Method eval

ppq/samples/QAT/trainer.py:102–132

Do Evaluation process on given dataloader. Split your dataset into training and evaluation dataset at first, then use eval function to monitor model performance on evaluation dataset. Here are some options to prevent overfitting, which helps improve the model pe

(self, dataloader: Iterable)

            return pred, loss.item()

    def eval(self, dataloader: Iterable) -> float:
        """Do Evaluation process on given dataloader.
        
        Split your dataset into training and evaluation dataset at first, then
            use eval function to monitor model performance on evaluation dataset.

        Here are some options to prevent overfitting, which helps improve the model performance.
        1. Train with more data.
        2. Data augmentation.
        3. Addition of noise to the input data.
        """
        total_pred, total_correct = 0, 0
        with torch.no_grad():
            for bidx, batch in enumerate(tqdm(dataloader, desc='Eval: ')):
                if type(batch) not in {tuple, list}:
                    raise TypeError('Feeding Data is invalid, expect a Tuple or List like [data, label], '
                                    f'however {type(batch)} was given. To feed customized data, you have to rewrite '
                                    '"eval" function of PPQ Trainer.')
                if len(batch) != 2:
                    raise ValueError('Unrecognized data format, '
                                     'your dataloader should contains batched data like [data, label]')
                data, label = batch
                data, label = data.cuda(), label.cuda()

                pred, _ = self.step(data, label, False)
                pred_label = torch.argmax(pred, dim=-1)
                total_correct += torch.sum(pred_label == label).item()
                total_pred    += pred_label.numel()

        print(f'Classification Accuracy: {total_correct / total_pred * 100:.3f}%\n')
        return total_correct / total_pred

    def save(self, file_path: str):
        """ Save model to given path.