Didn't set model to eval mode when validating in chapter 8

[tianlianghai]

In the book and code when we first met resnet, the model has some batch norm layers.
But in validate function, it didn't call model.eval(), I don't think this is true.

as a result, in my experiment, while I call model.eval() in the validating function, the result accuracy is lower and loss is higher

def val(model, val_loader, loss_fn):
    model.eval()
    losses = []
    correct = 0
    total = 0 
    for x, y in val_loader:
        with torch.no_grad():
            x, y = x.to(device), y.to(device)
            
            out = model(x)
            loss = loss_fn(out, y)
            losses.append(loss.item())
            _, pred = torch.max(out, dim=1)
            correct += int(torch.sum(pred == y))
            total += y.shape[0]
    loss = sum(losses) / len(losses)
    acc = correct/total
    print(f"val loss:{loss:.3f}, acc:{acc:.3f}")
    model.train()

if I add the model.val() line in the code , the validation accuracy even goes done during training, almost always 50%

while I don't add the model.eval() line the test accuracy actually reached about 85%.
Shouldn't we set model.eval() in validating?

[evenbrenden]

I noticed this too and found https://livebook.manning.com/forum?product=huang&comment=576297. It looks like the model implementation for NetResDeep is incorrect, which is masked by not calling model.eval(). Fixing the model and calling model.eval() before validation should give the true accuracies, which are lower than the ones in the book, at least for that particular model.