#120 Making predictions using the first CNN

[niktheod]

When we wrote the train and test step functions to calculate the accuracy we passed the raw logits only through torch.argmax() and not through torch.softmax(). This code is a part from the train_step function from Daniel's notebook and you can see that we pass as an argument to the accuracy function, the y_pred.argmax(dim=1), where y_pred is actually the y_logits.

loss = loss_fn(y_pred, y)
train_loss += loss # accumulate train loss
train_acc += accuracy_fn(y_true=y,
                         y_pred=y_pred.argmax(dim=1)) # go from logits -> prediction labels

Later at 18:47:30 when we wrote a function to make predictions with the already trained CNN model we wrote the following code:

def make_predictions(model: torch.nn.Module,
                 data: list,
                 device: torch.device = device):
pred_probs = []
model.to(device)
model.eval()
with torch.inference_mode():
  for sample in data:
    # Prepare the sample (add a batch dimension and pass to target device)
    sample = torch.unsqueeze(sample, dim=0).to(device)

    # Forward pass (model outputs raw logits)
    pred_logit = model(sample)

    # Get prediction probability (logit -> prediction probability)
    pred_prob = torch.softmax(pred_logit.squeeze(), dim=0)

    # Get pred_prob off the GPU for further calculations
    pred_probs.append(pred_prob.cpu())

# Stack the pred_probs to turn list into a tensor
return torch.stack(pred_probs)

Here you can see that we passed the logits first from torch.softmax() and then later to make the predictions we passed it also through torch.argmax()
Why is this? Why in the first case we didn't use torch.softmax() but then we did?
Btw I made predictions by passing the raw outcome of the model only through torch.argmax() and it worked very well.

[niktheod]

Nvm I just saw that someone else asked the same thing rexently.