Is your feature request related to a problem? Please describe.
If the input into the PyTorch classifier wrapper has variable length samples (e.g., the batch is a 1D numpy object array, where the samples (objects) of np arrays of different length/size).
I have not tested it on Keras or TensorFlow, but I am guessing that they may have similar issues.
Describe the solution you'd like
Enable input samples to classifier/estimator wrappers to be of variable size (variable length for audio, variable sizes for images, variables frames and sizes for video); this would also need to extend into the calculation of loss and class gradients for use in attacks.
davidslater
commented
4 years ago
Is your feature request related to a problem? Please describe. If the input into the PyTorch classifier wrapper has variable length samples (e.g., the batch is a 1D numpy object array, where the samples (objects) of np arrays of different length/size).
Places that currently throw errors: https://github.com/IBM/adversarial-robustness-toolbox/blob/master/art/classifiers/pytorch.py#L147 https://github.com/IBM/adversarial-robustness-toolbox/blob/master/art/classifiers/pytorch.py#L193-L194 (I think it also throws errors in gradient calculations as well - basically anywhere
torch.from_numpyshows up.)I have not tested it on Keras or TensorFlow, but I am guessing that they may have similar issues.
Describe the solution you'd like Enable input samples to classifier/estimator wrappers to be of variable size (variable length for audio, variable sizes for images, variables frames and sizes for video); this would also need to extend into the calculation of loss and class gradients for use in attacks.