Pytorch IF project

To run fenchel conjugate on GMM or CIFAR

python fenchel_GMM_main.py or python fenchel_CIFAR_main.py

• change the os.chdir() in the main file accordingly
• change YAML_PATH for hyperparameter tuning

TO run the baselines (under construction)

First run src/train.py to train your model, it will be saved to checkpoints
Then Run main.py to compute IF

problems with tracIn

• What should the save_step, max_iter be?

problemswith Hvp

• r, recursion_depth, scale, damp in hvp_grad need to be fixed
  • see ryokamoi/pytorch_influence_functions#2

Debug commands

x = torch.cuda.FloatTensor([[-1],[0],[1],[-1],[0],[1]]) y = torch.cuda.LongTensor([0,0,0,1,1,1]) self._influence_model(x,y)

import numpy as np idx = np.argwhere(train_labels.cpu() == 3) self._influence_model(train_inputs, train_labels)[idx] np.where(self._influence_model(train_inputs, train_labels).squeeze().cpu() > 0.2)