'save_file': 'pen_10.pkl.gz',
'pretraining_epochs': 50,
'pretrain_lr': 0.01,
'mu': 0.9,
'finetune_lr': 0.01,
'training_epochs': 50,
'dataset': dataset,
'batch_size': 20,
'nClass': K,
'hidden_dim': [50, 16, 10],
'diminishing': False
}
results = []
for i in range(trials):
res_metrics = test_SdC(**config)
results.append(res_metrics)
results_SAEKM = np.zeros((trials, 3))
results_DCN = np.zeros((trials, 3))
N = config['training_epochs'] / 5
for i in range(trials):
results_SAEKM[i] = results[i][0]
results_DCN[i] = results[i][N]
SAEKM_mean = np.mean(results_SAEKM, axis=0)
SAEKM_std = np.std(results_SAEKM, axis=0)
DCN_mean = np.mean(results_DCN, axis=0)
DCN_std = np.std(results_DCN, axis=0)
print >> sys.stderr, (
ari_sc[n] = metrics.adjusted_rand_score(train_y, ypred)
train_set = train_x, train_y
dataset = [train_set, train_set, train_set]
f = gzip.open('toy.pkl.gz','wb')
cPickle.dump(dataset, f, protocol=2)
f.close()
## Perform non-joint SAE+KM
nmi_nj[n], ari_nj[n] = test_SdC_NJ(lbd = 0, finetune_lr= .01, mu = 0.9, pretraining_epochs=50,
pretrain_lr=.01, training_epochs=100,
dataset='toy.pkl.gz', batch_size=20, nClass = nClass, hidden_dim = [100, 50, 10, 2])
## Perform proposed
nmi_dc[n], ari_dc[n] = test_SdC(lbd = 0.2, finetune_lr= .01, mu = 0.9, pretraining_epochs=50,
pretrain_lr=0.01, training_epochs=100,
dataset='toy.pkl.gz', batch_size=20, nClass = nClass,
hidden_dim = [100, 50, 10, 2])
result = np.concatenate((np.mean(nmi_km, keepdims=True), np.mean(ari_km, keepdims=True), np.mean(nmi_sc, keepdims=True),
np.mean(ari_sc, keepdims=True), np.mean(nmi_nj, keepdims=True), np.mean(ari_nj, keepdims=True),
np.mean(nmi_dc, keepdims=True), np.mean(ari_dc, keepdims=True)) )
f = gzip.open('MC_results.pkl.gz','wb')
cPickle.dump(result, f, protocol=2)
f.close()
from multi_layer_km import test_SdC
res_metrics = test_SdC(dataset='../data/MNIST/toy.pkl.gz',
save_file='toy_4.pkl.gz')
ari_km = metrics.adjusted_rand_score(train_y, ypred)
print >> sys.stderr, ('NMI for Kmeans: %.2f' % (nmi_km))
print >> sys.stderr, ('ARI for Kmeans: %.2f' % (ari_km))
train_set = train_x, train_y
dataset = [train_set, train_set, train_set]
f = gzip.open('toy.pkl.gz', 'wb')
cPickle.dump(dataset, f, protocol=2)
f.close()
nmi_dc, ari_dc = test_SdC(lbd=.1,
finetune_lr=.05,
mu=0.9,
pretraining_epochs=50,
pretrain_lr=0.01,
training_epochs=100,
dataset='toy.pkl.gz',
batch_size=20,
nClass=nClass,
hidden_dim=[100, 50, 10, 2])
#
print >> sys.stderr, ('NMI for spectral clustering: %.2f' % (nmi_sc))
print >> sys.stderr, ('ARI for spectral clustering: %.2f' % (ari_sc))
print >> sys.stderr, ('NMI for deep clustering: %.2f' % (nmi_dc))
print >> sys.stderr, ('ARI for deep clustering: %.2f' % (ari_dc))
#nmi_nj, ari_nj = test_SdC_NJ(lbd = 0, finetune_lr= .01, mu = 0.9, pretraining_epochs=50,
# pretrain_lr=.01, training_epochs=100,
# dataset='toy.pkl.gz', batch_size=20, nClass = nClass, hidden_dim = [100, 50, 10, 2])
#
'output_dir': 'Pendigits',
'save_file': 'pen_10.pkl.gz',
'pretraining_epochs': 50,
'pretrain_lr': 0.01,
'mu': 0.9,
'finetune_lr': 0.01,
'training_epochs': 50,
'dataset': dataset,
'batch_size': 20,
'nClass': K,
'hidden_dim': [50, 16, 10],
'diminishing': False}
results = []
for i in range(trials):
res_metrics = test_SdC(**config)
results.append(res_metrics)
results_SAEKM = np.zeros((trials, 3))
results_DCN = np.zeros((trials, 3))
N = config['training_epochs']/5
for i in range(trials):
results_SAEKM[i] = results[i][0]
results_DCN[i] = results[i][N]
SAEKM_mean = np.mean(results_SAEKM, axis = 0)
SAEKM_std = np.std(results_SAEKM, axis = 0)
DCN_mean = np.mean(results_DCN, axis = 0)
DCN_std = np.std(results_DCN, axis = 0)
print >> sys.stderr, ('KM avg. NMI = {0:.2f}, ARI = {1:.2f}, ACC = {2:.2f}'.format(KM_mean[0],
