Python keras.backend.sum() Examples

def optimizer(self):
        action = K.placeholder(shape=[None, 5])
        discounted_rewards = K.placeholder(shape=[None, ])
        
        # 크로스 엔트로피 오류함수 계산
        action_prob = K.sum(action * self.model.output, axis=1)
        cross_entropy = K.log(action_prob) * discounted_rewards
        loss = -K.sum(cross_entropy)
        
        # 정책신경망을 업데이트하는 훈련함수 생성
        optimizer = Adam(lr=self.learning_rate)
        updates = optimizer.get_updates(self.model.trainable_weights,[],
                                        loss)
        train = K.function([self.model.input, action, discounted_rewards], [],
                           updates=updates)

        return train

    # 정책신경망으로 행동 선택 

def rpn_loss_regr(num_anchors):
	def rpn_loss_regr_fixed_num(y_true, y_pred):
		if K.image_dim_ordering() == 'th':
			x = y_true[:, 4 * num_anchors:, :, :] - y_pred
			x_abs = K.abs(x)
			x_bool = K.less_equal(x_abs, 1.0)
			return lambda_rpn_regr * K.sum(
				y_true[:, :4 * num_anchors, :, :] * (x_bool * (0.5 * x * x) + (1 - x_bool) * (x_abs - 0.5))) / K.sum(epsilon + y_true[:, :4 * num_anchors, :, :])
		else:
			x = y_true[:, :, :, 4 * num_anchors:] - y_pred
			x_abs = K.abs(x)
			x_bool = K.cast(K.less_equal(x_abs, 1.0), tf.float32)

			return lambda_rpn_regr * K.sum(
				y_true[:, :, :, :4 * num_anchors] * (x_bool * (0.5 * x * x) + (1 - x_bool) * (x_abs - 0.5))) / K.sum(epsilon + y_true[:, :, :, :4 * num_anchors])

	return rpn_loss_regr_fixed_num 

def optimizer(self):
        action = K.placeholder(shape=[None, 5])
        discounted_rewards = K.placeholder(shape=[None, ])
        
        # 크로스 엔트로피 오류함수 계산
        action_prob = K.sum(action * self.model.output, axis=1)
        cross_entropy = K.log(action_prob) * discounted_rewards
        loss = -K.sum(cross_entropy)
        
        # 정책신경망을 업데이트하는 훈련함수 생성
        optimizer = Adam(lr=self.learning_rate)
        updates = optimizer.get_updates(self.model.trainable_weights,[],
                                        loss)
        train = K.function([self.model.input, action, discounted_rewards], [],
                           updates=updates)

        return train

    # 정책신경망으로 행동 선택 

def gradient_penalty_loss(self, y_true, y_pred, averaged_samples):
        """
        Computes gradient penalty based on prediction and weighted real / fake samples
        """
        gradients = K.gradients(y_pred, averaged_samples)[0]
        # compute the euclidean norm by squaring ...
        gradients_sqr = K.square(gradients)
        #   ... summing over the rows ...
        gradients_sqr_sum = K.sum(gradients_sqr,
                                  axis=np.arange(1, len(gradients_sqr.shape)))
        #   ... and sqrt
        gradient_l2_norm = K.sqrt(gradients_sqr_sum)
        # compute lambda * (1 - ||grad||)^2 still for each single sample
        gradient_penalty = K.square(1 - gradient_l2_norm)
        # return the mean as loss over all the batch samples
        return K.mean(gradient_penalty) 

def actor_optimizer(self):
        action = K.placeholder(shape=(None, self.action_size))
        advantages = K.placeholder(shape=(None, ))

        policy = self.actor.output

        good_prob = K.sum(action * policy, axis=1)
        eligibility = K.log(good_prob + 1e-10) * K.stop_gradient(advantages)
        loss = -K.sum(eligibility)

        entropy = K.sum(policy * K.log(policy + 1e-10), axis=1)

        actor_loss = loss + 0.01*entropy

        optimizer = Adam(lr=self.actor_lr)
        updates = optimizer.get_updates(self.actor.trainable_weights, [], actor_loss)
        train = K.function([self.actor.input, action, advantages], [], updates=updates)
        return train

    # make loss function for Value approximation 

def audio_discriminate_loss2(gamma=0.1,beta = 2*0.1,num_speaker=2):
    def loss_func(S_true,S_pred,gamma=gamma,beta=beta,num_speaker=num_speaker):
        sum_mtr = K.zeros_like(S_true[:,:,:,:,0])
        for i in range(num_speaker):
            sum_mtr += K.square(S_true[:,:,:,:,i]-S_pred[:,:,:,:,i])
            for j in range(num_speaker):
                if i != j:
                    sum_mtr -= gamma*(K.square(S_true[:,:,:,:,i]-S_pred[:,:,:,:,j]))

        for i in range(num_speaker):
            for j in range(i+1,num_speaker):
                #sum_mtr -= beta*K.square(S_pred[:,:,:,i]-S_pred[:,:,:,j])
                #sum_mtr += beta*K.square(S_true[:,:,:,:,i]-S_true[:,:,:,:,j])
                pass
        #sum = K.sum(K.maximum(K.flatten(sum_mtr),0))

        loss = K.mean(K.flatten(sum_mtr))

        return loss
    return loss_func 

def optimizer(self):
        a = K.placeholder(shape=(None, ), dtype='int32')
        y = K.placeholder(shape=(None, ), dtype='float32')

        py_x = self.model.output

        a_one_hot = K.one_hot(a, self.action_size)
        q_value = K.sum(py_x * a_one_hot, axis=1)
        error = K.abs(y - q_value)

        quadratic_part = K.clip(error, 0.0, 1.0)
        linear_part = error - quadratic_part
        loss = K.mean(0.5 * K.square(quadratic_part) + linear_part)

        optimizer = RMSprop(lr=0.00025, epsilon=0.01)
        updates = optimizer.get_updates(self.model.trainable_weights, [], loss)
        train = K.function([self.model.input, a, y], [loss], updates=updates)

        return train

    # approximate Q function using Convolution Neural Network
    # state is input and Q Value of each action is output of network
    # dueling network's Q Value is sum of advantages and state value 

def actor_optimizer(self):
        action = K.placeholder(shape=[None, self.action_size])
        advantages = K.placeholder(shape=[None, ])

        policy = self.actor.output

        good_prob = K.sum(action * policy, axis=1)
        eligibility = K.log(good_prob + 1e-10) * advantages
        actor_loss = -K.sum(eligibility)

        entropy = K.sum(policy * K.log(policy + 1e-10), axis=1)
        entropy = K.sum(entropy)

        loss = actor_loss + 0.01*entropy
        optimizer = RMSprop(lr=self.actor_lr, rho=0.99, epsilon=0.01)
        updates = optimizer.get_updates(self.actor.trainable_weights, [], loss)
        train = K.function([self.actor.input, action, advantages], [loss], updates=updates)

        return train

    # make loss function for Value approximation 

def reverse_generator(generator, X_sample, y_sample, title):
    """Gradient descent to map images back to their latent vectors."""

    latent_vec = np.random.normal(size=(1, 100))

    # Function for figuring out how to bump the input.
    target = K.placeholder()
    loss = K.sum(K.square(generator.outputs[0] - target))
    grad = K.gradients(loss, generator.inputs[0])[0]
    update_fn = K.function(generator.inputs + [target], [grad])

    # Repeatedly apply the update rule.
    xs = []
    for i in range(60):
        print('%d: latent_vec mean=%f, std=%f'
              % (i, np.mean(latent_vec), np.std(latent_vec)))
        xs.append(generator.predict_on_batch([latent_vec, y_sample]))
        for _ in range(10):
            update_vec = update_fn([latent_vec, y_sample, X_sample])[0]
            latent_vec -= update_vec * update_rate

    # Plots the samples.
    xs = np.concatenate(xs, axis=0)
    plot_as_gif(xs, X_sample, title) 

def build_generator():
    """Builds the big generator model."""

    latent = keras.layers.Input((100,), name='latent')

    image_class = keras.layers.Input((10,), dtype='float32',
                                     name='image_class')
    d = keras.layers.Dense(100)(image_class)
    merged = keras.layers.merge([latent, d], mode='sum')

    hidden = keras.layers.Dense(512)(merged)
    hidden = keras.layers.LeakyReLU()(hidden)

    hidden = keras.layers.Dense(512)(hidden)
    hidden = keras.layers.LeakyReLU()(hidden)

    output_layer = keras.layers.Dense(28 * 28,
        activation='tanh')(hidden)
    fake_image = keras.layers.Reshape((28, 28, 1))(output_layer)

    return keras.models.Model(input=[latent, image_class],
                              output=fake_image) 

def gen_adv_loss(logits, y, loss='logloss', mean=False):
    """
    Generate the loss function.
    """

    if loss == 'training':
        # use the model's output instead of the true labels to avoid
        # label leaking at training time
        y = K.cast(K.equal(logits, K.max(logits, 1, keepdims=True)), "float32")
        y = y / K.sum(y, 1, keepdims=True)
        out = K.categorical_crossentropy(y, logits, from_logits=True)
    elif loss == 'logloss':
        out = K.categorical_crossentropy(y, logits, from_logits=True)
    else:
        raise ValueError("Unknown loss: {}".format(loss))

    if mean:
        out = K.mean(out)
    # else:
    #     out = K.sum(out)
    return out 

def gen_adv_loss(logits, y, loss='logloss', mean=False):
    """
    Generate the loss function.
    """

    if loss == 'training':
        # use the model's output instead of the true labels to avoid
        # label leaking at training time
        y = K.cast(K.equal(logits, K.max(logits, 1, keepdims=True)), "float32")
        y = y / K.sum(y, 1, keepdims=True)
        out = K.categorical_crossentropy(logits, y, from_logits=True)
    elif loss == 'logloss':
        # out = K.categorical_crossentropy(logits, y, from_logits=True)
        out = tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=y)
        out = tf.reduce_mean(out)
    else:
        raise ValueError("Unknown loss: {}".format(loss))

    if mean:
        out = tf.mean(out)
    # else:
    #     out = K.sum(out)
    return out 

def call(self, x, mask=None):
        # computes a probability distribution over the timesteps
        # uses 'max trick' for numerical stability
        # reshape is done to avoid issue with Tensorflow
        # and 1-dimensional weights
        logits = K.dot(x, self.W)
        x_shape = K.shape(x)
        logits = K.reshape(logits, (x_shape[0], x_shape[1]))
        ai = K.exp(logits - K.max(logits, axis=-1, keepdims=True))

        # masked timesteps have zero weight
        if mask is not None:
            mask = K.cast(mask, K.floatx())
            ai = ai * mask
        att_weights = ai / (K.sum(ai, axis=1, keepdims=True) + K.epsilon())
        weighted_input = x * K.expand_dims(att_weights)
        result = K.sum(weighted_input, axis=1)
        if self.return_attention:
            return [result, att_weights]
        return result 

def step(self, x, states):  
        h = states[0]
        # states[1] necessary?
        
        # comes from the constants
        X_static = states[-2]
        # equals K.dot(static_x, self._W1) + self._b2 with X.shape=[bs, L, static_input_dim]
        total_x_static_prod = states[-1]

        # expand dims to add the vector which is only valid for this time step
        # to total_x_prod which is valid for all time steps
        hw = K.expand_dims(K.dot(h, self._W2), 1)
        additive_atn = total_x_static_prod + hw
        attention = K.softmax(K.dot(additive_atn, self._V), axis=1)
        static_x_weighted = K.sum(attention * X_static, [1])
        
        x = K.dot(K.concatenate([x, static_x_weighted], 1), self._W3) + self._b3

        h, new_states = self.layer.cell.call(x, states[:-2])
        
        # append attention to the states to "smuggle" it out of the RNN wrapper
        attention = K.squeeze(attention, -1)
        h = K.concatenate([h, attention])

        return h, new_states 

def step(self, x, states):   
        h = states[0]
        # states[1] necessary?

        # equals K.dot(X, self._W1) + self._b2 with X.shape=[bs, T, input_dim]
        total_x_prod = states[-1]
        # comes from the constants (equals the input sequence)
        X = states[-2]
        
        # expand dims to add the vector which is only valid for this time step
        # to total_x_prod which is valid for all time steps
        hw = K.expand_dims(K.dot(h, self._W2), 1)
        additive_atn = total_x_prod + hw
        attention = K.softmax(K.dot(additive_atn, self._V), axis=1)
        x_weighted = K.sum(attention * X, [1])

        x = K.dot(K.concatenate([x, x_weighted], 1), self._W3) + self._b3
        
        h, new_states = self.layer.cell.call(x, states[:-2])
        
        return h, new_states 

def optimizer(self):
        a = K.placeholder(shape=(None,), dtype='int32')
        y = K.placeholder(shape=(None,), dtype='float32')

        py_x = self.model.output

        a_one_hot = K.one_hot(a, self.action_size)
        q_value = K.sum(py_x * a_one_hot, axis=1)
        error = K.abs(y - q_value)

        quadratic_part = K.clip(error, 0.0, 1.0)
        linear_part = error - quadratic_part
        loss = K.mean(0.5 * K.square(quadratic_part) + linear_part)

        optimizer = RMSprop(lr=0.00025, epsilon=0.01)
        updates = optimizer.get_updates(self.model.trainable_weights, [], loss)
        train = K.function([self.model.input, a, y], [loss], updates=updates)

        return train

    # approximate Q function using Convolution Neural Network
    # state is input and Q Value of each action is output of network 

def cor(self,y1, y2, lamda):
        y1_mean = K.mean(y1, axis=0)
        y1_centered = y1 - y1_mean
        y2_mean = K.mean(y2, axis=0)
        y2_centered = y2 - y2_mean
        corr_nr = K.sum(y1_centered * y2_centered, axis=0)
        corr_dr1 = K.sqrt(T.sum(y1_centered * y1_centered, axis=0) + 1e-8)
        corr_dr2 = K.sqrt(T.sum(y2_centered * y2_centered, axis=0) + 1e-8)
        corr_dr = corr_dr1 * corr_dr2
        corr = corr_nr / corr_dr
        return K.sum(corr) * lamda 

def bbalpha_softmax_cross_entropy_with_mc_logits(alpha):
    alpha = K.cast_to_floatx(alpha)
    def loss(y_true, mc_logits):
        # log(p_ij), p_ij = softmax(logit_ij)
        #assert mc_logits.ndim == 3
        mc_log_softmax = mc_logits - K.max(mc_logits, axis=2, keepdims=True)
        mc_log_softmax = mc_log_softmax - K.log(K.sum(K.exp(mc_log_softmax), axis=2, keepdims=True))
        mc_ll = K.sum(y_true * mc_log_softmax, -1)  # N x K
        K_mc = mc_ll.get_shape().as_list()[1]	# only for tensorflow
        return - 1. / alpha * (logsumexp(alpha * mc_ll, 1) + K.log(1.0 / K_mc))
    return loss


###################################################################
# the model 

def test_MC_dropout(model, X, Y):
    pred = model.predict(X)  # N x K x D
    pred = np.mean(pred, 1)
    acc = np.mean(np.argmax(pred, axis=-1) == np.argmax(Y, axis=-1))
    ll = np.sum(np.log(np.sum(pred * Y, -1)))
    return acc, ll 

def logsumexp(x, axis=None):
    x_max = K.max(x, axis=axis, keepdims=True)
    return K.log(K.sum(K.exp(x - x_max), axis=axis, keepdims=True)) + x_max 

def compile_gradient_function(input_model, category_index, layer_name):
    model = Sequential()
    model.add(input_model)

    num_classes = model.output_shape[1]
    target_layer = lambda x: target_category_loss(x, category_index, num_classes)
    model.add(Lambda(target_layer,
                     output_shape = target_category_loss_output_shape))

    loss = K.sum(model.layers[-1].output)
    conv_output = model.layers[0].get_layer(layer_name).output
    gradients = normalize(K.gradients(loss, conv_output)[0])
    gradient_function = K.function([model.layers[0].input, K.learning_phase()],
                                                    [conv_output, gradients])
    return gradient_function 

def compile_gradient_function(input_model, category_index, layer_name):
    model = Sequential()
    model.add(input_model)

    num_classes = model.output_shape[1]
    target_layer = lambda x: target_category_loss(x, category_index, num_classes)
    model.add(Lambda(target_layer,
                     output_shape=target_category_loss_output_shape))

    loss = K.sum(model.layers[-1].output)
    conv_output = model.layers[0].get_layer(layer_name).output
    gradients = normalize(K.gradients(loss, conv_output)[0])
    gradient_function = K.function([model.layers[0].input, K.learning_phase()],
                                   [conv_output, gradients])
    return gradient_function 

def sum_corr(model):
    view1 = np.load("MFCC_Test.npy")
    view2 = np.load("XRMB_Test.npy")
    x = project(model,[view1,np.zeros_like(view2)])
    y = project(model,[np.zeros_like(view1),view2])
    print ("test correlation")
    corr = 0
    for i in range(0,len(x[0])):
        x1 = x[:,i] - (np.ones(len(x))*(sum(x[:,i])/len(x)))
        x2 = y[:,i] - (np.ones(len(y))*(sum(y[:,i])/len(y)))
        nr = sum(x1 * x2)/(math.sqrt(sum(x1*x1))*math.sqrt(sum(x2*x2)))
        corr+=nr
    print (corr) 

def call(self, x, mask=None):
        h1 = x[0]
        h2 = x[1]    
        dif = K.sum(K.abs(h1-h2),axis=1)  
        h = K.exp(-dif)
        #print h.shape
        h=K.clip(h,1e-7,1.0-1e-7)
        h = K.reshape(h, (h.shape[0],1))
        return h 

def compute_euclidean_match_score(l_r):
    l, r = l_r
    denominator = 1. + K.sqrt(
        -2 * K.batch_dot(l, r, axes=[2, 2]) +
        K.expand_dims(K.sum(K.square(l), axis=2), 2) +
        K.expand_dims(K.sum(K.square(r), axis=2), 1)
    )
    denominator = K.maximum(denominator, K.epsilon())
    return 1. / denominator 

def sum_corr(model):
    view1 = np.load("test_v1.npy")
    view2 = np.load("test_v2.npy")
    x = project(model,[view1,np.zeros_like(view1)])
    y = project(model,[np.zeros_like(view2),view2])
    print "test correlation"
    corr = 0
    for i in range(0,len(x[0])):
		x1 = x[:,i] - (np.ones(len(x))*(sum(x[:,i])/len(x)))
		x2 = y[:,i] - (np.ones(len(y))*(sum(y[:,i])/len(y)))
		nr = sum(x1 * x2)/(math.sqrt(sum(x1*x1))*math.sqrt(sum(x2*x2)))
		corr+=nr
    print corr 

def call(self,x):
        y_pred=x[0]
        y_recont_gt=x[1] 
        y_prob_pred=tf.squeeze(x[2],axis=3) 
        y_prob_gt=x[3]
        visible = tf.cast(y_prob_gt > 0.5,y_pred.dtype)
        visible = tf.squeeze(visible,axis=3) 
        #generate transformed values using sym
        if(len(self.sym)>1):
            #if(True):
            for sym_id,transform in enumerate(self.sym): #3x3 matrix
                tf_mat=tf.convert_to_tensor(transform,y_recont_gt.dtype)
                y_gt_transformed = tf.transpose(tf.matmul(tf_mat,tf.transpose(tf.reshape(y_recont_gt,[-1,3]))))
                y_gt_transformed = tf.reshape(y_gt_transformed,[-1,128,128,3])
                loss_xyz_temp = K.sum(K.abs(y_gt_transformed-y_pred),axis=3)/3
                loss_sum=K.sum(loss_xyz_temp,axis=[1,2])
                if(sym_id>0):
                    loss_sums = tf.concat([loss_sums,tf.expand_dims(loss_sum,axis=0)],axis=0)
                    loss_xyzs=  tf.concat([loss_xyzs,tf.expand_dims(loss_xyz_temp,axis=0)],axis=0)
                else:
                    loss_sums = tf.expand_dims(loss_sum,axis=0) 
                    loss_xyzs = tf.expand_dims(loss_xyz_temp,axis=0)
            
            min_values = tf.reduce_min(loss_sums,axis=0,keepdims=True) 
            loss_switch = tf.cast(tf.equal(loss_sums,min_values),y_pred.dtype)
            loss_xyz = tf.expand_dims(tf.expand_dims(loss_switch,axis=2),axis=3)*loss_xyzs
            loss_xyz = K.sum(loss_xyz,axis=0) 
        else:
            loss_xyz = K.sum(K.abs(y_recont_gt-y_pred),axis=3)/3
        prob_loss = K.square(y_prob_pred-K.minimum(loss_xyz,1)) 
        loss_invisible = (1-visible)*loss_xyz
        loss_visible = visible*loss_xyz
        loss = loss_visible*3 + loss_invisible+ 0.5*prob_loss 
        loss = K.mean(loss,axis=[1,2])
        return loss 

def build_model(self):
        input = Input(shape=self.state_size)
        shared = Conv2D(32, (8, 8), strides=(4, 4), activation='relu')(input)
        shared = Conv2D(64, (4, 4), strides=(2, 2), activation='relu')(shared)
        shared = Conv2D(64, (3, 3), strides=(1, 1), activation='relu')(shared)
        flatten = Flatten()(shared)

        # network separate state value and advantages
        advantage_fc = Dense(512, activation='relu')(flatten)
        advantage = Dense(self.action_size)(advantage_fc)
        advantage = Lambda(lambda a: a[:, :] - K.mean(a[:, :], keepdims=True),
                           output_shape=(self.action_size,))(advantage)

        value_fc = Dense(512, activation='relu')(flatten)
        value =  Dense(1)(value_fc)
        value = Lambda(lambda s: K.expand_dims(s[:, 0], -1),
                       output_shape=(self.action_size,))(value)

        # network merged and make Q Value
        q_value = merge([value, advantage], mode='sum')
        model = Model(inputs=input, outputs=q_value)
        model.summary()

        return model

    # after some time interval update the target model to be same with model 

def call(self, x):
        uit = dot_product(x, self.W)

        if self.bias:
            uit += self.b

        uit = K.tanh(uit)
        ait = dot_product(uit, self.u)

        a = K.softmax(ait)
        a = K.expand_dims(a)
        weighted_input = x * a

        return K.sum(weighted_input, axis=1) 

def optimizer(self):
        a = K.placeholder(shape=(None, ), dtype='int32')
        y = K.placeholder(shape=(None, ), dtype='float32')

        py_x = self.model.output

        a_one_hot = K.one_hot(a, self.action_size)
        q_value = K.sum(py_x * a_one_hot, axis=1)
        error = K.abs(y - q_value)

        quadratic_part = K.clip(error, 0.0, 1.0)
        linear_part = error - quadratic_part
        loss = K.mean(0.5 * K.square(quadratic_part) + linear_part)

        optimizer = RMSprop(lr=0.00025, epsilon=0.01)
        updates = optimizer.get_updates(self.model.trainable_weights, [], loss)
        train = K.function([self.model.input, a, y], [loss], updates=updates)

        return train

    # approximate Q function using Convolution Neural Network
    # state is input and Q Value of each action is output of network