Hi,
I want to use label_smoothing_factor parameter when training my PEGASUS model, however, setting this argument results in the following error. Removing this argument helps to avoid the error.
Traceback (most recent call last):
File "optimize_hp_pegasus.py", line 158, in <module>
compute_objective=compute_objective
File "/home/jovyan/.imgenv-hyperopt-pegasus-0/lib/python3.7/site-packages/transformers/trainer.py", line 1791, in hyperparameter_search
best_run = backend_dict[backend](self, n_trials, direction, **kwargs)
File "/home/jovyan/.imgenv-hyperopt-pegasus-0/lib/python3.7/site-packages/transformers/integrations.py", line 160, in run_hp_search_optuna
study.optimize(_objective, n_trials=n_trials, timeout=timeout, n_jobs=n_jobs)
File "/home/user/conda/lib/python3.7/site-packages/optuna/study/study.py", line 409, in optimize
show_progress_bar=show_progress_bar,
File "/home/user/conda/lib/python3.7/site-packages/optuna/study/_optimize.py", line 76, in _optimize
progress_bar=progress_bar,
File "/home/user/conda/lib/python3.7/site-packages/optuna/study/_optimize.py", line 163, in _optimize_sequential
trial = _run_trial(study, func, catch)
File "/home/user/conda/lib/python3.7/site-packages/optuna/study/_optimize.py", line 264, in _run_trial
raise func_err
File "/home/user/conda/lib/python3.7/site-packages/optuna/study/_optimize.py", line 213, in _run_trial
value_or_values = func(trial)
File "/home/jovyan/.imgenv-hyperopt-pegasus-0/lib/python3.7/site-packages/transformers/integrations.py", line 150, in _objective
trainer.train(resume_from_checkpoint=checkpoint, trial=trial)
File "/home/jovyan/.imgenv-hyperopt-pegasus-0/lib/python3.7/site-packages/transformers/trainer.py", line 1325, in train
tr_loss_step = self.training_step(model, inputs)
File "/home/jovyan/.imgenv-hyperopt-pegasus-0/lib/python3.7/site-packages/transformers/trainer.py", line 1884, in training_step
loss = self.compute_loss(model, inputs)
File "/home/jovyan/.imgenv-hyperopt-pegasus-0/lib/python3.7/site-packages/transformers/trainer.py", line 1916, in compute_loss
outputs = model(**inputs)
File "/home/jovyan/.imgenv-hyperopt-pegasus-0/lib/python3.7/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
result = self.forward(*input, **kwargs)
File "/home/jovyan/.imgenv-hyperopt-pegasus-0/lib/python3.7/site-packages/transformers/models/pegasus/modeling_pegasus.py", line 1405, in forward
return_dict=return_dict,
File "/home/jovyan/.imgenv-hyperopt-pegasus-0/lib/python3.7/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
result = self.forward(*input, **kwargs)
File "/home/jovyan/.imgenv-hyperopt-pegasus-0/lib/python3.7/site-packages/transformers/models/pegasus/modeling_pegasus.py", line 1262, in forward
return_dict=return_dict,
File "/home/jovyan/.imgenv-hyperopt-pegasus-0/lib/python3.7/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
result = self.forward(*input, **kwargs)
File "/home/jovyan/.imgenv-hyperopt-pegasus-0/lib/python3.7/site-packages/transformers/models/pegasus/modeling_pegasus.py", line 1014, in forward
raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
ValueError: You have to specify either decoder_input_ids or decoder_inputs_embeds
The code to reproduce the bug:
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
import numpy as np
from pathlib import Path
import transformers as ts
from datasets import load_dataset, load_metric
import torch
import nltk
from shutil import rmtree
Path("hyperopt").mkdir(exist_ok=True)
os.chdir("hyperopt")
# Upload the data
data = load_dataset("gigaword", revision="master", cache_dir='cache/data')#, download_mode="force_redownload")
train_split = data['train'].train_test_split(train_size=1_000, shuffle=True, seed=42)
train = train_split['train']
valid_split = data['validation'].train_test_split(train_size=500, shuffle=True, seed=42)
valid = valid_split['train']
# Upload the tokenizer
tokenizer = ts.AutoTokenizer.from_pretrained('google/pegasus-large', cache_dir='cache/tokenizer')
def tokenizing_fn(instances):
encoded = tokenizer(instances["document"], truncation=True)
with tokenizer.as_target_tokenizer():
labels = tokenizer(instances["summary"], truncation=True)
encoded["labels"] = labels["input_ids"]
return encoded
tokenized_train = train.map(tokenizing_fn, batched=True).remove_columns(['document', 'summary'])
tokenized_val = valid.map(tokenizing_fn, batched=True).remove_columns(['document', 'summary'])
data_collator = ts.DataCollatorForSeq2Seq(tokenizer = tokenizer, padding="longest")
metric = load_metric("rouge", cache_dir="cache/metric")
additional_metrics = [load_metric("sacrebleu", cache_dir="cache/metric")]
def compute_metrics(eval_preds):
# in this task we suppose that golden labels are reference sentences
# and predictions are summaries generated by model, in human-readable format
predictions, labels = eval_preds
decoded_preds = tokenizer.batch_decode(
predictions, skip_special_tokens=True
)
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
# Rouge expects a newline after each sentence
decoded_preds = [
"\n".join(nltk.sent_tokenize(pred.strip())) for pred in decoded_preds
]
decoded_labels = [
"\n".join(nltk.sent_tokenize(label.strip())) for label in decoded_labels
]
result = metric.compute(
predictions=decoded_preds, references=decoded_labels, use_stemmer=True
)
# Extract a few results
result = {key: value.mid.fmeasure * 100 for key, value in result.items()}
# Add mean generated length
prediction_lens = [
np.count_nonzero(pred != tokenizer.pad_token_id) for pred in predictions
]
result["gen_len"] = np.mean(prediction_lens)
for add_metric in additional_metrics:
result.update(
add_metric.compute(predictions=predictions, references=labels)
)
return result
def init_model():
return ts.AutoModelForSeq2SeqLM.from_pretrained('google/pegasus-large', cache_dir = 'cache/model')
training_args = ts.Seq2SeqTrainingArguments(
output_dir='pegasus_train',
# Batch size args
num_train_epochs=5,
per_device_train_batch_size=4,
per_device_eval_batch_size=75,
# Optimizer args
learning_rate=5e-4,
weight_decay=0.1,
max_grad_norm=1.,
# label_smoothing_factor=0.1,
# Scheduler args
warmup_ratio=0.1,
# Eval args
metric_for_best_model='eval_rouge1',
# load_best_model_at_end=True,
evaluation_strategy='epoch',
logging_strategy='epoch',
save_strategy='epoch',
save_total_limit=0,
eval_accumulation_steps=1,
# Seq2Seq args
generation_max_length=42,
predict_with_generate=True,
generation_num_beams=4,
# General args
seed=42,
fp16=True,
fp16_full_eval=False,
)
callbacks = []#[ts.EarlyStoppingCallback(early_stopping_patience=3)]
trainer = ts.Seq2SeqTrainer(
model_init=init_model,
args=training_args,
data_collator=data_collator,
train_dataset=tokenized_train,
eval_dataset=tokenized_val,
callbacks=callbacks,
compute_metrics=compute_metrics
)
trainer.train()
Libraries versions:
transformers==4.17.0
datasets==1.18.3, I installed them from git (impossible to load gigaword otherwise): pip install git+https://github.com/huggingface/datasets#egg=datasets