AttributeError: 'NoneType' object has no attribute 'get'

[karan842]

I am trying to run below code from @tomaarsen's HuggingFace blog on Sentence-Transformer-V3

Code:

from datasets import load_dataset

data = load_dataset('sentence-transformers/stsb')

train_data = data['train'].select(range(100))
val_data = data['validation'].select(range(100, 30))

from torch.utils.data import DataLoader
from sentence_transformers import(
    SentenceTransformer, models,
    losses, util,
    InputExample, evaluation,
    SentenceTransformerTrainer,
    SentenceTransformerTrainingArguments
)

from sentence_transformers.training_args import BatchSamplers
from torch.optim import AdamW
from accelerate import Accelerator

def main():
  word_embedding_model = models.Transformer('mixedbread-ai/mxbai-embed-large-v1')
  pooling_model = models.Pooling(word_embedding_model.get_word_embedding_dimension())
  model = SentenceTransformer(modules=[word_embedding_model, pooling_model])

  loss = losses.AnglELoss(model)

  evaluator = evaluation.EmbeddingSimilarityEvaluator(
      sentences1=val_data['sentence1'],
      sentences2=val_data['sentence2'],
      scores=val_data['score'],
      main_similarity=evaluation.SimilarityFunction.COSINE,
      name="sts-dev"
  )

  training_args = SentenceTransformerTrainingArguments(
      output_dir='./sbert-checkpoint',
      num_train_epochs=10,
      seed=33,
      per_device_train_batch_size=8,
      per_device_eval_batch_size=8,
      learning_rate=2e-5,
      # fp16=True
      warmup_ratio=0.1,
      eval_strategy="steps",
      eval_steps=2,
      save_total_limit=2,
      load_best_model_at_end=True,
      metric_for_best_model='spearman_cosine',
      greater_is_better=True

  )

  trainer = SentenceTransformerTrainer(
      model=model,
      evaluator=evaluator,
      args=training_args,
      train_dataset=train_data,
      eval_dataset=val_data,
      loss=loss
  )
  trainer.train()

main()

This code was working on pre-release==3.0.0 version but not on both sentence-transformers[train], sentence-transformera[dev]

Error:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
[<ipython-input-5-1470a0756fb4>](https://localhost:8080/#) in <cell line: 43>()
     41     loss=loss
     42 )
---> 43 trainer.train()

7 frames
[/usr/local/lib/python3.10/dist-packages/transformers/trainer.py](https://localhost:8080/#) in train(self, resume_from_checkpoint, trial, ignore_keys_for_eval, **kwargs)
   1883                 hf_hub_utils.enable_progress_bars()
   1884         else:
-> 1885             return inner_training_loop(
   1886                 args=args,
   1887                 resume_from_checkpoint=resume_from_checkpoint,

[/usr/local/lib/python3.10/dist-packages/transformers/trainer.py](https://localhost:8080/#) in _inner_training_loop(self, batch_size, args, resume_from_checkpoint, trial, ignore_keys_for_eval)
   2289                     self.control = self.callback_handler.on_step_end(args, self.state, self.control)
   2290 
-> 2291                     self._maybe_log_save_evaluate(tr_loss, grad_norm, model, trial, epoch, ignore_keys_for_eval)
   2292                 else:
   2293                     self.control = self.callback_handler.on_substep_end(args, self.state, self.control)

[/usr/local/lib/python3.10/dist-packages/transformers/trainer.py](https://localhost:8080/#) in _maybe_log_save_evaluate(self, tr_loss, grad_norm, model, trial, epoch, ignore_keys_for_eval)
   2719         metrics = None
   2720         if self.control.should_evaluate:
-> 2721             metrics = self.evaluate(ignore_keys=ignore_keys_for_eval)
   2722             self._report_to_hp_search(trial, self.state.global_step, metrics)
   2723 

[/usr/local/lib/python3.10/dist-packages/sentence_transformers/trainer.py](https://localhost:8080/#) in evaluate(self, eval_dataset, ignore_keys, metric_key_prefix)
    367         if isinstance(eval_dataset, DatasetDict) and isinstance(self.loss, dict):
    368             eval_dataset = self.add_dataset_name_column(eval_dataset)
--> 369         return super().evaluate(eval_dataset, ignore_keys, metric_key_prefix)
    370 
    371     def evaluation_loop(

[/usr/local/lib/python3.10/dist-packages/transformers/trainer.py](https://localhost:8080/#) in evaluate(self, eval_dataset, ignore_keys, metric_key_prefix)
   3570 
   3571         eval_loop = self.prediction_loop if self.args.use_legacy_prediction_loop else self.evaluation_loop
-> 3572         output = eval_loop(
   3573             eval_dataloader,
   3574             description="Evaluation",

[/usr/local/lib/python3.10/dist-packages/sentence_transformers/trainer.py](https://localhost:8080/#) in evaluation_loop(self, dataloader, description, prediction_loss_only, ignore_keys, metric_key_prefix)
    377         metric_key_prefix: str = "eval",
    378     ) -> EvalLoopOutput:
--> 379         output = super().evaluation_loop(
    380             dataloader=dataloader,
    381             description=description,

[/usr/local/lib/python3.10/dist-packages/transformers/trainer.py](https://localhost:8080/#) in evaluation_loop(self, dataloader, description, prediction_loss_only, ignore_keys, metric_key_prefix)
   3755 
   3756             # Prediction step
-> 3757             loss, logits, labels = self.prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys)
   3758             main_input_name = getattr(self.model, "main_input_name", "input_ids")
   3759             inputs_decode = self._prepare_input(inputs[main_input_name]) if args.include_inputs_for_metrics else None

[/usr/local/lib/python3.10/dist-packages/transformers/trainer.py](https://localhost:8080/#) in prediction_step(self, model, inputs, prediction_loss_only, ignore_keys)
   3926         # If `return_loss` is not specified or being `None` in `inputs`, we check if the default value of `return_loss`
   3927         # is `True` in `model.forward`.
-> 3928         return_loss = inputs.get("return_loss", None)
   3929         if return_loss is None:
   3930             return_loss = self.can_return_loss

AttributeError: 'NoneType' object has no attribute 'get'

Thanks, Karan Shingde

[tomaarsen]

Hello!

I've narrowed the issue down to this line:

val_data = data['validation'].select(range(100, 30))

If I print that dataset, I get:

Dataset({
    features: ['sentence1', 'sentence2', 'score'],
    num_rows: 0
})

I would recommend updating your code to:

val_data = data['validation'].select(range(100, 130))

and then training works like expected!

• Tom Aarsen

[karan842]

This is working on Kaggle but when I am running locally, Running with:

accelerate launch --multi-gpu --num_processes=4 main.py

It is saying,

TypeError: SentenceTransformerTrainingArguments.__init__() got an unexpected keyword argument 'eval_strategy'   

I tried with evaluation_strategy and now getting this error:

trainer.train()trainer.train()

  File "/data/projects/llm-env/lib/python3.10/site-packages/transformers/trainer.py", line 1780, in train
  File "/data/projects/llm-env/lib/python3.10/site-packages/transformers/trainer.py", line 1780, in train
        return inner_training_loop(return inner_training_loop(

  File "/data/projects/llm-env/lib/python3.10/site-packages/transformers/trainer.py", line 2193, in _inner_training_loop
  File "/data/projects/llm-env/lib/python3.10/site-packages/transformers/trainer.py", line 2193, in _inner_training_loop
    self._maybe_log_save_evaluate(tr_loss, grad_norm, model, trial, epoch, ignore_keys_for_eval)    
self._maybe_log_save_evaluate(tr_loss, grad_norm, model, trial, epoch, ignore_keys_for_eval)
  File "/data/projects/llm-env/lib/python3.10/site-packages/transformers/trainer.py", line 2577, in _maybe_log_save_evaluate
  File "/data/projects/llm-env/lib/python3.10/site-packages/transformers/trainer.py", line 2577, in _maybe_log_save_evaluate
    metrics = self.evaluate(ignore_keys=ignore_keys_for_eval)    
metrics = self.evaluate(ignore_keys=ignore_keys_for_eval)
  File "/data/projects/llm-env/lib/python3.10/site-packages/sentence_transformers/trainer.py", line 369, in evaluate
  File "/data/projects/llm-env/lib/python3.10/site-packages/sentence_transformers/trainer.py", line 369, in evaluate
        return super().evaluate(eval_dataset, ignore_keys, metric_key_prefix)return super().evaluate(eval_dataset, ignore_keys, metric_key_prefix)

  File "/data/projects/llm-env/lib/python3.10/site-packages/transformers/trainer.py", line 3365, in evaluate
  File "/data/projects/llm-env/lib/python3.10/site-packages/transformers/trainer.py", line 3365, in evaluate
        output = eval_loop(output = eval_loop(

  File "/data/projects/llm-env/lib/python3.10/site-packages/sentence_transformers/trainer.py", line 379, in evaluation_loop
  File "/data/projects/llm-env/lib/python3.10/site-packages/sentence_transformers/trainer.py", line 379, in evaluation_loop
        output = super().evaluation_loop(output = super().evaluation_loop(

  File "/data/projects/llm-env/lib/python3.10/site-packages/transformers/trainer.py", line 3554, in evaluation_loop
  File "/data/projects/llm-env/lib/python3.10/site-packages/transformers/trainer.py", line 3554, in evaluation_loop
    loss, logits, labels = self.prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys)
  File "/data/projects/llm-env/lib/python3.10/site-packages/transformers/trainer.py", line 3728, in prediction_step
    loss, logits, labels = self.prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys)
  File "/data/projects/llm-env/lib/python3.10/site-packages/transformers/trainer.py", line 3728, in prediction_step
    return_loss = inputs.get("return_loss", None)
AttributeError: 'NoneType' object has no attribute 'get'    
return_loss = inputs.get("return_loss", None)
AttributeError: 'NoneType' object has no attribute 'get'

My dependencies are:

Name: accelerate Version: 0.29.2

Name: transformers Version: 4.39.3

Name: sentence-transformers Version: 3.0.0

[tomaarsen]

The evaluation_strategy fix was indeed necessary because older versions of transformers do not yet support the new name of eval_strategy (they renamed it).

On Kaggle, can you please print the dataset that you're feeding to eval_dataset in the Trainer? I think it's empty.

• Tom Aarsen

[karan842]

I got it I was giving very less amount of data range and running on 4 GPUs. I increased the data size, For train_data=200. val_data=90

Now working on multi-gpu

[tomaarsen]

Oh, it is indeed possible that using less samples than the number of GPUs causes this issue, even if the number of evaluation samples isn't strictly 0.

[karan842]

What Evaluator can we use for (anchor, positive) type of data?

My task is embedding similarity, but it needed label score

[tomaarsen]

You could use InformationRetrievalEvaluator:

queries = dict(enumerate(dataset["anchor"]))
corpus = dict(enumerate(dataset["positive"]))
relevant_docs = {idx: idx for idx in range(len(dataset))}

ir_evaluator = InformationRetrievalEvaluator(
    queries=queries,
    corpus=corpus,
    relevant_docs=relevant_docs,
    name="...",
)

Here the "positive" is the relevant document and all other texts are seen as not relevant.

You can get roughly the same behaviour with the TranslationEvaluator:

evaluator = TranslationEvaluator(
    source_sentences=dataset["anchor"],
    target_sentences=dataset["positive"],
    name="...",
)

This computes what percentage of time the anchor at index $i$ is most similar to the positive at index $i$ out of all positive texts. The bigger the dataset, the harder both of these tasks get.

• Tom Aarsen

[karan842]

Cool! Thanks

Lastly what is best way to store model locally?

save_safetensors or save_only_model

I just need model locally which I will load into another script for testing

[tomaarsen]

model.save_pretrained("local_path") is the recommended way.

[karan842]

1. Do we need to mention DataLoader and collate_fn explicitly for lazy loading?

2. Are these arguments, doing same thing? per_device_train_batch_size=64, per_device_eval_batch_size=64,

3. I am training on huge data where number of records in train and validation are hundreds of millions. I have 4 GPU 15 GB each. I am new to large data training in prod. How can I run on this machine configuration?

Final Code:

def main():
    data_files = split_data(data_dir)
    # print(data_files)
    data = load_dataset("parquet", data_dir=data_dir, data_files=data_files)
    train_data, val_data = data['train'], data['val']
    print(train_data,"\n\n")
    print(val_data,"\n\n")

    word_embedding_model = models.Transformer('sentence-transformers/all-MiniLM-L6-v2')
    pooling_model = models.Pooling(word_embedding_model.get_word_embedding_dimension())
    model = SentenceTransformer(modules=[word_embedding_model, pooling_model])

    loss = losses.MultipleNegativesRankingLoss(model)

    training_args = SentenceTransformerTrainingArguments(
        output_dir='sbert-output-dir',
        num_train_epochs=1000,
        seed=33,
        per_device_train_batch_size=64,
        per_device_eval_batch_size=64,
        learning_rate=2e-5,
        warmup_ratio=0.2,
        save_only_model=True,
        fp16=True,
        evaluation_strategy="steps",
        eval_steps=50,
        save_total_limit=50,
        load_best_model_at_end=True,
        metric_for_best_model='spearman_cosine',
        greater_is_better=True

    )  

    trainer = SentenceTransformerTrainer(
        model=model,
        args=training_args,
        train_dataset=train_data,
        eval_dataset=val_data,
        loss=loss
    )
    trainer.train()
    model.save_pretrained('sbert-model')

if __name__ == '__main__':
    main()

• Thanks