heard u liek to cnn

rcnnchar.py

# import csv
# import os
# import numpy as np

# from cervantes.box import WordVectorBox
# from cervantes.language import OneLevelEmbedding
# from cervantes.nn.models import RNNClassifier
# from cervantes.box import EnglishCharBox

from keras.models import Model, Sequential

from keras.layers import Input, Reshape, \
                         Embedding, GRU, \
                         Dense, Highway, \
                         Convolution1D,  \
                         Dropout, merge, \
                         TimeDistributed,\
                         Lambda

import keras.backend as K

NGRAMS = [1, 2, 3, 4, 5]
NFILTERS = 32 * 3

CHARACTERS_PER_WORD = 15
WORDS_PER_DOCUMENT = 300
NUMBER_CHARACTERS = 30 + 2
EMBEDDING_DIM = 100
INPUT_SHAPE = (CHARACTERS_PER_WORD * WORDS_PER_DOCUMENT, )
EMBEDDING_SHAPE = (WORDS_PER_DOCUMENT, CHARACTERS_PER_WORD, EMBEDDING_DIM)

doc = Input(shape=(INPUT_SHAPE[0], ), dtype='int32')

embedded = Sequential([
        Embedding(
            input_dim=NUMBER_CHARACTERS, 
            output_dim=EMBEDDING_DIM, 
            input_length=INPUT_SHAPE[0]
            ), 
        Reshape(EMBEDDING_SHAPE)
    ])(doc)

def sub_model(n):
    return Sequential([
            Convolution1D(NFILTERS, n, 
                activation='relu', 
                input_shape=EMBEDDING_SHAPE[1:]
                ), 
            Lambda(
                lambda x: K.max(x, axis=1), 
                output_shape=(NFILTERS,)
                )
        ])


rep = Dropout(0.5)(
    merge(
        [TimeDistributed(sub_model(n))(embedded) for n in NGRAMS], 
        mode='concat', 
        concat_axis=-1
    )
)


rep = Dropout(0.5)(
    merge(
        [GRU(90)(rep), GRU(90, go_backwards=True)(rep)], 
        mode='concat', 
        concat_axis=-1
    )
)

output = Highway(activation='relu')(rep)

output = Dropout(0.5)(output)

output = Dense(64, activation='relu')(output)

output = Dropout(0.4)(output)

output = Dense(1, activation='sigmoid')(output)

nn = Model(input=doc, output=output)