Feed CNN features to LSTM - tensorflow

Question

I want to build an end-to-end trainable model with the following proprieties:
CNN to extract features from image
The features is reshaped to a matrix
Each row of this matrix is then fed to LSTM1
Each column of this matrix is then fed to LSTM2
The output of LSTM1 and LSTM2 are concatenated for the final output
(it's more or less similar to Figure 2 in this paper: https://arxiv.org/pdf/1611.07890.pdf)
My problem now is after the reshape, how can I feed the values of feature matrix to LSTM with Keras or Tensorflow?
This is my code so far with VGG16 net (also a link to Keras issues):
# VGG16
model = Sequential()
model.add(Conv2D(64, (3, 3), activation='relu', padding='same', input_shape=(224, 224, 3)))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D((2, 2)))
# block 2
model.add(Conv2D(128, (3, 3), activation='relu', padding='same'))
model.add(Conv2D(128, (3, 3), activation='relu'))
model.add(MaxPooling2D((2, 2)))
# block 3
model.add(Conv2D(256, (3, 3), activation='relu', padding='same'))
model.add(Conv2D(256, (3, 3), activation='relu'))
model.add(Conv2D(256, (3, 3), activation='relu'))
model.add(MaxPooling2D((2, 2)))
# block 4
model.add(Conv2D(512, (3, 3), activation='relu', padding='same'))
model.add(Conv2D(512, (3, 3), activation='relu'))
model.add(Conv2D(512, (3, 3), activation='relu'))
model.add(MaxPooling2D((2, 2)))
# block 5
model.add(Conv2D(512, (3, 3), activation='relu', padding='same'))
model.add(Conv2D(512, (3, 3), activation='relu'))
model.add(Conv2D(512, (3, 3), activation='relu'))
model.add(MaxPooling2D((2, 2)))
# block 6
model.add(Flatten())
model.add(Dense(4096, activation='relu'))
model.add(Dense(4096, activation='relu'))
# reshape the feature 4096 = 64 * 64
model.add(Reshape((64, 64)))
# How to feed each row of this to LSTM?
# This is my first solution but it doesn’t look correct:
# model.add(LSTM(256, input_shape=(64, 1))) # 256 hidden units, sequence length = 64, feature dim = 1

Answer 1

Consider building your CNN model with Conv2D and MaxPool2D layers, until you reach your Flatten layer, because the vectorized output from the Flatten layer will be you input data to the LSTM part of your structure.
So, build your CNN model like this:
model_cnn = Sequential()
model_cnn.add(Conv2D...)
model_cnn.add(MaxPooling2D...)
...
model_cnn.add(Flatten())
Now, this is an interesting point, the current version of Keras has some incompatibility with some TensorFlow structures that will not let you stack your entire layers in just one Sequential object.
So it's time to use the Keras Model Object to complete you neural network with a trick:
input_lay = Input(shape=(None, ?, ?, ?)) #dimensions of your data
time_distribute = TimeDistributed(Lambda(lambda x: model_cnn(x)))(input_lay) # keras.layers.Lambda is essential to make our trick work :)
lstm_lay = LSTM(?)(time_distribute)
output_lay = Dense(?, activation='?')(lstm_lay)
And finally, now it's time to put together our 2 separated models:
model = Model(inputs=[input_lay], outputs=[output_lay])
model.compile(...)
OBS: Note that you can substitute my model_cnn example by your VGG without including the top layers, once the vectorized output from the VGG Flatten layer will be the input of the LSTM model.