def transfer(args):
load_model = args.load_model if args.mode == 'train' else True
class Dummy:
pass
train_env = Dummy()
infer_env = Dummy()
_, infer_graph = model(args, train_env, infer_env)
args.data = 'overnight'
args.load_data = True
X_tran, y_tran = load_data(args)
args.data = 'overnight_set'
tran_sets = load_data(args)
model2load = 'model/{}'.format(args.subset)
sess = tf.InteractiveSession(graph=infer_graph)
infer_env.infer_saver.restore(sess, model2load)
print('========subset transfer set========')
subsets = ['basketball', 'calendar', 'housing', 'recipes', 'restaurants']
for subset, (X_tran_subset, y_tran_subset) in zip(subsets, tran_sets):
print('---------' + subset + '---------')
tran_em = decode_data(sess,
infer_env,
args,
X_tran_subset,
y_tran_subset,
filename=str(subset + '.txt'))
print('===========transfer set============')
tran_em = decode_data(sess, infer_env, args, X_tran, y_tran)
return
def inferrence(args):
args.load_model = True
class Dummy:
pass
train_env = Dummy()
infer_env = Dummy()
_, infer_graph = model(args, train_env, infer_env)
args.data = 'wikisql'
args.load_data = True
X_train, y_train, X_test, y_test, X_dev, y_dev = load_data(args)
model2load = 'model/{}'.format(args.subset)
sess = tf.InteractiveSession(graph=infer_graph)
infer_env.infer_saver.restore(sess, model2load)
print('===========dev set============')
decode_data(sess, infer_env, args, X_dev, y_dev)
em = decode_data_recover(sess, infer_env, args, X_dev, y_dev, 'dev')
print('==========test set===========')
decode_data(sess, infer_env, args, X_test, y_test)
test_em = decode_data_recover(sess, infer_env, args, X_test, y_test,
'test')
return
def train_model(args):
class Dummy:
pass
train_env = Dummy()
infer_env = Dummy()
train_graph, infer_graph = model(args, train_env, infer_env)
args.data = 'geo'
args.load_data = True
args.load_model = False
X_train, y_train, X_test, y_test, X_dev, y_dev = load_data(args)
model2load = 'model/{}'.format(args.subset)
max_em, global_test_em, best_base = -1, -1, -1
acc = 0
sess1 = tf.InteractiveSession(graph=train_graph)
sess1.run(tf.global_variables_initializer())
sess1.run(tf.local_variables_initializer())
sess2 = tf.InteractiveSession(graph=infer_graph)
sess2.run(tf.global_variables_initializer())
sess2.run(tf.global_variables_initializer())
for base in range(args.total_epochs / args.epochs):
print('\nIteration: %d (%d epochs)' % (base, args.epochs))
model2load, acc = train(
sess1,
train_env,
X_train,
y_train,
epochs=args.epochs,
load=args.load_model,
name=args.subset,
batch_size=args.batch_size,
base=base,
acc=acc,
model2load=model2load)
args.load_model = True
if acc > 0:
infer_env.infer_saver.restore(sess2, model2load)
print('===========dev set============')
dev_em = decode_data(sess2, infer_env, args, X_dev, y_dev)
#dev_em = decode_data_recover(sess, infer_env, args, X_dev, y_dev,
# 'dev')
print('==========test set===========')
test_em = decode_data(sess2, infer_env, args, X_test, y_test)
#test_em = decode_data_recover(sess, infer_env, args, X_test, y_test,
# 'test')
if dev_em > max_em:
max_em = dev_em
global_test_em = test_em
best_base = base
print('\n Saving model for best testing')
train_env.saver.save(sess1, 'model/{0}-{1}-{2}'.format(args.subset, base, max_em))
print('Max EM acc: %.4f during %d iteration.' % (max_em, best_base))
print('test EM acc: %.4f ' % global_test_em)
return
def train_model(args):
class Dummy:
pass
train_env = Dummy()
infer_env = Dummy()
train_graph, infer_graph = model(args, train_env, infer_env)
args.data = 'wikisql'
args.load_data = True
args.load_model = False
X_train, y_train, X_test, y_test, X_dev, y_dev = load_data(args)
model2load = 'model/{}'.format(args.subset)
max_em, global_test_em, best_base = -1, -1, -1
for base in range(args.total_epochs / args.epochs):
print('\nIteration: %d (%d epochs)' % (base, args.epochs))
sess = tf.InteractiveSession(graph=train_graph)
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
model2load = train(sess,
train_env,
X_train,
y_train,
epochs=args.epochs,
load=args.load_model,
name=args.subset,
batch_size=args.batch_size,
base=base,
model2load=model2load)
args.load_model = True
break
sess = tf.InteractiveSession(graph=infer_graph)
infer_env.infer_saver.restore(sess, model2load)
print('===========dev set============')
decode_data(sess, infer_env, X_dev, y_dev)
dev_em = decode_data_recover(sess, infer_env, args, X_dev, y_dev,
'dev')
print('==========test set===========')
decode_data(sess, infer_env, X_test, y_test)
test_em = decode_data_recover(sess, infer_env, args, X_test, y_test,
'test')
if em > max_em:
max_em = em
global_test_em = test_em
best_base = base
print('Max EM acc: %.4f during %d iteration.' % (max_em, best_base))
print('test EM acc: %.4f ' % global_test_em)
return