = torch.device(f"cuda:{cuda_device}")
else:
raise ValueError(
"'use_cuda' set to True when cuda is unavailable."
" Make sure CUDA is available or set use_cuda=False."
)
else:
self.device = "cpu"
self.results = {}
if not use_cuda:
self.args.fp16 = False
self.args.model_name = model_name
self.args.model_type = model_type
config_class, model_class, tokenizer_class = MODEL_CLASSES[model_type]
self.tokenizer_class = tokenizer_class
new_tokenizer = False
if self.args.tokenizer_name:
self.tokenizer = tokenizer_class.from_pretrained(self.args.tokenizer_name, cache_dir=self.args.cache_dir)
elif self.args.model_name:
if self.args.model_name == "electra":
self.tokenizer = tokenizer_class.from_pretrained(
generator_name, cache_dir=self.args.cache_dir, **kwargs
)
self.args.tokenizer_name = self.args.model_name
else:
self.tokenizer = tokenizer_class.from_pretrained(model_name, cache_dir=self.args.cache_dir, **kwargs)
self.args.tokenizer_name = self.args.model_name
else:
if not train_files:
raise ValueError(
"model_name and tokenizer_name are not specified."
"You must specify train_files to train a Tokenizer."
)
else:
self.train_tokenizer(train_files)
new_tokenizer = True
if self.args.config_name:
self.config = config_class.from_pretrained(self.args.config_name, cache_dir=self.args.cache_dir)
elif self.args.model_name and self.args.model_name != "electra":
self.config = config_class.from_pretrained(model_name, cache_dir=self.args.cache_dir, **kwargs)
else:
self.config = config_class(**self.args.config, **kwargs)
if self.args.vocab_size:
self.config.vocab_size = self.args.vocab_size
if new_tokenizer:
self.config.vocab_size = len(self.tokenizer)
if self.args.model_type == "electra":
if generator_name:
self.generator_config = ElectraConfig.from_pretrained(generator_name)
elif self.args.model_name:
self.generator_config = ElectraConfig.from_pretrained(
os.path.join(self.args.model_name, "generator_config"), **kwargs,
)
else:
self.generator_config = ElectraConfig(**self.args.generator_config, **kwargs)
if new_tokenizer:
self.generator_config.vocab_size = len(self.tokenizer)
if discriminator_name:
self.discriminator_config = ElectraConfig.from_pretrained(discriminator_name)
elif self.args.model_name:
self.discriminator_config = ElectraConfig.from_pretrained(
os.path.join(self.args.model_name, "discriminator_config"), **kwargs,
)
else:
self.discriminator_config = ElectraConfig(**self.args.discriminator_config, **kwargs)
if new_tokenizer:
self.discriminator_config.vocab_size = len(self.tokenizer)
if self.args.block_size <= 0:
self.args.block_size = min(self.args.max_seq_length, self.tokenizer.max_len)
else:
self.args.block_size = min(self.args.block_size, self.tokenizer.max_len, self.args.max_seq_length)
if self.args.model_name:
if self.args.model_type == "electra":
if self.args.model_name == "electra":
generator_model = ElectraForMaskedLM.from_pretrained(generator_name)
discriminator_model = ElectraForPreTraining.from_pretrained(discriminator_name)
self.model = ElectraForLanguageModelingModel(
config=self.config,
generator_model=generator_model,
discriminator_model=discriminator_model,
generator_config=self.generator_config,
discriminator_config=self.discriminator_config,
tie_generator_and_discriminator_embeddings=self.args.tie_generator_and_discriminator_embeddings,
)
model_to_resize = (
self.model.generator_model.module
if hasattr(self.model.generator_model, "module")
else self.model.generator_model
)
model_to_resize.resize_token_embeddings(len(self.tokenizer))
model_to_resize = (
self.model.discriminator_model.module
if hasattr(self.model.discriminator_model, "module")
else self.model.discriminator_model
)
model_to_resize.resize_token_embeddings(len(self.tokenizer))
self.model.generator_model = generator_model
self.model.discriminator_model = discriminator_model
else:
self.model = model_class.from_pretrained(
model_name,
config=self.config,
cache_dir=self.args.cache_dir,
generator_config=self.generator_config,
discriminator_config=self.discriminator_config,
**kwargs,
)
self.model.load_state_dict(torch.load(os.path.join(self.args.model_name, "pytorch_model.bin")))
else:
self.model = model_class.from_pretrained(
model_name, config=self.config, cache_dir=self.args.cache_dir, **kwargs,
)
else:
logger.info(" Training language model from scratch")
if self.args.model_type == "electra":
generator_model = ElectraForMaskedLM(config=self.generator_config)
discriminator_model = ElectraForPreTraining(config=self.discriminator_config)
self.model = ElectraForLanguageModelingModel(
config=self.config,
generator_model=generator_model,
discriminator_model=discriminator_model,
generator_config=self.generator_config,
discriminator_config=self.discriminator_config,
tie_generator_and_discriminator_embeddings=self.args.tie_generator_and_discriminator_embeddings,
)
model_to_resize = (
self.model.generator_model.module
if hasattr(self.model.generator_model, "module")
else self.model.generator_model
)
model_to_resize.resize_token_embeddings(len(self.tokenizer))
model_to_resize = (
self.model.discriminator_model.module
if hasattr(self.model.discriminator_model, "module")
else self.model.discriminator_model
)
model_to_resize.resize_token_embeddings(len(self.tokenizer))
else:
self.model = model_class(config=self.config)
model_to_resize = self.model.module if hasattr(self.model, "module") else self.model
model_to_resize.resize_token_embeddings(len(self.tokenizer))
if model_type in ["camembert", "xlmroberta"]:
warnings.warn(
f"use_multiprocessing automatically disabled as {model_type}"
" fails when using multiprocessing for feature conversion."
)
self.args.use_multiprocessing = False
if self.args.wandb_project and not wandb_available:
warnings.warn("wandb_project specified but wandb is not available. Wandb disabled.")
self.args.wandb_project = None
def get_glue_learner(task, run_name=None, inference=False):
# Num_epochs
if task in ['rte', 'stsb']: num_epochs = 10
else: num_epochs = 3
# Dataloaders
dls = glue_dls[task]
if isinstance(c.device, str): dls.to(torch.device(c.device))
elif isinstance(c.device, list): dls.to(torch.device('cuda', c.device[0]))
else: dls.to(torch.device('cuda:0'))
# Load pretrained model
if not c.pretrained_checkpoint:
discriminator = ElectraForPreTraining.from_pretrained(f"google/electra-{c.size}-discriminator")
else:
discriminator = ModelForDiscriminator(hparam) if c.my_model else ElectraForPreTraining(electra_config)
load_part_model(c.pretrained_ckp_path, discriminator, 'discriminator')
# Create finetuning model
if task=='wnli' and c.wsc_trick:
model = ELECTRAWSCTrickModel(discriminator, hf_tokenizer.pad_token_id)
else:
model = SentencePredictor(discriminator.electra, electra_config.hidden_size, num_class=NUM_CLASS[task])
# Discriminative learning rates
splitter = partial( hf_electra_param_splitter, wsc_trick=(task=='wnli' and c.wsc_trick) )
layer_lrs = get_layer_lrs(lr=c.lr,
decay_rate=c.layer_lr_decay,
num_hidden_layers=electra_config.num_hidden_layers,)
# Optimizer
if c.adam_bias_correction: opt_func = partial(Adam, eps=1e-6, mom=0.9, sqr_mom=0.999, wd=0.01)
else: opt_func = partial(Adam_no_bias_correction, eps=1e-6, mom=0.9, sqr_mom=0.999, wd=0.01)
# Learner
learn = Learner(dls, model,
loss_func=LOSS_FUNC[task],
opt_func=opt_func,
metrics=[eval(f'{metric}()') for metric in METRICS[task]],
splitter=splitter if not inference else trainable_params,
lr=layer_lrs if not inference else defaults.lr,
path='./checkpoints',
model_dir='glue',)
# Multi gpu
if isinstance(c.device, list) or c.device is None:
learn.model = nn.DataParallel(learn.model, device_ids=c.device)
# Gradient clip
learn.add_cb(GradientClipping(1.0))
# Logging
if run_name and not inference:
neptune.create_experiment(name=run_name, params={'task':task, **c, **hparam_update})
learn.add_cb(SimplerNeptuneCallback(False))
# Learning rate schedule
if c.schedule == 'one_cycle':
return learn, partial(learn.fit_one_cycle, n_epoch=num_epochs, lr_max=layer_lrs)
elif c.schedule == 'adjusted_one_cycle':
return learn, partial(learn.fit_one_cycle, n_epoch=num_epochs, lr_max=layer_lrs, div=1e5, pct_start=0.1)
else:
lr_shed_func = linear_warmup_and_then_decay if c.schedule=='separate_linear' else linear_warmup_and_decay
lr_shedule = ParamScheduler({'lr': partial(lr_shed_func,
lr_max=np.array(layer_lrs),
warmup_pct=0.1,
total_steps=num_epochs*(len(dls.train)))})
return learn, partial(learn.fit, n_epoch=num_epochs, cbs=[lr_shedule])
def get_glue_learner(task, run_name=None, inference=False):
is_wsc_trick = task == "wnli" and c.wsc_trick
# Num_epochs
if task in ["rte", "stsb"]:
num_epochs = 10
else:
num_epochs = 3
# Dataloaders
dls = glue_dls[task]
if isinstance(c.device, str):
dls.to(torch.device(c.device))
elif isinstance(c.device, list):
dls.to(torch.device("cuda", c.device[0]))
else:
dls.to(torch.device("cuda:0"))
# Load pretrained model
if not c.pretrained_checkpoint:
discriminator = ElectraForPreTraining.from_pretrained(
f"google/electra-{c.size}-discriminator")
else:
discriminator = (ModelForDiscriminator(hparam) if c.my_model else
ElectraForPreTraining(electra_config))
load_part_model(c.pretrained_ckp_path, discriminator, "discriminator")
# Seeds & PyTorch benchmark
torch.backends.cudnn.benchmark = True
if c.seeds:
dls[0].rng = random.Random(c.seeds[i]) # for fastai dataloader
random.seed(c.seeds[i])
np.random.seed(c.seeds[i])
torch.manual_seed(c.seeds[i])
# Create finetuning model
if is_wsc_trick:
model = ELECTRAWSCTrickModel(discriminator, hf_tokenizer.pad_token_id)
else:
model = SentencePredictor(discriminator.electra,
electra_config.hidden_size,
num_class=NUM_CLASS[task])
# Discriminative learning rates
splitter = partial(hf_electra_param_splitter, wsc_trick=is_wsc_trick)
layer_lrs = get_layer_lrs(
lr=c.lr,
decay_rate=c.layer_lr_decay,
num_hidden_layers=electra_config.num_hidden_layers,
)
# Optimizer
if c.adam_bias_correction:
opt_func = partial(Adam,
eps=1e-6,
mom=0.9,
sqr_mom=0.999,
wd=c.weight_decay)
else:
opt_func = partial(Adam_no_bias_correction,
eps=1e-6,
mom=0.9,
sqr_mom=0.999,
wd=c.weight_decay)
# Learner
learn = Learner(
dls,
model,
loss_func=LOSS_FUNC[task],
opt_func=opt_func,
metrics=METRICS[task],
splitter=splitter if not inference else trainable_params,
lr=layer_lrs if not inference else defaults.lr,
path="./checkpoints/glue",
model_dir=c.group_name,
)
# Multi gpu
if isinstance(c.device, list) or c.device is None:
learn.create_opt()
learn.model = nn.DataParallel(learn.model, device_ids=c.device)
# Mixed precision
learn.to_native_fp16(init_scale=2.0**14)
# Gradient clip
learn.add_cb(GradientClipping(1.0))
# Logging
# Logging
if run_name and not inference:
if c.logger == "neptune":
neptune.create_experiment(name=run_name,
params={
"task": task,
**c,
**hparam_update
})
learn.add_cb(LightNeptuneCallback(False))
elif c.logger == "wandb":
wandb_run = wandb.init(
name=run_name,
project="electra_glue",
config={
"task": task,
**c,
**hparam_update
},
reinit=True,
)
learn.add_cb(LightWandbCallback(wandb_run))
# Learning rate schedule
if c.schedule == "one_cycle":
return learn, partial(learn.fit_one_cycle,
n_epoch=num_epochs,
lr_max=layer_lrs)
elif c.schedule == "adjusted_one_cycle":
return learn, partial(
learn.fit_one_cycle,
n_epoch=num_epochs,
lr_max=layer_lrs,
div=1e5,
pct_start=0.1,
)
else:
lr_shed_func = (linear_warmup_and_then_decay if c.schedule
== "separate_linear" else linear_warmup_and_decay)
lr_shedule = ParamScheduler({
"lr":
partial(
lr_shed_func,
lr_max=np.array(layer_lrs),
warmup_pct=0.1,
total_steps=num_epochs * (len(dls.train)),
)
})
return learn, partial(learn.fit, n_epoch=num_epochs, cbs=[lr_shedule])
def train(rank, args):
#######################
## distributed
if args.distributed_enabled:
torch.distributed.init_process_group(
backend='nccl',
init_method='env://',
world_size=args.distributed_world_size,
rank=rank)
if args.gpu_enabled:
device = torch.device('cuda:{}'.format(rank))
else:
device = torch.device('cpu')
is_master = True if not args.distributed_enabled else args.distributed_enabled and rank == 0
#######################
## preamble
set_gpus(rank)
set_seed(rank)
set_cuda(deterministic=args.gpu_deterministic)
output_dir = f'{args.output_dir}/{rank}'
os.makedirs(output_dir, exist_ok=False)
setup_logging(filename=f'{output_dir}/output.log', console=is_master)
#######################
## dataset
tokenizer = new_tokenizer(vocab_file=args.data_vocab_file)
vocab_size = len(tokenizer.vocab)
ds_train = wrap_example_builder(
dataset=load_owt(owt_dir=args.data_dir,
n_tensors_per_file=args.data_n_tensors_per_file),
vocab=tokenizer.vocab,
max_length=args.data_max_seq_length)
pad_token_id = tokenizer.vocab['[PAD]']
mask_token_id = tokenizer.vocab['[MASK]']
cls_token_id = tokenizer.vocab['[CLS]']
sep_token_id = tokenizer.vocab['[SEP]']
assert pad_token_id == 0
assert cls_token_id == 101
assert sep_token_id == 102
assert mask_token_id == 103
def collate_batch(examples):
input_ids = torch.nn.utils.rnn.pad_sequence(
[example['input_ids'] for example in examples],
batch_first=True,
padding_value=pad_token_id)
input_mask = torch.nn.utils.rnn.pad_sequence(
[example['input_mask'] for example in examples],
batch_first=True,
padding_value=pad_token_id)
segment_ids = torch.nn.utils.rnn.pad_sequence(
[example['segment_ids'] for example in examples],
batch_first=True,
padding_value=pad_token_id)
return input_ids, input_mask, segment_ids
def cycle(iterable):
while True:
for x in iterable:
yield x
ds_train_loader = iter(
cycle(
DataLoader(ds_train,
batch_size=args.opt_batch_size,
collate_fn=collate_batch)))
#######################
## model
def to_distributed_model(model):
return model if not args.distributed_enabled else torch.nn.parallel.DistributedDataParallel(
model, device_ids=[rank], find_unused_parameters=True)
def tie_weights(generator, discriminator):
generator.electra.embeddings.word_embeddings = discriminator.electra.embeddings.word_embeddings
generator.electra.embeddings.position_embeddings = discriminator.electra.embeddings.position_embeddings
generator.electra.embeddings.token_type_embeddings = discriminator.electra.embeddings.token_type_embeddings
class LogitsAdapter(torch.nn.Module):
def __init__(self, adaptee):
super().__init__()
self.adaptee = adaptee
def forward(self, *args, **kwargs):
return self.adaptee(*args, **kwargs)[0]
from transformers import AutoConfig, ElectraForMaskedLM, ElectraForPreTraining
generator = ElectraForMaskedLM(
AutoConfig.from_pretrained(args.model_generator))
discriminator = ElectraForPreTraining(
AutoConfig.from_pretrained(args.model_discriminator))
tie_weights(generator, discriminator)
model = to_distributed_model(
Electra(LogitsAdapter(generator),
LogitsAdapter(discriminator),
num_tokens=vocab_size,
mask_token_id=mask_token_id,
pad_token_id=pad_token_id,
mask_prob=args.model_mask_prob,
mask_ignore_token_ids=[
tokenizer.vocab['[CLS]'], tokenizer.vocab['[SEP]']
],
random_token_prob=0.0).to(device))
#######################
## optimizer
def get_linear_schedule_with_warmup(optimizer,
num_warmup_steps,
num_training_steps,
last_epoch=-1):
def lr_lambda(current_step):
learning_rate = max(
0.0, 1. - (float(current_step) / float(num_training_steps)))
learning_rate *= min(1.0,
float(current_step) / float(num_warmup_steps))
return learning_rate
return LambdaLR(optimizer, lr_lambda, last_epoch)
def get_params_without_weight_decay_ln(named_params, weight_decay):
no_decay = ['bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{
'params': [
p for n, p in named_params
if not any(nd in n for nd in no_decay)
],
'weight_decay':
weight_decay,
},
{
'params': [
p for n, p in named_params
if any(nd in n for nd in no_decay)
],
'weight_decay':
0.0,
},
]
return optimizer_grouped_parameters
optimizer = torch.optim.AdamW(get_params_without_weight_decay_ln(
model.named_parameters(), weight_decay=0.1),
lr=args.opt_lr,
betas=(0.9, 0.999),
eps=1e-08)
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=args.opt_warmup_steps,
num_training_steps=args.opt_num_training_steps)
scaler = torch.cuda.amp.GradScaler(enabled=args.gpu_mixed_precision)
#######################
## train
t, steps_s, eta_m = time(), 0., 0
for step in range(args.opt_num_training_steps + 1):
input_ids, input_mask, segment_ids = next(ds_train_loader)
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
assert input_ids.shape[1] <= args.data_max_seq_length
optimizer.zero_grad()
with torch.cuda.amp.autocast(enabled=args.gpu_mixed_precision):
loss, loss_mlm, loss_disc, acc_gen, acc_disc, disc_labels, disc_pred = model(
input_ids,
attention_mask=input_mask,
token_type_ids=segment_ids)
scaler.scale(loss).backward()
scaler.unscale_(optimizer)
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
scaler.step(optimizer)
scaler.update()
scheduler.step()
metrics = {
'step': (step, '{:8d}'),
'loss': (loss.item(), '{:8.5f}'),
'loss_mlm': (loss_mlm.item(), '{:8.5f}'),
'loss_disc': (loss_disc.item(), '{:8.5f}'),
'acc_gen': (acc_gen.item(), '{:5.3f}'),
'acc_disc': (acc_disc.item(), '{:5.3f}'),
'lr': (scheduler.get_last_lr()[0], '{:8.7f}'),
'steps': (steps_s, '{:4.1f}/s'),
'eta': (eta_m, '{:4d}m'),
}
if step % args.step_log == 0:
sep = ' ' * 2
logger.info(
sep.join([
f'{k}: {v[1].format(v[0])}' for (k, v) in metrics.items()
]))
if step > 0 and step % 100 == 0:
t2 = time()
steps_s = 100. / (t2 - t)
eta_m = int(((args.opt_num_training_steps - step) / steps_s) // 60)
t = t2
if step % 200 == 0:
logger.info(
np.array2string(disc_labels[0].cpu().numpy(),
threshold=sys.maxsize,
max_line_width=sys.maxsize))
logger.info(
np.array2string(disc_pred[0].cpu().numpy(),
threshold=sys.maxsize,
max_line_width=sys.maxsize))
if step > 0 and step % args.step_ckpt == 0 and is_master:
discriminator.electra.save_pretrained(
f'{args.output_dir}/ckpt/{step}')
