My Task is to fine-tune a Multi-class text classification model with close to 400 labels. As recommended in various NLP blogs, I decided to fine tune BertForSequenceClassification on custom dataset using Accelerate library from Huggingface. Code is written and executed on AWS SageMaker with 4 GPUs and 24 GB RAM on each GPU. Training hyper-parameters: batch size:64 which on 4 GPUs would mean a batch size of 16 per GPU; learning rate: 1e-5, optimizer: torch.optim.AdamW and linear learning rate scheduler.
I tested the code for 1 step i.e. executed the code end to end, printed training loss and validation loss for sanity check and the execution was successful with all loss/accuracy values printed.
But now when I run the code on full dataset (3500k samples), the notebook cell does not produce any error but doesnot execute also. At times, accelerate doesnot identify all. the 4 GPUs else simply quits without even executing one step of training/validation.
I have these basic issues:
a) Why does accelerate() not recognize all the GPUs ( I can see them printed using nvidia-smi)
b) notebook-launcher quits execution with out any error, are there execution logs being saved that can help identify teh problem
c) I am fine tuning BertForSequenceClassification in Multi-class setting, ideally loss should be computed using CrossEntropyLossFunction, but when I printed the loss for sanity check, it printed the value: tensor(6.06, NllLossBackward0) why doesit show NegativeLogLikelihood in the loss values.
Heres. the relevant code snippet:
def training_loop():
model_path = "models/category_classifier_bert.pt"
accelerator = Accelerator(mixed_precision="fp16")
device = accelerator.device
model = BertForSequenceClassification.from_pretrained("bert-base-uncased",num_labels=NUM_LABELS)
print("Loading model ")
learning_rate = 1e-5
optimizer = AdamW(model.parameters(), lr=learning_rate)
print("Loading optimizer")
train_dl, val_dl = get_dataloaders()
total_steps = int((len(train_dl) * EPOCHS)/4*BATCH_SIZE)
print("total_steps : ",total_steps)
# Set up the learning rate scheduler
scheduler = get_linear_schedule_with_warmup(optimizer,
num_warmup_steps=100, # Default value
num_training_steps=total_steps)
print("Loading model on accelerators .... ")
model, optimizer, train_dl, val_dl, scheduler = accelerator.prepare(model, optimizer, train_dl, val_dl, scheduler)
#val_dl = accelerator.prepare(val_dl)
metric = Accuracy(task="multiclass", num_classes=NUM_LABELS).to(device)
# MulticlassAccuracy(num_classes=NUM_LABELS).to(device)
print("Training ... ")
for epoch in range(EPOCHS):
model.train()
for step,batch in enumerate(train_dl):
input_ids = batch["input_ids"].to(device)
targets = batch["label"].to(device)
attention_mask = batch["attention_mask"].to(device)
optimizer.zero_grad()
output = model(input_ids, attention_mask)
loss = CrossEntropyLoss(output.logits, targets)
accelerator.backward(loss)
optimizer.step()
if step%5000 == 0:
print(step, loss)
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
accelerator.save(unwrapped_model.state_dict(), model_path)
model.eval()
# accuracy, num_elements = 0, 0
val_accuracy = []
print("Model validation .... ")
for steps,batch in enumerate(val_dl):
preds = model(batch["input_ids"], batch["attention_mask"])
targets = batch["label"].to(device)
all_predictions = accelerator.gather(preds.logits)#.detach().cpu()
all_targets = accelerator.gather(targets)
loss = CrossEntropyLoss(preds.logits, targets)
# print("all_predictions : ",type(all_predictions))
# print("all_targets : ",type(all_targets))
step_accuracy = metric(all_predictions, all_targets)
val_accuracy.append(step_accuracy.items())
if steps%10000 ==0:
print(f"Loss : {loss}")
# accuracy_preds = all_predictions == all_targets
# print("all_predictions : ",type(accuracy_preds))
# print("preds ",preds.logits.shape)
# num_elements += BATCH_SIZE
# accuracy += accuracy_preds.sum()
# print("accuracy : ",accuracy)
# We can calculate the overall accuracy
#eval_metric = accuracy.item() / num_elements
# And we use `accelerator.print` to print on only the main process
accelerator.print(f'epoch {epoch}: accuracy - {100 * mean(accuracy):.2f}')
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
torch.save(unwrapped_model.state_dict(), model_path)
break
from accelerate import notebook_launcher
notebook_launcher(training_loop, num_processes=4)
Heres the output generated after executing. the above cell
Already tried Restarting the kernel and shutting down/restarting the server.
