Trainer bug? Loss and logits are “nan” when fine-tuning NLI model (both RoBERTa/BART)

[MoritzLaurer]

Environment info

• transformers version: 4.0.1 (also reproduced the same issue with 3.5.1)
• Platform: Google Colab
• Python version: 3.6
• PyTorch version (GPU?): 1.7.0+cu101
• Tensorflow version (GPU?): no
• Using GPU in script?: Yes
• Using distributed or parallel set-up in script?: (don't know. Probably not)

Who can help

@sgugger

Information

The problem arises when using:

• [ x] my own modified scripts: (give details below)

The tasks I am working on is:

• [ x] my own task or dataset: (give details below)

Description: I’m trying to fine-tune a pre-trained NLI model (ynie/roberta-large-snli_mnli_fever_anli_R1_R2_R3-nli) on a dataset of around 276.000 hypothesis-premise pairs. I’m following the instructions from the docs here and here. When I run the training, it seems like the fine-tuning works (it does the training and saves the checkpoints), but trainer.train() and trainer.evaluate() return "nan" as loss value.

What I've tried:

• I tried using both ynie/roberta-large-snli_mnli_fever_anli_R1_R2_R3-nli and facebook/bart-large-mnli to make sure that it's not linked to specific model, but I get the issue for both models
• I tried following the advice in this related github issue, but adding num_labels=3 to the config file does not solve the issue. (I think my issue is different because the models are already fine-tuned on NLI in my case)
• I tried many different ways of changing my input data because I suspected that there could be an issue with my input data, but I also couldn't solve it that way.
• The probable source of the issue: I inspected the prediction output from the model during training and the weird thing is that the prediction value always seems to be "0" (entailment) in 100% of cases (see printed output at the bottom of the code below). This cannot be right. Even weirder: When I first run the model to predict a test sequences before running the trainer, I get normal logits as output. When I run the exact same code block again at the end after having run the trainer, I get tensor([[nan, nan, nan]] as output (see code below).
• I suspect that the source for the 'only 0 prediction output' is that the logits the model returns during training are possibly always torch.tensor([[np.nan, np.nan, np.nan]]). torch.tensor([[np.nan, np.nan, np.nan]]).argmax(-1) returns torch.tensor(0) without triggering an error. The big mystery for me is why the logits would become "nan", because the model does not do that when I use the same input data only outside of the trainer, but something changes once I've run the trainer.

=> I would be very thankful for any help on this! (I've been trying to solve this since two days now) Thanks a lot in advance.

To reproduce

Here is my code:

### load model & tokenize
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

max_length = 256
hg_model_hub_name = "ynie/roberta-large-snli_mnli_fever_anli_R1_R2_R3-nli"
# also tried: hg_model_hub_name = "facebook/bart-large-mnli"
tokenizer = AutoTokenizer.from_pretrained(hg_model_hub_name)
model = AutoModelForSequenceClassification.from_pretrained(hg_model_hub_name)
model.config

device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Device: {device}")
if device == "cuda":
  model = model.half()
model.to(device)
model.train();

Running a test inference with the model at this point works fine:

test_enc = tokenizer(nli_train[0]["premise"], nli_train[0]["hypothesis"], return_tensors="pt", max_length=max_length,
                            return_token_type_ids=True, truncation=True, padding=True)
model.eval();
test_output_loss = model(test_enc["input_ids"].to(device), attention_mask=test_enc["attention_mask"].to(device), token_type_ids=test_enc["token_type_ids"].to(device), labels=torch.tensor(2).to(device))
print(test_output_loss)
#output: SequenceClassifierOutput(loss=tensor(2.2168, device='cuda:0', dtype=torch.float16, grad_fn=<NllLossBackward>), logits=tensor([[ 0.4075,  0.8511, -0.7549]], device='cuda:0', dtype=torch.float16,
       grad_fn=<AddmmBackward>), hidden_states=None, attentions=None)

Then I continue with preprocessing and training:

... some data preprocessing

encodings_train = tokenizer(premise_train, hypothesis_train, return_tensors="pt", max_length=max_length,
                            return_token_type_ids=True, truncation=False, padding=True)
encodings_val = tokenizer(premise_val, hypothesis_val, return_tensors="pt", max_length=max_length,
                          return_token_type_ids=True, truncation=False, padding=True)
encodings_test = tokenizer(premise_test, hypothesis_test, return_tensors="pt", max_length=max_length,
                           return_token_type_ids=True, truncation=False, padding=True)

### create pytorch dataset object
class XDataset(torch.utils.data.Dataset):
    def __init__(self, encodings, labels):
        self.encodings = encodings
        self.labels = labels
    def __getitem__(self, idx):
        item = {key: torch.as_tensor(val[idx]) for key, val in self.encodings.items()}
        #item = {key: torch.as_tensor(val[idx]).to(device) for key, val in self.encodings.items()}
        item['labels'] = torch.as_tensor(self.labels[idx])
        #item['labels'] = self.labels[idx]
        return item
    def __len__(self):
        return len(self.labels)

dataset_train = XDataset(encodings_train, label_train)
dataset_val = XDataset(encodings_val, label_val)
dataset_test = XDataset(encodings_test, label_test)

# compute metrics with trainer
from sklearn.metrics import accuracy_score, precision_recall_fscore_support
def compute_metrics(pred):
    labels = pred.label_ids
    print(labels)
    preds = pred.predictions.argmax(-1)
    print(preds)
    precision, recall, f1, _ = precision_recall_fscore_support(labels, preds, average='binary', pos_label=0)
    acc = accuracy_score(labels, preds)
    return {
        'accuracy': acc,
        'f1': f1,
        'precision': precision,
        'recall': recall
    }

## training
from transformers import Trainer, TrainingArguments

# https://huggingface.co/transformers/main_classes/trainer.html#transformers.TrainingArguments
training_args = TrainingArguments(
    output_dir='./results',          # output directory
    num_train_epochs=1,              # total number of training epochs
    per_device_train_batch_size=8,  # batch size per device during training
    per_device_eval_batch_size=8,   # batch size for evaluation
    warmup_steps=500,                # number of warmup steps for learning rate scheduler
    weight_decay=0.01,               # strength of weight decay
    logging_dir='./logs',            # directory for storing logs
    logging_steps=100,
)

trainer = Trainer(
    model=model,                         # the instantiated 🤗 Transformers model to be trained
    args=training_args,                  # training arguments, defined above
    train_dataset=dataset_train,         # training dataset
    eval_dataset=dataset_val             # evaluation dataset
)

trainer.train()
# output: TrainOutput(global_step=181, training_loss=nan)
trainer.evaluate()
# output: 
[2 2 2 0 0 2 2 2 0 2 0 0 2 2 2 2 0 2 0 2 2 2 2 0 2 0 2 0 0 2 0 0 2 0 0 0 2
 0 2 0 0 0 0 0 2 0 0 2 2 2 0 2 2 2 2 2 0 0 0 0 2 0 0 0 2 2 0 0 0 2 0 0 0 2
 2 0 2 0 0 2 2 2 0 2 2 0 0 0 0 0 0 0 2 0 0 0 0 2 0 2 2 0 2 0 0 2 2 2 2 2 2
 2 0 0 0 0 2 0 0 2 0 0 0 0 2 2 2 0 0 0 0 0 2 0 0 2 0 2 0 2 0 2 0 0 2 2 0 0
 2 2 2 2 2 2 0 0 2 2 2 2 0 2 0 0 2 2 2 0 0 2 0 2 0 2 0 0 0 0 0 0 2 0 0 2 2
 0 2 2 2 0 2 2 0 2 2 2 2 2 2 0 0 2 0 0 2 2 0 0 0 2 0 2 2 2 0 0 0 0 0 0 0 0
 2 0 2 2 2 0 2 0 0 2 0 2 2 0 0 0 0 2 2 2 0 0 0 2 2 2 2 0 2 0 2 2 2]
[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]

{'epoch': 1.0,
 'eval_accuracy': 0.5137254901960784,
 'eval_f1': 0.6787564766839378,
 'eval_loss': nan,
 'eval_precision': 0.5137254901960784,
 'eval_recall': 1.0}

Test running the model again after training, returns tensor([[nan, nan, nan]] for some reason:

test_enc = tokenizer(nli_train[0]["premise"], nli_train[0]["hypothesis"], return_tensors="pt", max_length=max_length,
                            return_token_type_ids=True, truncation=True, padding=True)
model.eval();
test_output_loss = model(test_enc["input_ids"].to(device), attention_mask=test_enc["attention_mask"].to(device), token_type_ids=test_enc["token_type_ids"].to(device), labels=torch.tensor(2).to(device))
print(test_output_loss)
#output: SequenceClassifierOutput(loss=tensor(nan, device='cuda:0', dtype=torch.float16, grad_fn=<NllLossBackward>), logits=tensor([[nan, nan, nan]], device='cuda:0', dtype=torch.float16,
       grad_fn=<AddmmBackward>), hidden_states=None, attentions=None)

Expected behavior

Model should not return "nan" for logits and return a loss value.

[MoritzLaurer]

Update: I reran the training in native PyTorch with the following code and I did not get the same issue. This means that there is some issue with the trainer?

import torch

class XDataset(torch.utils.data.Dataset):
    def __init__(self, encodings, labels):
        self.encodings = encodings
        self.labels = labels
    def __getitem__(self, idx):
        #item = {key: torch.as_tensor(val[idx]) for key, val in self.encodings.items()}
        item = {key: torch.as_tensor(val[idx]) for key, val in self.encodings.items()}
        #item = {key: torch.as_tensor(val[idx]).to(device) for key, val in self.encodings.items()}
        item['labels'] = torch.as_tensor(self.labels[idx])
        #item['labels'] = torch.LongTensor(self.labels[idx])
        #item['labels'] = self.labels[idx]
        return item
    def __len__(self):
        return len(self.labels)

dataset_train = XDataset(encodings_train, label_train)
dataset_val = XDataset(encodings_val, label_val)
dataset_test = XDataset(encodings_test, label_test)

from torch.utils.data import DataLoader
from transformers import AdamW

device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
model.to(device)
model.train()

train_loader = DataLoader(dataset_train, batch_size=16, shuffle=True)

optim = AdamW(model.parameters(), lr=5e-5)

for epoch in range(1):
    for batch in train_loader:
        optim.zero_grad()
        input_ids = batch['input_ids'].to(device)
        attention_mask = batch['attention_mask'].to(device)
        token_type_ids = batch['token_type_ids'].to(device)
        labels = batch['labels'].to(device)
        print(labels)
        outputs = model(input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids, labels=labels)
        loss = outputs[0] # outputs.loss
        print(loss)
        loss.backward()
        optim.step()
# Output: it prints the labels and the loss correctly!
#tensor([2, 0, 2, 2, 2, 2, 0, 0, 0, 2, 2, 2, 0, 0, 0, 2], device='cuda:0')
#tensor(0.6895, device='cuda:0', dtype=torch.float16, grad_fn=<NllLossBackward>) ....

When I rerun the model for a test inference after this native pytorch training step, it also returns logits and loss as expected (no "nan").

[sgugger]

In the first snippet of code you convert your whole model to FP16 with model.half() (this is not in your second snippet of code). This is not how mixed-precision training works and you should pass the flag fp16=True to your TrainingArguments.

[MoritzLaurer]

thanks, I don't know much about mixed-precision training (the only reason why I added model.half() is because I understood that it reduces memory usage). Now, when I add fp16=True, i get the error: ValueError: Attempting to unscale FP16 gradients. when running trainer.train()

training_args = TrainingArguments(
    output_dir='./results',          # output directory
    num_train_epochs=1,              # total number of training epochs
    per_device_train_batch_size=8,   # batch size per device during training
    per_device_eval_batch_size=8,    # batch size for evaluation
    warmup_steps=500,                # number of warmup steps for learning rate scheduler
    weight_decay=0.01,               # strength of weight decay
    logging_dir='./logs',            # directory for storing logs
    logging_steps=30,
    fp16=True
)

[MoritzLaurer]

Cool, but when I remove the model.half(), it does return the loss, that's great!

[sgugger]

Yes you have to remove that line, that's what I was saying :-)

[MoritzLaurer]

Great, so I understand that I can use mixed precision training by simply passing the flag fp16=True without manual modifications to the model. Is there actually any good reason not to pass "fp16=True"? The articles on mixed precision training I've found seem to be very positive about it.

In any case, thanks for solving my issue! :)

[sgugger]

There are no reason not to use, no. Sometimes for debugging purposes or there may be one of the exotic models that don't support FP16, but in general, it's a good way to speed up training and saving GPU memory.

Closing the issues since it's solved!

[artmatsak]

there may be one of the exotic models that don't support FP16

That was my case with ltgoslo/norbert producing the nan loss with FP16. Setting fp16 to False solved the issue, thanks!