lfxai.utils.influence module

This code is adapted from https://github.com/nimarb/pytorch_influence_functions/

get_numpy_parameters(model)

Recovers the parameters of a pytorch model in numpy format

hessian_vector_product(loss, model, v)

Multiplies the Hessians of the loss of a model with respect to its parameters by a vector v. Adapted from: https://github.com/kohpangwei/influence-release

This function uses a backproplike approach to compute the product between the Hessian and another vector efficiently, which even works for large Hessians with O(p) compelxity for p parameters.

Parameters

• loss – scalar/tensor, for example the output of the loss function

• model – the model for which the Hessian of the loss is evaluated

• v – list of torch tensors, rnn.parameters(), will be multiplied with the Hessian

Returns

list of torch tensors, contains product of Hessian and v.

Return type

return_grads

stack_torch_tensors(input_tensors)

Takes a list of tensors and stacks them into one tensor