CNN 和 LSTM 的输入形状答案

【问题标题】：Input shape for CNN and LSTMCNN 和 LSTM 的输入形状
【发布时间】：2020-06-15 01:08:07
【问题描述】：

我想使用此代码训练 CNN + LSTM 模型以使用 CNN + LSTM 驾驶汽车

import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, MaxPooling2D, GlobalMaxPool2D, Flatten, Dropout, Dense, TimeDistributed, GRU, LSTM
from tensorflow.keras.applications.vgg16 import VGG16

def vgg(input_shape, num_classes):

    # create a VGG16 "model", we will use
    # image with shape (224, 224, 3)
    vgg = VGG16(
        include_top=False,
        weights='imagenet',
        input_shape=(200, 160, 3)
    )
    # do not train first layers, I want to only train
    # the 4 last layers (my own choice, up to you)
    for layer in vgg.layers[:-4]:
        layer.trainable = False
    # create a Sequential model
    model = Sequential()
    # add vgg model for 4 input images (keeping the right shape
    model.add(
        TimeDistributed(vgg, input_shape=(4, 200, 160, 3))
    )
    # now, flatten on each output to send 5 
    # outputs with one dimension to LSTM
    model.add(
        TimeDistributed(
            Flatten()
        )
    )
    model.add(LSTM(256, activation='relu', return_sequences=False))
    # finalize with standard Dense, Dropout...
    model.add(Dense(128, activation='relu'))
    model.add(Dropout(.5))
    model.add(Dense(num_classes, activation='softmax'))
    model.compile('adam', loss='categorical_crossentropy', metrics=['accuracy'])

    return model

这是我的主要代码

# Splitting data into a training set and test set
train_data = np.load("/mydrive/Mister_car/training_data-1.npy", allow_pickle=True)
train = train_data[:]
test = train_data[:]

print(np.shape(train[5]))

X = np.array([i[0] for i in train]).reshape(-1, 4, WIDTH, HEIGHT, 3)
Y = np.array([i[1] for i in train])#.reshape(-1, 9)

x_test = np.array([i[0] for i in test]).reshape(-1, 4, WIDTH, HEIGHT, 3)
y_test = np.array([i[1] for i in test])#.reshape(-1, 9)

# start training
model.fit(x=X,
          y=Y,
          epochs=10,
          validation_data=(x_test, y_test)
          )

# save the whole model
model.save(MODEL_DIR)

我有一个数组，这个数组中的每个元素都有 2 个元素：数组 [0] 是 4 个形状为 (200,160,3) 的图像的序列 array 1 是一个有 9 个元素的数组

但在模型保存时的最后一个纪元后出现以下错误

【问题讨论】：

标签： python tensorflow keras conv-neural-network

【解决方案1】：

我认为您需要删除 vgg 周围的 TimeDistributed() 包装器，因为它所做的是添加另一个维度。因此，您最终得到 5 而不是 4，这导致了错误。同时删除这一行中的 -1：

X = np.array([i[0] for i in train]).reshape(-1, 4, WIDTH, HEIGHT, 3)

除此之外，您的代码的某些部分丢失了。如果您提供它们，我将能够更确定地提供帮助。

【讨论】：