def main():
# Load Data
data = np.loadtxt("zipcombo.dat")
X, y = data_split(data, y_col=0)
# Investigate effect of k values
k_list = [1, 3, 7, 9, 27]
xval_k_val = validate_k(X, y, k_list)
np.savetxt('knn_variable_selection.csv',
xval_k_val,
delimiter=',',
fmt='%10.20f')
# Solve quesiton 1
q1_errors = q1_regularised(X, y, k_list)
np.savetxt('knn_q1.csv', q1_errors, delimiter=',', fmt='%10.20f')
n = 20
xval_k_stats, k_max_index = q2_regularised(X, y, n, k_list)
np.savetxt('knn_variable_selection_n_{}.csv'.format(n),
xval_k_stats,
delimiter=',',
fmt='%10.20f')
np.savetxt('knn_k_max_index.csv'.format(n),
k_max_index,
delimiter=',',
fmt='%10.20f')
def main():
# Load Data
data = np.loadtxt("zipcombo.dat")
X, y = data_split(data, y_col=0)
gauss_c = list(range(-6, 3))
gauss_c = [3**i for i in gauss_c]
d_errors, full_errors = kernel_d_selection(VectorisedKernelPerceptron,
X,
y,
k=5,
d_vals=gauss_c,
epochs=8)
q1_train, q1_test = q1(VectorisedKernelPerceptron, X, y, gauss_c, 0.2, 10,
1, 20)
np.savetxt('q1_train_errors_gauss.csv',
q1_train,
delimiter=',',
fmt='%10.20f')
np.savetxt('q1_test_errors_gauss.csv',
q1_test,
delimiter=',',
fmt='%10.20f')
d_prime_errors, confusion = d_hyperparameter_selection(
VectorisedKernelPerceptron,
X,
y,
gauss_c,
k=5,
epochs=10,
seed=1,
runs=20)
np.savetxt('confusion_gauss.csv', confusion, delimiter=',', fmt='%i')
np.savetxt('d_prime_errors_gauss.csv',
d_prime_errors,
delimiter=',',
fmt='%10.20f')
np.savetxt('d_errors_gauss.csv', d_errors, delimiter=',', fmt='%10.20f')
np.savetxt('full_errors_gauss.csv',
full_errors,
delimiter=',',
fmt='%10.20f')
def main():
# Load Data
data = np.loadtxt("zipcombo.dat")
X, y = data_split(data, y_col=0)
poly_d = list(range(1, 8))
d_errors, full_errors = kernel_d_selection(onevsonePerceptron,
X,
y,
k=5,
d_vals=list(range(1, 8)),
epochs=10)
q1_train, q1_test = q1(onevsonePerceptron,
X,
y,
poly_d,
percentage=0.2,
epochs=10,
seed=0,
runs=20)
np.savetxt('q1_train_errors_1v1.csv',
q1_train,
delimiter=',',
fmt='%10.20f')
np.savetxt('q1_test_errors_1v1.csv', q1_test, delimiter=',', fmt='%10.20f')
d_prime_errors = d_hyperparameter_selection(onevsonePerceptron,
X,
y,
d_vals=poly_d,
k=5,
epochs=10,
seed=0,
runs=20)
np.savetxt('d_prime_errors_1v1.csv',
d_prime_errors,
delimiter=',',
fmt='%10.20f')
np.savetxt('d_errors_1v1.csv', d_errors, delimiter=',', fmt='%10.20f')
np.savetxt('full_errors_1v1.csv',
full_errors,
delimiter=',',
fmt='%10.20f')
def train_model(self,
epochs=32,
val_ratio=0.1,
train_batch_size=256,
val_batch_size=None):
# train_batch_size
if val_batch_size is None:
val_batch_size = 2 * train_batch_size
x_train, y_train, x_val, y_val = data_split(self.data,
val_ratio=val_ratio)
train_len = x_train.shape[0]
steps_per_epoch = math.ceil(train_len / train_batch_size)
input_x = self.inputParams['input_ids']
input_x_mask = self.inputParams['input_mask']
input_x_seg = self.inputParams['segment_ids']
input_y = self.inputParams['input_labels']
keep_prob = self.inputParams['keep_prob']
# learning_rate = self.inputParams['learning_rate']
num_train_steps = self.inputParams['num_train_steps']
loss = self.outputParams['loss']
accuracy = self.outputParams['accuracy']
train_op = self.outputParams['train_op']
merged_summary = self.summaryParams['merged_summary']
# 获取模型中所有的训练参数。
tvars = tf.trainable_variables()
# 加载BERT模型
(assignment_map, initialized_variable_names
) = modeling.get_assignment_map_from_checkpoint(tvars, BERT_CKPT)
tf.train.init_from_checkpoint(BERT_CKPT, assignment_map)
gen_train_batch = gen_batch_data(x_train, y_train, train_batch_size)
gen_val_batch = gen_batch_data(x_val, y_val, val_batch_size)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
max_acc = 0
noChangeEpoches = 0
save_saver = tf.train.Saver()
lr = 0.01
for epoch in range(epochs):
print("epoch:{0}".format(epoch))
noChangedSteps = 0
for j in range(steps_per_epoch):
x_train_batch, y_train_batch = next(gen_train_batch)
# processor_x 返回的是list 构成的np.arry
x_train_batch = self.data.processor_x(x_train_batch)
y_train_batch = self.data.processor_y(y_train_batch)
input_ids_batch, input_mask_batch, segment_ids_batch \
= conver2Input(x_train_batch,max_seq_len=256)
fetches = [loss, accuracy, train_op]
feed_dict = {
input_x: input_ids_batch,
input_x_mask: input_mask_batch,
input_x_seg: segment_ids_batch,
input_y: y_train_batch,
keep_prob: 0.8,
# learning_rate:lr,
num_train_steps: epochs * steps_per_epoch,
}
loss_, accuracy_, _ = sess.run(fetches,
feed_dict=feed_dict)
print('当前批次: {}/{}/{} | 当前训练损失: {} | 当前训练准确率： {}'.format(
j,
steps_per_epoch,
epoch,
loss_,
accuracy_,
))
if j % 100 == 0 or j == self.data.get_step() - 1:
x_val_batch, y_val_batch = next(gen_val_batch)
x_val_batch = self.data.processor_x(x_val_batch)
y_val_batch = self.data.processor_y(y_val_batch)
input_ids_val_batch, input_mask_val_batch, segment_ids_val_batch \
= conver2Input(x_val_batch, max_seq_len=256)
val_loss, val_acc = sess.run(
[loss, accuracy],
feed_dict={
input_x: input_ids_val_batch,
input_x_mask: input_mask_val_batch,
input_x_seg: segment_ids_val_batch,
input_y: y_val_batch,
keep_prob: 1,
})
print('当前批次: {}/{}/{} | 当前验证集损失： {} | 当前验证集准确率： {}'.
format(
j,
steps_per_epoch,
epoch,
val_loss,
val_acc,
))
if val_acc > max_acc:
noChangedSteps = 0
max_acc = val_acc
if not os.path.exists(MODEL_PATH):
os.makedirs(MODEL_PATH)
save_path = save_saver.save(sess, self.model_path)
print("Model saved in path: %s" % save_path)
else:
noChangedSteps += 100
# lr = 0.9*lr
if noChangedSteps >= steps_per_epoch:
noChangeEpoches += 1
else:
noChangeEpoches == 0
if noChangeEpoches >= 3:
break
'''
flyai库中的提供的数据处理方法
传入整个数据训练多少轮，每批次批大小
'''
dataset = Dataset(epochs=args.EPOCHS,
batch=args.BATCH,
val_batch=args.VAL_BATCH)
mymodel = Model(dataset)
print("number of train examples:%d" % dataset.get_train_length())
print("number of validation examples:%d" % dataset.get_validation_length())
QA_model = mymodel.QA_model
print(QA_model)
x_train, y_train, x_val, y_val = data_split(dataset, val_ratio=0.1)
train_len = x_train.shape[0]
val_len = x_val.shape[0]
test_x_data = x_val[-args.BATCH:]
test_y_data = y_val[-args.BATCH:]
def show_result(model, que_data, ans_data):
score = 0.0
for i, que in enumerate(que_data):
predict = model.predict(**que)
print("预测结果：%s" % predict)
print("实际答案：%s" % ans_data[i]["ans_text"])
score += sentence_bleu([jieba.lcut(predict)],
jieba.lcut(ans_data[i]["ans_text"]),