Two small issues about the paper and code about learning-data-manipulation HOT 2 CLOSED

Thanks for carefully reading our code.

Let's start from augmentation_main.py-line94, generator.finetune_batch() is the function to update \phi (paper Figure 1).

generator.py-line128: get augmented examples for a training data example (paper Equation 9)

classifier.finetune_generator() includes

• getting \theta'(\phi) = \theta + \nabla_{\theta} L(\theta, \phi) (line 136-142 is computing grads = \nabla_{\theta} L(\theta, \phi)).
• getting gradients of \phi with dev data (line 150-158).

generator.py-line 131: update \phi according collected gradients of \phi from a batch of examples.

As for the magic_module.py, it is for us to get L(\theta'(\phi)) and gradients of \phi:

• grads = \nabla_{\theta} L(\theta, \phi) -- classifier.py-line140
• \theta'(\phi) = \theta + grads -- classifier.py-line148
• L(\theta'(\phi)) -- classifier.py-line154
• \nabla_{\phi} L(\theta'(\phi)) -- classifier.py-line158

Hope these can help you. Please let me know if any part is still confusing you.

from learning-data-manipulation.

Thanks for your help.

from learning-data-manipulation.