def main():
experiment_num, sub_num, rule_num, data_name = 10, 10, 32, "ecoli"
data, target, att_dim, res_dim = dataset_numeric_classification(
data_name, 1)
Ot = 0.5 * (np.nanmax(data, 0) + np.nanmin(data, 0))
Dt = 0.5 * (np.nanmax(data, 0) - np.nanmin(data, 0))
data = ((data - Ot) / Dt).astype(np.float64)
acc, cnt = tf.metrics.Accuracy(), 0
for en in range(experiment_num):
for train_data, train_target, test_data, test_target in kfold(
data, target, 10, 'numeric', random_state=en):
train_target_one_hot = tf.one_hot(train_target,
res_dim,
dtype=tf.float64)
sub_models, preds = [], []
model = training(rule_num, att_dim, res_dim, train_data,
train_target_one_hot, 0)
sub_models.append(model), preds.append(model.output(train_data))
cce = tf.keras.losses.categorical_crossentropy(
train_target_one_hot, tf.nn.softmax(preds[0]))
mean_cce, now_mask = tf.reduce_mean(cce).numpy(), goss(
tf.argsort(tf.negative(cce)))
tf.summary.scalar('cce', mean_cce, 0)
now_train, now_target = train_data[now_mask], (
train_target_one_hot -
model.predict(train_data)).numpy()[now_mask]
pre_mean = mean_cce
for _ in range(sub_num - 1):
model = training(rule_num, att_dim, res_dim, now_train,
now_target, 1, 64, 500)
preds.append(0.7 * model.output(train_data))
now_pred = tf.nn.softmax(tf.reduce_sum(preds, 0))
cce = tf.keras.losses.categorical_crossentropy(
train_target_one_hot, now_pred)
mean_cce = tf.reduce_mean(cce).numpy()
tf.summary.scalar('cce', mean_cce, _ + 1)
if mean_cce > pre_mean:
break
if mean_cce < 0.05:
break
sub_models.append(model)
pre_mean, now_mask = mean_cce, goss(
tf.argsort(tf.negative(cce)))
now_train, now_target = train_data[now_mask], (
train_target_one_hot - now_pred).numpy()[now_mask]
test_pred = tf.math.argmax(
tf.nn.softmax(
tf.reduce_sum(
[(1.0 - (mi != 0) * 0.3) * model.output(test_data)
for mi, model in enumerate(sub_models)], 0)), -1)
acc.update_state(test_target, test_pred)
tf.summary.scalar('acc_%s' % data_name, acc.result().numpy(), cnt)
cnt += 1
def main():
experiment_num = 10
data, target, att_dim, res_dim = dataset_numeric_classification('ecoli', 1)
acc = tf.metrics.Accuracy()
cnt = 0
for en in range(experiment_num):
for train_data, train_target, test_data, test_target in kfold(data, target, 10, 'numeric', random_state=en):
O, D = 0.5*(np.min(train_data, 0) + np.max(train_data, 0)), 0.5 * np.ptp(train_data, 0)
D = np.where(D == 0.0, 0.5, D)
model = training(32, 4, 16, att_dim, 4, res_dim, (train_data - O) / D, np.eye(res_dim)[train_target])
test_pred = tf.math.argmax(model((test_data - O) / D), -1)
acc.update_state(test_target, test_pred)
with sw.as_default():
tf.summary.scalar('ecoli_16*4_32_64', acc.result().numpy(), cnt)
cnt += 1
def main():
experiment_num = 20
data, target, att_dim, res_dim = dataset_numeric_classification('ecoli', 1)
# mass = pandas.read_csv('mass.csv').to_numpy()
# data, target, att_dim, res_dim = mass[:, :5], mass[:, -1], 5, 2
data = StandardScaler().fit_transform(data).astype(ntype)
acc, cnt = tf.metrics.Accuracy(), 0
for en in range(experiment_num):
for train_data, train_target, test_data, test_target in kfold(
data, target, 10, 'numeric', random_state=en):
model = training(32, att_dim, res_dim, train_data,
tf.one_hot(train_target, res_dim))
acc.update_state(test_target, model.evaluate(test_data))
tf.summary.scalar('acc', acc.result().numpy(), cnt)
cnt += 1
def main():
experiment_num, data_name = 20, "ecoli"
rule_num, epoch, batch_size = 32, 4000, 128
data, target, att_dim, res_dim = dataset_numeric_classification(
data_name, 1)
data = StandardScaler().fit_transform(data).astype(ndtype)
sum_acc, sum_cnt = 0.0, 0
for en in range(experiment_num):
for train_data, train_target, test_data, test_target in kfold(
data, target, 10, 'numeric', random_state=en):
model = training(train_data, train_target, rule_num, att_dim,
res_dim, epoch, batch_size, en)
_, acc = model.evaluate(test_data, test_target)
sum_acc += acc
sum_cnt += 1
tf.summary.scalar('acc_each_%s' % data_name, acc, sum_cnt)
tf.summary.scalar('acc_%s' % data_name, sum_acc / sum_cnt, sum_cnt)
def main():
experiment_num = 10
data, target, att_dim, res_dim = dataset_numeric_classification('yeast', 1)
acc, cnt = tf.metrics.Accuracy(), 0
for en in range(experiment_num):
for train_data, train_target, test_data, test_target in kfold(data, target, 10, 'numeric', random_state=en):
Ot = 0.5 * (np.nanmax(train_data, 0) + np.nanmin(train_data, 0))
Dt = 0.5 * (np.nanmax(train_data, 0) - np.nanmin(train_data, 0))
Dt = np.where(Dt == 0.0, 0.5, Dt)
train_data, test_data = (train_data - Ot) / Dt, (test_data - Ot) / Dt
sub_models, ros = [], RandomOverSampler(random_state=en)
for rd in range(res_dim):
x, y = ros.fit_resample(train_data, train_target == rd)
sub_models.append(training(16, att_dim, 2, x, np.eye(2)[y.astype(int)]))
modelz = trainingz(32, att_dim, res_dim, res_dim, sub_models, train_data, np.eye(res_dim)[train_target])
test_pred = evaluating(modelz, sub_models, test_data)
acc.update_state(test_pred, test_target)
tf.summary.scalar('acc', acc.result().numpy(), cnt)
cnt += 1
def main():
experiment_num, rule_num, data_name = 50, 32, 'thyroid-new'
data, target, att_dim, res_dim = dataset_numeric_classification(
data_name, 1)
data = StandardScaler().fit_transform(data).astype(np.float32)
acc, cnt = tf.metrics.Accuracy(), 0
missing_rate = 0.5
for en in range(experiment_num):
for train_data, train_target, test_data, test_target in kfold(
data, target, 5, 'numeric', en):
model = training(rule_num,
att_dim,
res_dim,
train_data,
tf.one_hot(train_target, res_dim),
missing_rate=missing_rate)
acc.update_state(
test_target,
model.evaluate(
random_replace_with_nan(test_data,
missing_rate).astype(np.float32)))
tf.summary.scalar('acc_%s' % data_name, acc.result().numpy(), cnt)
cnt += 1
def main():
data_name = [
'iris', 'wine', 'diabetes', 'ecoli', 'glass', 'seeds', 'yeast',
'thyroid-new', 'blood-transfusion-service-center'
]
res = []
missing_rate = 0.0
for dn in data_name:
data, target, att_dim, res_dim = dataset_numeric_classification(dn, 1)
experiment_num, acc = 50, 0
for en in range(experiment_num):
p = Pool()
r = [
p.apply_async(run, args=(datai, missing_rate))
for datai in kfold(data, target, 5, 'numeric', en)
]
p.close(), p.join()
acc += np.mean([_.get() for _ in r]) / experiment_num
print('now_round_%d' % en, end='\r')
with open('res', 'a') as f:
f.writelines(dn + str(acc) + '\n')
res.append((dn, acc))
print(res)