def createmodel():
    img = layers.Input(shape=imgshape)  # Get image as an input of size 50,200,1
    conv1 = layers.Conv2D(16, (3, 3), padding='same', activation='relu')(img)  # 50*200
    bn1 = layers.BatchNormalization()(conv1)
    # dn1_1 = layers.Dropout(0.2)(bn1)
    mp1 = layers.MaxPooling2D(pool_size=(2, 2),padding='same')(bn1)  # 25*100
    conv2 = layers.Conv2D(32, (3, 3), padding='same', activation='relu')(mp1)
    bn2 = layers.BatchNormalization()(conv2)
    # dn2_2 = layers.Dropout(0.2)(bn2)
    mp2 = layers.MaxPooling2D(pool_size=(2, 2),padding='same')(bn2)  # 13*50
    bn2_2 = layers.BatchNormalization()(mp2)
    conv3 = layers.Conv2D(64, (3, 3), padding='same', activation='relu')(bn2_2)
    bn3 = layers.BatchNormalization()(conv3)  # to improve the stability of model
    # dn3_3 = layers.Dropout(0.2)(bn3)
    mp3 = layers.MaxPooling2D(pool_size=(2, 2),padding='same')(bn3)  # 7*25
    bn3_3 = layers.BatchNormalization()(mp3)  # t
    conv4 = layers.Conv2D(128, (3, 3), padding='same', activation='relu')(bn3_3)
    bn4 = layers.BatchNormalization()(conv4)  # to improve the stability of model
    # dn4_4 = layers.Dropout(0.2)(bn4)
    mp4 = layers.MaxPooling2D(pool_size=(2, 2), padding='same')(bn4)
    # bn4 = layers.BatchNormalization()(mp4)
    bn4_4 = layers.BatchNormalization()(mp4)
    conv5 = layers.Conv2D(256, (3, 3), padding='same', activation='relu')(bn4_4)
    bn5 = layers.BatchNormalization()(conv5)
    mp5 = layers.MaxPooling2D((2, 2), padding='same')(bn5)
    # conv6 = layers.Conv2D(512, (3, 3), padding='same', activation='relu')(mp5)
    # bn6 = layers.BatchNormalization()(conv6)
    # mp6 = layers.MaxPooling2D((2, 2), padding='same')(bn6)
    # final = layers.Dropout(0.3)(mp5)
    # bn5_5 = layers.BatchNormalization()(mp5)
    # conv6 = layers.Conv2D(512, (3, 3), padding='same', activation='relu')(bn5_5)
    # bn6 = layers.BatchNormalization()(conv6)
    # mp6 = layers.MaxPooling2D((2, 2), padding='same')(bn6)
    # bn6_6 = layers.BatchNormalization()(mp6)
    # drop_la = layers.Dropout(0.5)(bn5_5)
    # full = layers.Dropout(0.5)(mp5)
    # bn4 = layers.BatchNormalization()(conv4)  # to improve the stability of model
    # mp4 = layers.MaxPooling2D((2,2), padding='same')(bn4)  # 7*25

    flat = layers.Flatten()(mp5)  # convert the layer into 1-D

    outs = []
    for _ in range(5):  # for 5 letters of captcha
        dens1 = layers.Dense(512, activation='relu')(flat)
        drop = layers.BatchNormalization()(dens1)
        #1 - sigmoid
        #2 - softmax
        res = layers.Dense(nchar, activation='softmax')(drop)
        # rea_ = layers.Dropout(0.2)(res)  # drops 0.5 fraction of nodes
        # layers.Dropout(0.5)
        outs.append(res)  # result of layers

    # Compile model and return it
    # img = img.reshape(1000, 50, 200, 1)
    model = Model(img, outs)  # create model
    # optimizer = keras.optimizers.Adam(learning_rate=0.01)
    model.compile(loss='categorical_crossentropy', optimizer='Adam', metrics=["accuracy"]*5)
    return model