Keras to build a CNN with variable input size

Question

I would like to build a CNN from the following diagram

CNN

I presume that each time the sentence can have various lengths and it is still possible to build it.

The below is a something I found on the web but kind of got lost

ngram_filters = [2, 3, 4] conv_filters = [] for n_gram in ngram_filters: conv_filters.append(Sequential()) conv_filters[-1].add(Conv2D(1, 1, n_gram, 5)) conv_filters[-1].add(MaxPooling2D(poolsize=(nb_tokens - n_gram + 1, 1))) conv_filters[-1].add(Flatten()) model = Sequential() model.add(Merge(conv_filters, mode='concat'))

The picture comes from the following blog - http://www.wildml.com/2015/11/understanding-convolutional-neural-networks-for-nlp/

Take a look at this article - richliao.github.io/supervised/classification/2016/11/26/… — Wasi Ahmad
– Wasi Ahmad, Commented Mar 16, 2017 at 20:11

Lior Magen · Accepted Answer · 2017-04-03 08:54:01Z

You don't need to use a variable input size, you just need to pad your input with zeros (make sure to mask these zeros). You can define your input size as a really big one and just pad the difference between each sentence length and the input size.

There is a built function in Keras', you can find it in Sequence Preprocessing and the function is "pad_sequences". Make sure your model is aware of the padding value by masking these values using the Masking class.

You can see an example of the implementation here.

Collectives™ on Stack Overflow

Keras to build a CNN with variable input size

1 Answer 1

Comments

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Related