def compare_acc_f1(old_list, new_list, y_true, stepC, stepV):
pred = compare_pred(old_list, new_list)
# print (" pred: %s" % (pred))
acc = accuracy(y_true, pred)
f1_score = expand_dims_f1(y_true, pred)
# print("concat hc=%s,hv=%s, max accuracy: %f, f1: %f" % (stepC, stepV, acc, f1_score))
return f1_score
def model_predict(model, data):
if data.train_enable:
pred = model.predict(data.get_train_data())
loss = cross_entropy_loss(pred, data.train_y)
print(" pred--loss: %s" % (loss))
acc = accuracy(data.train_y, pred)
print(" pred--acc: %s" % (acc))
f1_score = expand_dims_f1(data.train_y, pred)
print(" pred--f1: %s" % (f1_score))
if data.test_enable:
pred = model.predict(data.get_train_data())
loss = cross_entropy_loss(pred, data.test_y)
print(" pred--loss: %s" % (loss))
acc = accuracy(data.test_y, pred)
print(" pred--acc: %s" % (acc))
f1_score = expand_dims_f1(data.test_y, pred) # 用f1有问题
print(" pred--f1: %s" % (f1_score))
def run_block_model():
data = LoadData(sample_size=None, train_enable=True, test_enable=False)
# data = LoadData(sample_size=None, train_enable=False, test_enable=True)
data.show_data_shape()
### model
# 初始化logging
debug = False
# 训练用旧指标(mean)
# 预测用混合指标, 新指标(max)为False才用旧指标(mean)
# mean='mean'
# mean='max'
mean = 'max_mean'
# 保存比较值(np), 用于外部调参
# 开debug, 代码走call
save_data = True
model = create_model(data.max_vocab_len,
data.max_seq_len,
data.max_modes_len,
h5_file=h5_file,
debug=debug,
mean=mean,
save_data=save_data)
if data.train_enable:
if debug:
# 这样才能打logging
pred = model([data.get_train_data()])
else:
# 速度快
pred = model.predict(data.get_train_data())
loss = cross_entropy_loss(pred, data.train_y)
print(" pred--loss: %s" % (loss))
acc = accuracy(data.train_y, pred)
print(" pred--acc: %s" % (acc))
f1_score = expand_dims_f1(data.train_y, pred)
print(" pred--f1: %s" % (f1_score))
res_string = 'loss=%s, acc=%s, f1_score=%s' % (
np.array(loss), np.array(acc), np.array(f1_score))
save_pred_result(data.train_t1,
data.train_t2,
pred,
data.train_y,
res_string,
name='train_calc')
def old_acc_f1(p_list, y_true, HP_m=None):
pred = old_pred(p_list)
# print (" pred: %s" % (pred))
acc = accuracy(y_true, pred)
f1_score = expand_dims_f1(y_true, pred)
return acc, f1_score
def check_run_stop(step, data, model):
global check_stop_acc, check_stop_f1, cur_stop_step
# 批大小不一样, 单独计算
if data.test_enable:
pred = model.predict(data.get_test_data()) # 感觉这个要快
loss = cross_entropy_loss(pred, data.train_y)
acc = accuracy(data.train_y, pred)
f1_score = expand_dims_f1(data.train_y, pred) # 用f1有问题
if step < 0:
# 训练结束, 显示总指标, 传参数-1:
print("----------------\n predict: accuracy: %f, f1: %f" %
(acc, f1_score))
elif acc > check_stop_acc and f1_score > check_stop_f1:
# 更新f1, 用于判断下次是否保存:
# 不更新acc, 大于0.6就行
check_stop_f1 = f1_score
print("----------------\n update stop step: %s, acc: %s, f1: %s" %
(step + diff_stop_step, check_stop_acc, check_stop_f1))
# 更新weights, 加大结束步长
# 第1步不保存, 效果未知, 避免覆盖上次训练的结果
if cur_stop_step > 0:
print("--------save_weights")
model.save_weights(h5_file.format(data.max_vocab_len),
overwrite=True)
# 渐进6步后退出:
cur_stop_step = step + diff_stop_step
else:
# 退出条件
if cur_stop_step > 0 and step > cur_stop_step:
return 1
elif data.train_enable:
pred = model.predict(data.get_train_data()) # 感觉这个要快
loss = cross_entropy_loss(pred, data.train_y)
acc = accuracy(data.train_y, pred)
f1_score = expand_dims_f1(data.train_y, pred) # 用f1有问题
if step < 0:
# 训练结束, 显示总指标, 传参数-1:
print("----------------\n predict: accuracy: %f, f1: %f" %
(acc, f1_score))
elif acc > check_stop_acc and f1_score > check_stop_f1:
# 更新f1, 判断下次是否保存:
# 不更新acc, 大于0.6就行
check_stop_f1 = f1_score
print(
"----------------\n update stop step: %s, acc > %s, f1 > %s" %
(step + diff_stop_step, check_stop_acc, check_stop_f1))
# 更新weights, 加大结束步长
# 第1步不保存, 效果未知, 避免覆盖上次训练的结果
if cur_stop_step > 0:
print("--------save_weights")
model.save_weights(h5_file.format(data.max_vocab_len),
overwrite=True)
# 渐进6步后退出:
cur_stop_step = step + diff_stop_step
else:
# 退出条件
if cur_stop_step > 0 and step > cur_stop_step:
return 1
return 0
continue
check_ret = 0
if is_use_check_stop and step > start_check_step:
check_ret = check_run_stop(step, data, model)
if check_ret > 0:
break
else:
# Run the optimization to update W and b values.
pred, loss = run_optimization(x1, x2, m1, m2, mi1, mi2, n1, n2, cnt1,
cnt2, y)
if step % display_step == 0 or step == 1:
#loss = cross_entropy_loss(pred, y)
acc = accuracy(y, pred)
f1_score = expand_dims_f1(y, pred) # 用f1有问题
print("step: %i, loss: %f, accuracy: %f, f1: %f" %
(step, loss, acc, f1_score))
if is_use_check_stop == False:
print(" save: %s" % (h5_file.format(data.max_vocab_len)))
model.save_weights(h5_file.format(data.max_vocab_len), overwrite=True)
# 总的acc,f1
check_run_stop(-1, data, model)
#w = new_model.get_weights()
#logging.getLogger().info("--weights\n %s" % (tf.shape(w)))
#logging.getLogger().info("--weights\n %s" % (w))