代码 :
import numpy as np import pandas as pdimport osfrom tqdm import tqdm# 固定种子from numpy.random import seedseed(639)from tensorflow import set_random_seedset_random_seed(5944)# 导入float_data = pd.read_csv("train.csv", dtype={"acoustic_data": np.float32, "time_to_failure": np.float32}).values# 数据生成器的辅助函数。每时间步提取均值、标准差和分位数。# 可以轻松扩展。期望是一个二维数组。def extract_features(z): return np.c_[z.mean(axis=1), z.min(axis=1), z.max(axis=1), z.std(axis=1)]# 对于给定的结束位置"last_index"，我们将"x"的最后150,000个值# 分为150个长度为1000的片段。因此n_steps * step_length应等于150,000。# 从每个片段中提取一组特征。这将生成一个特征矩阵，# 其维度为(150个时间步 x 特征)。 def create_X(x, last_index=None, n_steps=150, step_length=1000): if last_index == None: last_index=len(x) assert last_index - n_steps * step_length >= 0 # 重塑和近似标准化，均值为5，标准差为3。 temp = (x[(last_index - n_steps * step_length):last_index].reshape(n_steps, -1) - 5 ) / 3 # 提取完整长度为1000的序列的特征，最后100个值的特征，以及最后10个观测值的特征。 return np.c_[extract_features(temp), extract_features(temp[:, -step_length // 10:]), extract_features(temp[:, -step_length // 100:])]# 查询"create_X"以确定特征数量n_features = create_X(float_data[0:150000]).shape[1]print("我们的RNN基于%i个特征"% n_features)# 生成器无休止地选择"batch_size"个子时间序列的结束位置。对于每个结束位置，# "time_to_failure"作为目标，而特征由函数"create_X"创建。def generator(data, min_index=0, max_index=None, batch_size=16, n_steps=150, step_length=1000): if max_index is None: max_index = len(data) - 1 while True: # 选择结束位置的索引 rows = np.random.randint(min_index + n_steps * step_length, max_index, size=batch_size) # 初始化特征矩阵和目标 samples = np.zeros((batch_size, n_steps, n_features)) targets = np.zeros(batch_size, ) for j, row in enumerate(rows): samples[j] = create_X(data[:, 0], last_index=row, n_steps=n_steps, step_length=step_length) targets[j] = data[row - 1, 1] yield samples, targetsbatch_size = 64# 第二个（共16个）地震的位置。用于在训练和验证之间进行清晰的分割# 第二个地震second_earthquake = 50085877float_data[second_earthquake, 1]# 初始化生成器strain_gen = generator(float_data, batch_size=batch_size) # 用于获得更好的分数# train_gen = generator(float_data, batch_size=batch_size, min_index=second_earthquake + 1)valid_gen = generator(float_data, batch_size=batch_size, max_index=second_earthquake)# 定义模型from keras.models import Sequentialfrom keras.layers import Densefrom keras.optimizers import adamfrom keras.callbacks import ModelCheckpointfrom keras.models import Modelfrom keras.callbacks import ModelCheckpointfrom keras.models import Sequentialfrom keras.layers import Dense, Activation, Flattenfrom sklearn.model_selection import train_test_splitfrom sklearn.ensemble import RandomForestRegressorfrom sklearn.metrics import mean_absolute_error from matplotlib import pyplot as pltimport matplotlib.pyplot as pltimport pandas as pdimport numpy as npimport warnings model = Sequential()model.add(Dense(10, activation='relu'))model.add(Dense(1))cb = [ModelCheckpoint("model.hdf5", save_best_only=True, period=3)]# 编译并拟合模型model = model.compile(optimizer=adam(lr=0.0005), loss="mae")history = model.fit_generator(train_gen, steps_per_epoch=1000, epochs=30, verbose=0, callbacks=cb, validation_data=valid_gen, validation_steps=200)model.summary()# 可视化准确率import matplotlib.pyplot as pltdef perf_plot(history, what = 'loss'): x = history.history[what] val_x = history.history['val_' + what] epochs = np.asarray(history.epoch) + 1 plt.plot(epochs, x, 'bo', label = "训练 " + what) plt.plot(epochs, val_x, 'b', label = "验证 " + what) plt.title("训练和验证 " + what) plt.xlabel("轮次") plt.legend() plt.show() return Noneperf_plot(history)# 加载提交文件submission = pd.read_csv('sample_submission.csv', index_col='seg_id', dtype={"time_to_failure": np.float32})# 加载每个测试数据，创建特征矩阵，获取数值预测for i, seg_id in enumerate(tqdm(submission.index)): # print(i) seg = pd.read_csv('../test/' + seg_id + '.csv') x = seg['acoustic_data'].values submission.time_to_failure[i] = model.predict(np.expand_dims(create_X(x), 0))submission.head()# 保存submission.to_csv('submissionearth.csv')我遇到的错误 :
Traceback (most recent call last):
File “”, line 1, in model.fit_generator(train_gen,
AttributeError: ‘NoneType’ object has no attribute ‘fit_generator’
我还导入了Keras.models，其中包含fit_generator，并尝试使用fit代替fit_generator，但仍然无法解决问题。
期待一些帮助！
回答：
你的问题在这里：
model = model.compile(optimizer=adam(lr=0.0005), loss="mae")history = model.fit_generator(train_gen, steps_per_epoch=1000, epochs=30, verbose=0, callbacks=cb, validation_data=valid_gen, validation_steps=200)你不应该将model.compile(..)赋值给model，因为它不返回任何东西，取而代之的是那行代码应该只写成model.compile(optimizer=adam(lr=0.0005), loss="mae")，所以就改成这样
model.compile(optimizer=adam(lr=0.0005), loss="mae")history = model.fit_generator(train_gen, steps_per_epoch=1000, epochs=30, verbose=0, callbacks=cb, validation_data=valid_gen, validation_steps=200)