Character Embedding convolution1d: Input incompatible with layer, expected ndim=3, found ndim=4

[abaheti95]

I have a character to integer mapping and a word to integer mapping. Similarly, I have two different character and word embedding layers.

In the model, I want to send one word index and a sequence of its character index at a time to the network. The word can be of variable length, therefore the character sequence has to be variable. For characters, I want to pass the sequence through an embedding layer to get char embedding sequence. Then this embedding sequence will go through a Convolution1D layer followed by GlobalMaxPooling1D layer.

But after the embedding layer when I send the char embedding sequence to Convolution1D layer it shows ValueError: Input 0 is incompatible with layer convolution1d_1: expected ndim=3, found ndim=4

My code snippet:

word_index = Input(shape=(1,))
# List of characters with unknown sequence length but one input at a time
characters = Input(shape=(None, 1))

shared_char_embedding_layer = Embedding(input_dim=(G.char_vocab_size+2), output_dim=G.char_embedding_dimension, mask_zero=True)
char_embeddings = shared_char_embedding_layer(characters)
print K.int_shape(char_embeddings)
print K.ndim(char_embeddings)
# CNN on top of char embeddings
convolution_trichar_embedding = Convolution1D(G.charCNN_dimension, 3, border_mode='valid', input_dim=G.char_embedding_dimension)(char_embeddings)

first_trichar = Flatten(convolution_trichar_embedding[:,0,:])
last_trichar = Flatten(convolution_trichar_embedding[:,-1,:])

max_pool_trichar = GlobalMaxPooling1D(convolution_trichar_embedding)

How to do I take variable length character sequence as input (what should be it's shape)?

Also, is the way of extracting first and last element of the convolution output correct?

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs, but feel free to re-open it if needed.