return (np.argmax(tpy, axis=1) == np.argmax(
ty, axis=1)).sum() * 1.0 / tpy.shape[0]
# load the data: x, y, tx, ty, graph
NAMES = ['x', 'y', 'tx', 'ty', 'graph']
objects = {}
for name in NAMES:
data = pkl.load(open("data/trans.{}.{}".format(DATASET, name), 'rb'),
encoding='latin1')
objects[name] = data
# objects.append(cPickle.load(open("data/trans.{}.{}".format(DATASET, name))))
# x, y, tx, ty, graph = tuple(objects)
# initialize the model
m = model(args)
# add data
m.add_data(objects['x'], objects['y'], objects['graph'])
# build the model
m.build()
# pre-training
m.init_train(init_iter_label=2000, init_iter_graph=70)
iter_cnt, max_accu = 0, 0
for _ in range(10000):
# perform a training step
m.step_train(max_iter=1, iter_graph=0, iter_inst=1, iter_label=0)
# predict the dev set
tpy = m.predict(objects['tx'])
# compute the accuracy on the dev set
accu = comp_accu(tpy, objects['ty'])
print(iter_cnt, accu, max_accu)
def comp_accu(tpy, ty):
import numpy as np
return (np.argmax(tpy, axis=1) == np.argmax(
ty, axis=1)).sum() * 1.0 / tpy.shape[0]
# load the data: x, y, tx, ty, graph
NAMES = ['x', 'y', 'tx', 'ty', 'graph']
OBJECTS = []
for i in range(len(NAMES)):
OBJECTS.append(
cPickle.load(open("data/trans.{}.{}".format(DATASET, NAMES[i]))))
x, y, tx, ty, graph = tuple(OBJECTS)
m = model(args) # initialize the model
m.add_data(x, y, graph) # add data
m.build() # build the model
m.init_train(init_iter_label=2000, init_iter_graph=70) # pre-training
iter_cnt, max_accu = 0, 0
while True:
m.step_train(max_iter=1, iter_graph=0, iter_inst=1,
iter_label=0) # perform a training step
tpy = m.predict(tx) # predict the dev set
accu = comp_accu(tpy, ty) # compute the accuracy on the dev set
print iter_cnt, accu, max_accu
iter_cnt += 1
if accu > max_accu:
m.store_params() # store the model if better result is obtained
max_accu = max(max_accu, accu)
parser.add_argument('--use_feature', help = 'whether use input features', type = bool, default = True)
parser.add_argument('--update_emb', help = 'whether update embedding when optimizing supervised loss', type = bool, default = True)
parser.add_argument('--layer_loss', help = 'whether incur loss on hidden layers', type = bool, default = True)
args = parser.parse_args()
def comp_accu(tpy, ty):
import numpy as np
return (np.argmax(tpy, axis = 1) == np.argmax(ty, axis = 1)).sum() * 1.0 / tpy.shape[0]
# load the data: x, y, tx, ty, graph
NAMES = ['x', 'y', 'tx', 'ty', 'graph']
OBJECTS = []
for i in range(len(NAMES)):
OBJECTS.append(cPickle.load(open("data/trans.{}.{}".format(DATASET, NAMES[i]))))
x, y, tx, ty, graph = tuple(OBJECTS)
m = model(args) # initialize the model
m.add_data(x, y, graph) # add data
m.build() # build the model
m.init_train(init_iter_label = 2000, init_iter_graph = 70) # pre-training
iter_cnt, max_accu = 0, 0
while True:
m.step_train(max_iter = 1, iter_graph = 0, iter_inst = 1, iter_label = 0) # perform a training step
tpy = m.predict(tx) # predict the dev set
accu = comp_accu(tpy, ty) # compute the accuracy on the dev set
print iter_cnt, accu, max_accu
iter_cnt += 1
if accu > max_accu:
m.store_params() # store the model if better result is obtained
max_accu = max(max_accu, accu)