Python backend.log方法代码示例

# 需要导入模块: from keras import backend [as 别名]
# 或者: from keras.backend import log [as 别名]
def actor_optimizer(self):
        action = K.placeholder(shape=[None, self.action_size])
        advantages = K.placeholder(shape=[None, ])

        policy = self.actor.output

        # 정책 크로스 엔트로피 오류함수
        action_prob = K.sum(action * policy, axis=1)
        cross_entropy = K.log(action_prob + 1e-10) * advantages
        cross_entropy = -K.sum(cross_entropy)

        # 탐색을 지속적으로 하기 위한 엔트로피 오류
        entropy = K.sum(policy * K.log(policy + 1e-10), axis=1)
        entropy = K.sum(entropy)

        # 두 오류함수를 더해 최종 오류함수를 만듬
        loss = cross_entropy + 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

    # 가치신경망을 업데이트하는 함수 

# 需要导入模块: from keras import backend [as 别名]
# 或者: from keras.backend import log [as 别名]
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 

# 需要导入模块: from keras import backend [as 别名]
# 或者: from keras.backend import log [as 别名]
def build_model(self):
        input = Input(shape=self.state_size)
        conv = Conv2D(16, (8, 8), strides=(4, 4), activation='relu')(input)
        conv = Conv2D(32, (4, 4), strides=(2, 2), activation='relu')(conv)
        conv = Flatten()(conv)
        fc = Dense(256, activation='relu')(conv)
        policy = Dense(self.action_size, activation='softmax')(fc)
        value = Dense(1, activation='linear')(fc)

        actor = Model(inputs=input, outputs=policy)
        critic = Model(inputs=input, outputs=value)

        actor._make_predict_function()
        critic._make_predict_function()

        actor.summary()
        critic.summary()

        return actor, critic

    # make loss function for Policy Gradient
    # [log(action probability) * advantages] will be input for the back prop
    # we add entropy of action probability to loss 

# 需要导入模块: from keras import backend [as 别名]
# 或者: from keras.backend import log [as 别名]
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 

# 需要导入模块: from keras import backend [as 别名]
# 或者: from keras.backend import log [as 别名]
def optimizer(self):
        action = K.placeholder(shape=[None, 5])
        discounted_rewards = K.placeholder(shape=[None, ])

        # Calculate cross entropy error function
        action_prob = K.sum(action * self.model.output, axis=1)
        cross_entropy = K.log(action_prob) * discounted_rewards
        loss = -K.sum(cross_entropy)

        # create training function
        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

    # get action from policy network 

# 需要导入模块: from keras import backend [as 别名]
# 或者: from keras.backend import log [as 别名]
def build_model(self):
        state = Input(batch_shape=(None,  self.state_size))
        shared = Dense(self.hidden1, input_dim=self.state_size, activation='relu', kernel_initializer='glorot_uniform')(state)

        actor_hidden = Dense(self.hidden2, activation='relu', kernel_initializer='glorot_uniform')(shared)
        action_prob = Dense(self.action_size, activation='softmax', kernel_initializer='glorot_uniform')(actor_hidden)

        value_hidden = Dense(self.hidden2, activation='relu', kernel_initializer='he_uniform')(shared)
        state_value = Dense(1, activation='linear', kernel_initializer='he_uniform')(value_hidden)

        actor = Model(inputs=state, outputs=action_prob)
        critic = Model(inputs=state, outputs=state_value)

        actor._make_predict_function()
        critic._make_predict_function()

        actor.summary()
        critic.summary()

        return actor, critic

    # make loss function for Policy Gradient
    # [log(action probability) * advantages] will be input for the back prop
    # we add entropy of action probability to loss 

# 需要导入模块: from keras import backend [as 别名]
# 或者: from keras.backend import log [as 别名]
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 

# 需要导入模块: from keras import backend [as 别名]
# 或者: from keras.backend import log [as 别名]
def __init__(self,
                 kind,
                 reduce_loss=True,
                 clip_prob=1e-6,
                 regularize=False,
                 location=None,
                 growth=None):

        self.kind = kind
        self.reduce_loss = reduce_loss
        self.clip_prob = clip_prob

        if regularize == True or location is not None or growth is not None:
            raise DeprecationWarning('Directly penalizing beta has been found \
                                      to be unneccessary when using bounded activation \
                                      and clipping of log-likelihood.\
                                      Use this method instead.') 

# 需要导入模块: from keras import backend [as 别名]
# 或者: from keras.backend import log [as 别名]
def loss_function(self, y_true, y_pred):

        y, u, a, b = _keras_split(y_true, y_pred)
        if self.kind == 'discrete':
            loglikelihoods = loglik_discrete(y, u, a, b)
        elif self.kind == 'continuous':
            loglikelihoods = loglik_continuous(y, u, a, b)

        if self.clip_prob is not None:
            loglikelihoods = K.clip(loglikelihoods, 
                log(self.clip_prob), log(1 - self.clip_prob))
        if self.reduce_loss:
            loss = -1.0 * K.mean(loglikelihoods, axis=-1)
        else:
            loss = -loglikelihoods

        return loss

# For backwards-compatibility 

# 需要导入模块: from keras import backend [as 别名]
# 或者: from keras.backend import log [as 别名]
def sort4minibatches(xvals, evals, tvals, batchsize):
    ntot = len(xvals)
    indices = np.arange(ntot)
    np.random.shuffle(indices)
    start_idx=0
    esall = []
    for end_idx in list(range(batchsize, batchsize*(ntot//batchsize)+1, batchsize))+[ntot]:
        excerpt = indices[start_idx:end_idx]
        sort_idx = np.argsort(tvals[excerpt])[::-1]
        es = excerpt[sort_idx]
        esall += list(es)
        start_idx = end_idx
    return (xvals[esall], evals[esall], tvals[esall], esall)


#Define Cox PH partial likelihood function loss.
#Arguments: E (censoring status), risk (risk [log hazard ratio] predicted by network) for batch of input subjects
#As defined, this function requires that all subjects in input batch must be sorted in descending order of survival/censoring time (i.e. arguments E and risk will be in this order) 

# 需要导入模块: from keras import backend [as 别名]
# 或者: from keras.backend import log [as 别名]
def crossentropy_reed_wrap(_beta):
    def crossentropy_reed_core(y_true, y_pred):
        """
        This loss function is proposed in:
        Reed et al. "Training Deep Neural Networks on Noisy Labels with Bootstrapping", 2014

        :param y_true:
        :param y_pred:
        :return:
        """

        # hyper param
        print(_beta)
        y_pred = K.clip(y_pred, K.epsilon(), 1)

        # (1) dynamically update the targets based on the current state of the model: bootstrapped target tensor
        # use predicted class proba directly to generate regression targets
        y_true_update = _beta * y_true + (1 - _beta) * y_pred

        # (2) compute loss as always
        _loss = -K.sum(y_true_update * K.log(y_pred), axis=-1)

        return _loss
    return crossentropy_reed_core 

# 需要导入模块: from keras import backend [as 别名]
# 或者: from keras.backend import log [as 别名]
def softmax_loss(y_true, y_pred):
    """Compute cross entropy loss aka softmax loss.

    # Arguments
        y_true: Ground truth targets,
            tensor of shape (?, num_boxes, num_classes).
        y_pred: Predicted logits,
            tensor of shape (?, num_boxes, num_classes).

    # Returns
        softmax_loss: Softmax loss, tensor of shape (?, num_boxes).
    """
    eps = K.epsilon()
    y_pred = K.clip(y_pred, eps, 1. - eps)
    softmax_loss = -tf.reduce_sum(y_true * tf.log(y_pred), axis=-1)
    return softmax_loss 

# 需要导入模块: from keras import backend [as 别名]
# 或者: from keras.backend import log [as 别名]
def focal_loss(y_true, y_pred, gamma=2, alpha=0.25):
    """Compute focal loss.
    
    # Arguments
        y_true: Ground truth targets,
            tensor of shape (?, num_boxes, num_classes).
        y_pred: Predicted logits,
            tensor of shape (?, num_boxes, num_classes).
    
    # Returns
        focal_loss: Focal loss, tensor of shape (?, num_boxes).

    # References
        https://arxiv.org/abs/1708.02002
    """
    #y_pred /= K.sum(y_pred, axis=-1, keepdims=True)
    eps = K.epsilon()
    y_pred = K.clip(y_pred, eps, 1. - eps)
    
    pt = tf.where(tf.equal(y_true, 1), y_pred, 1 - y_pred)
    focal_loss = -tf.reduce_sum(alpha * K.pow(1. - pt, gamma) * K.log(pt), axis=-1)
    return focal_loss 

# 需要导入模块: from keras import backend [as 别名]
# 或者: from keras.backend import log [as 别名]
def kl_divergence(y_true, y_pred):
    max_y_pred = K.repeat_elements(K.expand_dims(K.repeat_elements(K.expand_dims(K.max(K.max(y_pred, axis=2), axis=2)), 
                                                                   shape_r_out, axis=-1)), shape_c_out, axis=-1)
    y_pred /= max_y_pred

    sum_y_true = K.repeat_elements(K.expand_dims(K.repeat_elements(K.expand_dims(K.sum(K.sum(y_true, axis=2), axis=2)), 
                                                                   shape_r_out, axis=-1)), shape_c_out, axis=-1)
    sum_y_pred = K.repeat_elements(K.expand_dims(K.repeat_elements(K.expand_dims(K.sum(K.sum(y_pred, axis=2), axis=2)), 
                                                                   shape_r_out, axis=-1)), shape_c_out, axis=-1)
    y_true /= (sum_y_true + K.epsilon())
    y_pred /= (sum_y_pred + K.epsilon())

    return 10 * K.sum(K.sum(y_true * K.log((y_true / (y_pred + K.epsilon())) + K.epsilon()), axis=-1), axis=-1)


# Correlation Coefficient Loss 

# 需要导入模块: from keras import backend [as 别名]
# 或者: from keras.backend import log [as 别名]
def sparse_categorical_crossentropy(gt_ids, pred_one_hot_post_softmax):
    """
    K.sparse_categorical_crossentropyだと結果がNaNになる。。。
    0割り算が発生しているかも。
    https://qiita.com/4Ui_iUrz1/items/35a8089ab0ebc98061c1
    対策として、微少値を用いてlog(0)にならないよう調整した本関数を作成。
    """
    gt_ids = log.tfprint(gt_ids, "cross:gt_ids:")
    pred_one_hot_post_softmax = log.tfprint(pred_one_hot_post_softmax,
                                            "cross:pred_one_hot_post_softmax:")

    gt_one_hot = K.one_hot(gt_ids, K.shape(pred_one_hot_post_softmax)[-1])
    gt_one_hot = log.tfprint(gt_one_hot, "cross:gt_one_hot:")

    epsilon = K.epsilon()  # 1e-07
    loss = -K.sum(
        gt_one_hot * K.log(
            tf.clip_by_value(pred_one_hot_post_softmax, epsilon, 1 - epsilon)),
        axis=-1)
    loss = log.tfprint(loss, "cross:loss:")
    return loss