update_freq=args.update_freq)
criterion = torch.nn.CrossEntropyLoss()
best_acc = 0.0
for epoch in range(args.epochs):
track.debug("Starting epoch %d" % epoch)
args.lr = adjust_learning_rate(epoch, optimizer, args.lr,
args.schedule, args.gamma)
train_loss, train_acc = train(trainloader, model, criterion, optimizer,
epoch, args.cuda)
test_loss, test_acc = test(testloader, model, criterion, epoch,
args.cuda)
track.debug('Finished epoch %d... | train loss %.3f | train acc %.3f '
'| test loss %.3f | test acc %.3f' %
(epoch, train_loss, train_acc, test_loss, test_acc))
# Save model
model_fname = os.path.join(track.trial_dir(),
"model{}.ckpt".format(epoch))
torch.save(model, model_fname)
if test_acc > best_acc:
best_acc = test_acc
best_fname = os.path.join(track.trial_dir(), "best.ckpt")
track.debug("New best score! Saving model")
torch.save(model, best_fname)
if __name__ == '__main__':
skeletor.supply_args(add_train_args)
skeletor.execute(do_training)
"""
loads the model and trial data and runs the specified experiment(s)!
"""
# load the project from track
proj = track.Project(args.results_dir)
# create the ensemble
model, trial_df = load_trial(proj, args.start_epoch, args.end_epoch,
args.noise_scale)
# register svhn so we can load it in OOD
sk.datasets.add_dataset(svhn)
# run the experiment
def _run(experiment):
track.debug('Starting to run experiment: %s' % experiment)
experiment_module = 'sgld.experiments.' + experiment
runner = getattr(importlib.import_module(experiment_module), 'run')
runner(model, trial_df, **vars(args))
if args.mode == 'all':
for experiment in EXPERIMENTS:
_run(experiment)
else:
_run(args.mode)
if __name__ == '__main__':
sk.supply_args(make_args)
sk.execute(main)
num_workers=2,
)
if args.fname:
print("Loading model from %s" % args.fname)
model = torch.load(args.fname, map_location="cpu").cuda()
else:
model = build_model("ResNet18", num_classes=10)
criterion = torch.nn.CrossEntropyLoss()
eigenvals, eigenvecs = compute_hessian_eigenthings(
model,
testloader,
criterion,
args.num_eigenthings,
mode=args.mode,
# power_iter_steps=args.num_steps,
max_samples=args.max_samples,
# momentum=args.momentum,
full_dataset=args.full_dataset,
use_gpu=args.cuda,
)
print("Eigenvecs:")
print(eigenvecs)
print("Eigenvals:")
print(eigenvals)
# track.metric(iteration=0, eigenvals=eigenvals)
if __name__ == "__main__":
skeletor.supply_args(extra_args)
skeletor.execute(main)
track.debug("FINAL STATS:%s" % _debug_stats_str(final_stats))
def _add_args(parser):
parser.add_argument('--alg', default='deepq', help='agent to train',
choices=list(ALG_LEARN_FNS.keys()))
parser.add_argument('--env', default='CartPole-v0',
help='Gym environment name to train on')
parser.add_argument('--attack', default='fgsm',
choices=list(ATTACKS.keys()),
help='attack method to run')
parser.add_argument('--network', default='mlp', type=str,
help='policy network arhitecture')
parser.add_argument('--attack_ord', default=2, type=int,
help="norm we use to constrain perturbation size")
parser.add_argument('--num_rollouts', default=5, type=int,
help='how many episodes to run for each attack')
parser.add_argument('--eps', default=.1, type=float,
help='perturbation magnitude')
parser.add_argument('--model_dir', default='./models', type=str,
help='where to look for model pkls by default')
parser.add_argument('--load_path', default='',
help='location of model .pkl with correct policy')
parser.add_argument('--render', action='store_true',
help='if true, render the actual gym env on-screen')
if __name__ == '__main__':
skeletor.supply_args(_add_args)
skeletor.execute(main)