model.print_model_eval()
cur_params, cur_margins = model.sess.run(
[model.params, model.margin], feed_dict=model.all_train_feed_dict)
cur_influences = model.get_influence_on_test_loss(
test_indices=[test_idx],
train_idx=np.arange(num_train),
force_refresh=False)
params[counter, :] = np.concatenate(cur_params)
margins[counter, :] = cur_margins
influences[counter, :] = cur_influences
if temp == 0:
actual_loss_diffs[counter, :], predicted_loss_diffs[
counter, :], indices_to_remove[
counter, :] = experiments.test_retraining(
model,
test_idx,
iter_to_load=0,
force_refresh=False,
num_steps=2000,
remove_type='maxinf',
num_to_remove=num_to_remove)
np.savez('output/hinge_results',
temps=temps,
indices_to_remove=indices_to_remove,
actual_loss_diffs=actual_loss_diffs,
predicted_loss_diffs=predicted_loss_diffs,
influences=influences)
data_sets=data_sets,
initial_learning_rate=initial_learning_rate,
keep_probs=keep_probs,
decay_epochs=decay_epochs,
mini_batch=False,
train_dir='output',
log_dir='log',
model_name='spam_logreg_lbfgs')
tf_model.train()
test_idx = 8
actual_loss_diffs, predicted_loss_diffs_cg, indices_to_remove = experiments.test_retraining(
tf_model,
test_idx,
iter_to_load=0,
force_refresh=False,
num_to_remove=500,
remove_type='maxinf',
random_seed=0)
# LiSSA
np.random.seed(17)
predicted_loss_diffs_lissa = tf_model.get_influence_on_test_loss(
[test_idx],
indices_to_remove,
approx_type='cg',
approx_params={
'scale': 25,
'recursion_depth': 5000,
'damping': 0,
'batch_size': 1,
train_dir='output',
log_dir='log',
model_name='cifar_all_cnn_c')
num_steps = 500000
model.train(num_steps=num_steps,
iter_to_switch_to_batch=10000000,
iter_to_switch_to_sgd=10000000)
iter_to_load = num_steps - 1
test_idx = 6
actual_loss_diffs, predicted_loss_diffs, indices_to_remove = experiments.test_retraining(
model,
test_idx=test_idx,
iter_to_load=iter_to_load,
num_to_remove=100,
num_steps=30000,
remove_type='maxinf',
force_refresh=True)
np.savez('output/cifar_all_cnn_c_iter-500k_retraining-100.npz',
actual_loss_diffs=actual_loss_diffs,
predicted_loss_diffs=predicted_loss_diffs,
indices_to_remove=indices_to_remove)
# Load the trained model
model.load_checkpoint(499999)
# compute influence values for the set of test points
test_indices = [6]
