app.py

import streamlit as st
from keras.layers import LSTM, Dropout, Bidirectional, Dense,Embedding,Flatten,Maximum,Activation,Conv2D,LayerNormalization,add\
, BatchNormalization, SpatialDropout1D ,Input,Layer,Multiply,Reshape ,Add, GRU,Concatenate,Conv1D,TimeDistributed,ZeroPadding1D,concatenate,MaxPool1D,GlobalMaxPooling1D
import keras.backend as K
from keras import initializers, regularizers, constraints, activations
from keras.initializers import Constant
from keras import Model
import sys
import json
import pandas as pd
import numpy as np

with open('CHAR_TYPES_MAP.json') as json_file:
    CHAR_TYPES_MAP = json.load(json_file)
with open('CHARS_MAP.json') as json_file:
    CHARS_MAP = json.load(json_file)
with open('CHAR_TYPE_FLATTEN.json') as json_file:
    CHAR_TYPE_FLATTEN = json.load(json_file)


class TimestepDropout(Dropout):

    def __init__(self, rate, **kwargs):
        super(TimestepDropout, self).__init__(rate, **kwargs)

    def _get_noise_shape(self, inputs):
        input_shape = K.shape(inputs)
        noise_shape = (input_shape[0], input_shape[1], 1)
        return noise_shape


def model_(n_gram = 21):
    
    input1 = Input(shape=(21,),dtype='float32',name = 'char_input')
    input2 = Input(shape=(21,),dtype='float32',name = 'type_input')

    a = Embedding(178, 32,input_length=21)(input1)
    a = SpatialDropout1D(0.15)(a)
    #a = TimestepDropout(0.05)(a)
    char_input = BatchNormalization()(a)

    a_concat = []
    filters = [[1,200],[2,200],[3,200],[4,200],[5,200],[6,200],[8,200],[11,150],[12,100]]
    #filters = [[1,200],[2,200],[3,200],[4,200],[5,200],[6,200],[7,200],[8,200],[9,150],[10,150],[11,150],[12,100]]
    
    for (window_size, filters_size) in filters:
        convs = Conv1D(filters=filters_size, kernel_size=window_size, strides=1)(char_input)
        convs = Activation('elu')(convs)
        convs = TimeDistributed(Dense(5, input_shape=(21, filters_size)))(convs)
        convs = ZeroPadding1D(padding=(0, window_size-1))(convs)
        a_concat.append(convs)
    token_max = Maximum()(a_concat)
    lstm_char = Bidirectional(LSTM(128 ,return_sequences=True,kernel_regularizer=regularizers.L2(0.0000001),bias_regularizer=regularizers.L2(0.0000001)))(char_input)
    lstm_char = Dense(64, activation='elu')(lstm_char)
    #lstm_char = Bidirectional(LSTM(64 ,return_sequences=True))(lstm_char)
    #lstm_char = Attention(return_sequences=True)(lstm_char)
    
    b = Embedding(12, 12, input_length=21)(input2)
    type_inputs = SpatialDropout1D(0.15)(b)
    #type_inputs = TimestepDropout(0.05)(b)

    x = Concatenate()([type_inputs, char_input, lstm_char, token_max])
    x = BatchNormalization()(x)
    
    x = Flatten()(x)
    x = Dense(100, activation='elu')(x)
    x = Dropout(0.2)(x)
    out = Dense(1, activation='sigmoid',dtype = 'float32',kernel_regularizer=regularizers.L2(0.01),bias_regularizer=regularizers.L2(0.01))(x)
    

    model = Model(inputs=[input1, input2], outputs=out)
   
    return model


def create_feature_array(text, n_pad=21):

    n = len(text)
    n_pad_2 = int((n_pad - 1)/2)
    text_pad = [' '] * n_pad_2  + [t for t in text] + [' '] * n_pad_2
    x_char, x_type = [], []
    for i in range(n_pad_2, n_pad_2 + n):
        char_list = text_pad[i + 1: i + n_pad_2 + 1] + \
                    list(reversed(text_pad[i - n_pad_2: i])) + \
                    [text_pad[i]]
        char_map = [CHARS_MAP.get(c, 179) for c in char_list]
        char_type = [CHAR_TYPES_MAP.get(CHAR_TYPE_FLATTEN.get(c, 'o'), 4)
                     for c in char_list]
        x_char.append(char_map)
        x_type.append(char_type)
    x_char = np.array(x_char).astype(float)
    x_type = np.array(x_type).astype(float)
    return x_char, x_type

def tokenize(text):
        n_pad = 21

        if not text:
            return ['']

        if isinstance(text, str) and sys.version_info.major == 2:
            text = text.decode('utf-8')

        x_char, x_type = create_feature_array(text, n_pad=n_pad)
        word_end = []
 
        y_predict = model.predict([x_char, x_type], batch_size = 512)
        y_predict = (y_predict.ravel() > 0.46542968749999997).astype(int)
        word_end = y_predict[1:].tolist() + [1]
        
        tokens = []
        word = ''
        for char, w_e in zip(text, word_end):
            word += char
            if w_e:
                tokens.append(word)
                word = ''
        return tokens


model = model_()
model.load_weights("cutto_tf2.h5")

text = st.text_area("Enter original text!")
words = tokenize(text)

st.write('|'.join(words))