我正在尝试使用Keras实现一个LSTM网络，以预测动作后边界框的运动。
我正在使用的数据集看起来像这样：（要预测的边界框是从列b0，b1，b2，b3中得出的）
act b0 b1 b2 b3 id score
0 85 238 129 256 69 0.9289865493774414
1 -1 -1 -1 -1 -1 -1
1 -1 -1 -1 -1 -1 -1
2 -1 -1 -1 -1 -1 -1
3 -1 -1 -1 -1 -1 -1
0 46 136 256 245 73 0.9369892477989197
1 18 18 256 252 73 0.8203921318054199
1 144 212 169 223 10 0.9630357623100281
1 13 9 252 199 73 0.9374213814735413
3 -1 -1 -1 -1 -1 -1
0 215 141 255 233 72 0.9941028952598572
2 199 116 243 183 74 0.8685483932495117
3 215 141 255 233 72 0.9941184520721436
1 189 78 215 95 76 0.8610376119613647
3 206 50 255 169 72 0.8224002122879028
0 -1 -1 -1 -1 -1 -1
3 19 129 249 253 73 0.8635225892066956
2 -1 -1 -1 -1 -1 -1
2 0 78 13 91 10 0.9488454461097717
3 -1 -1 -1 -1 -1 -1
0 206 123 255 189 62 0.9980332255363464
2 221 197 256 255 62 0.9782524704933167
2 -1 -1 -1 -1 -1 -1
2 -1 -1 -1 -1 -1 -1
1 -1 -1 -1 -1 -1 -1
0 184 78 243 169 72 0.9953457713127136
2 191 139 254 246 72 0.9929528832435608
3 184 78 243 169 72 0.9953963160514832
3 197 1 254 91 72 0.9956125020980835
2 184 78 243 169 72 0.9953963160514832
我正在创建一个新列，以使边框具有以下单个输入：
b_box = ['b0', 'b1', 'b2', 'b3']
df['b_box'] = df[b.box].apply(tuple, axis=1).apply(list)
这样做是因为我传递给LSTM网络的值的形状是一个无法使用的4D数组(3994, 5, 1, 4)。
import pandas as pd
import os
from sklearn.model_selection import train_test_split
import numpy as np
PATH = "./csv/"
filename = os.path.join(PATH, "test.csv")
names = ['act', 'b0', 'b1', 'b2', 'b3', 'id', 'score']
df = pd.read_csv(filename, sep=',')
# print("Starting file:")
# print(df[0:10])
# print("Ending file:")
# print(df[-10:])
b_box = ['b0', 'b1', 'b2', 'b3']
df['b_box'] = df[b_box].apply(tuple, axis=1).apply(list)
# print(df.head())
train, test = train_test_split(df, test_size=0.2)
boxes_train = train['b_box'].tolist()
boxes_test = test['b_box'].tolist()
# print("Training set has {} observations".format(len(boxes_train)))
# print("Test set has {} observations".format(len(boxes_test)))
def to_sequences(seq_size, obs):
x = []
y = []
for i in range(len(obs)-SEQUENCE_SIZE-1):
window = obs[i:(i+SEQUENCE_SIZE)]
after_window = obs[i+SEQUENCE_SIZE]
window = [[x] for x in window]
x.append(window)
y.append(after_window)
return np.array(x), np.array(y)
SEQUENCE_SIZE = 5
x_train, y_train = to_sequences(SEQUENCE_SIZE, boxes_train)
x_test, y_test = to_sequences(SEQUENCE_SIZE, boxes_test)
print(x_train)
# print("Shape of training set: {}".format(x_train.shape))
# print("Shape of test set: {}".format(x_test.shape))
from keras.preprocessing import sequence
from keras.models import Sequential
from keras.layers import Dense, Embedding
from keras.layers import LSTM
from keras.datasets import imdb
print("Build model...")
model = Sequential()
model.add(LSTM(64, dropout=0.2, recurrent_dropout=0.2, input_dim=1))
model.add(Dense(1))
model.compile(loss='mean_squared_error', optimizer='adam')
print('Train...')
model.fit(x_train, y_train, validation_data=(x_test, y_test), verbose=2, epochs=100)
答案 0 :(得分：0)
您需要使用TimeSeriesGenerator准备数据。这是一个具有2个功能的快速示例。
import numpy as np
from keras.models import Sequential
from keras.layers import Dense
from keras.preprocessing.sequence import TimeseriesGenerator
dim = 1000
# 2 features
x1 = np.random.normal(size=dim).reshape(1000,1)
x2 = np.random.normal(size=dim).reshape(1000,1)
data = np.hstack((x1,x2))
# target is simply the sum
targets = x1 + x2
targets = np.insert(targets,0,0)[0:dim]
data_gen = TimeseriesGenerator(data, targets,length=1,batch_size=1)
model = Sequential()
model.add(LSTM(128, input_shape=(1,2), dropout=0.2, recurrent_dropout=0.2))
model.add(Dense(1))
model.compile(loss='mse',optimizer='adam')
# Please note 'fit_generator' instead of 'fit'
model.fit_generator(generator=data_gen,steps_per_epoch=32,epochs=5)
出局：
Epoch 1/5
32/32 [==============================] - 3s 89ms/step - loss: 2.6368
Epoch 2/5
32/32 [==============================] - 0s 4ms/step - loss: 1.8943
Epoch 3/5
32/32 [==============================] - 0s 3ms/step - loss: 1.0667
Epoch 4/5
32/32 [==============================] - 0s 3ms/step - loss: 1.1214
Epoch 5/5
32/32 [==============================] - 0s 3ms/step - loss: 0.5821
在您的情况下，因为您使用的是熊猫，所以您只需提取特征和目标（我假设“分数”为目标），然后将它们输入TimeseriesGenerator。 insert 步骤对于将来将标签移动一个步骤是必需的。另外参数 length 和batch_size均为1，因为我想每个时间步长都被标记了（即，您没有使用整个时间步长序列来进行每个预测）。检查文档中的参数是否确实是模型所需的。
dim = 1000
# the columns are the same but I filled the df with random values
train_cols = ['act','b0','b1','b2','b3','id']
df = pd.DataFrame(np.random.randint(0,100,size=(dim, 6)), columns=train_cols)
df['score'] = df.mean(axis=1)
targets = np.insert(np.asarray(df['score']),0,0)[0:dim]
data_gen = TimeseriesGenerator(df[train_cols].values, targets,length=1,batch_size=1)