CODEWITHSUNDEEP
tensorflowkeras
mkuse
mkuse

Reputation: 2488

Keras TimeDistributed layer without LSTM

I am kind of new to keras. Here is what I am trying to achieve. I have a keras-model which takes as input an image to produce 512 vector. I create this as:

enter image description here

input_img = keras.layers.Input( shape=(240, 320, 3 ) )
cnn = make_vgg( input_img )
out = NetVLADLayer(num_clusters = 16)( cnn )
model = keras.models.Model( inputs=input_img, outputs=out )

Now, for the training, each of my samples are actually 13 images. Say I have 2500 samples then my data's dimensions are 2500x13x240x320x3. I want the model to be applied independently to the 13 images. I came across the TimeDistributed layer in keras and wondering how can I use it to achieve my objective. 

t_input = Input( shape=(13,240,320,3) )
# How to use TimeDistributed with model? 
t_out = TimeDistributed( out )
t_model = Model( inputs=t_input, outputs=t_out )

I am expecting t_out of size: None,13,512. The above code, however, throws a ValueError. Can anyone help my understanding?

Upvotes: 1

Views: 335

Answers (1)

rvinas
rvinas

Reputation: 11895

The error occurs in this line:

t_out = TimeDistributed(out)

It happens because out is a tensor, but TimeDistributed expects a layer as argument. This layer will be applied to every temporal slice (dimension of index one) of the input. You could instead do the following:

t_input = Input(shape=(13, 240, 320, 3))
t_out = TimeDistributed(model)(t_input)
t_model = Model(inputs=t_input, outputs=t_out)

Upvotes: 2

Related Questions