Python keras.optimizers 模块，RMSprop() 实例源码

def build_network(deepest=False):
    dropout = [0., 0.1, 0.2, 0.3, 0.4]
    conv = [(64, 3, 3), (128, 3, 3), (256, 3, 3), (512, 3, 3), (512, 2, 2)]
    input= Input(shape=(3, 32, 32))
    output = fractal_net(
        c=3, b=5, conv=conv,
        drop_path=0.15, dropout=dropout,
        deepest=deepest)(input)
    output = Flatten()(output)
    output = Dense(NB_CLASSES, init='he_normal')(output)
    output = Activation('softmax')(output)
    model = Model(input=input, output=output)
    optimizer = SGD(lr=LEARN_START, momentum=MOMENTUM)
    #optimizer = RMSprop(lr=LEARN_START)
    #optimizer = Adam()
    #optimizer = Nadam()
    model.compile(optimizer=optimizer, loss='categorical_crossentropy', metrics=['accuracy'])
    plot(model, to_file='model.png')
    return model

def main():
    game_width = 12
    game_height = 9
    nb_frames = 4
    actions = ((-1, 0), (1, 0), (0, -1), (0, 1), (0, 0))

    # Recipe of deep reinforcement learning model
    model = Sequential()
    model.add(Convolution2D(
        16,
        nb_row=3,
        nb_col=3,
        activation='relu',
        input_shape=(nb_frames, game_height, game_width)))
    model.add(Convolution2D(32, nb_row=3, nb_col=3, activation='relu'))
    model.add(Flatten())
    model.add(Dense(256, activation='relu'))
    model.add(Dense(len(actions)))
    model.compile(RMSprop(), 'MSE')

    agent = Agent(
        model, nb_frames, snake_game, actions, size=(game_width, game_height))
    agent.train(nb_epochs=10000, batch_size=64, gamma=0.8, save_model=True)
    agent.play(nb_rounds=10)

def __init__(self, input_shape, lr=0.01, n_layers=2, n_hidden=8, rate_dropout=0.2, loss='risk_estimation'):
        print("initializing..., learing rate %s, n_layers %s, n_hidden %s, dropout rate %s." %(lr, n_layers, n_hidden, rate_dropout))
        self.model = Sequential()
        self.model.add(Dropout(rate=rate_dropout, input_shape=(input_shape[0], input_shape[1])))
        for i in range(0, n_layers - 1):
            self.model.add(LSTM(n_hidden * 4, return_sequences=True, activation='tanh',
                                recurrent_activation='hard_sigmoid', kernel_initializer='glorot_uniform',
                                recurrent_initializer='orthogonal', bias_initializer='zeros',
                                dropout=rate_dropout, recurrent_dropout=rate_dropout))
        self.model.add(LSTM(n_hidden, return_sequences=False, activation='tanh',
                                recurrent_activation='hard_sigmoid', kernel_initializer='glorot_uniform',
                                recurrent_initializer='orthogonal', bias_initializer='zeros',
                                dropout=rate_dropout, recurrent_dropout=rate_dropout))
        self.model.add(Dense(1, kernel_initializer=initializers.glorot_uniform()))
        # self.model.add(BatchNormalization(axis=-1, moving_mean_initializer=Constant(value=0.5),
        #               moving_variance_initializer=Constant(value=0.25)))
        self.model.add(BatchRenormalization(axis=-1, beta_init=Constant(value=0.5)))
        self.model.add(Activation('relu_limited'))
        opt = RMSprop(lr=lr)
        self.model.compile(loss=loss,
                      optimizer=opt,
                      metrics=['accuracy'])

def simple_cnn(agent, env, dropout=0, learning_rate=1e-3, **args):
  with tf.device("/cpu:0"):
    state = tf.placeholder('float', [None, agent.input_dim])
    S = Input(shape=[agent.input_dim])
    h = Reshape( agent.input_dim_orig )(S)
    h = TimeDistributed( Convolution2D(16, 8, 8, subsample=(4, 4), border_mode='same', activation='relu', dim_ordering='tf'))(h)
#    h = Dropout(dropout)(h)
    h = TimeDistributed( Convolution2D(32, 4, 4, subsample=(2, 2), border_mode='same', activation='relu', dim_ordering='tf'))(h)
    h = Flatten()(h)
#    h = Dropout(dropout)(h)
    h = Dense(256, activation='relu')(h)
#    h = Dropout(dropout)(h)
    h = Dense(128, activation='relu')(h)
    V = Dense(env.action_space.n, activation='linear',init='zero')(h)
    model = Model(S, V)
    model.compile(loss='mse', optimizer=RMSprop(lr=learning_rate) )
    return state, model

def init_model():
    start_time = time.time()
    print 'Compiling Model ... '
    model = Sequential()
    model.add(Dense(500, input_dim=784))
    model.add(Activation('relu'))
    model.add(Dropout(0.4))
    model.add(Dense(300))
    model.add(Activation('relu'))
    model.add(Dropout(0.4))
    model.add(Dense(10))
    model.add(Activation('softmax'))

    rms = RMSprop()
    model.compile(loss='categorical_crossentropy', optimizer=rms,
      metrics=['accuracy'])
    print 'Model compiled in {0} seconds'.format(time.time() - start_time)
    return model

def init_model():
    """
    """
    start_time = time.time()
    print 'Compiling model...'
    model = Sequential()

    model.add(Convolution2D(64, 3,3, border_mode='valid', input_shape=INPUT_SHAPE))
    model.add(Activation('relu'))
    model.add(MaxPooling2D(pool_size=(2,2)))
    model.add(Dropout(.25))

    model.add(Flatten())

    model.add(Dense(10))
    model.add(Activation('softmax'))

    rms = RMSprop()
    model.compile(loss='categorical_crossentropy', optimizer=rms,
      metrics=['accuracy'])
    print 'Model compiled in {0} seconds'.format(time.time() - start_time)

    model.summary()
    return model

def test_fuctional_model_saving():
    input = Input(shape=(3,))
    x = Dense(2)(input)
    output = Dense(3)(x)

    model = Model(input, output)
    model.compile(loss=objectives.MSE,
                  optimizer=optimizers.RMSprop(lr=0.0001),
                  metrics=[metrics.categorical_accuracy])
    x = np.random.random((1, 3))
    y = np.random.random((1, 3))
    model.train_on_batch(x, y)

    out = model.predict(x)
    _, fname = tempfile.mkstemp('.h5')
    save_model(model, fname)

    model = load_model(fname)
    os.remove(fname)

    out2 = model.predict(x)
    assert_allclose(out, out2, atol=1e-05)

def __init__(self):
        super().__init__()
        self._learning = True
        self._learning_rate = .1
        self._discount = .1
        self._epsilon = .9

        # Create Model
        model = Sequential()

        model.add(Dense(2, init='lecun_uniform', input_shape=(2,)))
        model.add(Activation('relu'))

        model.add(Dense(10, init='lecun_uniform'))
        model.add(Activation('relu'))

        model.add(Dense(4, init='lecun_uniform'))
        model.add(Activation('linear'))

        rms = RMSprop()
        model.compile(loss='mse', optimizer=rms)

        self._model = model

def model(X_train, X_test, Y_train, Y_test):
    model = Sequential()
    model.add(Dense(512, input_shape=(784,)))
    model.add(Activation('relu'))
    model.add(Dropout({{uniform(0, 1)}}))
    model.add(Dense({{choice([400, 512, 600])}}))
    model.add(Activation('relu'))
    model.add(Dropout({{uniform(0, 1)}}))
    model.add(Dense(10))
    model.add(Activation('softmax'))

    rms = RMSprop()
    model.compile(loss='categorical_crossentropy', optimizer=rms, metrics=['accuracy'])

    nb_epoch = 10
    batch_size = 128

    model.fit(X_train, Y_train,
              batch_size=batch_size, nb_epoch=nb_epoch,
              verbose=2,
              validation_data=(X_test, Y_test))

    score, acc = model.evaluate(X_test, Y_test, verbose=0)

    return {'loss': -acc, 'status': STATUS_OK, 'model': model}

def model(X_train, Y_train, X_test, Y_test):
    model = Sequential()
    model.add(Dense(50, input_shape=(784,)))
    model.add(Activation('relu'))
    model.add(Dropout({{uniform(0, 1)}}))
    model.add(Dense({{choice([20, 30, 40])}}))
    model.add(Activation('relu'))
    model.add(Dropout({{uniform(0, 1)}}))
    model.add(Dense(10))
    model.add(Activation('softmax'))

    rms = RMSprop()
    model.compile(loss='categorical_crossentropy', optimizer=rms, metrics=['accuracy'])

    model.fit(X_train, Y_train,
              batch_size={{choice([64, 128])}},
              nb_epoch=1,
              verbose=2,
              validation_data=(X_test, Y_test))
    score, acc = model.evaluate(X_test, Y_test, verbose=0)
    print('Test accuracy:', acc)
    return {'loss': -acc, 'status': STATUS_OK, 'model': model}

def ensemble_model(X_train, X_test, Y_train, Y_test):
    model = Sequential()
    model.add(Dense(512, input_shape=(784,)))
    model.add(Activation('relu'))
    model.add(Dropout({{uniform(0, 1)}}))
    model.add(Dense({{choice([400, 512, 600])}}))
    model.add(Activation('relu'))
    model.add(Dropout({{uniform(0, 1)}}))
    model.add(Dense(10))
    model.add(Activation('softmax'))

    rms = RMSprop()
    model.compile(loss='categorical_crossentropy', optimizer=rms, metrics=['accuracy'])

    nb_epoch = 10
    batch_size = 128

    model.fit(X_train, Y_train,
              batch_size=batch_size, nb_epoch=nb_epoch,
              verbose=2,
              validation_data=(X_test, Y_test))

    score, acc = model.evaluate(X_test, Y_test, verbose=0)

    return {'loss': -acc, 'status': STATUS_OK, 'model': model}

def model_masking(discrete_time, init_alpha, max_beta):
    model = Sequential()

    model.add(Masking(mask_value=mask_value,
                      input_shape=(n_timesteps, n_features)))
    model.add(TimeDistributed(Dense(2)))
    model.add(Lambda(wtte.output_lambda, arguments={"init_alpha": init_alpha,
                                                    "max_beta_value": max_beta}))

    if discrete_time:
        loss = wtte.loss(kind='discrete', reduce_loss=False).loss_function
    else:
        loss = wtte.loss(kind='continuous', reduce_loss=False).loss_function

    model.compile(loss=loss, optimizer=RMSprop(
        lr=lr), sample_weight_mode='temporal')
    return model

def get_optimizer(self):

        if self.opt == 'sgd':
            return k_opt.SGD(lr=self.learning_rate, momentum=self.momentum)

        if self.opt == 'rmsprop':
            return k_opt.RMSprop(lr=self.learning_rate)

        if self.opt == 'adagrad':
            return k_opt.Adagrad(lr=self.learning_rate)

        if self.opt == 'adadelta':
            return k_opt.Adadelta(lr=self.learning_rate)

        if self.opt == 'adam':
            return k_opt.Adam(lr=self.learning_rate)

        raise Exception('Invalid optimization function - %s' % self.opt)

def get_dense_model():
    """Make keras model"""

    learning_rate=1e-4
    inp = Input(shape=(80*80,))
    h = Dense(200, activation='relu')(inp)
    out = Dense(1, activation='sigmoid')(h)
    model = Model(inp, out)
    optim = RMSprop(learning_rate)
    model.compile(optim, 'binary_crossentropy')
    try:
        model.load_weights('mod_weights_binary.h5')
        print('weights loaded')
    except:
        pass
    return model

def __init__(self):
        filters1 = [16, 32, 64]  # filters1 = [4, 8, 16, 32, 64, 128, 256]
        filters2 = [16, 32, 64]  # filters2 = [4, 8, 16, 32, 64, 128, 256]
        losses1 = [losses.MSE, losses.MAE, losses.hinge, losses.categorical_crossentropy]  # losses1 = [losses.MSE, losses.MAE, losses.hinge, losses.categorical_crossentropy]
        optimizers1 = [optimizers.Adam()]  # optimizers1 = [optimizers.Adadelta(), optimizers.Adagrad(), optimizers.Adam(), optimizers.Adamax(), optimizers.SGD(), optimizers.RMSprop()]
        units1 = [16, 32, 64]  # units1 = [4, 8, 16, 32, 64, 128, 256]
        kernel_sizes1 = [(3, 3)]  # kernel_sizes = [(3, 3), (5, 5)]
        dropouts1 = [0.25]  # dropouts1 = [0.25, 0.5, 0.75]
        dropouts2 = [0.5]  # dropouts2 = [0.25, 0.5, 0.75]
        pool_sizes1 = [(2, 2)]  # pool_sizes1 = [(2, 2)]

        # create standard experiments structure
        self.experiments = {"filters1": filters1,
                            "filters2": filters2,
                            "losses1": losses1,
                            "units1": units1,
                            "optimizers1": optimizers1,
                            "kernel_sizes1": kernel_sizes1,
                            "dropouts1": dropouts1,
                            "dropouts2": dropouts2,
                            "pool_sizes1": pool_sizes1}

def get_optimizer(name='Adadelta'):
    if name == 'SGD':
        return optimizers.SGD(clipnorm=1.)
    if name == 'RMSprop':
        return optimizers.RMSprop(clipnorm=1.)
    if name == 'Adagrad':
        return optimizers.Adagrad(clipnorm=1.)
    if name == 'Adadelta':
        return optimizers.Adadelta(clipnorm=1.)
    if name == 'Adam':
        return optimizers.Adam(clipnorm=1.)
    if name == 'Adamax':
        return optimizers.Adamax(clipnorm=1.)
    if name == 'Nadam':
        return optimizers.Nadam(clipnorm=1.)

    return optimizers.Adam(clipnorm=1.)

def test_fuctional_model_saving():
    input = Input(shape=(3,))
    x = Dense(2)(input)
    output = Dense(3)(x)

    model = Model(input, output)
    model.compile(loss=objectives.MSE,
                  optimizer=optimizers.RMSprop(lr=0.0001),
                  metrics=[metrics.categorical_accuracy])
    x = np.random.random((1, 3))
    y = np.random.random((1, 3))
    model.train_on_batch(x, y)

    out = model.predict(x)
    _, fname = tempfile.mkstemp('.h5')
    save_model(model, fname)

    model = load_model(fname)
    os.remove(fname)

    out2 = model.predict(x)
    assert_allclose(out, out2, atol=1e-05)

def test_saving_lambda_custom_objects():
    input = Input(shape=(3,))
    x = Lambda(lambda x: square_fn(x), output_shape=(3,))(input)
    output = Dense(3)(x)

    model = Model(input, output)
    model.compile(loss=objectives.MSE,
                  optimizer=optimizers.RMSprop(lr=0.0001),
                  metrics=[metrics.categorical_accuracy])
    x = np.random.random((1, 3))
    y = np.random.random((1, 3))
    model.train_on_batch(x, y)

    out = model.predict(x)
    _, fname = tempfile.mkstemp('.h5')
    save_model(model, fname)

    model = load_model(fname, custom_objects={'square_fn': square_fn})
    os.remove(fname)

    out2 = model.predict(x)
    assert_allclose(out, out2, atol=1e-05)

def init_model():
    start_time = time.time()
    print 'Compiling Model ... '
    model = Sequential()
    model.add(Dense(500, input_dim=784))
    model.add(Activation('relu'))
    model.add(Dropout(0.4))
    model.add(Dense(300))
    model.add(Activation('relu'))
    model.add(Dropout(0.4))
    model.add(Dense(10))
    model.add(Activation('softmax'))

    rms = RMSprop()
    model.compile(loss='categorical_crossentropy', optimizer=rms, metrics=['accuracy'])
    print 'Model compield in {0} seconds'.format(time.time() - start_time)
    return model

def createNetwork(self):

        model = Sequential()

        firstlayer = True
        for l in self.layers:
            if firstlayer:
                model.add(Dense(l.size, input_shape=self.input_shape))
                firstlayer = False
            else:
                model.add(Dense(l.size))
            model.add(Activation(l.activation))
            if l.dropout > 0:
                model.add(Dropout(l.dropout))

        # final part 
        model.add(Dense(self.noutputs))
        if Config.task_type == "classification":
            model.add(Activation('softmax'))

        model.compile(loss=Config.loss,
                      optimizer=RMSprop())

        return model

def create_model(self, epsilon):
        """Return a compiled model and the state and action input
        layers with the given epsilon for numerical stability.
        """
        inputs = Input(shape=(self.state_shape,))
        action_input = Input(shape=(self.action_shape,))
        x1 = Dense(self.neurons_per_layer[0], activation='relu')(inputs)
        x1 = Dense(self.neurons_per_layer[1], activation='relu')(x1)
        x2 = Dense(self.neurons_per_layer[1], activation='relu')(action_input)
        x = add([x1, x2])
        for n in self.neurons_per_layer[2:]:
            x = Dense(n, activation='relu')(x)
        outputs = Dense(self.action_shape)(x)

        model = Model(inputs=[inputs, action_input], outputs=outputs)

        assert self.optimizer_choice in ['adam', 'rmsprop']
        if self.optimizer_choice == 'adam':
            opti = Adam(lr=self.alpha, epsilon=epsilon)
        else:
            opti = RMSprop(lr=self.alpha, epsilon=epsilon)
        model.compile(optimizer=opti, loss='mse')
        return model, inputs, action_input

def build_model(layers):
    model = Sequential()

    model.add(Dense(layers[1], input_shape=(20,), activation='relu'))
    model.add(Dropout(0.2))  # Dropout overfitting

    # model.add(Dense(layers[2],activation='tanh'))
    # model.add(Dropout(0.2))  # Dropout overfitting

    model.add(Dense(layers[2], activation='relu'))
    model.add(Dropout(0.2))  # Dropout overfitting

    model.add(Dense(output_dim=layers[3]))
    model.add(Activation("softmax"))

    model.summary()

    start = time.time()
    # sgd = SGD(lr=0.5, decay=1e-6, momentum=0.9, nesterov=True)
    # model.compile(loss="mse", optimizer=sgd)
    model.compile(loss="categorical_crossentropy", optimizer="adam", metrics=['accuracy']) # Nadam RMSprop()
    print "Compilation Time : ", time.time() - start
    return model

def createModel(self, inputs, outputs, hiddenLayers, activationType):
        model = Sequential()
        if len(hiddenLayers) == 0: 
            model.add(Dense(self.output_size, input_shape=(self.input_size,), init='lecun_uniform'))
            model.add(Activation("linear"))
        else :
            model.add(Dense(hiddenLayers[0], input_shape=(self.input_size,), init='lecun_uniform'))
            if (activationType == "LeakyReLU") :
                model.add(LeakyReLU(alpha=0.01))
            else :
                model.add(Activation(activationType))

            for index in range(1, len(hiddenLayers)-1):
                layerSize = hiddenLayers[index]
                model.add(Dense(layerSize, init='lecun_uniform'))
                if (activationType == "LeakyReLU") :
                    model.add(LeakyReLU(alpha=0.01))
                else :
                    model.add(Activation(activationType))
            model.add(Dense(self.output_size, init='lecun_uniform'))
            model.add(Activation("linear"))
        optimizer = optimizers.RMSprop(lr=1, rho=0.9, epsilon=1e-06)
        model.compile(loss="mse", optimizer=optimizer)
        return model

def createModel(self, inputs, outputs, hiddenLayers, activationType):
        model = Sequential()
        if len(hiddenLayers) == 0: 
            model.add(Dense(self.output_size, input_shape=(self.input_size,), init='lecun_uniform'))
            model.add(Activation("linear"))
        else :
            model.add(Dense(hiddenLayers[0], input_shape=(self.input_size,), init='lecun_uniform'))
            if (activationType == "LeakyReLU") :
                model.add(LeakyReLU(alpha=0.01))
            else :
                model.add(Activation(activationType))

            for index in range(1, len(hiddenLayers)-1):
                layerSize = hiddenLayers[index]
                model.add(Dense(layerSize, init='lecun_uniform'))
                if (activationType == "LeakyReLU") :
                    model.add(LeakyReLU(alpha=0.01))
                else :
                    model.add(Activation(activationType))
            model.add(Dense(self.output_size, init='lecun_uniform'))
            model.add(Activation("linear"))
        optimizer = optimizers.RMSprop(lr=1, rho=0.9, epsilon=1e-06)
        model.compile(loss="mse", optimizer=optimizer)
        return model

def __init__(self, action_space, batch_size=32, screen=(84, 84), swap_freq=200):
        from keras.optimizers import RMSprop        
        # -----
        self.screen = screen
        self.input_depth = 1
        self.past_range = 3
        self.observation_shape = (self.input_depth * self.past_range,) + self.screen
        self.batch_size = batch_size

        _, _, self.train_net, adventage = build_network(self.observation_shape, action_space.n)

        self.train_net.compile(optimizer=RMSprop(epsilon=0.1, rho=0.99),
                               loss=[value_loss(), policy_loss(adventage, args.beta)])

        self.pol_loss = deque(maxlen=25)
        self.val_loss = deque(maxlen=25)
        self.values = deque(maxlen=25)
        self.entropy = deque(maxlen=25)
        self.swap_freq = swap_freq
        self.swap_counter = self.swap_freq
        self.unroll = np.arange(self.batch_size)
        self.targets = np.zeros((self.batch_size, action_space.n))
        self.counter = 0

def __init__(self, action_space, batch_size=32, screen=(84, 84), swap_freq=200):
        from keras.optimizers import RMSprop
        # -----
        self.screen = screen
        self.input_depth = 1
        self.past_range = 3
        self.observation_shape = (self.input_depth * self.past_range,) + self.screen
        self.batch_size = batch_size

        self.action_value = build_network(self.observation_shape, action_space.n)
        self.action_value.compile(optimizer=RMSprop(clipnorm=1.), loss='mse')

        self.losses = deque(maxlen=25)
        self.q_values = deque(maxlen=25)
        self.swap_freq = swap_freq
        self.swap_counter = self.swap_freq
        self.unroll = np.arange(self.batch_size)
        self.frames = 0

def __init__(self, action_space, screen=(84, 84), n_step=8, discount=0.99):
        from keras.optimizers import RMSprop
        # -----
        self.screen = screen
        self.input_depth = 1
        self.past_range = 3
        self.observation_shape = (self.input_depth * self.past_range,) + self.screen

        self.action_value = build_network(self.observation_shape, action_space.n)
        self.action_value.compile(optimizer=RMSprop(clipnorm=1.), loss='mse')  # clipnorm=1.

        self.action_space = action_space
        self.observations = np.zeros(self.observation_shape)
        self.last_observations = np.zeros_like(self.observations)
        # -----
        self.n_step_observations = deque(maxlen=n_step)
        self.n_step_actions = deque(maxlen=n_step)
        self.n_step_rewards = deque(maxlen=n_step)
        self.n_step = n_step
        self.discount = discount
        self.counter = 0

def get_optimizer(args):

    clipvalue = 0
    clipnorm = 10

    if args.algorithm == 'rmsprop':
        optimizer = opt.RMSprop(lr=0.001, rho=0.9, epsilon=1e-06, clipnorm=clipnorm, clipvalue=clipvalue)
    elif args.algorithm == 'sgd':
        optimizer = opt.SGD(lr=0.01, momentum=0.0, decay=0.0, nesterov=False, clipnorm=clipnorm, clipvalue=clipvalue)
    elif args.algorithm == 'adagrad':
        optimizer = opt.Adagrad(lr=0.01, epsilon=1e-06, clipnorm=clipnorm, clipvalue=clipvalue)
    elif args.algorithm == 'adadelta':
        optimizer = opt.Adadelta(lr=1.0, rho=0.95, epsilon=1e-06, clipnorm=clipnorm, clipvalue=clipvalue)
    elif args.algorithm == 'adam':
        optimizer = opt.Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, clipnorm=clipnorm, clipvalue=clipvalue)
    elif args.algorithm == 'adamax':
        optimizer = opt.Adamax(lr=0.002, beta_1=0.9, beta_2=0.999, epsilon=1e-08, clipnorm=clipnorm, clipvalue=clipvalue)

    return optimizer

def main():
    from keras.optimizers import Adam, RMSprop, SGD
    from metric import dice_coef, dice_coef_loss
    import numpy as np
    img_rows = IMG_ROWS
    img_cols = IMG_COLS

    optimizer = RMSprop(lr=0.045, rho=0.9, epsilon=1.0)
    model = get_unet(Adam(lr=1e-5))
    model.compile(optimizer=optimizer, loss=dice_coef_loss, metrics=[dice_coef])

    x = np.random.random((1, 1,img_rows,img_cols))
    res = model.predict(x, 1)
    print res
    #print 'res', res[0].shape
    print 'params', model.count_params()
    print 'layer num', len(model.layers)
    #

def setOptimizer(self, **kwargs):

        """
        Sets a new optimizer for the Translation_Model.
        :param **kwargs:
        """

        # compile differently depending if our model is 'Sequential' or 'Graph'
        if self.verbose > 0:
            logging.info("Preparing optimizer and compiling.")
        if self.params['OPTIMIZER'].lower() == 'adam':
            optimizer = Adam(lr=self.params['LR'], clipnorm=self.params['CLIP_C'])
        elif self.params['OPTIMIZER'].lower() == 'rmsprop':
            optimizer = RMSprop(lr=self.params['LR'], clipnorm=self.params['CLIP_C'])
        elif self.params['OPTIMIZER'].lower() == 'nadam':
            optimizer = Nadam(lr=self.params['LR'], clipnorm=self.params['CLIP_C'])
        elif self.params['OPTIMIZER'].lower() == 'adadelta':
            optimizer = Adadelta(lr=self.params['LR'], clipnorm=self.params['CLIP_C'])
        elif self.params['OPTIMIZER'].lower() == 'sgd':
            optimizer = SGD(lr=self.params['LR'], clipnorm=self.params['CLIP_C'])
        else:
            logging.info('\tWARNING: The modification of the LR is not implemented for the chosen optimizer.')
            optimizer = eval(self.params['OPTIMIZER'])
        self.model.compile(optimizer=optimizer, loss=self.params['LOSS'],
                           sample_weight_mode='temporal' if self.params['SAMPLE_WEIGHTS'] else None)

def test_fuctional_model_saving():
    input = Input(shape=(3,))
    x = Dense(2)(input)
    output = Dense(3)(x)

    model = Model(input, output)
    model.compile(loss=objectives.MSE,
                  optimizer=optimizers.RMSprop(lr=0.0001),
                  metrics=[metrics.categorical_accuracy])
    x = np.random.random((1, 3))
    y = np.random.random((1, 3))
    model.train_on_batch(x, y)

    out = model.predict(x)
    _, fname = tempfile.mkstemp('.h5')
    save_model(model, fname)

    model = load_model(fname)
    os.remove(fname)

    out2 = model.predict(x)
    assert_allclose(out, out2, atol=1e-05)

def test_saving_lambda_custom_objects():
    input = Input(shape=(3,))
    x = Lambda(lambda x: square_fn(x), output_shape=(3,))(input)
    output = Dense(3)(x)

    model = Model(input, output)
    model.compile(loss=objectives.MSE,
                  optimizer=optimizers.RMSprop(lr=0.0001),
                  metrics=[metrics.categorical_accuracy])
    x = np.random.random((1, 3))
    y = np.random.random((1, 3))
    model.train_on_batch(x, y)

    out = model.predict(x)
    _, fname = tempfile.mkstemp('.h5')
    save_model(model, fname)

    model = load_model(fname, custom_objects={'square_fn': square_fn})
    os.remove(fname)

    out2 = model.predict(x)
    assert_allclose(out, out2, atol=1e-05)

def _init_MLP(self, mlp_layers):
        """Initialize the MLP that corresponds to the Q function.

        Parameters
        ----------
        state_dim : The state dimensionality
        nb_actions : The number of possible actions
        mlp_layers : A list consisting of an integer number of neurons for each
                     hidden layer. Default = [20, 20]
        """
        model = Sequential()
        for i in range(len(mlp_layers)):
            if i == 0:
                model.add(Dense(mlp_layers[i],
                          input_dim=self.state_dim + self.nb_actions))
            else:
                model.add(Dense(mlp_layers[i]))
            model.add(Activation('relu'))
        model.add(Dense(1))
        model.add(Activation('relu'))
        rmsprop = RMSprop()
        model.compile(loss='mean_squared_error', optimizer=rmsprop)
        return model

def make_discriminator(self):
        # TODO just to have something, 5 layers vgg-like
        inputs = Input(shape=self.img_shape)
        enc1 = self.downsampling_block_basic(inputs, 64, 7)
        enc2 = self.downsampling_block_basic(enc1,   64, 7)
        enc3 = self.downsampling_block_basic(enc2,   92, 7)
        enc4 = self.downsampling_block_basic(enc3,  128, 7)
        enc5 = self.downsampling_block_basic(enc4,  128, 7)
        flat = Flatten()(enc5)
        dense1 = Dense(512, activation='sigmoid')(flat)
        dense2 = Dense(512, activation='sigmoid')(dense1)
        fake = Dense(1, activation='sigmoid', name='generation')(dense2)
        # Dense(2,... two classes : real and fake
        # change last activation to softmax ?
        discriminator = kmodels.Model(input=inputs, output=fake)

        lr = 1e-04
        optimizer = RMSprop(lr=lr, rho=0.9, epsilon=1e-8, clipnorm=10)
        print ('   Optimizer discriminator: rmsprop. Lr: {}. Rho: 0.9, epsilon=1e-8, '
               'clipnorm=10'.format(lr))

        discriminator.compile(loss='binary_crossentropy', optimizer=optimizer)
        # TODO metrics=metrics,
        return discriminator

def build_functions(self):
        S = Input(shape=self.state_size)
        NS = Input(shape=self.state_size)
        A = Input(shape=(1,), dtype='int32')
        R = Input(shape=(1,), dtype='float32')
        T = Input(shape=(1,), dtype='int32')
        self.build_model()
        self.value_fn = K.function([S], self.model(S))

        VS = self.model(S)
        VNS = disconnected_grad(self.model(NS))
        future_value = (1-T) * VNS.max(axis=1, keepdims=True)
        discounted_future_value = self.discount * future_value
        target = R + discounted_future_value
        cost = ((VS[:, A] - target)**2).mean()
        opt = RMSprop(0.0001)
        params = self.model.trainable_weights
        updates = opt.get_updates(params, [], cost)
        self.train_fn = K.function([S, NS, A, R, T], cost, updates=updates)

def image_caption_model(vocab_size=2500, embedding_matrix=None, lang_dim=100,
            max_caplen=28, img_dim=2048, clipnorm=1):
    print('generating vocab_history model v5')
    # text: current word
    lang_input = Input(shape=(1,))
    img_input = Input(shape=(img_dim,))
    seq_input = Input(shape=(max_caplen,))
    vhist_input = Input(shape=(vocab_size,))

    if embedding_matrix is not None:
        x = Embedding(output_dim=lang_dim, input_dim=vocab_size, init='glorot_uniform', input_length=1, weights=[embedding_matrix])(lang_input)
    else:
        x = Embedding(output_dim=lang_dim, input_dim=vocab_size, init='glorot_uniform', input_length=1)(lang_input)

    lang_embed = Reshape((lang_dim,))(x)
    lang_embed = merge([lang_embed, seq_input], mode='concat', concat_axis=-1)
    lang_embed = Dense(lang_dim)(lang_embed)
    lang_embed = Dropout(0.25)(lang_embed)

    merge_layer = merge([img_input, lang_embed, vhist_input], mode='concat', concat_axis=-1)
    merge_layer = Reshape((1, lang_dim+img_dim+vocab_size))(merge_layer)

    gru_1 = GRU(img_dim)(merge_layer)
    gru_1 = Dropout(0.25)(gru_1)
    gru_1 = Dense(img_dim)(gru_1)
    gru_1 = BatchNormalization()(gru_1)
    gru_1 = Activation('softmax')(gru_1)

    attention_1 = merge([img_input, gru_1], mode='mul', concat_axis=-1)
    attention_1 = merge([attention_1, lang_embed, vhist_input], mode='concat', concat_axis=-1)
    attention_1 = Reshape((1, lang_dim + img_dim + vocab_size))(attention_1)
    gru_2 = GRU(1024)(attention_1)
    gru_2 = Dropout(0.25)(gru_2)
    gru_2 = Dense(vocab_size)(gru_2)
    gru_2 = BatchNormalization()(gru_2)
    out = Activation('softmax')(gru_2)

    model = Model(input=[img_input, lang_input, seq_input, vhist_input], output=out)
    model.compile(loss='categorical_crossentropy', optimizer=RMSprop(lr=0.0001, clipnorm=1.))
    return model

def run_benchmark(self, gpus=0):
        input_dim_1 = 40
        input_dim_2 = 60

        input_shape = (self.num_samples, input_dim_1, 60)
        x, y = generate_text_input_data(input_shape)

        # build the model: a single LSTM
        model = Sequential()
        model.add(LSTM(128, input_shape=(input_dim_1, input_dim_2)))
        model.add(Dense(input_dim_2), activation='softmax')

        optimizer = RMSprop(lr=0.01)

        if keras.backend.backend() is "tensorflow" and gpus > 1:
            model = multi_gpu_model(model, gpus=gpus)

        model.compile(loss='categorical_crossentropy', optimizer=optimizer)

        # create a distributed trainer for cntk
        if keras.backend.backend() is "cntk" and gpus > 1:
            start, end = cntk_gpu_mode_config(model, x.shape[0])
            x = x[start: end]
            y = y[start: end]

        time_callback = timehistory.TimeHistory()

        model.fit(x, y,
                  batch_size=self.batch_size,
                  epochs=self.epochs,
                  callbacks=[time_callback])

        self.total_time = 0
        for i in range(1, self.epochs):
            self.total_time += time_callback.times[i]

def run_benchmark(self, gpus=0):
        num_classes = 10

        # Generate random input data
        input_shape = (self.num_samples, 28, 28)
        x_train, y_train = generate_img_input_data(input_shape)

        x_train = x_train.reshape(self.num_samples, 784)
        x_train = x_train.astype('float32')
        x_train /= 255

        # convert class vectors to binary class matrices
        y_train = keras.utils.to_categorical(y_train, num_classes)

        model = Sequential()
        model.add(Dense(512, activation='relu', input_shape=(784,)))
        model.add(Dropout(0.2))
        model.add(Dense(512, activation='relu'))
        model.add(Dropout(0.2))
        model.add(Dense(num_classes, activation='softmax'))

        if keras.backend.backend() is "tensorflow" and gpus > 1:
            model = multi_gpu_model(model, gpus=gpus)

        model.compile(loss='categorical_crossentropy',
                      optimizer=RMSprop(),
                      metrics=['accuracy'])

        # create a distributed trainer for cntk
        if keras.backend.backend() is "cntk" and gpus > 1:
            start, end = cntk_gpu_mode_config(model, x_train.shape[0])
            x_train = x_train[start: end]
            y_train = y_train[start: end]

        time_callback = timehistory.TimeHistory()
        model.fit(x_train, y_train, batch_size=self.batch_size,
                  epochs=self.epochs, verbose=1, callbacks=[time_callback])

        self.total_time = 0
        for i in range(1, self.epochs):
            self.total_time += time_callback.times[i]

def init_model_1():
    start_time = time.time()
    print 'Compiling model...'
    model = Sequential()
    model.add(Convolution2D(64, 3,3, border_mode='valid',input_shape=INPUT_SHAPE))
    model.add(Activation('relu'))

    model.add(Convolution2D(64, 3,3, border_mode='valid'))
    model.add(Activation('relu'))

    model.add(MaxPooling2D(pool_size=(2,2)))
    model.add(Dropout(.25))

    model.add(Flatten())

    model.add(Dense(128))
    model.add(Activation('relu'))
    model.add(Dropout(.5))

    model.add(Dense(10))
    model.add(Activation('softmax'))

    rms = RMSprop()
    model.compile(loss='categorical_crossentropy', optimizer=rms,
      metrics=['accuracy'])
    print 'Model compiled in {0} seconds'.format(time.time() - start_time)
    return model

def __init__(self, data_dir, model_filename, weights_filename, categories_filename, history_filename, train_test_split_percentage, num_training_epochs, verbose):
        self.sample_size = 64 # input image size (will rescale data to this size)
        self.optimizer = RMSprop(lr=0.001, rho=0.9, epsilon=1e-08)
        self.train_test_split_percentage = train_test_split_percentage
        self.data_dir = data_dir
        self.model_filename = model_filename
        self.weights_filename = weights_filename
        self.categories_filename = categories_filename
        self.history_filename = history_filename
        self.num_training_epochs = num_training_epochs
        self.verbose = verbose

def __init__(self, data_dir, model_architecture, model_filename, weights_filename, categories_filename, history_filename, train_test_split_percentage, num_training_epochs, verbose):
        self.sample_size = 224 # this needs to match the input size of the pre-trained network, or you need to remove the input layer from the current network and replace it with the input size you would like
        self.optimizer = RMSprop(lr=0.001, rho=0.9, epsilon=1e-08)
        self.train_test_split_percentage = train_test_split_percentage
        self.data_dir = data_dir
        self.model_architecture = model_architecture
        self.model_filename = model_filename
        self.weights_filename = weights_filename
        self.categories_filename = categories_filename
        self.history_filename = history_filename
        self.num_training_epochs = num_training_epochs
        self.verbose = verbose

def test_rmsprop():
    _test_optimizer(RMSprop())
    _test_optimizer(RMSprop(decay=1e-3))

def test_sequential_model_saving():
    model = Sequential()
    model.add(Dense(2, input_dim=3))
    model.add(RepeatVector(3))
    model.add(TimeDistributed(Dense(3)))
    model.compile(loss=objectives.MSE,
                  optimizer=optimizers.RMSprop(lr=0.0001),
                  metrics=[metrics.categorical_accuracy],
                  sample_weight_mode='temporal')
    x = np.random.random((1, 3))
    y = np.random.random((1, 3, 3))
    model.train_on_batch(x, y)

    out = model.predict(x)
    _, fname = tempfile.mkstemp('.h5')
    save_model(model, fname)

    new_model = load_model(fname)
    os.remove(fname)

    out2 = new_model.predict(x)
    assert_allclose(out, out2, atol=1e-05)

    # test that new updates are the same with both models
    x = np.random.random((1, 3))
    y = np.random.random((1, 3, 3))
    model.train_on_batch(x, y)
    new_model.train_on_batch(x, y)
    out = model.predict(x)
    out2 = new_model.predict(x)
    assert_allclose(out, out2, atol=1e-05)

def _compile(self):
        """ compile self.q_vals
        """
        if (self._update_rule=="sgd"):
            optimizer = SGD(lr=self._lr, momentum=self._momentum, nesterov=False)
        elif (self._update_rule=="rmsprop"):
            optimizer = RMSprop(lr=self._lr, rho=self._rho, epsilon=self._rms_epsilon)
        else:
            raise Exception('The update_rule '+self._update_rule+' is not implemented.')

        self.q_vals.compile(optimizer=optimizer, loss='mse')

def build_stateful_lstm_model(batch_size,time_step,input_dim,output_dim,dropout=0.2,rnn_layer_num=2,hidden_dim=128,hidden_num=0,rnn_type='LSTM'):

    model = Sequential()
    # may use BN for accelerating speed
    # add first LSTM
    if rnn_type == 'LSTM':
        rnn_cell = LSTM
    elif rnn_type == 'GRU':
        rnn_cell = GRU
    elif rnn_type == 'SimpleRNN':
        rnn_cell = SimpleRNN
    else:
        raise ValueError('Option rnn_type could only be configured as LSTM, GRU or SimpleRNN')
    model.add(rnn_cell(hidden_dim,return_sequences=True,batch_input_shape=(batch_size,time_step,input_dim)))

    for _ in range(rnn_layer_num-2):
        model.add(rnn_cell(hidden_dim, return_sequence=True))
        # prevent over fitting
        model.add(Dropout(dropout))



    model.add(rnn_cell(hidden_dim,return_sequences=False))

    # add hidden layer

    for _ in range(hidden_num):
        model.add(Dense(hidden_dim))

    model.add(Dropout(dropout))

    model.add(Dense(output_dim))

    rmsprop = RMSprop(lr=0.01)
    adam = Adam(lr=0.01)


    model.compile(loss='mse',metrics=['acc'],optimizer=rmsprop)

    return model

def discriminator_model(self):
        if self.DM:
            return self.DM
        optimizer = RMSprop(lr=0.0002, decay=6e-8)
        self.DM = Sequential()
        self.DM.add(self.discriminator())
        self.DM.compile(loss='binary_crossentropy', optimizer=optimizer,\
            metrics=['accuracy'])
        return self.DM

def adversarial_model(self):
        if self.AM:
            return self.AM
        optimizer = RMSprop(lr=0.0001, decay=3e-8)
        self.AM = Sequential()
        self.AM.add(self.generator())
        self.AM.add(self.discriminator())
        self.AM.compile(loss='binary_crossentropy', optimizer=optimizer,\
            metrics=['accuracy'])
        return self.AM