pertubation_rank.py

def perturbation_rank(model,x,y,names):
    errors = []

    X_saved = x
    y = y.flatten()

    with torch.no_grad():
        model.eval()
        for i in range(x.shape[1]):

            # Convert to numpy, shuffle, convert back to tensor, predict
            x = x.detach().numpy()
            np.random.shuffle(x[:,i])
            x = torch.from_numpy(x).float().to(device)
            pred = model(x)

            # log_loss requires (classification target, probabilities)
            pred = pred.cpu().detach().numpy()
            error = metrics.log_loss(y, pred)
            errors.append(error)

            # Reset x to saved tensor matrix
            x = X_saved
    
    max_error = np.max(errors)
    importance = [e/max_error for e in errors]
    
    data = {'name':names,'error':errors,'importance':importance}
    result = pd.DataFrame(data,columns = ['name','error','importance'])
    result.sort_values(by=['importance'],ascending=[0],inplace=True)
    result.reset_index(inplace=True,drop=True)
    return result

names = list(fractional_diff_data.columns)
names.remove('label_barrier')
rank = perturbation_rank(model_NN1, 
                         torch.from_numpy(X_test).float(),
                         torch.from_numpy(y_test.astype(int)).long(),  
                         names
                         )

display(rank)