Keras Flatten Conv3D 出现 ValueError 错误：输入的形状未完全定义。

Question

我正在尝试使用基于Keras和Tensorflow后端的Marcin的PS3示例构建可变长度序列分类模型，链接如下：https://dev59.com/P1gQ5IYBdhLWcg3whEVn#42635571 我遇到了一个错误： ValueError: The shape of the input to "Flatten" is not fully defined (got (None, 1, 1, 32). Make sure to pass a complete "input_shape" or "batch_input_shape" argument to the first layer in your model. 我尝试在Inception层上加入输入形状，但错误仍然存在。我该如何纠正这个问题？

复现步骤如下：

import numpy as np
import keras
from keras.utils import to_categorical
from keras.layers import TimeDistributed, Conv3D, Input, Flatten, Dense
from keras.applications.inception_v3 import InceptionV3
from random import randint
from keras.models import Model

HEIGHT = 224
WIDTH = 224
NDIMS = 3
NUM_CLASSES = 4

def input_generator():
    while True:
        nframes = randint(1,5)
        label = randint(0,NUM_CLASSES-1)
        x = np.random.random((nframes, HEIGHT, WIDTH, NDIMS))
        x = np.expand_dims(x, axis=0)
        y = keras.utils.to_categorical(label, num_classes=NUM_CLASSES)
        yield (x, y)

def make_model():
    layers = 32
    inp = Input(shape=(None, HEIGHT, WIDTH, NDIMS))
    cnn = InceptionV3(include_top=False, weights='imagenet')
    # cnn = InceptionV3(include_top=False, weights='imagenet', input_shape=(HEIGHT, WIDTH, NDIMS)) # same result
    td = TimeDistributed(cnn)(inp)
    c3da = Conv3D(layers, 3,3,3)(td)
    c3db = Conv3D(layers, 3,3,3)(c3da)
    flat = Flatten()(c3db)
    out = Dense(NUM_CLASSES, activation="softmax")(flat)
    model = Model(input=(None, HEIGHT, WIDTH, NDIMS), output=out)
    model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])
    return model

if __name__ == '__main__':
    model = make_model()
    model.fit_generator(input_generator(), samples_per_epoch=5, nb_epoch=2, verbose=1)

- Clark Kent

Answer 1

无法将变长张量展平。如果可以的话，Keras 如何知道最后一个全连接层的输入单元数？模型的参数数量需要在图形创建时定义。

有两种可能的解决方案：

a）固定帧数：

inp = Input(shape=(NFRAMES, HEIGHT, WIDTH, NDIMS))

b) 在展平层之前聚合帧的维度。例如:

from keras.layers import Lambda
import keras.backend as K    

def make_model():
    layers = 32
    inp = Input(shape=(None, HEIGHT, WIDTH, NDIMS))
    cnn = InceptionV3(include_top=False, weights='imagenet')
    # cnn = InceptionV3(include_top=False, weights='imagenet', input_shape=(HEIGHT, WIDTH, NDIMS)) # same result
    td = TimeDistributed(cnn)(inp)
    c3da = Conv3D(layers, 3,3,3)(td)
    c3db = Conv3D(layers, 3,3,3)(c3da)
    aggregated = Lambda(lambda x: K.sum(x, axis=1))(c3db)
    flat = Flatten()(aggregated)
    out = Dense(NUM_CLASSES, activation="softmax")(flat)
    model = Model(input=inp, output=out)
    model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])
    return model

注意 1：可能有更好的策略来聚合帧的尺寸。

注意 2：keras.utils.to_categorical的输入应该是标签列表：

y = keras.utils.to_categorical([label], num_classes=NUM_CLASSES)

- rvinas