ctctoyproblem.py

#LICENSE MIT#
from keras.models import Model
from keras.layers import Input, Dense, Merge, Recurrent

from keras.layers.recurrent import SimpleRNN, GRU,LSTM
from keras.layers.embeddings import Embedding

from keras.layers.wrappers import TimeDistributed

import random
import numpy as np


maxStringLength=20
spreadlength = 4 * maxStringLength

def SpreadWithZeros(x, length):
    out = np.zeros( (x.shape[0],length), dtype='uint16')
    mask = np.zeros((x.shape[0], length), dtype='uint16')
    for i in range( x.shape[0] ):
        co = 0
        ind = 0
        while ind < x.shape[1] and co < length:
            out[i,co] = x[i,ind]
            ind = ind + 1
            if( ind < x.shape[1]):
                co += random.randint(1,3)

        for j in range(co+1):
            mask[i,j]=1
    return out,mask


def BuildSpreadModel(Nclass,loss):
    text = Input(batch_shape=(None, spreadlength),dtype='uint16')
    emb = Embedding(Nclass, 5, input_length=spreadlength)(text)
    #SimpleRNN shouldn't be even needed to solve this toy problem
    enc = SimpleRNN( 100, return_sequences=True,activation='tanh' )(emb)
    dectext = TimeDistributed(Dense(Nclass+1, activation="softmax"))(enc)
    spreadModel = Model(text, dectext)
    #ctc_cost_precise
    #ctc_cost_for_train
    spreadModel.compile(loss=loss, optimizer='adam', sample_weight_mode='temporal')
    return spreadModel


def TrainSpreadModel(model,nbepoch,batch_size):
    for i in range(nbepoch):
        X = np.random.randint(1, 30, (batch_size, maxStringLength))
        xcat, mask = SpreadWithZeros(X, spreadlength)

        res = model.train_on_batch(x=xcat, y=X, sm_mask=mask, return_sm=True)
        print(res[0])
        #resultseqs = CTC.best_path_decode_batch(res[1], np.ones((X.shape[0], X.shape[1])))
        # for j in range(len(resultseqs)):
        #    print(resultseqs[j])
        #lengths = np.array([len(item) for item in resultseqs])
        #print("length " + str(np.mean(lengths)))

modelTrain = BuildSpreadModel(33,"ctc_cost_for_train")
TrainSpreadModel(modelTrain,10,50)

modelPrecise = BuildSpreadModel(33,"ctc_cost_precise")
TrainSpreadModel(modelPrecise,10,50)