Function eval_model

pytorch-model/regression_eval.py:23–89

Evaluates model over validation set.

(model, val_dataloader, criterion, args)

def eval_model(model, val_dataloader, criterion, args):
    """
    Evaluates model over validation set.
    """
    
    # --- Buckets for seeing how close the classification *would* be ---
    bins = [-1800, -15, -5, 0, 5, 15, 1800]
    total_correct = 0

    print("\n" + ("-" * 30) + " Evaluating model! " + ("-" * 30))
    total_loss = 0
    total_residual = 0
    total_examples = 0

    avg_loss = None
    avg_residual = None

    model.eval()

    with torch.no_grad():
        for idx, (x, y) in tqdm(enumerate(val_dataloader), total=len(val_dataloader)):

            # --- Move to GPU ---
            # x = x.cuda(constants.GPU, non_blocking=True)
            # y = y.cuda(constants.GPU, non_blocking=True)

            # --- Compute logits ---
            print(x.shape)
            print(x[0])
            print(x)
            print(y)
            logits = model(x)
            loss = criterion(y, logits)

            # --- Bookkeeping ---
            residuals = torch.abs(y - logits)

            total_loss += loss.item()
            total_residual += torch.sum(residuals).item()
            total_examples += y.shape[0]

            avg_loss = total_loss / total_examples
            avg_residual = total_residual / total_examples
            
            # --- Computing the presumed classification accuracy ---
            label_diff_buckets = list()
            for label in y:
                for bin_idx in range(len(bins) - 1):
                    if label >= bins[bin_idx] and label < bins[bin_idx + 1]:
                        label_diff_buckets.append(bin_idx)
                        break
            label_diff_buckets = torch.from_numpy(np.asarray(label_diff_buckets))

            output_diff_buckets = list()
            for pred_price_diff in logits:
                print(pred_price_diff)
                for bin_idx in range(len(bins) - 1):
                    if pred_price_diff >= bins[bin_idx] and pred_price_diff < bins[bin_idx + 1]: