getting no grad set error during backpropagation.

[RishabhDotasara]

RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn

i'm getting this error with the following training loop, the grads must have been set by the sequential itself, but it's saying no grads.

"""Training"""
Epochs = 100


for epoch in range(Epochs):
    model.train()

    train_logits = model(X_train)
    train_preds_probs = torch.softmax(train_logits,dim=1).argmax(dim=1).type(torch.float32)
    loss = loss_fn(train_preds_probs,y_train)
    train_accu = accuracy(y_train,train_preds_probs)
    print(train_preds_probs)
    optimiser.zero_grad()

    loss.backward()

    optimiser.step()

    #training
    model.eval()
    with torch.inference_mode():
        test_logits = model(X_test)
        test_preds = torch.softmax(test_logits.type(torch.float32),dim=1).argmax(dim=1)
        test_loss = loss_fn(test_preds,y_train)
        test_acc = accuracy(y_test,test_preds)

    
    if epoch%10 == 0:
        print(f'Epoch:{epoch} | Train loss: {loss} |Taining acc:{train_accu} | Test Loss: {test_loss} | Test accu: {test_acc}')

[mrdbourke]

Hi @RishabhDotasara ,

It's hard to tell what's happening without the rest of your code.

From here the training loop looks to be good to me.

Have you inspected your model/training data and seen if the issue is there?