audio_adversarial_examples
audio_adversarial_examples copied to clipboard
The cost output seems to be incorrect.
I'm getting nan cost output during running attack.py. The docker image aae_deepspeech_041_gpu is used and the test is performed on GeForce RTX 3090 with cuda version 11.2.
You might need to upgrade the CUDA version that is used inside the docker container in order for it to work on the RTX 3090.
I have the same error. Could you please give me a hint on how to upgrade the CUDA version inside the docker? Thanks a lot...
In the docker directory in the gpu dockerfile, there is a line at the top which specifies the version number of the tensorflow docker base image. If you change that number to a newer version it will not only use a newer version of tensorflow, but also a newer cuda version if you increase the version number far enough. However increasing the version number will eventually break the code, so I'm not sure that this will work. I would fix it myself, but I don't have a 3090 to test.
It may be very complicated to upgrade the TensorFlow version to 2.5.
Because DeepSpeech
has no TensorFlow 2 support, we will meet a lot of problems if we want to upgrade TensorFlow version.
Jumping on this late... nvidia-tensorflow is patching TF v1.x for updates to CUDA.
The nvcr.io/nvidia/tensorflow:22.03-tf1-py3 image contains the latest patched version, but requires an NGC account to pull the image (sign up with email from what I remember, no charges).
YMMV as I don't have a 3090 to test, and the NVIDIA images are massive, so expect to end up with a very large image!