Interesting, I thought averaging rmse of all batches will give the same results as store the predictions and then calculate the rmse. But I was wrong. I just updated the check_accuracy and hopefully fix the mistake.

This is the updated check accuracy, can you review it and see if I'm checking the accuracy the right way:

def check_acc_update(loader, model):    
    all_preds = np.array([]); all_groud_truth = np.array([])

    model.eval()

    with torch.no_grad():

        for x,y in tqdm(loader):

            x = x.to('cuda').to(torch.float32)

            y = y.to(torch.float).unsqueeze(1)

            all_preds       = np.append(all_preds, (model(x).cpu()))

            all_groud_truth = np.append(all_groud_truth, (y.cpu()))

    loss = np.sqrt(mean_squared_error(all_preds, all_groud_truth))

    print(f'Loss function: {loss}')

    model.train()

    return loss

Thanks for the feadback @Reacher.

yes it is ok now !

hi @TAUIL_Abdelilah ,

your function was false before because you tend to return the loss of only the last batch, and it is the same to return the average of each batch loss or the loss of all predictions with a reduction="mean".

and this function does't return the accuracy (the fraction of correctly predicted divided by the total of predictions), here you are looking for the validation loss (!=train loss: you can't use numpy to calculate loss for a grading train function).

I didn't get semi_data.csv, I don't use any bounding box in the notebook.

I use only the data that zindi provide.