Error running librispeech_asr_imperceptible_undefended.json

[lcadalzo]

File "/opt/conda/lib/python3.8/site-packages/art/attacks/evasion/imperceptible_asr/imperceptible_asr_pytorch.py", line 567, in _attack_2nd_stage[best_loss_2nd_stage[local_batch_size_idx] = loss_2nd_stage[local_batch_size_idx].numpy()
		RuntimeError^[[0m:^[[1m Can't call numpy() on Tensor that requires grad. Use tensor.detach().numpy() instead

python -m armory run scenario_configs/librispeech_asr_imperceptible_undefended.json

[lcadalzo]

Note: I'm unable to reproduce this error manually after observing it from the logs of the batch of runs

[davidslater]

I think this error only shows up when using GPUs. I was able to reproduce.

It looks like it's an ART error in
https://github.com/Trusted-AI/adversarial-robustness-toolbox/blob/main/art/attacks/evasion/imperceptible_asr/imperceptible_asr_pytorch.py#L567

best_loss_2nd_stage[local_batch_size_idx] = loss_2nd_stage[local_batch_size_idx].numpy()

should be

best_loss_2nd_stage[local_batch_size_idx] = loss_2nd_stage[local_batch_size_idx].detach().cpu().numpy()

Looks like there could be a similar issue here at https://github.com/Trusted-AI/adversarial-robustness-toolbox/blob/main/art/attacks/evasion/imperceptible_asr/imperceptible_asr_pytorch.py#L412 , where

                        max_local_delta = np.max(np.abs(local_delta[local_batch_size_idx].detach().numpy()))

should be

                        max_local_delta = np.max(np.abs(local_delta[local_batch_size_idx].detach().cpu().numpy()))

[davidslater]

PR submitted to ART: Trusted-AI/adversarial-robustness-toolbox#1677