def test_mbart_fast_forward(self):
config = MBartConfig(
vocab_size=99,
d_model=24,
encoder_layers=2,
decoder_layers=2,
encoder_attention_heads=2,
decoder_attention_heads=2,
encoder_ffn_dim=32,
decoder_ffn_dim=32,
max_position_embeddings=48,
add_final_layer_norm=True,
return_dict=True,
)
lm_model = MBartForConditionalGeneration(config).to(torch_device)
context = torch.Tensor([[71, 82, 18, 33, 46, 91, 2],
[68, 34, 26, 58, 30, 2,
1]]).long().to(torch_device)
summary = torch.Tensor([[82, 71, 82, 18, 2],
[58, 68, 2, 1, 1]]).long().to(torch_device)
result = lm_model(input_ids=context,
decoder_input_ids=summary,
labels=summary)
expected_shape = (*summary.shape, config.vocab_size)
self.assertEqual(result.logits.shape, expected_shape)
def prepare_config_and_inputs(self):
input_ids = np.clip(
ids_tensor([self.batch_size, self.seq_length - 1],
self.vocab_size), 3, self.vocab_size)
input_ids = np.concatenate((input_ids, 2 * np.ones(
(self.batch_size, 1), dtype=np.int64)), -1)
decoder_input_ids = shift_tokens_right(input_ids, 1)
config = MBartConfig(
vocab_size=self.vocab_size,
d_model=self.hidden_size,
encoder_layers=self.num_hidden_layers,
decoder_layers=self.num_hidden_layers,
encoder_attention_heads=self.num_attention_heads,
decoder_attention_heads=self.num_attention_heads,
encoder_ffn_dim=self.intermediate_size,
decoder_ffn_dim=self.intermediate_size,
dropout=self.hidden_dropout_prob,
attention_dropout=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
eos_token_id=self.eos_token_id,
bos_token_id=self.bos_token_id,
pad_token_id=self.pad_token_id,
decoder_start_token_id=self.decoder_start_token_id,
initializer_range=self.initializer_range,
use_cache=False,
)
inputs_dict = prepare_mbart_inputs_dict(config, input_ids,
decoder_input_ids)
return config, inputs_dict
def prepare_config_and_inputs(self):
input_ids = ids_tensor([self.batch_size, self.seq_length],
self.vocab_size)
input_ids = ids_tensor([self.batch_size, self.seq_length],
self.vocab_size).clamp(3, )
input_ids[:, -1] = self.eos_token_id # Eos Token
decoder_input_ids = ids_tensor([self.batch_size, self.seq_length],
self.vocab_size)
config = MBartConfig(
vocab_size=self.vocab_size,
d_model=self.hidden_size,
encoder_layers=self.num_hidden_layers,
decoder_layers=self.num_hidden_layers,
encoder_attention_heads=self.num_attention_heads,
decoder_attention_heads=self.num_attention_heads,
encoder_ffn_dim=self.intermediate_size,
decoder_ffn_dim=self.intermediate_size,
dropout=self.hidden_dropout_prob,
attention_dropout=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
eos_token_id=self.eos_token_id,
bos_token_id=self.bos_token_id,
pad_token_id=self.pad_token_id,
)
inputs_dict = prepare_mbart_inputs_dict(config, input_ids,
decoder_input_ids)
return config, inputs_dict
def prepare_config_and_inputs(self):
input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
attention_mask = None
if self.use_attention_mask:
attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2)
lm_labels = None
if self.use_labels:
lm_labels = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
config = MBartConfig(
vocab_size=self.vocab_size,
d_model=self.d_model,
decoder_layers=self.decoder_layers,
decoder_ffn_dim=self.decoder_ffn_dim,
encoder_attention_heads=self.encoder_attention_heads,
decoder_attention_heads=self.decoder_attention_heads,
eos_token_id=self.eos_token_id,
bos_token_id=self.bos_token_id,
use_cache=self.use_cache,
pad_token_id=self.pad_token_id,
decoder_start_token_id=self.decoder_start_token_id,
max_position_embeddings=self.max_position_embeddings,
is_encoder_decoder=self.is_encoder_decoder,
)
return (
config,
input_ids,
attention_mask,
lm_labels,
)
def get_configs(model):
original_config = model.config
encoder_config = DonutSwinConfig(
image_size=original_config.input_size,
patch_size=4,
depths=original_config.encoder_layer,
num_heads=[4, 8, 16, 32],
window_size=original_config.window_size,
embed_dim=128,
)
decoder_config = MBartConfig(
is_decoder=True,
is_encoder_decoder=False,
add_cross_attention=True,
decoder_layers=original_config.decoder_layer,
max_position_embeddings=original_config.max_position_embeddings,
vocab_size=len(
model.decoder.tokenizer
), # several special tokens are added to the vocab of XLMRobertaTokenizer, see repo on the hub (added_tokens.json)
scale_embedding=True,
add_final_layer_norm=True,
)
return encoder_config, decoder_config
def __init__(self, parent):
self.config = MBartConfig(
vocab_size=99,
d_model=24,
encoder_layers=2,
decoder_layers=2,
encoder_attention_heads=2,
decoder_attention_heads=2,
encoder_ffn_dim=32,
decoder_ffn_dim=32,
max_position_embeddings=48,
add_final_layer_norm=True,
)
def get_config(self):
return MBartConfig(
vocab_size=self.vocab_size,
d_model=self.hidden_size,
encoder_layers=self.num_hidden_layers,
decoder_layers=self.num_hidden_layers,
encoder_attention_heads=self.num_attention_heads,
decoder_attention_heads=self.num_attention_heads,
encoder_ffn_dim=self.intermediate_size,
decoder_ffn_dim=self.intermediate_size,
dropout=self.hidden_dropout_prob,
attention_dropout=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
eos_token_id=self.eos_token_id,
bos_token_id=self.bos_token_id,
pad_token_id=self.pad_token_id,
)
def test_lm_uneven_forward(self):
config = MBartConfig(
vocab_size=self.vocab_size,
d_model=14,
encoder_layers=2,
decoder_layers=2,
encoder_attention_heads=2,
decoder_attention_heads=2,
encoder_ffn_dim=8,
decoder_ffn_dim=8,
max_position_embeddings=48,
)
lm_model = FlaxMBartForConditionalGeneration(config)
context = np.array([[71, 82, 18, 33, 46, 91, 2], [68, 34, 26, 58, 30, 2, 1]], dtype=np.int64)
summary = np.array([[82, 71, 82, 18, 2], [58, 68, 2, 1, 1]], dtype=np.int64)
outputs = lm_model(input_ids=context, decoder_input_ids=summary)
expected_shape = (*summary.shape, config.vocab_size)
self.assertEqual(outputs["logits"].shape, expected_shape)
def train_MBart(data_path,tokenizer,output_path):
model_config = MBartConfig(vocab_size=300,d_model=10,encoder_layers=1,decoder_layers=1,encoder_attention_heads=1,decoder_attention_heads=1,encoder_ffn_dim=10,decoder_ffn_dim=10,max_position_embeddings=512)
model = MBartModel(config=model_config)
sentences = {} #associates lang_id with list of sentences
#read data files and separate language data into different lists
lang_id = 0 #counter for languages in dataset
for sentence_file in os.listdir(data_path):
with open(data_path+sentence_file,'r') as data:
sentences[lang_id] = []
for line in data:
sentences[lang_id].append(line)
lang_id += 1
#create token sequences to pass into model
src_lang,tgt_lang = (sentences[lang_id] for lang_id in sentences)
batch = tokenizer.prepare_seq2seq_batch(src_texts=src_lang,tgt_texts=tgt_lang,return_tensors='pt')
model(input_ids=batch['input_ids'],decoder_input_ids=batch['labels'])
model.save_pretrained(output_path)
def _get_config_and_data(self):
input_ids = np.array(
[
[71, 82, 18, 33, 46, 91, 2],
[68, 34, 26, 58, 30, 82, 2],
[5, 97, 17, 39, 94, 40, 2],
[76, 83, 94, 25, 70, 78, 2],
[87, 59, 41, 35, 48, 66, 2],
[55, 13, 16, 58, 5, 2, 1], # note padding
[64, 27, 31, 51, 12, 75, 2],
[52, 64, 86, 17, 83, 39, 2],
[48, 61, 9, 24, 71, 82, 2],
[26, 1, 60, 48, 22, 13, 2],
[21, 5, 62, 28, 14, 76, 2],
[45, 98, 37, 86, 59, 48, 2],
[70, 70, 50, 9, 28, 0, 2],
],
dtype=np.int64,
)
batch_size = input_ids.shape[0]
config = MBartConfig(
vocab_size=self.vocab_size,
d_model=24,
encoder_layers=2,
decoder_layers=2,
encoder_attention_heads=2,
decoder_attention_heads=2,
encoder_ffn_dim=32,
decoder_ffn_dim=32,
max_position_embeddings=48,
eos_token_id=2,
pad_token_id=1,
bos_token_id=0,
)
return config, input_ids, batch_size
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
check_output_dir(training_args)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if training_args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.parallel_mode == ParallelMode.DISTRIBUTED),
training_args.fp16,
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed
set_seed(training_args.seed)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = MBartConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
extra_model_params = ("encoder_layerdrop", "decoder_layerdrop", "dropout", "attention_dropout")
for p in extra_model_params:
if getattr(training_args, p, None):
assert hasattr(config, p), f"({config.__class__.__name__}) doesn't have a `{p}` attribute"
setattr(config, p, getattr(training_args, p))
tokenizer = MBartTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
# model = MBartForConditionalGeneration.from_pretrained(
# model_args.model_name_or_path,
# from_tf=".ckpt" in model_args.model_name_or_path,
# config=config,
# cache_dir=model_args.cache_dir,
# )
# model = MBartForConditionalGeneration(config)
# model = MBartForConditionalGeneration.from_pretrained(model_args.config_name)
model_config = MBartConfig(vocab_size=300,d_model=10,encoder_layers=1,decoder_layers=1,encoder_attention_heads=1,decoder_attention_heads=1,encoder_ffn_dim=10,decoder_ffn_dim=10,max_position_embeddings=512)
model = MBartModel(config=model_config)
# use task specific params
use_task_specific_params(model, data_args.task)
# set num_beams for evaluation
if data_args.eval_beams is None:
data_args.eval_beams = model.config.num_beams
# set decoder_start_token_id for MBart
if model.config.decoder_start_token_id is None and isinstance(tokenizer, MBartTokenizer):
assert (
data_args.tgt_lang is not None and data_args.src_lang is not None
), "mBart requires --tgt_lang and --src_lang"
model.config.decoder_start_token_id = tokenizer.lang_code_to_id[data_args.tgt_lang]
if model_args.freeze_embeds:
freeze_embeds(model)
if model_args.freeze_encoder:
freeze_params(model.get_encoder())
assert_all_frozen(model.get_encoder())
dataset_class = Seq2SeqDataset
# Get datasets
train_dataset = (
dataset_class(
tokenizer,
type_path="train",
data_dir=data_args.data_dir,
n_obs=data_args.n_train,
max_target_length=data_args.max_target_length,
max_source_length=data_args.max_source_length,
prefix=model.config.prefix or "",
)
if training_args.do_train
else None
)
eval_dataset = (
dataset_class(
tokenizer,
type_path="val",
data_dir=data_args.data_dir,
n_obs=data_args.n_val,
max_target_length=data_args.val_max_target_length,
max_source_length=data_args.max_source_length,
prefix=model.config.prefix or "",
)
if training_args.do_eval or training_args.evaluation_strategy != EvaluationStrategy.NO
else None
)
test_dataset = (
dataset_class(
tokenizer,
type_path="test",
data_dir=data_args.data_dir,
n_obs=data_args.n_test,
max_target_length=data_args.test_max_target_length,
max_source_length=data_args.max_source_length,
prefix=model.config.prefix or "",
)
if training_args.do_predict
else None
)
# Initialize our Trainer
compute_metrics_fn = (
build_compute_metrics_fn(data_args.task, tokenizer) if training_args.predict_with_generate else None
)
trainer = Seq2SeqTrainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
data_collator=Seq2SeqDataCollator(tokenizer, data_args, training_args.tpu_num_cores),
compute_metrics=compute_metrics_fn,
tokenizer=tokenizer,
)
all_metrics = {}
# Training
if training_args.do_train:
logger.info("*** Train ***")
train_result = trainer.train(
model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
)
metrics = train_result.metrics
metrics["train_n_objs"] = data_args.n_train
trainer.save_model() # this also saves the tokenizer
if trainer.is_world_process_zero():
handle_metrics("train", metrics, training_args.output_dir)
all_metrics.update(metrics)
# Need to save the state, since Trainer.save_model saves only the tokenizer with the model
trainer.state.save_to_json(os.path.join(training_args.output_dir, "trainer_state.json"))
# For convenience, we also re-save the tokenizer to the same directory,
# so that you can share your model easily on huggingface.co/models =)
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate(
metric_key_prefix="val", max_length=data_args.val_max_target_length, num_beams=data_args.eval_beams
)
metrics["val_n_objs"] = data_args.n_val
metrics["val_loss"] = round(metrics["val_loss"], 4)
if trainer.is_world_process_zero():
handle_metrics("val", metrics, training_args.output_dir)
all_metrics.update(metrics)
if training_args.do_predict:
logger.info("*** Predict ***")
test_output = trainer.predict(
test_dataset=test_dataset,
metric_key_prefix="test",
max_length=data_args.val_max_target_length,
num_beams=data_args.eval_beams,
)
metrics = test_output.metrics
metrics["test_n_objs"] = data_args.n_test
if trainer.is_world_process_zero():
metrics["test_loss"] = round(metrics["test_loss"], 4)
handle_metrics("test", metrics, training_args.output_dir)
all_metrics.update(metrics)
if training_args.predict_with_generate:
test_preds = tokenizer.batch_decode(
test_output.predictions, skip_special_tokens=True, clean_up_tokenization_spaces=True
)
test_preds = lmap(str.strip, test_preds)
write_txt_file(test_preds, os.path.join(training_args.output_dir, "test_generations.txt"))
if trainer.is_world_process_zero():
save_json(all_metrics, os.path.join(training_args.output_dir, "all_results.json"))
return all_metrics
TGT_DATA = "./data_tgt_de.txt"
SRC_DATA = "./data_source_hsb.txt"
from transformers import MBartForConditionalGeneration, MBartTokenizer, MBartModel, MBartConfig
#Read from the data files
src_txts = []
tgt_txts = []
with open(SRC_DATA) as f:
for line in f:
src_txts.append(line)
with open(TGT_DATA) as f:
for line in f:
tgt_txts.append(line)
tokenizer = MBartTokenizer.from_pretrained('./tokenizer_de_hsb.model')
batch = tokenizer.prepare_seq2seq_batch(src_texts=src_txts,
src_lang="en_XX",
tgt_texts=tgt_txts,
tgt_lang="ro_RO",
return_tensors="pt")
config = MBartConfig()
model = MBartModel(config)
model(input_ids=batch['input_ids'],
decoder_input_ids=batch['labels']) # forward pass
model.save_pretrained('./trained_model')
