action_3d_model.py

from keras.models import Model
from keras.layers import Input, Dense, LSTM, multiply, concatenate, Activation, Masking, Reshape
from keras.layers import Conv1D, BatchNormalization, GlobalAveragePooling1D, Permute, Dropout

from utils.constants import MAX_NB_VARIABLES, NB_CLASSES_LIST, MAX_TIMESTEPS_LIST
from utils.keras_utils import train_model, evaluate_model, set_trainable
from utils.layer_utils import AttentionLSTM

DATASET_INDEX = 3

MAX_TIMESTEPS = MAX_TIMESTEPS_LIST[DATASET_INDEX]
MAX_NB_VARIABLES = MAX_NB_VARIABLES[DATASET_INDEX]
NB_CLASS = NB_CLASSES_LIST[DATASET_INDEX]

TRAINABLE = True


def generate_model():
    ip = Input(shape=(MAX_NB_VARIABLES, MAX_TIMESTEPS))

    x = Masking()(ip)
    x = LSTM(8)(x)
    x = Dropout(0.8)(x)

    y = Permute((2, 1))(ip)
    y = Conv1D(128, 8, padding='same', kernel_initializer='he_uniform')(y)
    y = BatchNormalization()(y)
    y = Activation('relu')(y)
    y = squeeze_excite_block(y)

    y = Conv1D(256, 5, padding='same', kernel_initializer='he_uniform')(y)
    y = BatchNormalization()(y)
    y = Activation('relu')(y)
    y = squeeze_excite_block(y)

    y = Conv1D(128, 3, padding='same', kernel_initializer='he_uniform')(y)
    y = BatchNormalization()(y)
    y = Activation('relu')(y)

    y = GlobalAveragePooling1D()(y)

    x = concatenate([x, y])

    out = Dense(NB_CLASS, activation='softmax')(x)

    model = Model(ip, out)
    model.summary()

    # add load model code here to fine-tune

    return model


def generate_model_2():
    ip = Input(shape=(MAX_NB_VARIABLES, MAX_TIMESTEPS))

    ''' sabsample timesteps to prevent OOM due to Attention LSTM '''
    stride = 2

    x = Permute((2, 1))(ip)
    x = Conv1D(MAX_NB_VARIABLES // stride, 8, strides=stride, padding='same', activation='relu', use_bias=False,
               kernel_initializer='he_uniform')(x) # (None, variables / stride, timesteps)
    x = Permute((2, 1))(x)

    x = Masking()(x)
    x = AttentionLSTM(128)(x)
    x = Dropout(0.8)(x)

    y = Permute((2, 1))(ip)
    y = Conv1D(128, 8, padding='same', kernel_initializer='he_uniform')(y)
    y = BatchNormalization()(y)
    y = Activation('relu')(y)
    y = squeeze_excite_block(y)

    y = Conv1D(256, 5, padding='same', kernel_initializer='he_uniform')(y)
    y = BatchNormalization()(y)
    y = Activation('relu')(y)
    y = squeeze_excite_block(y)

    y = Conv1D(128, 3, padding='same', kernel_initializer='he_uniform')(y)
    y = BatchNormalization()(y)
    y = Activation('relu')(y)

    y = GlobalAveragePooling1D()(y)

    x = concatenate([x, y])

    out = Dense(NB_CLASS, activation='softmax')(x)

    model = Model(ip, out)
    model.summary()

    # add load model code here to fine-tune

    return model

def generate_model_3():
    ip = Input(shape=(MAX_NB_VARIABLES, MAX_TIMESTEPS))

    x = Masking()(ip)
    x = LSTM(8)(x)
    x = Dropout(0.8)(x)

    y = Permute((2, 1))(ip)
    y = Conv1D(128, 8, padding='same', kernel_initializer='he_uniform')(y)
    y = BatchNormalization()(y)
    y = Activation('relu')(y)
    #y = squeeze_excite_block(y)

    y = Conv1D(256, 5, padding='same', kernel_initializer='he_uniform')(y)
    y = BatchNormalization()(y)
    y = Activation('relu')(y)
    #y = squeeze_excite_block(y)

    y = Conv1D(128, 3, padding='same', kernel_initializer='he_uniform')(y)
    y = BatchNormalization()(y)
    y = Activation('relu')(y)

    y = GlobalAveragePooling1D()(y)

    x = concatenate([x, y])

    out = Dense(NB_CLASS, activation='softmax')(x)

    model = Model(ip, out)
    model.summary()

    # add load model code here to fine-tune

    return model


def generate_model_4():
    ip = Input(shape=(MAX_NB_VARIABLES, MAX_TIMESTEPS))
    # stride = 3
    #
    # x = Permute((2, 1))(ip)
    # x = Conv1D(MAX_NB_VARIABLES // stride, 8, strides=stride, padding='same', activation='relu', use_bias=False,
    #            kernel_initializer='he_uniform')(x)  # (None, variables / stride, timesteps)
    # x = Permute((2, 1))(x)

    x = Masking()(ip)
    x = AttentionLSTM(8)(x)
    x = Dropout(0.8)(x)

    y = Permute((2, 1))(ip)
    y = Conv1D(128, 8, padding='same', kernel_initializer='he_uniform')(y)
    y = BatchNormalization()(y)
    y = Activation('relu')(y)
    #y = squeeze_excite_block(y)

    y = Conv1D(256, 5, padding='same', kernel_initializer='he_uniform')(y)
    y = BatchNormalization()(y)
    y = Activation('relu')(y)
    #y = squeeze_excite_block(y)

    y = Conv1D(128, 3, padding='same', kernel_initializer='he_uniform')(y)
    y = BatchNormalization()(y)
    y = Activation('relu')(y)

    y = GlobalAveragePooling1D()(y)

    x = concatenate([x, y])

    out = Dense(NB_CLASS, activation='softmax')(x)

    model = Model(ip, out)
    model.summary()

    # add load model code here to fine-tune

    return model


def squeeze_excite_block(input):
    ''' Create a squeeze-excite block
    Args:
        input: input tensor
        filters: number of output filters
        k: width factor

    Returns: a keras tensor
    '''
    filters = input._keras_shape[-1] # channel_axis = -1 for TF

    se = GlobalAveragePooling1D()(input)
    se = Reshape((1, filters))(se)
    se = Dense(filters // 16,  activation='relu', kernel_initializer='he_normal', use_bias=False)(se)
    se = Dense(filters, activation='sigmoid', kernel_initializer='he_normal', use_bias=False)(se)
    se = multiply([input, se])
    return se


if __name__ == "__main__":
    model = generate_model_2()

    #train_model(model, DATASET_INDEX, dataset_prefix='action_3d', epochs=600, batch_size=128)

    evaluate_model(model, DATASET_INDEX, dataset_prefix='action_3d', batch_size=128)