Why is my test accuracy so low even though Daniel's so so high?

[GodRishUniverse]

This is somewhere between chapter 79 and 84 of his course. My accuracy is:

While Daniel's is:

Here is my code for accuracy:

def accuracy_fn(y_true,y_pred):
  # torch.eq tells how many samples in y_true and y_preds are equal - sums gets the total and item gets it as a single value
  correct = torch.eq(y_true,y_pred).sum().item()
  acc = (correct/(len(y_preds)))*100
  return acc

And here is my code for the Model:

torch.manual_seed(42)
torch.cuda.manual_seed(42)

X_train, y_train, X_test, y_test = X_train.to(device), y_train.to(device), X_test.to(device), y_test.to(device)

epochs =1000

for epoch in range(epoch):
  # Training
  model_3.train()

  #Forward pass
  y_logits= model_3(X_train).squeeze()
  y_preds = torch.round(torch.sigmoid(y_logits))  # logits -> prediction probabilities -> prediction labels

  #computing the loss 
  loss = loss_fn(y_logits,y_train) # BECEwithlogitsloss takes logits as the first input
  #Calculate the Accuracy
  acc = accuracy_fn(y_true= y_train,
                    y_pred= y_preds)

  #Optimizer zero grad
  optimizer.zero_grad()

  #backpropogation
  loss.backward()

  #Optimizer step
  optimizer.step()

  #Training loop 
  model_3.eval()
  with torch.inference_mode():
    test_logits = model_3(X_test).squeeze()
    test_preds = torch.round(torch.sigmoid(test_logits))

    tets_loss = loss_fn(test_logits, y_test)
    test_acc = accuracy_fn(y_true= y_test,
                    y_pred= test_preds)


  if epoch %100 ==0:
    print(f"Epoch: {epoch} | Loss: {loss:.5f} , Acc: {acc:.2f}% | Test Loss: {test_loss:.5f} , Test Acc: {test_acc:.2f}%")

[Yer-Marti]

I'm still learning too so I'm not very confident in myself yet, but I'm pretty sure the code there is fine. You should check whether the optimizer has the right params of your model_3 or model class implementation is okay, as well as previous steps. I recall having the same issue once and it was simply a matter of rerunning the model implementation cells and everything went smooth.

[GodRishUniverse]

I have re-run this multiple times and it is the same. I even checked the parameters entered but nothing is working - the model's graph shows very high accuracy but not these print statements

[Yer-Marti]

I think I got it. In the torch.inference_mode() context for testing you have a typo at tets_loss = .... I think that might be it. What a silly thing right? I hope thats all there is to it :)

[GodRishUniverse]

@Yer-Marti I changed that too but still it shows the same error

[GodRishUniverse]

Nevermind, I resolved it. The problem was in the accuracy function - typo in y_pred:

Corrected version:

def accuracy_fn(y_true,y_pred):
   correct = torch.eq(y_true,y_pred).sum().item()
  acc = (correct/(len(y_pred)))*100
  return acc

Thanks for the help!

[Yer-Marti]

No problem, I'm glad you spotted it