Python keras.backend.switch() Examples

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for BG/FG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Cross entropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.
    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Cross entropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Cross entropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Crossentropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def time_distributed_nonzero_max_pooling(x):
    """
    Computes maximum along the first (time) dimension.
    It ignores the mask m.

    In:
        x - input; a 3D tensor
        mask_value - value to mask out, if None then no masking; 
            by default 0.0, 
    """

    import theano.tensor as T

    mask_value=0.0
    x = T.switch(T.eq(x, mask_value), -numpy.inf, x)
    masked_max_x = x.max(axis=1)
    # replace infinities with mask_value
    masked_max_x = T.switch(T.eq(masked_max_x, -numpy.inf), 0, masked_max_x)
    return masked_max_x 

def _rpn_loss_regr(y_true, y_pred):
    """
    smooth L1 loss

    y_ture [1][HXWX10][3] (class,regr)
    y_pred [1][HXWX10][2] (reger)
    """

    sigma = 9.0

    cls = y_true[0, :, 0]
    regr = y_true[0, :, 1:3]
    regr_keep = tf.where(K.equal(cls, 1))[:, 0]
    regr_true = tf.gather(regr, regr_keep)
    regr_pred = tf.gather(y_pred[0], regr_keep)
    diff = tf.abs(regr_true - regr_pred)
    less_one = tf.cast(tf.less(diff, 1.0 / sigma), 'float32')
    loss = less_one * 0.5 * diff ** 2 * sigma + tf.abs(1 - less_one) * (diff - 0.5 / sigma)
    loss = K.sum(loss, axis=1)

    return K.switch(tf.size(loss) > 0, K.mean(loss), K.constant(0.0)) 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Crossentropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def clip_norm(g, c, n):
    if c > 0:
        if K.backend() == 'tensorflow':
            import tensorflow as tf
            import copy
            condition = n >= c
            then_expression = tf.scalar_mul(c / n, g)
            else_expression = g

            if hasattr(then_expression, 'get_shape'):
                g_shape = copy.copy(then_expression.get_shape())
            elif hasattr(then_expression, 'dense_shape'):
                g_shape = copy.copy(then_expression.dense_shape)
            if condition.dtype != tf.bool:
                condition = tf.cast(condition, 'bool')
            g = K.tensorflow_backend.control_flow_ops.cond(
                condition, lambda: then_expression, lambda: else_expression)
            if hasattr(then_expression, 'get_shape'):
                g.set_shape(g_shape)
            elif hasattr(then_expression, 'dense_shape'):
                g._dense_shape = g_shape
        else:
            g = K.switch(n >= c, g * c / n, g)
    return g 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Crossentropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Crossentropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def f1_score_keras(y_true, y_pred):
    #convert probas to 0,1
    y_ppred = K.zeros_like(y_true)
    y_pred_ones = K.T.set_subtensor(y_ppred[K.T.arange(y_true.shape[0]), K.argmax(y_pred, axis=-1)], 1)

    #where y_ture=1 and y_pred=1 -> true positive
    y_true_pred = K.sum(y_true*y_pred_ones, axis=0)

    #for each class: how many where classified as said class
    pred_cnt = K.sum(y_pred_ones, axis=0)

    #for each class: how many are true members of said class
    gold_cnt = K.sum(y_true, axis=0)

    #precision for each class
    precision = K.T.switch(K.T.eq(pred_cnt, 0), 0, y_true_pred/pred_cnt)

    #recall for each class
    recall = K.T.switch(K.T.eq(gold_cnt, 0), 0, y_true_pred/gold_cnt)

    #f1 for each class
    f1_class = K.T.switch(K.T.eq(precision + recall, 0), 0, 2*(precision*recall)/(precision+recall))

    #return average f1 score over all classes
    return K.mean(f1_class) 

def f1_score_taskB(y_true, y_pred):
    #convert probas to 0,1
    y_pred_ones = K.zeros_like(y_true)
    y_pred_ones[:, K.argmax(y_pred, axis=-1)] = 1

    #where y_ture=1 and y_pred=1 -> true positive
    y_true_pred = K.sum(y_true*y_pred_ones, axis=0)

    #for each class: how many where classified as said class
    pred_cnt = K.sum(y_pred_ones, axis=0)

    #for each class: how many are true members of said class
    gold_cnt = K.sum(y_true, axis=0)

    #precision for each class
    precision = K.switch(K.equal(pred_cnt, 0), 0, y_true_pred/pred_cnt)

    #recall for each class
    recall = K.switch(K.equal(gold_cnt, 0), 0, y_true_pred/gold_cnt)

    #f1 for each class
    f1_class = K.switch(K.equal(precision + recall, 0), 0, 2*(precision*recall)/(precision+recall))

    #return average f1 score over all classes
    return f1_class 

def f1_score_semeval(y_true, y_pred):
    # convert probas to 0,1
    y_ppred = K.zeros_like(y_true)
    y_pred_ones = K.T.set_subtensor(y_ppred[K.T.arange(y_true.shape[0]), K.argmax(y_pred, axis=-1)], 1)

    # where y_ture=1 and y_pred=1 -> true positive
    y_true_pred = K.sum(y_true * y_pred_ones, axis=0)

    # for each class: how many where classified as said class
    pred_cnt = K.sum(y_pred_ones, axis=0)

    # for each class: how many are true members of said class
    gold_cnt = K.sum(y_true, axis=0)

    # precision for each class
    precision = K.T.switch(K.T.eq(pred_cnt, 0), 0, y_true_pred / pred_cnt)

    # recall for each class
    recall = K.T.switch(K.T.eq(gold_cnt, 0), 0, y_true_pred / gold_cnt)

    # f1 for each class
    f1_class = K.T.switch(K.T.eq(precision + recall, 0), 0, 2 * (precision * recall) / (precision + recall))

    #return average f1 score over all classes
    return (f1_class[0] + f1_class[2])/2.0 

def f1_score_taskB(y_true, y_pred):
    # convert probas to 0,1
    y_pred_ones = K.zeros_like(y_true)
    y_pred_ones[:, K.argmax(y_pred, axis=-1)] = 1

    # where y_ture=1 and y_pred=1 -> true positive
    y_true_pred = K.sum(y_true * y_pred_ones, axis=0)

    # for each class: how many where classified as said class
    pred_cnt = K.sum(y_pred_ones, axis=0)

    # for each class: how many are true members of said class
    gold_cnt = K.sum(y_true, axis=0)

    # precision for each class
    precision = K.switch(K.equal(pred_cnt, 0), 0, y_true_pred / pred_cnt)

    # recall for each class
    recall = K.switch(K.equal(gold_cnt, 0), 0, y_true_pred / gold_cnt)

    # f1 for each class
    f1_class = K.switch(K.equal(precision + recall, 0), 0, 2 * (precision * recall) / (precision + recall))

    # return average f1 score over all classes
    return f1_class 

def time_distributed_masked_max(x, m):
    """
    Computes max along the first (time) dimension.

    In:
        x - input; a 3D tensor
        m - mask
        m_value - value for masking
    """
    # place infinities where mask is off
    m_value = 0.0
    tmp = K.switch(K.equal(m, 0.0), -numpy.inf, 0.0)
    x_with_inf = x + K.expand_dims(tmp)
    x_max = K.max(x_with_inf, axis=1) 
    r = K.switch(K.equal(x_max, -numpy.inf), m_value, x_max)
    return r 


## classes  ##

# Transforms existing layers to masked layers 

def labels_loss(gt, pred):
    """ラベル分類の損失関数

    gt: 正解
        [N, R]
        2軸目はラベルを示すID
    pred: 予測値(softmax済み)
        [N, R, labels].
    """

    # 交差エントロピー誤差
    # バッチ毎の計算ではなく、全体の平均値でOK。
    # 論文に以下の記載がある。
    #    In our current implementation (as in the released code),
    #    the cls term in Eqn.(1) is normalized by the mini-batch size
    #    (i.e., Ncls = 256) and the reg term is normalized by the number of
    #    anchor locations (i.e., Nreg ∼ 2, 400).
    gt = K.cast(gt, 'int32')
    loss = K.switch(tf.size(gt) > 0,
                    sparse_categorical_crossentropy(gt, pred), K.constant(0.0))
    loss = K.mean(loss)
    return loss 

def offsets_loss(gt_offsets, pred_offsets, dump=False):
    """オフセット回帰の損失関数
    positive（gt_fg > 0）データのみ評価対象とする

    Args:
        gt_offsets: 正解オフセット
            [R, 4]
            3軸目は領域提案とアンカーのオフセット（中心、幅、高さ）。
                (tx, ty, th, tw)
        pred_offsets: 予測値
            [R, 4].

    Note:
        この関数の呼び出し元はrpn_offsets_lossとhead_offsets_loss。
        RPNでのRoI予測が外れると全てNegativeなBBoxとなり、結果的にhead_offsets_lossへ渡される正解データのラベルが全てNegativeとなる。
        その場合、head_offsets_lossで得られる損失は0となるが、rpn_offsets_lossで得られる損失は大きくなるはずなので、
        損失全体(rpn_offsets_loss + head_offsets_loss)で評価すれば適切な損失になるはず。
    """
    loss = K.switch(tf.size(gt_offsets) > 0,
                    smooth_l1(gt_offsets, pred_offsets), tf.constant(0.0))
    loss = K.mean(loss)
    return loss 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Cross entropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    '''RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    '''
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Cross entropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Crossentropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def rpn_class_loss_graph(rpn_match, rpn_class_logits):
    """RPN anchor classifier loss.

    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG.
    """
    # Squeeze last dim to simplify
    rpn_match = tf.squeeze(rpn_match, -1)
    # Get anchor classes. Convert the -1/+1 match to 0/1 values.
    anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32)
    # Positive and Negative anchors contribute to the loss,
    # but neutral anchors (match value = 0) don't.
    indices = tf.where(K.not_equal(rpn_match, 0))
    # Pick rows that contribute to the loss and filter out the rest.
    rpn_class_logits = tf.gather_nd(rpn_class_logits, indices)
    anchor_class = tf.gather_nd(anchor_class, indices)
    # Cross entropy loss
    loss = K.sparse_categorical_crossentropy(target=anchor_class,
                                             output=rpn_class_logits,
                                             from_logits=True)
    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss 

def precision_keras(y_true, y_pred):
    #convert probas to 0,1
    y_pred_ones = K.zeros_like(y_true)
    y_pred_ones[:, K.argmax(y_pred, axis=-1)] = 1

    #where y_ture=1 and y_pred=1 -> true positive
    y_true_pred = K.sum(y_true*y_pred_ones, axis=0)

    #for each class: how many where classified as said class
    pred_cnt = K.sum(y_pred_ones, axis=0)

    #precision for each class
    precision = K.switch(K.equal(pred_cnt, 0), 0, y_true_pred/pred_cnt)
    
    #return average f1 score over all classes
    return K.mean(precision) 

def f1_score_taskB(y_true, y_pred):
    #convert probas to 0,1
    y_pred_ones = K.zeros_like(y_true)
    y_pred_ones[:, K.argmax(y_pred, axis=-1)] = 1

    #where y_ture=1 and y_pred=1 -> true positive
    y_true_pred = K.sum(y_true*y_pred_ones, axis=0)

    #for each class: how many where classified as said class
    pred_cnt = K.sum(y_pred_ones, axis=0)

    #for each class: how many are true members of said class
    gold_cnt = K.sum(y_true, axis=0)

    #precision for each class
    precision = K.switch(K.equal(pred_cnt, 0), 0, y_true_pred/pred_cnt)

    #recall for each class
    recall = K.switch(K.equal(gold_cnt, 0), 0, y_true_pred/gold_cnt)

    #f1 for each class
    f1_class = K.switch(K.equal(precision + recall, 0), 0, 2*(precision*recall)/(precision+recall))

    #return average f1 score over all classes
    return f1_class 

def precision_keras(y_true, y_pred):
    # convert probas to 0,1
    y_pred_ones = K.zeros_like(y_true)
    y_pred_ones[:, K.argmax(y_pred, axis=-1)] = 1

    # where y_ture=1 and y_pred=1 -> true positive
    y_true_pred = K.sum(y_true * y_pred_ones, axis=0)

    # for each class: how many where classified as said class
    pred_cnt = K.sum(y_pred_ones, axis=0)

    # precision for each class
    precision = K.switch(K.equal(pred_cnt, 0), 0, y_true_pred / pred_cnt)

    # return average f1 score over all classes
    return K.mean(precision) 

def precision_keras(y_true, y_pred):
    #convert probas to 0,1
    y_pred_ones = K.zeros_like(y_true)
    y_pred_ones[:, K.argmax(y_pred, axis=-1)] = 1

    #where y_ture=1 and y_pred=1 -> true positive
    y_true_pred = K.sum(y_true*y_pred_ones, axis=0)

    #for each class: how many where classified as said class
    pred_cnt = K.sum(y_pred_ones, axis=0)

    #precision for each class
    precision = K.switch(K.equal(pred_cnt, 0), 0, y_true_pred/pred_cnt)
    
    #return average f1 score over all classes
    return K.mean(precision) 

def side_regress_loss(predict_deltas, deltas, indices):
    """
    侧边改善回归目标
    :param predict_deltas: 预测的x周偏移回归目标，(batch_num, anchors_num, 1)
    :param deltas: 真实的回归目标，(batch_num, ctpn_train_anchors, 3+1), 最后一位为tag, tag=0 为padding
    :param indices: 正负样本索引，(batch_num, ctpn_train_anchors, (idx,tag))，
             idx:指定anchor索引位置，最后一位为tag, tag=0 为padding; 1为正样本，-1为负样本
    :return:
    """
    # 去除padding和负样本
    positive_indices = tf.where(tf.equal(indices[:, :, -1], 1))
    deltas = tf.gather_nd(deltas[..., 2:3], positive_indices)  # (n,(dy,dh,dx,tag))  取 dx
    true_positive_indices = tf.gather_nd(indices[..., 0], positive_indices)  # 一维，正anchor索引

    # batch索引
    batch_indices = positive_indices[:, 0]
    # 正样本anchor的2维索引
    train_indices_2d = tf.stack([batch_indices, tf.cast(true_positive_indices, dtype=tf.int64)], axis=1)
    # 正样本anchor预测的回归类型
    predict_deltas = tf.gather_nd(predict_deltas, train_indices_2d, name='ctpn_regress_loss_predict_side_deltas')

    # Smooth-L1 # 非常重要，不然报NAN
    loss = K.switch(tf.size(deltas) > 0,
                    smooth_l1_loss(deltas, predict_deltas),
                    tf.constant(0.0))
    loss = K.mean(loss)
    return loss 

def mrcnn_bbox_loss_graph(target_bbox, target_class_ids, pred_bbox):
    """Loss for Mask R-CNN bounding box refinement.

    target_bbox: [batch, num_rois, (dy, dx, log(dh), log(dw))]
    target_class_ids: [batch, num_rois]. Integer class IDs.
    pred_bbox: [batch, num_rois, num_classes, (dy, dx, log(dh), log(dw))]
    """
    # Reshape to merge batch and roi dimensions for simplicity.
    target_class_ids = K.reshape(target_class_ids, (-1,))
    target_bbox = K.reshape(target_bbox, (-1, 4))
    pred_bbox = K.reshape(pred_bbox, (-1, K.int_shape(pred_bbox)[2], 4))

    # Only positive ROIs contribute to the loss. And only
    # the right class_id of each ROI. Get their indices.
    positive_roi_ix = tf.where(target_class_ids > 0)[:, 0]
    positive_roi_class_ids = tf.cast(
        tf.gather(target_class_ids, positive_roi_ix), tf.int64)
    indices = tf.stack([positive_roi_ix, positive_roi_class_ids], axis=1)

    # Gather the deltas (predicted and true) that contribute to loss
    target_bbox = tf.gather(target_bbox, positive_roi_ix)
    pred_bbox = tf.gather_nd(pred_bbox, indices)

    # Smooth-L1 Loss
    loss = K.switch(tf.size(target_bbox) > 0,
                    smooth_l1_loss(y_true=target_bbox, y_pred=pred_bbox),
                    tf.constant(0.0))
    loss = K.mean(loss)
    return loss 

def get_initial_state(self, inputs):
        if isinstance(inputs, list):
            assert len(inputs) == 2  # inputs == [encoder_outputs, y_true]
            encoder_outputs = inputs[0]
        else:
            encoder_outputs = inputs

        memory_shape = K.shape(encoder_outputs)

        # apply the matrix on the first time step to get the initial s0.
        s0 = activations.tanh(K.dot(encoder_outputs[:, 0], self.W_s))

        y0 = K.zeros((memory_shape[0],), dtype='int64') + self.start_token
        t0 = K.zeros((memory_shape[0],), dtype='int64')

        initial_states = [y0, s0, t0]
        if self.is_monotonic:
            # initial attention has form: [1, 0, 0, ..., 0] for each sample in batch
            alpha0 = K.ones((memory_shape[0], 1))
            alpha0 = K.switch(K.greater(memory_shape[1], 1),
                              lambda: K.concatenate([alpha0, K.zeros((memory_shape[0], memory_shape[1] - 1))], axis=-1),
                              alpha0)
            # like energy, attention is stored in shape (samples, time, 1)
            alpha0 = K.expand_dims(alpha0, -1)
            initial_states.append(alpha0)

        return initial_states 

def metric_func_maker(metric_name="f", beta=1):

        return_recall=False
        return_precision=False

        func_name = metric_name
        if metric_name == "recall":
            return_recall = True
        elif metric_name == "precision":
            return_precision = True
        else:
            func_name = "f_%s" % beta

        def f_beta_score(y, y_pred):
            ''' for convienence '''
            y_pred_binary = K.round(y_pred)
            num_true = K.sum(y)
            num_pred = K.sum(y_pred_binary)
            tp = K.sum(y * y_pred_binary)

            recall = K.switch(num_true>0, tp / num_true, 0.0)
            if return_recall:
                return recall

            precision = K.switch(num_pred>0, tp / num_pred, 0.0)
            if return_precision:
                return precision

            precision_recall_sum = recall + (beta*precision)

            return K.switch(precision_recall_sum>0,
                             (beta+1)*((precision*recall)/(precision_recall_sum)), 0.0)


        f_beta_score.__name__ = func_name
        return f_beta_score 

def rpn_bbox_loss_graph(config, target_bbox, rpn_match, rpn_bbox):
    """Return the RPN bounding box loss graph.

    config: the model config object.
    target_bbox: [batch, max positive anchors, (dy, dx, log(dh), log(dw))].
        Uses 0 padding to fill in unsed bbox deltas.
    rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive,
               -1=negative, 0=neutral anchor.
    rpn_bbox: [batch, anchors, (dy, dx, log(dh), log(dw))]
    """
    # Positive anchors contribute to the loss, but negative and
    # neutral anchors (match value of 0 or -1) don't.
    rpn_match = K.squeeze(rpn_match, -1)
    indices = tf.where(K.equal(rpn_match, 1))

    # Pick bbox deltas that contribute to the loss
    rpn_bbox = tf.gather_nd(rpn_bbox, indices)

    # Trim target bounding box deltas to the same length as rpn_bbox.
    batch_counts = K.sum(K.cast(K.equal(rpn_match, 1), tf.int32), axis=1)
    target_bbox = batch_pack_graph(target_bbox, batch_counts,
                                   config.IMAGES_PER_GPU)

    # TODO: use smooth_l1_loss() rather than reimplementing here
    #       to reduce code duplication
    diff = K.abs(target_bbox - rpn_bbox)
    less_than_one = K.cast(K.less(diff, 1.0), "float32")
    loss = (less_than_one * 0.5 * diff**2) + (1 - less_than_one) * (diff - 0.5)

    loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0))
    return loss