AttributeError occurred at amp.scale_loss

Problem

I tried to use amp, and this error occurred

Traceback (most recent call last):
  File "train.py", line 75, in <module>
    main(args)
  File "train.py", line 54, in main
    [model_checkpoint, metrics_logger])
  File "/home/dwaydwaydway/adl/adl-hw1-example-code/src/base_predictor.py", line 75, in fit_dataset
    log_train = self._run_epoch(dataloader, True)
  File "/home/dwaydwaydway/adl/adl-hw1-example-code/src/base_predictor.py", line 182, in _run_epoch
    with amp.scale_loss(batch_loss, self.optimizer) as scaled_loss:
  File "/home/dwaydwaydway/anaconda3/lib/python3.7/contextlib.py", line 112, in __enter__
    return next(self.gen)
  File "/home/dwaydwaydway/anaconda3/lib/python3.7/site-packages/apex/amp/handle.py", line 59, in scale_loss
    if not _amp_state.opt_properties.enabled:
AttributeError: 'AmpState' object has no attribute 'opt_properties'
> /home/dwaydwaydway/anaconda3/lib/python3.7/site-packages/apex/amp/handle.py(59)scale_loss()
-> if not _amp_state.opt_properties.enabled:

This is my code

def _run_epoch(self, dataloader, training):
        # set model training/evaluation mode
        self.model.train(training)

        # run batches for train
        loss = 0

        # reset metric accumulators
        for metric in self.metrics:
            metric.reset()

        if training:
            iter_in_epoch = min(len(dataloader), self.max_iters_in_epoch)
            description = 'training'
        else:
            iter_in_epoch = len(dataloader)
            description = 'evaluating'

        # run batches
        trange = tqdm(enumerate(dataloader),
                      total=iter_in_epoch,
                      desc=description)
        for i, batch in trange:
            if training and i >= iter_in_epoch:
                break

            if training:
                output, batch_loss = \
                    self._run_iter(batch, training)

                batch_loss /= self.grad_accumulate_steps

                # accumulate gradient - zero_grad
                if i % self.grad_accumulate_steps == 0:
                    # TODO: call zero gradient here
                    self.optimizer.zero_grad()

                # TODO: Call backward on `batch_loss` here.
########################################################
#Error Here
########################################################
                with amp.scale_loss(batch_loss, self.optimizer) as scaled_loss:
                    scaled_loss.backward()
#                 batch_loss.backward()
                
                # accumulate gradient - step
                if (i + 1) % self.grad_accumulate_steps == 0:
                    # TODO: update gradient here
                    self.optimizer.step()
            else:
                with torch.no_grad():
                    output, batch_loss = \
                        self._run_iter(batch, training)

            # accumulate loss and metric scores
            loss += batch_loss.item()
            for metric in self.metrics:
                metric.update(output, batch)
            trange.set_postfix(
                loss=loss / (i + 1),
                **{m.name: m.print_score() for m in self.metrics})

        # calculate averate loss and metrics
        loss /= iter_in_epoch

        epoch_log = {}
        epoch_log['loss'] = float(loss)
        for metric in self.metrics:
            score = metric.get_score()
            print('{}: {} '.format(metric.name, score))
            epoch_log[metric.name] = score
        print('loss=%f\n' % loss)
        return epoch_log

Environment

• Ubuntu 18.04
• python version : 3.7.1
• pytorch 1.0.1
• conda version : 4.6.8
• CUDA version: 10.0
• CUDA driver version: 410.48
• GPU: GeForce RTX 2070

Did I install it incorrectly?

$ git clone https://github.com/NVIDIA/apex.git
$ cd apex
$ pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" .

Reproduce the traceback in the provided Python, PyTorch, CUDA, and Apex environment. Start with apex/amp/handle.py at scale_loss and the call in src/base_predictor.py, then compare the AMP setup with the installation command shown. Done means identifying the cause of the missing AmpState attribute and documenting a verified correction or installation path.