def save_clf(args, Z, acc):
""" gross """
if args.dataset == 'mnist':
import models.mnist_clf as models
model = models.Small2().cuda()
elif args.dataset == 'cifar':
import models.cifar_clf as models
model = models.MedNet().cuda()
""" end gross """
state = model.state_dict()
layers = zip(args.stat['layer_names'], Z)
for i, (name, params) in enumerate(layers):
name = name + '.weight'
loader = state[name]
state[name] = params.detach()
assert state[name].equal(loader) == False
model.load_state_dict(state)
#import cifar
#ac, loss = cifar.test(args, model, 0)
#print ('acc: {}, loss: {}'.format(ac, loss))
path = 'exp_models/hyper{}_clf_{}_{}.pt'.format(args.dataset, args.exp, acc)
if args.scratch:
path = '/scratch/eecs-share/ratzlafn/HyperGAN/' + path
print ('saving hypernet to {}'.format(path))
torch.save({'state_dict': model.state_dict()}, path)
def get_network(args):
if args.net == 'small':
model = models.Small().cuda()
elif args.net == 'small2':
model = models.Small2().cuda()
else:
raise NotImplementedError
return model
pop_mean = pop_outputs.mean(0).view(10000, 10)
ent = entropy(pop_mean.cpu().numpy().T)
return ent
def get_stats(l1, l2, l3, inspect):
norms = []
for i in range(len(inspect)):
norms.append(np.linalg.norm(inspect[i].detach()))
m, s = np.array(norms).mean(), np.array(norms).std()
return m, s
args = arg.load_mnist_args()
model = mnist_clf.Small2().cuda()
modeldef = netdef.nets()[args.target]
names = modeldef['layer_names']
paths = natsorted(glob('saved_models/mnist/noadds/*.pt'))
paths = sorted(paths, key=lambda i: float(i[44:-3]))
accs, losses, cv = [], [], []
for path in paths:
netE, netD, W1, W2, W3 = utils.load_hypernet_mnist(args, path)
l1, l2, l3, codes = utils.sample_hypernet_mnist(args, [netE.eval(), W1.eval(), W2.eval(), W3.eval()], 1)
acc, loss = test_mnist(args, [l1, l2, l3], names, model, 1)
m, s = get_stats(l1, l2, l3, inspect=l2)
print ('Acc: {}, Loss: {} Nmean: {} Nstd: {}\n'.format(acc, loss, m, s))
accs.append(acc)
losses.append(loss)
cv.append(s/m)
accs = np.array(accs)