Python keras.backend.dtype() Examples

def yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape):
    '''Get corrected boxes'''
    box_yx = box_xy[..., ::-1]
    box_hw = box_wh[..., ::-1]
    input_shape = K.cast(input_shape, K.dtype(box_yx))
    image_shape = K.cast(image_shape, K.dtype(box_yx))
    new_shape = K.round(image_shape * K.min(input_shape/image_shape))
    offset = (input_shape-new_shape)/2./input_shape
    scale = input_shape/new_shape
    box_yx = (box_yx - offset) * scale
    box_hw *= scale

    box_mins = box_yx - (box_hw / 2.)
    box_maxes = box_yx + (box_hw / 2.)
    boxes =  K.concatenate([
        box_mins[..., 0:1],  # y_min
        box_mins[..., 1:2],  # x_min
        box_maxes[..., 0:1],  # y_max
        box_maxes[..., 1:2]  # x_max
    ])

    # Scale boxes back to original image shape.
    boxes *= K.concatenate([image_shape, image_shape])
    return boxes 

def yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape):
    '''Get corrected boxes'''
    box_yx = box_xy[..., ::-1]
    box_hw = box_wh[..., ::-1]
    input_shape = K.cast(input_shape, K.dtype(box_yx))
    image_shape = K.cast(image_shape, K.dtype(box_yx))
    new_shape = K.round(image_shape * K.min(input_shape/image_shape))
    offset = (input_shape-new_shape)/2./input_shape
    scale = input_shape/new_shape
    box_yx = (box_yx - offset) * scale
    box_hw *= scale

    box_mins = box_yx - (box_hw / 2.)
    box_maxes = box_yx + (box_hw / 2.)
    boxes =  K.concatenate([
        box_mins[..., 0:1],  # y_min
        box_mins[..., 1:2],  # x_min
        box_maxes[..., 0:1],  # y_max
        box_maxes[..., 1:2]  # x_max
    ])

    # Scale boxes back to original image shape.
    boxes *= K.concatenate([image_shape, image_shape])
    return boxes 

def yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape):
    '''Get corrected boxes'''
    box_yx = box_xy[..., ::-1]
    box_hw = box_wh[..., ::-1]
    input_shape = K.cast(input_shape, K.dtype(box_yx))
    image_shape = K.cast(image_shape, K.dtype(box_yx))
    new_shape = K.round(image_shape * K.min(input_shape/image_shape))
    offset = (input_shape-new_shape)/2./input_shape
    scale = input_shape/new_shape
    box_yx = (box_yx - offset) * scale
    box_hw *= scale

    box_mins = box_yx - (box_hw / 2.)
    box_maxes = box_yx + (box_hw / 2.)
    boxes =  K.concatenate([
        box_mins[..., 0:1],  # y_min
        box_mins[..., 1:2],  # x_min
        box_maxes[..., 0:1],  # y_max
        box_maxes[..., 1:2]  # x_max
    ])

    # Scale boxes back to original image shape.
    boxes *= K.concatenate([image_shape, image_shape])
    return boxes 

def yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape):
    '''Get corrected boxes'''
    box_yx = box_xy[..., ::-1]
    box_hw = box_wh[..., ::-1]
    input_shape = K.cast(input_shape, K.dtype(box_yx))
    image_shape = K.cast(image_shape, K.dtype(box_yx))
    new_shape = K.round(image_shape * K.min(input_shape/image_shape))
    offset = (input_shape-new_shape)/2./input_shape
    scale = input_shape/new_shape
    box_yx = (box_yx - offset) * scale
    box_hw *= scale

    box_mins = box_yx - (box_hw / 2.)
    box_maxes = box_yx + (box_hw / 2.)
    boxes =  K.concatenate([
        box_mins[..., 0:1],  # y_min
        box_mins[..., 1:2],  # x_min
        box_maxes[..., 0:1],  # y_max
        box_maxes[..., 1:2]  # x_max
    ])

    # Scale boxes back to original image shape.
    boxes *= K.concatenate([image_shape, image_shape])
    return boxes 

def yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape):
    '''Get corrected boxes'''
    box_yx = box_xy[..., ::-1]
    box_hw = box_wh[..., ::-1]
    input_shape = K.cast(input_shape, K.dtype(box_yx))
    image_shape = K.cast(image_shape, K.dtype(box_yx))
    new_shape = K.round(image_shape * K.min(input_shape/image_shape))
    offset = (input_shape-new_shape)/2./input_shape
    scale = input_shape/new_shape
    box_yx = (box_yx - offset) * scale
    box_hw *= scale

    box_mins = box_yx - (box_hw / 2.)
    box_maxes = box_yx + (box_hw / 2.)
    boxes =  K.concatenate([
        box_mins[..., 0:1],  # y_min
        box_mins[..., 1:2],  # x_min
        box_maxes[..., 0:1],  # y_max
        box_maxes[..., 1:2]  # x_max
    ])

    # Scale boxes back to original image shape.
    boxes *= K.concatenate([image_shape, image_shape])
    return boxes 

def __init__(self, input_dim, output_dim, fixed_weights, embeddings_initializer='uniform', 
                 input_length=None, **kwargs):
        kwargs['dtype'] = 'int32'
        if 'input_shape' not in kwargs:
            if input_length:
                kwargs['input_shape'] = (input_length,)
            else:
                kwargs['input_shape'] = (None,)
        super(Embedding2, self).__init__(**kwargs)
    
        self.input_dim = input_dim
        self.output_dim = output_dim
        self.embeddings_initializer = embeddings_initializer
        self.fixed_weights = fixed_weights
        self.num_trainable = input_dim - len(fixed_weights)
        self.input_length = input_length
        
        w_mean = fixed_weights.mean(axis=0)
        w_std = fixed_weights.std(axis=0)
        self.variable_weights = w_mean + w_std*np.random.randn(self.num_trainable, output_dim) 

def __init__(self, input_dim, output_dim, fixed_weights, embeddings_initializer='uniform', 
                 input_length=None, **kwargs):
        kwargs['dtype'] = 'int32'
        if 'input_shape' not in kwargs:
            if input_length:
                kwargs['input_shape'] = (input_length,)
            else:
                kwargs['input_shape'] = (None,)
        super(Embedding2, self).__init__(**kwargs)
    
        self.input_dim = input_dim
        self.output_dim = output_dim
        self.embeddings_initializer = embeddings_initializer
        self.fixed_weights = fixed_weights
        self.num_trainable = input_dim - len(fixed_weights)
        self.input_length = input_length
        
        w_mean = fixed_weights.mean(axis=0)
        w_std = fixed_weights.std(axis=0)
        self.variable_weights = w_mean + w_std*np.random.randn(self.num_trainable, output_dim) 

def __init__(self, lr=0.001, beta_1=0.9, beta_2=0.999,
                 epsilon=None, decay=0.,
                 terminal_bound=0.1, lower_bound=0., upper_bound=None, **kwargs):
        super(AdaBound, self).__init__(**kwargs)
        with K.name_scope(self.__class__.__name__):
            self.iterations = K.variable(0, dtype='int64', name='iterations')
            self.lr = K.variable(lr, name='lr')
            self.beta_1 = K.variable(beta_1, name='beta_1')
            self.beta_2 = K.variable(beta_2, name='beta_2')
            self.decay = K.variable(decay, name='decay')
            if upper_bound is None:
                upper_bound = terminal_bound * 2.
            self.terminal_bound = K.variable(terminal_bound, name='terminal_bound')
            self.lower_bound = K.variable(lower_bound, name='lower_bound')
            self.upper_bound = K.variable(upper_bound, name='upper_bound')
        if epsilon is None:
            epsilon = K.epsilon()
        self.epsilon = epsilon
        self.initial_decay = decay 

def __init__(self, lr=0.001, beta_1=0.9, beta_2=0.999,
                 epsilon=None, decay=0.,
                 terminal_bound=0.1, lower_bound=0., upper_bound=None, **kwargs):
        super(AdaBound, self).__init__(**kwargs)
        with K.name_scope(self.__class__.__name__):
            self.iterations = K.variable(0, dtype='int64', name='iterations')
            self.lr = K.variable(lr, name='lr')
            self.beta_1 = K.variable(beta_1, name='beta_1')
            self.beta_2 = K.variable(beta_2, name='beta_2')
            self.decay = K.variable(decay, name='decay')
            if upper_bound is None:
                upper_bound = terminal_bound * 2.
            self.terminal_bound = K.variable(terminal_bound, name='terminal_bound')
            self.lower_bound = K.variable(lower_bound, name='lower_bound')
            self.upper_bound = K.variable(upper_bound, name='upper_bound')
        if epsilon is None:
            epsilon = K.epsilon()
        self.epsilon = epsilon
        self.initial_decay = decay 

def probs_to_word_ix(pk, is_first):
	if is_first:
		pk[0] = 0.0
		pk /= np.sum(pk)
	else:
		pk *= pk
		pk /= np.sum(pk)
		#for i in range(3):
		#	max_val = np.amax(pk)
		#	if max_val > 0.5:
		#		break
		#	pk *= pk
		#	pk /= np.sum(pk)

	xk = np.arange(pk.shape[0], dtype=np.int32)
	custm = stats.rv_discrete(name='custm', values=(xk, pk))
	return custm.rvs() 

def pred_text(model, context, max_len=64):
	output = []
	context = np.expand_dims(context, axis=0)
	if MAKE_STATEFUL:
		past_sample = np.zeros((1,), dtype=np.int32)
	else:
		past_sample = np.zeros((SEQ_SIZE,), dtype=np.int32)
	while len(output) < max_len:
		pk = model.predict([context, np.expand_dims(past_sample, axis=0)], batch_size=1)[-1]
		if MAKE_STATEFUL:
			pk = pk[0]
		else:
			past_sample = np.roll(past_sample, 1 if IS_REVERSE else -1)
		new_sample = probs_to_word_ix(pk, len(output) == 0)
		past_sample[0 if IS_REVERSE else -1] = new_sample
		if new_sample == 0:
			break
		output.append(new_sample)

	model.reset_states()
	return output

#Load Keras and Theano 

def yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape):
    '''Get corrected boxes'''
    box_yx = box_xy[..., ::-1]
    box_hw = box_wh[..., ::-1]
    input_shape = K.cast(input_shape, K.dtype(box_yx))
    image_shape = K.cast(image_shape, K.dtype(box_yx))
    new_shape = K.round(image_shape * K.min(input_shape/image_shape))
    offset = (input_shape-new_shape)/2./input_shape
    scale = input_shape/new_shape
    box_yx = (box_yx - offset) * scale
    box_hw *= scale

    box_mins = box_yx - (box_hw / 2.)
    box_maxes = box_yx + (box_hw / 2.)
    boxes =  K.concatenate([
        box_mins[..., 0:1],  # y_min
        box_mins[..., 1:2],  # x_min
        box_maxes[..., 0:1],  # y_max
        box_maxes[..., 1:2]  # x_max
    ])

    # Scale boxes back to original image shape.
    boxes *= K.concatenate([image_shape, image_shape])
    return boxes 

def yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape):
    '''Get corrected boxes'''
    box_yx = box_xy[..., ::-1]
    box_hw = box_wh[..., ::-1]
    input_shape = K.cast(input_shape, K.dtype(box_yx))
    image_shape = K.cast(image_shape, K.dtype(box_yx))
    new_shape = K.round(image_shape * K.min(input_shape/image_shape))
    offset = (input_shape-new_shape)/2./input_shape
    scale = input_shape/new_shape
    box_yx = (box_yx - offset) * scale
    box_hw *= scale

    box_mins = box_yx - (box_hw / 2.)
    box_maxes = box_yx + (box_hw / 2.)
    boxes =  K.concatenate([
        box_mins[..., 0:1],  # y_min
        box_mins[..., 1:2],  # x_min
        box_maxes[..., 0:1],  # y_max
        box_maxes[..., 1:2]  # x_max
    ])

    # Scale boxes back to original image shape.
    boxes *= K.concatenate([image_shape, image_shape])
    return boxes 

def yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape):
    '''Get corrected boxes'''
    box_yx = box_xy[..., ::-1]
    box_hw = box_wh[..., ::-1]
    input_shape = K.cast(input_shape, K.dtype(box_yx))
    image_shape = K.cast(image_shape, K.dtype(box_yx))
    new_shape = K.round(image_shape * K.min(input_shape/image_shape))
    offset = (input_shape-new_shape)/2./input_shape
    scale = input_shape/new_shape
    box_yx = (box_yx - offset) * scale
    box_hw *= scale

    box_mins = box_yx - (box_hw / 2.)
    box_maxes = box_yx + (box_hw / 2.)
    boxes =  K.concatenate([
        box_mins[..., 0:1],  # y_min
        box_mins[..., 1:2],  # x_min
        box_maxes[..., 0:1],  # y_max
        box_maxes[..., 1:2]  # x_max
    ])

    # Scale boxes back to original image shape.
    boxes *= K.concatenate([image_shape, image_shape])
    return boxes 

def yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape):
    '''Get corrected boxes'''
    box_yx = box_xy[..., ::-1]
    box_hw = box_wh[..., ::-1]
    input_shape = K.cast(input_shape, K.dtype(box_yx))
    image_shape = K.cast(image_shape, K.dtype(box_yx))
    new_shape = K.round(image_shape * K.min(input_shape/image_shape))
    offset = (input_shape-new_shape)/2./input_shape
    scale = input_shape/new_shape
    box_yx = (box_yx - offset) * scale
    box_hw *= scale

    box_mins = box_yx - (box_hw / 2.)
    box_maxes = box_yx + (box_hw / 2.)
    boxes =  K.concatenate([
        box_mins[..., 0:1],  # y_min
        box_mins[..., 1:2],  # x_min
        box_maxes[..., 0:1],  # y_max
        box_maxes[..., 1:2]  # x_max
    ])

    # Scale boxes back to original image shape.
    boxes *= K.concatenate([image_shape, image_shape])
    return boxes 

def attention_call(self,
                       inputs,
                       cell_states,
                       attended,
                       attention_states,
                       attended_mask,
                       training=None):
        # there must be two attended sequences (verified in build)
        [attended, u] = attended
        attended_mask = attended_mask[0]
        h_cell_tm1 = cell_states[0]

        # compute attention weights
        w = K.repeat(K.dot(h_cell_tm1, self.W_a) + self.b_UW, K.shape(attended)[1])
        e = K.exp(K.dot(K.tanh(w + u), self.v_a) + self.b_v)

        if attended_mask is not None:
            e = e * K.cast(K.expand_dims(attended_mask, -1), K.dtype(e))

        a = e / K.sum(e, axis=1, keepdims=True)
        c = K.sum(a * attended, axis=1, keepdims=False)

        return c, [c], K.squeeze(a, -1) 

def yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape):
    '''Get corrected boxes'''
    box_yx = box_xy[..., ::-1]
    box_hw = box_wh[..., ::-1]
    input_shape = K.cast(input_shape, K.dtype(box_yx))
    image_shape = K.cast(image_shape, K.dtype(box_yx))
    new_shape = K.round(image_shape * K.min(input_shape/image_shape))
    offset = (input_shape-new_shape)/2./input_shape
    scale = input_shape/new_shape
    box_yx = (box_yx - offset) * scale
    box_hw *= scale

    box_mins = box_yx - (box_hw / 2.)
    box_maxes = box_yx + (box_hw / 2.)
    boxes =  K.concatenate([
        box_mins[..., 0:1],  # y_min
        box_mins[..., 1:2],  # x_min
        box_maxes[..., 0:1],  # y_max
        box_maxes[..., 1:2]  # x_max
    ])

    # Scale boxes back to original image shape.
    boxes *= K.concatenate([image_shape, image_shape])
    return boxes 

def yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape):
    '''Get corrected boxes'''
    box_yx = box_xy[..., ::-1]
    box_hw = box_wh[..., ::-1]
    input_shape = K.cast(input_shape, K.dtype(box_yx))
    image_shape = K.cast(image_shape, K.dtype(box_yx))
    new_shape = K.round(image_shape * K.min(input_shape/image_shape))
    offset = (input_shape-new_shape)/2./input_shape
    scale = input_shape/new_shape
    box_yx = (box_yx - offset) * scale
    box_hw *= scale

    box_mins = box_yx - (box_hw / 2.)
    box_maxes = box_yx + (box_hw / 2.)
    boxes =  K.concatenate([
        box_mins[..., 0:1],  # y_min
        box_mins[..., 1:2],  # x_min
        box_maxes[..., 0:1],  # y_max
        box_maxes[..., 1:2]  # x_max
    ])

    # Scale boxes back to original image shape.
    boxes *= K.concatenate([image_shape, image_shape])
    return boxes 

def __init__(self, lr=0.001, beta_1=0.9, beta_2=0.999,
                 epsilon=None, decay=0., amsgrad=False, accum_iters=1, **kwargs):
        if accum_iters < 1:
            raise ValueError('accum_iters must be >= 1')
        super(AdamAccumulate, self).__init__(**kwargs)
        with K.name_scope(self.__class__.__name__):
            self.iterations = K.variable(0, dtype='int64', name='iterations')
            self.lr = K.variable(lr, name='lr')
            self.beta_1 = K.variable(beta_1, name='beta_1')
            self.beta_2 = K.variable(beta_2, name='beta_2')
            self.decay = K.variable(decay, name='decay')
        if epsilon is None:
            epsilon = K.epsilon()
        self.epsilon = epsilon
        self.initial_decay = decay
        self.amsgrad = amsgrad
        self.accum_iters = K.variable(accum_iters, K.dtype(self.iterations))
        self.accum_iters_float = K.cast(self.accum_iters, K.floatx()) 

def _correct_boxes(
            self, box_xy, box_wh, input_shape, image_shape):
        """Get corrected boxes, which are scaled to original shape."""
        box_yx = box_xy[..., ::-1]
        box_hw = box_wh[..., ::-1]
        input_shape = K.cast(input_shape, K.dtype(box_yx))
        image_shape = K.cast(image_shape, K.dtype(box_yx))
        new_shape = K.round(image_shape * K.min(input_shape / image_shape))
        offset = (input_shape - new_shape) / 2. / input_shape
        scale = input_shape / new_shape
        box_yx = (box_yx - offset) * scale
        box_hw *= scale

        box_mins = box_yx - (box_hw / 2.)
        box_maxes = box_yx + (box_hw / 2.)
        boxes = K.concatenate([
            box_mins[..., 0:1],  # y_min
            box_mins[..., 1:2],  # x_min
            box_maxes[..., 0:1],  # y_max
            box_maxes[..., 1:2]  # x_max
        ])

        # Scale boxes back to original image shape.
        boxes *= K.concatenate([image_shape, image_shape])
        return boxes 

def yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape):

    box_yx = box_xy[..., ::-1]
    box_hw = box_wh[..., ::-1]
    input_shape = K.cast(input_shape, K.dtype(box_yx))
    image_shape = K.cast(image_shape, K.dtype(box_yx))
    new_shape = K.round(image_shape * K.min(input_shape/image_shape))
    offset = (input_shape-new_shape)/2./input_shape
    scale = input_shape/new_shape
    box_yx = (box_yx - offset) * scale
    box_hw *= scale

    box_mins = box_yx - (box_hw / 2.)
    box_maxes = box_yx + (box_hw / 2.)
    boxes =  K.concatenate([
        box_mins[..., 0:1],
        box_mins[..., 1:2],
        box_maxes[..., 0:1],
        box_maxes[..., 1:2]
    ])


    boxes *= K.concatenate([image_shape, image_shape])
    return boxes 

def yolo_head(feats, anchors, num_classes, input_shape, calc_loss=False):
    """Convert final layer features to bounding box parameters."""
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3] # height, width
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
        [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
        [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])

    # Adjust preditions to each spatial grid point and anchor size.
    box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    box_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    if calc_loss == True:
        return grid, feats, box_xy, box_wh
    return box_xy, box_wh, box_confidence, box_class_probs 

def yolo_head(feats, anchors, num_classes, input_shape):
    """
    Convert final layer features to bounding box parameters.
    """
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3]  # height, width
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
                    [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
                    [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])

    # Adjust predictions to each spatial grid point and anchor size.
    box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    box_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    return box_xy, box_wh, box_confidence, box_class_probs 

def yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape):
    """
    Get corrected boxes according to the image's original shapes
    In other words, we scale the predicted bounding boxes to the input's original dimensions
    """
    box_yx = box_xy[..., ::-1]
    box_hw = box_wh[..., ::-1]
    input_shape = K.cast(input_shape, K.dtype(box_yx))
    image_shape = K.cast(image_shape, K.dtype(box_yx))
    new_shape = K.round(image_shape * K.min(input_shape / image_shape))
    offset = (input_shape - new_shape) / 2. / input_shape
    scale = input_shape / new_shape
    box_yx = (box_yx - offset) * scale
    box_hw *= scale

    box_mins = box_yx - (box_hw / 2.)
    box_maxes = box_yx + (box_hw / 2.)
    boxes = K.concatenate([
        box_mins[..., 0:1],  # y_min
        box_mins[..., 1:2],  # x_min
        box_maxes[..., 0:1],  # y_max
        box_maxes[..., 1:2]  # x_max
    ])

    # Scale boxes back to original image shape.
    boxes *= K.concatenate([image_shape, image_shape])
    return boxes 

def __init__(self, lr=0.001, beta_1=0.9, beta_2=0.999,
                 epsilon=None, weight_decay=0.01, decay=0., **kwargs):
        super(LAMBOptimizer, self).__init__(**kwargs)
        with K.name_scope(self.__class__.__name__):
            self.iterations = K.variable(0, dtype='int64', name='iterations')
            self.lr = K.variable(lr, name='lr')
            self.beta_1 = K.variable(beta_1, name='beta_1')
            self.beta_2 = K.variable(beta_2, name='beta_2')
            self.decay = K.variable(decay, name='decay')
        if epsilon is None:
            epsilon = 1e-6
        self.epsilon = epsilon
        self.initial_decay = decay
        self.weight_decay = weight_decay 

def yolo_head(feats, anchors, num_classes, input_shape, calc_loss=False):
    """Convert final layer features to bounding box parameters."""
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3] # height, width
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
        [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
        [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])

    # Adjust preditions to each spatial grid point and anchor size.
    box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    box_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    if calc_loss == True:
        return grid, feats, box_xy, box_wh
    return box_xy, box_wh, box_confidence, box_class_probs 

def yolo_eval(yolo_outputs,
              image_shape,
              max_boxes=10,
              score_threshold=.6,
              iou_threshold=.5):
    """Evaluate YOLO model on given input batch and return filtered boxes."""
    box_xy, box_wh, box_confidence, box_class_probs = yolo_outputs
    boxes = yolo_boxes_to_corners(box_xy, box_wh)
    boxes, scores, classes = yolo_filter_boxes(
        boxes, box_confidence, box_class_probs, threshold=score_threshold)

    # Scale boxes back to original image shape.
    height = image_shape[0]
    width = image_shape[1]
    image_dims = K.stack([height, width, height, width])
    image_dims = K.reshape(image_dims, [1, 4])
    boxes = boxes * image_dims

    # TODO: Something must be done about this ugly hack!
    max_boxes_tensor = K.variable(max_boxes, dtype='int32')
    K.get_session().run(tf.variables_initializer([max_boxes_tensor]))
    nms_index = tf.image.non_max_suppression(
        boxes, scores, max_boxes_tensor, iou_threshold=iou_threshold)
    boxes = K.gather(boxes, nms_index)
    scores = K.gather(scores, nms_index)
    classes = K.gather(classes, nms_index)
    return boxes, scores, classes 

def __init__(self, lr=0.001, beta_1=0.9, beta_2=0.999,
                 epsilon=1e-6, decay=0., weight_decay=0.01,
                 lower_bound=1e-3, upper_bound=10.0, **kwargs):
        super(Lamb, self).__init__(**kwargs)
        with K.name_scope(self.__class__.__name__):
            self.iterations = K.variable(0, dtype='int64', name='iterations')
            self.lr = K.variable(lr, name='lr')
            self.beta_1 = K.variable(beta_1, name='beta_1')
            self.beta_2 = K.variable(beta_2, name='beta_2')
            self.decay = K.variable(decay, name='decay')
            self.weight_decay = K.variable(weight_decay, name='weight_decay')
        self.epsilon = epsilon
        self.initial_decay = decay
        self.lower_bound = lower_bound
        self.upper_bound = upper_bound 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        t = K.cast(self.iterations, K.floatx()) + 1
        lr = self.lr
        if self.initial_decay > 0:
            lr = lr * (1. / (1. + self.decay * t))

        ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        self.weights = [self.iterations] + ms + vs

        for p, g, m, v in zip(params, grads, ms, vs):
            m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
            v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
            self.updates.append(K.update(m, m_t))
            self.updates.append(K.update(v, v_t))

            mhat_t = m_t / (1. - K.pow(self.beta_1, t))
            vhat_t = v_t / (1. - K.pow(self.beta_2, t))

            u_t = mhat_t / K.sqrt(vhat_t + self.epsilon) + self.weight_decay * p
            trust_ratio = K.sqrt(K.sum(K.square(p)) / K.sum(K.square(u_t)))
            trust_ratio = K.minimum(K.maximum(trust_ratio, self.lower_bound), self.upper_bound)

            lr_p = trust_ratio * lr
            new_p = p - lr_p * u_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, new_p))
        return self.updates 

def call(self, inputs):
        if len(inputs[1].shape) == 3:
            boolean_mask = K.any(K.not_equal(inputs[1], self.mask_value),
                                 axis=-1, keepdims=True)
        else:
            boolean_mask = K.expand_dims(K.not_equal(inputs[1], self.mask_value))

        return inputs[0] * K.cast(boolean_mask, K.dtype(inputs[0])) 

def call(self, inputs):
        if len(inputs[1].shape) == 3:
            boolean_mask = K.any(K.not_equal(inputs[1], self.mask_value),
                                 axis=-1, keepdims=True)
        else:
            boolean_mask = K.expand_dims(K.not_equal(inputs[1], self.mask_value))

        return inputs[0] * K.cast(boolean_mask, K.dtype(inputs[0])) 

def yolo_head(feats, anchors, num_classes, input_shape):
    """Convert final layer features to bounding box parameters."""
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3] # height, width
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
        [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
        [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])

    box_xy = K.sigmoid(feats[..., :2])
    box_wh = K.exp(feats[..., 2:4])
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    # Adjust preditions to each spatial grid point and anchor size.
    box_xy = (box_xy + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    box_wh = box_wh * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))

    return box_xy, box_wh, box_confidence, box_class_probs 

def yolo_head(feats, anchors, num_classes, input_shape, calc_loss=False):
    """Convert final layer features to bounding box parameters."""
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3] # height, width
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
        [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
        [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])

    # Adjust preditions to each spatial grid point and anchor size.
    box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    #print(K.dtype(feats))
    box_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    if calc_loss == True:
        return grid, feats, box_xy, box_wh
    return box_xy, box_wh, box_confidence, box_class_probs 

def yolo_head(feats, anchors, num_classes, input_shape, calc_loss=False):
    """Convert final layer features to bounding box parameters."""
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3] # height, width
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
        [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
        [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])

    # Adjust preditions to each spatial grid point and anchor size.
    box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    box_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    if calc_loss == True:
        return grid, feats, box_xy, box_wh
    return box_xy, box_wh, box_confidence, box_class_probs 

def call(self, inputs):
        if K.dtype(inputs) != 'int32':
            inputs = K.cast(inputs, 'int32')
        out = K.gather(self.embeddings, inputs)
        return out 

def call(self, inputs):
        if K.dtype(inputs) != 'int32':
            inputs = K.cast(inputs, 'int32')
        out = K.gather(self.embeddings, inputs)
        return out 

def __init__(self, learning_rate=0.001, beta_1=0.9, beta_2=0.999,
                 amsgrad=False, batch_size=32, total_iterations=0,
                 total_iterations_wd=None, use_cosine_annealing=False,
                 weight_decays=None, lr_multipliers=None, init_verbose=True,
                 eta_min=0, eta_max=1, t_cur=0, **kwargs):
        self.initial_decay = kwargs.pop('decay', 0.0)
        self.epsilon = kwargs.pop('epsilon', K.epsilon())
        learning_rate = kwargs.pop('lr', learning_rate)
        eta_t = kwargs.pop('eta_t', 1.)
        super(AdamW, self).__init__(**kwargs)

        with K.name_scope(self.__class__.__name__):
            self.iterations = K.variable(0, dtype='int64', name='iterations')
            self.learning_rate = K.variable(learning_rate, name='learning_rate')
            self.beta_1 = K.variable(beta_1, name='beta_1')
            self.beta_2 = K.variable(beta_2, name='beta_2')
            self.decay = K.variable(self.initial_decay, name='decay')
            self.batch_size = K.variable(batch_size, dtype='int64',
                                         name='batch_size')
            self.eta_min = K.constant(eta_min, name='eta_min')
            self.eta_max = K.constant(eta_max, name='eta_max')
            self.eta_t = K.variable(eta_t, dtype='float32', name='eta_t')
            self.t_cur = K.variable(t_cur, dtype='int64', name='t_cur')

        self.total_iterations = total_iterations
        self.total_iterations_wd = total_iterations_wd or total_iterations
        self.amsgrad = amsgrad
        self.lr_multipliers = lr_multipliers
        self.weight_decays = weight_decays or {}
        self.init_verbose = init_verbose
        self.use_cosine_annealing = use_cosine_annealing

        self._init_notified = False
        _check_args(total_iterations, use_cosine_annealing, self.weight_decays) 

def __init__(self, learning_rate=0.002, beta_1=0.9, beta_2=0.999,
                 batch_size=32, total_iterations=0,
                 total_iterations_wd=None, use_cosine_annealing=False,
                 weight_decays=None, lr_multipliers=None, init_verbose=True,
                 eta_min=0, eta_max=1, t_cur=0, **kwargs):
        self.schedule_decay = kwargs.pop('schedule_decay', 0.004)
        self.epsilon = kwargs.pop('epsilon', K.epsilon())
        learning_rate = kwargs.pop('lr', learning_rate)
        eta_t = kwargs.pop('eta_t', 1.)
        super(NadamW, self).__init__(**kwargs)

        with K.name_scope(self.__class__.__name__):
            self.iterations = K.variable(0, dtype='int64', name='iterations')
            self.m_schedule = K.variable(1., name='m_schedule')
            self.learning_rate = K.variable(learning_rate, name='learning_rate')
            self.beta_1 = K.variable(beta_1, name='beta_1')
            self.beta_2 = K.variable(beta_2, name='beta_2')
            self.batch_size = K.variable(batch_size, dtype='int64',
                                         name='batch_size')
            self.eta_min = K.constant(eta_min, name='eta_min')
            self.eta_max = K.constant(eta_max, name='eta_max')
            self.eta_t = K.variable(eta_t, dtype='float32', name='eta_t')
            self.t_cur = K.variable(t_cur, dtype='int64', name='t_cur')

        self.total_iterations = total_iterations
        self.total_iterations_wd = total_iterations_wd or total_iterations
        self.lr_multipliers = lr_multipliers
        self.weight_decays = weight_decays or {}
        self.use_cosine_annealing = use_cosine_annealing
        self.init_verbose = init_verbose

        self._init_notified = False
        _check_args(total_iterations, use_cosine_annealing, self.weight_decays) 

def __init__(self, learning_rate=0.01, momentum=0., nesterov=False,
                 batch_size=32, total_iterations=0,
                 total_iterations_wd=None, use_cosine_annealing=False,
                 weight_decays=None, lr_multipliers=None, init_verbose=True,
                 eta_min=0, eta_max=1, t_cur=0, **kwargs):
        self.initial_decay = kwargs.pop('decay', 0.0)
        learning_rate = kwargs.pop('lr', learning_rate)
        eta_t = kwargs.pop('eta_t', 1.)
        super(SGDW, self).__init__(**kwargs)

        with K.name_scope(self.__class__.__name__):
            self.iterations = K.variable(0, dtype='int64', name='iterations')
            self.learning_rate = K.variable(learning_rate, name='learning_rate')
            self.momentum = K.variable(momentum, name='momentum')
            self.decay = K.variable(self.initial_decay, name='decay')
            self.batch_size = K.variable(batch_size, dtype='int64',
                                         name='batch_size')
            self.eta_min = K.constant(eta_min, name='eta_min')
            self.eta_max = K.constant(eta_max, name='eta_max')
            self.eta_t = K.variable(eta_t, dtype='float32', name='eta_t')
            self.t_cur = K.variable(t_cur, dtype='int64', name='t_cur')

        self.total_iterations = total_iterations
        self.total_iterations_wd = total_iterations_wd or total_iterations
        self.nesterov = nesterov
        self.lr_multipliers = lr_multipliers
        self.weight_decays = weight_decays or {}
        self.init_verbose = init_verbose
        self.use_cosine_annealing = use_cosine_annealing

        self._init_notified = False
        _check_args(total_iterations, use_cosine_annealing, self.weight_decays) 

def __init__(self, lr=0.001, beta_1=0.9, beta_2=0.999,
                 amsgrad=False, epsilon=None, decay=0.0,
                 batch_size=32, total_iterations=0,
                 total_iterations_wd=None, use_cosine_annealing=False,
                 weight_decays=None, lr_multipliers=None, init_verbose=True,
                 eta_min=0, eta_max=1, t_cur=0, **kwargs):
        eta_t = kwargs.pop('eta_t', 1.)
        super(AdamW, self).__init__(**kwargs)

        with K.name_scope(self.__class__.__name__):
            self.iterations = K.variable(0, dtype='int64', name='iterations')
            self.lr = K.variable(lr, name='lr')
            self.beta_1 = K.variable(beta_1, name='beta_1')
            self.beta_2 = K.variable(beta_2, name='beta_2')
            self.decay = K.variable(decay, name='decay')
            self.batch_size = K.variable(batch_size, dtype='int64',
                                         name='batch_size')
            self.eta_min = K.constant(eta_min, name='eta_min')
            self.eta_max = K.constant(eta_max, name='eta_max')
            self.eta_t = K.variable(eta_t, dtype='float32', name='eta_t')
            self.t_cur = K.variable(t_cur, dtype='int64', name='t_cur')

        self.initial_decay = decay
        self.epsilon = epsilon or K.epsilon()
        self.total_iterations = total_iterations
        self.total_iterations_wd = total_iterations_wd or total_iterations
        self.amsgrad = amsgrad
        self.lr_multipliers = lr_multipliers
        self.weight_decays = weight_decays or {}
        self.init_verbose = init_verbose
        self.use_cosine_annealing = use_cosine_annealing

        self._init_notified = False
        _check_args(total_iterations, use_cosine_annealing, self.weight_decays) 

def __init__(self, lr=0.002, beta_1=0.9, beta_2=0.999,
                 schedule_decay=0.004, epsilon=None,
                 batch_size=32, total_iterations=0,
                 total_iterations_wd=None, use_cosine_annealing=False,
                 weight_decays=None, lr_multipliers=None, init_verbose=True,
                 eta_min=0, eta_max=1, t_cur=0, **kwargs):
        eta_t = kwargs.pop('eta_t', 1.)
        super(NadamW, self).__init__(**kwargs)

        with K.name_scope(self.__class__.__name__):
            self.iterations = K.variable(0, dtype='int64', name='iterations')
            self.m_schedule = K.variable(1., name='m_schedule')
            self.lr = K.variable(lr, name='lr')
            self.beta_1 = K.variable(beta_1, name='beta_1')
            self.beta_2 = K.variable(beta_2, name='beta_2')
            self.batch_size = K.variable(batch_size, dtype='int64',
                                         name='batch_size')
            self.eta_min = K.constant(eta_min, name='eta_min')
            self.eta_max = K.constant(eta_max, name='eta_max')
            self.eta_t = K.variable(eta_t, dtype='float32', name='eta_t')
            self.t_cur = K.variable(t_cur, dtype='int64', name='t_cur')

        self.epsilon = epsilon or K.epsilon()
        self.schedule_decay = schedule_decay
        self.total_iterations = total_iterations
        self.total_iterations_wd = total_iterations_wd or total_iterations
        self.lr_multipliers = lr_multipliers
        self.weight_decays = weight_decays or {}
        self.use_cosine_annealing = use_cosine_annealing
        self.init_verbose = init_verbose

        self._init_notified = False
        _check_args(total_iterations, use_cosine_annealing, self.weight_decays) 

def __init__(self, lr=0.01, momentum=0., nesterov=False, decay=0.0,
                 batch_size=32, total_iterations=0,
                 total_iterations_wd=None, use_cosine_annealing=False,
                 weight_decays=None, lr_multipliers=None, init_verbose=True,
                 eta_min=0, eta_max=1, t_cur=0, **kwargs):
        eta_t = kwargs.pop('eta_t', 1.)
        super(SGDW, self).__init__(**kwargs)

        with K.name_scope(self.__class__.__name__):
            self.iterations = K.variable(0, dtype='int64', name='iterations')
            self.lr = K.variable(lr, name='lr')
            self.momentum = K.variable(momentum, name='momentum')
            self.decay = K.variable(decay, name='decay')
            self.batch_size = K.variable(batch_size, dtype='int64',
                                         name='batch_size')
            self.eta_min = K.constant(eta_min, name='eta_min')
            self.eta_max = K.constant(eta_max, name='eta_max')
            self.eta_t = K.variable(eta_t, dtype='float32', name='eta_t')
            self.t_cur = K.variable(t_cur, dtype='int64', name='t_cur')

        self.initial_decay = decay
        self.total_iterations = total_iterations
        self.total_iterations_wd = total_iterations_wd or total_iterations
        self.nesterov = nesterov
        self.lr_multipliers = lr_multipliers
        self.weight_decays = weight_decays or {}
        self.init_verbose = init_verbose
        self.use_cosine_annealing = use_cosine_annealing

        self._init_notified = False
        _check_args(total_iterations, use_cosine_annealing, self.weight_decays) 

def new_sparse_categorical_accuracy(y_true, y_pred):
	y_pred_rank = ops.convert_to_tensor(y_pred).get_shape().ndims
	y_true_rank = ops.convert_to_tensor(y_true).get_shape().ndims
	# If the shape of y_true is (num_samples, 1), squeeze to (num_samples,)
	if (y_true_rank is not None) and (y_pred_rank is not None) and (len(K.int_shape(y_true)) == len(K.int_shape(y_pred))):
		y_true = array_ops.squeeze(y_true, [-1])
	y_pred = math_ops.argmax(y_pred, axis=-1)
	# If the predicted output and actual output types don't match, force cast them
	# to match.
	if K.dtype(y_pred) != K.dtype(y_true):
		y_pred = math_ops.cast(y_pred, K.dtype(y_true))
	return math_ops.cast(math_ops.equal(y_true, y_pred), K.floatx())

#Load the model 

def new_sparse_categorical_accuracy(y_true, y_pred):
	y_pred_rank = ops.convert_to_tensor(y_pred).get_shape().ndims
	y_true_rank = ops.convert_to_tensor(y_true).get_shape().ndims
	# If the shape of y_true is (num_samples, 1), squeeze to (num_samples,)
	if (y_true_rank is not None) and (y_pred_rank is not None) and (len(K.int_shape(y_true)) == len(K.int_shape(y_pred))):
		y_true = array_ops.squeeze(y_true, [-1])
	y_pred = math_ops.argmax(y_pred, axis=-1)
	# If the predicted output and actual output types don't match, force cast them
	# to match.
	if K.dtype(y_pred) != K.dtype(y_true):
		y_pred = math_ops.cast(y_pred, K.dtype(y_true))
	return math_ops.cast(math_ops.equal(y_true, y_pred), K.floatx())

#Build the training models 

def yolo_head(feats, anchors, num_classes, n):
    """Convert final layer features to bounding box parameters."""
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    conv_dims = K.shape(feats)[1:3]  # assuming channels last
    # In YOLO the height index is the inner most iteration.
    conv_height_index = K.arange(0, stop=conv_dims[0])
    conv_width_index = K.arange(0, stop=conv_dims[1])
    conv_height_index = K.tile(conv_height_index, [conv_dims[1]])

    conv_width_index = K.tile(K.expand_dims(conv_width_index, 0), [conv_dims[0], 1])
    conv_width_index = K.flatten(K.transpose(conv_width_index))
    conv_index = K.transpose(K.stack([conv_height_index, conv_width_index]))
    conv_index = K.reshape(conv_index, [1, conv_dims[0], conv_dims[1], 1, 2])
    conv_index = K.cast(conv_index, K.dtype(feats))

    feats = K.reshape(feats, [-1, conv_dims[0], conv_dims[1], num_anchors, num_classes + 5])
    conv_dims = K.cast(K.reshape(conv_dims, [1, 1, 1, 1, 2]), K.dtype(feats))

    box_xy = K.sigmoid(feats[..., :2])
    box_wh = K.exp(feats[..., 2:4])
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    # Adjust preditions to each spatial grid point and anchor size.
    # Note: YOLO iterates over height index before width index.
    # TODO: It works with +1, don't know why.
    box_xy = (box_xy + conv_index + 1) / conv_dims
    # TODO: Input layer size
    box_wh = box_wh * anchors_tensor / conv_dims / {0:32, 1:16, 2:8}[n]

    return [box_xy, box_wh, box_confidence, box_class_probs] 

def yolo_non_max_suppression(scores, boxes, classes, max_boxes = 10, iou_threshold = 0.5):
    max_boxes_tensor = K.variable(max_boxes, dtype='int32') # tensor to be used in tf.image.non_max_suppression()
    K.get_session().run(tf.variables_initializer([max_boxes_tensor])) # initialize variable max_boxes_tensor
    
    # Use tf.image.non_max_suppression() to get the list of indices corresponding to boxes you keep
    nms_indices = tf.image.non_max_suppression(boxes, scores, max_boxes, iou_threshold)
    
    # Use K.gather() to select only nms_indices from scores, boxes and classes
    scores = K.gather(scores, nms_indices)
    boxes = K.gather(boxes, nms_indices)
    classes = K.gather(classes, nms_indices)
    
    return scores, boxes, classes 

def yolo_eval(yolo_outputs,
              image_shape,
              max_boxes=10,
              score_threshold=.6,
              iou_threshold=.5):
    """Evaluate YOLO model on given input batch and return filtered boxes."""
    box_confidence, box_xy, box_wh, box_class_probs = yolo_outputs
    boxes = yolo_boxes_to_corners(box_xy, box_wh)
    boxes, scores, classes = yolo_filter_boxes(
        box_confidence, boxes, box_class_probs, threshold=score_threshold)
    
    # Scale boxes back to original image shape.
    height = image_shape[0]
    width = image_shape[1]
    image_dims = K.stack([height, width, height, width])
    image_dims = K.reshape(image_dims, [1, 4])
    boxes = boxes * image_dims

    # TODO: Something must be done about this ugly hack!
    max_boxes_tensor = K.variable(max_boxes, dtype='int32')
    K.get_session().run(tf.variables_initializer([max_boxes_tensor]))
    nms_index = tf.image.non_max_suppression(
        boxes, scores, max_boxes_tensor, iou_threshold=iou_threshold)
    boxes = K.gather(boxes, nms_index)
    scores = K.gather(scores, nms_index)
    classes = K.gather(classes, nms_index)
    
    return boxes, scores, classes 

def yolo_head(feats, anchors, num_classes, input_shape, calc_loss=False):
    """Convert final layer features to bounding box parameters."""
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3] # height, width
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
        [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
        [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])

    # Adjust preditions to each spatial grid point and anchor size.
    box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    box_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    if calc_loss == True:
        return grid, feats, box_xy, box_wh
    return box_xy, box_wh, box_confidence, box_class_probs 

def yolo_head(feats, anchors, num_classes, input_shape, calc_loss=False):
    """Convert final layer features to bounding box parameters."""
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3] # height, width
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
        [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
        [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])

    # Adjust preditions to each spatial grid point and anchor size.
    box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    box_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    if calc_loss == True:
        return grid, feats, box_xy, box_wh
    return box_xy, box_wh, box_confidence, box_class_probs 

def yolo_head(feats, anchors, num_classes, input_shape):
    """Convert final layer features to bounding box parameters."""
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3] # height, width
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
        [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
        [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])

    box_xy = K.sigmoid(feats[..., :2])
    box_wh = K.exp(feats[..., 2:4])
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    # Adjust preditions to each spatial grid point and anchor size.
    box_xy = (box_xy + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    box_wh = box_wh * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))

    return box_xy, box_wh, box_confidence, box_class_probs 

def yolo_head(feats, anchors, num_classes, input_shape):
    """Convert final layer features to bounding box parameters."""
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3] # height, width
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
        [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
        [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])

    box_xy = K.sigmoid(feats[..., :2])
    box_wh = K.exp(feats[..., 2:4])
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    # Adjust preditions to each spatial grid point and anchor size.
    box_xy = (box_xy + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    box_wh = box_wh * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))

    return box_xy, box_wh, box_confidence, box_class_probs 

def yolo_head(feats, anchors, num_classes, input_shape, calc_loss=False):
    """Convert final layer features to bounding box parameters."""
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3] # height, width
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
        [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
        [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])

    # Adjust preditions to each spatial grid point and anchor size.
    box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    box_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    if calc_loss == True:
        return grid, feats, box_xy, box_wh
    return box_xy, box_wh, box_confidence, box_class_probs 

def yolo_head(feats, anchors, num_classes, input_shape, calc_loss=False):
    """Convert final layer features to bounding box parameters."""
    num_anchors = len(anchors)
    # Reshape to batch, height, width, num_anchors, box_params.
    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3] # height, width
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
        [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
        [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])

    # Adjust preditions to each spatial grid point and anchor size.
    box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    box_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    if calc_loss == True:
        return grid, feats, box_xy, box_wh
    return box_xy, box_wh, box_confidence, box_class_probs 

def build(self, input_shape):
        self.embeddings = self.add_weight(
            shape=(self.input_dim, self.output_dim),
            initializer=self.embeddings_initializer,
            name='embeddings',
            regularizer=self.embeddings_regularizer,
            constraint=self.embeddings_constraint,
            dtype=self.dtype)
        
        self.u = self.add_weight(shape=tuple([1, self.embeddings.shape.as_list()[-1]]),
                         initializer=initializers.RandomNormal(0, 1),
                         name='sn',
                         trainable=False)
        
        self.built = True 

def call(self, inputs):
        if K.dtype(inputs) != 'int32':
            inputs = K.cast(inputs, 'int32')
            
        def _l2normalize(v, eps=1e-12):
            return v / (K.sum(v ** 2) ** 0.5 + eps)
        def power_iteration(W, u):
            #Accroding the paper, we only need to do power iteration one time.
            _u = u
            _v = _l2normalize(K.dot(_u, K.transpose(W)))
            _u = _l2normalize(K.dot(_v, W))
            return _u, _v
        W_shape = self.embeddings.shape.as_list()
        #Flatten the Tensor
        W_reshaped = K.reshape(self.embeddings, [-1, W_shape[-1]])
        _u, _v = power_iteration(W_reshaped, self.u)
        #Calculate Sigma
        sigma=K.dot(_v, W_reshaped)
        sigma=K.dot(sigma, K.transpose(_u))
        #normalize it
        W_bar = W_reshaped / sigma
        #reshape weight tensor
        if training in {0, False}:
            W_bar = K.reshape(W_bar, W_shape)
        else:
            with tf.control_dependencies([self.u.assign(_u)]):
                W_bar = K.reshape(W_bar, W_shape)
        self.embeddings = W_bar
            
        out = K.gather(self.embeddings, inputs)
        return out 

def yolo_head(feats, anchors, num_classes, input_shape, calc_loss=False):

    num_anchors = len(anchors)

    anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2])

    grid_shape = K.shape(feats)[1:3]
    grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),
        [1, grid_shape[1], 1, 1])
    grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),
        [grid_shape[0], 1, 1, 1])
    grid = K.concatenate([grid_x, grid_y])
    grid = K.cast(grid, K.dtype(feats))

    feats = K.reshape(
        feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5])


    box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
    box_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))
    box_confidence = K.sigmoid(feats[..., 4:5])
    box_class_probs = K.sigmoid(feats[..., 5:])

    if calc_loss == True:
        return grid, feats, box_xy, box_wh
    return box_xy, box_wh, box_confidence, box_class_probs 

def yolo_eval(yolo_outputs,
              anchors,
              num_classes,
              image_shape,
              max_boxes=20,
              score_threshold=.6,
              iou_threshold=.5):
    """Evaluate YOLO model on given input and return filtered boxes."""

    num_layers = len(yolo_outputs)

    anchor_mask = [[6,7,8], [3,4,5], [0,1,2]] if num_layers==3 else [[3,4,5], [1,2,3]] # default setting
    input_shape = K.shape(yolo_outputs[0])[1:3] * 32

    # print("yolo_outputs",yolo_outputs)
    boxes = []
    box_scores = []
    for l in range(num_layers):
        _boxes, _box_scores = yolo_boxes_and_scores(yolo_outputs[l],
            anchors[anchor_mask[l]], num_classes, input_shape, image_shape)
        boxes.append(_boxes)
        box_scores.append(_box_scores)
    boxes = K.concatenate(boxes, axis=0)
    box_scores = K.concatenate(box_scores, axis=0)

    mask = box_scores >= score_threshold
    max_boxes_tensor = K.constant(max_boxes, dtype='int32')
    boxes_ = []
    scores_ = []
    classes_ = []
    for c in range(num_classes):
        # TODO: use keras backend instead of tf.
        class_boxes = tf.boolean_mask(boxes, mask[:, c])
        class_box_scores = tf.boolean_mask(box_scores[:, c], mask[:, c])
        nms_index = tf.image.non_max_suppression(
            class_boxes, class_box_scores, max_boxes_tensor, iou_threshold=iou_threshold)
        class_boxes = K.gather(class_boxes, nms_index)
        class_box_scores = K.gather(class_box_scores, nms_index)
        classes = K.ones_like(class_box_scores, 'int32') * c
        boxes_.append(class_boxes)
        scores_.append(class_box_scores)
        classes_.append(classes)
    boxes_ = K.concatenate(boxes_, axis=0)
    scores_ = K.concatenate(scores_, axis=0)
    classes_ = K.concatenate(classes_, axis=0)

    return boxes_, scores_, classes_ 

def yolo_predict(yolo_outputs,
                 anchors,
                 num_classes,
                 image_shape,
                 max_boxes=20,
                 score_threshold=.6,
                 iou_threshold=.5):
    """
    Get prediction from YOLO model on given input and return filtered boxes.
    """
    num_layers = len(yolo_outputs)
    anchor_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]]

    input_shape = K.shape(yolo_outputs[0])[1:3] * 32
    boxes = []
    box_scores = []
    for l in range(num_layers):
        # Get the boxes and the confidences for each box
        box_xy, box_wh, box_confidence, box_class_probs = yolo_head(yolo_outputs[l],
                                                                    anchors[anchor_mask[l]],
                                                                    num_classes,
                                                                    input_shape)
        _boxes = yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape)
        _boxes = K.reshape(_boxes, [-1, 4])
        _box_scores = box_confidence * box_class_probs
        _box_scores = K.reshape(_box_scores, [-1, num_classes])

        boxes.append(_boxes)
        box_scores.append(_box_scores)

    boxes = K.concatenate(boxes, axis=0)
    box_scores = K.concatenate(box_scores, axis=0)

    mask = box_scores >= score_threshold
    max_boxes_tensor = K.constant(max_boxes, dtype='int32')
    boxes_ = []
    scores_ = []
    classes_ = []
    for c in range(num_classes):
        class_boxes = tf.boolean_mask(boxes, mask[:, c])
        class_box_scores = tf.boolean_mask(box_scores[:, c], mask[:, c])
        # TensorFlow function for non max suppression of detection candidates
        # This is for filtering overlapping bounding boxes
        nms_index = tf.image.non_max_suppression(
            class_boxes, class_box_scores, max_boxes_tensor, iou_threshold=iou_threshold)
        class_boxes = K.gather(class_boxes, nms_index)
        class_box_scores = K.gather(class_box_scores, nms_index)
        classes = K.ones_like(class_box_scores, 'int32') * c
        boxes_.append(class_boxes)
        scores_.append(class_box_scores)
        classes_.append(classes)
    boxes_ = K.concatenate(boxes_, axis=0)
    scores_ = K.concatenate(scores_, axis=0)
    classes_ = K.concatenate(classes_, axis=0)

    return boxes_, scores_, classes_ 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.lr
        if self.initial_decay > 0:
            lr = lr * (1. / (1. + self.decay * K.cast(self.iterations,
                                                      K.dtype(self.decay))))

        t = K.cast(self.iterations, K.floatx()) + 1

        ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        self.weights = [self.iterations] + ms + vs

        for p, g, m, v in zip(params, grads, ms, vs):
            m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
            v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)

            m_t_hat = m_t / (1. - K.pow(self.beta_1, t))
            v_t_hat = v_t / (1. - K.pow(self.beta_2, t))

            p_dash = m_t_hat / (K.sqrt(v_t_hat + self.epsilon))

            if self.weight_decay > 0.:
                wd = self.weight_decay * p
                p_dash = p_dash + wd

            r1 = K.sqrt(K.sum(K.square(p)))
            r2 = K.sqrt(K.sum(K.square(p_dash)))

            r = tf.where(tf.greater(r1, 0.),
                         tf.where(tf.greater(r2, 0.),
                                  r1 / r2,
                                  1.0),
                         1.0)
            # r = r1 / r2
            eta = r * lr

            p_t = p - eta * p_dash

            self.updates.append(K.update(m, m_t))
            self.updates.append(K.update(v, v_t))
            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, new_p))
        return self.updates 

def apply_gradients(self, grads_and_vars, global_step=None, name=None):
        """See base class."""
        assignments = []
        for (grad, param) in grads_and_vars:
            if grad is None or param is None:
                continue

            param_name = self._get_variable_name(param.name)

            m = tf.get_variable(
                name=param_name + "/adam_m",
                shape=param.shape.as_list(),
                dtype=tf.float32,
                trainable=False,
                initializer=tf.zeros_initializer())
            v = tf.get_variable(
                name=param_name + "/adam_v",
                shape=param.shape.as_list(),
                dtype=tf.float32,
                trainable=False,
                initializer=tf.zeros_initializer())

            # Standard Adam update.
            next_m = (
                    tf.multiply(self.beta_1, m) + tf.multiply(1.0 - self.beta_1, grad))
            next_v = (
                    tf.multiply(self.beta_2, v) + tf.multiply(1.0 - self.beta_2,
                                                              tf.square(grad)))

            update = next_m / (tf.sqrt(next_v) + self.epsilon)

            # Just adding the square of the weights to the loss function is *not*
            # the correct way of using L2 regularization/weight decay with Adam,
            # since that will interact with the m and v parameters in strange ways.
            #
            # Instead we want ot decay the weights in a manner that doesn't interact
            # with the m/v parameters. This is equivalent to adding the square
            # of the weights to the loss with plain (non-momentum) SGD.
            if self._do_use_weight_decay(param_name):
                update += self.weight_decay_rate * param

            update_with_lr = self.learning_rate * update

            next_param = param - update_with_lr

            assignments.extend(
                [param.assign(next_param),
                 m.assign(next_m),
                 v.assign(next_v)])
        return tf.group(*assignments, name=name) 

def yolo_eval(yolo_outputs,
              anchors,
              num_classes,
              image_shape,
              max_boxes=20,
              score_threshold=.6,
              iou_threshold=.5):
    """Evaluate YOLO model on given input and return filtered boxes."""
    num_layers = len(yolo_outputs)
    anchor_mask = [[6,7,8], [3,4,5], [0,1,2]] if num_layers==3 else [[3,4,5], [1,2,3]] # default setting
    input_shape = K.shape(yolo_outputs[0])[1:3] * 32
    boxes = []
    box_scores = []
    for l in range(num_layers):
        _boxes, _box_scores = yolo_boxes_and_scores(yolo_outputs[l],
            anchors[anchor_mask[l]], num_classes, input_shape, image_shape)
        boxes.append(_boxes)
        box_scores.append(_box_scores)
    boxes = K.concatenate(boxes, axis=0)
    box_scores = K.concatenate(box_scores, axis=0)

    mask = box_scores >= score_threshold
    max_boxes_tensor = K.constant(max_boxes, dtype='int32')
    boxes_ = []
    scores_ = []
    classes_ = []
    for c in range(num_classes):
        # TODO: use keras backend instead of tf.
        class_boxes = tf.boolean_mask(boxes, mask[:, c])
        class_box_scores = tf.boolean_mask(box_scores[:, c], mask[:, c])
        nms_index = tf.image.non_max_suppression(
            class_boxes, class_box_scores, max_boxes_tensor, iou_threshold=iou_threshold)
        class_boxes = K.gather(class_boxes, nms_index)
        class_box_scores = K.gather(class_box_scores, nms_index)
        classes = K.ones_like(class_box_scores, 'int32') * c
        boxes_.append(class_boxes)
        scores_.append(class_box_scores)
        classes_.append(classes)
    boxes_ = K.concatenate(boxes_, axis=0)
    scores_ = K.concatenate(scores_, axis=0)
    classes_ = K.concatenate(classes_, axis=0)

    return boxes_, scores_, classes_ 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.lr
        if self.initial_decay:
            lr = lr * (1. / (1. + self.decay * K.cast(
                self.iterations, K.dtype(self.decay)
            )))

        t = K.cast(self.iterations, K.floatx()) + 1.
        beta_1 = self.beta_1
        beta_2 = self.beta_2
        beta_1_t = K.pow(beta_1, t)
        beta_2_t = K.pow(beta_2, t)
        rho_inf = 2. / (1. - beta_2) - 1.
        rho_t = rho_inf - 2. * t * beta_2_t / (1. - beta_2_t)
        r_t = K.sqrt(
            K.relu(rho_t - 4.) * (rho_t - 2.) * rho_inf / (
                (rho_inf - 4.) * (rho_inf - 2.) * rho_t )
        )
        flag = K.cast(rho_t > 4., K.floatx())

        ms = [K.zeros(K.int_shape(p)) for p in params]
        vs = [K.zeros(K.int_shape(p)) for p in params]

        self.weights = [self.iterations] + ms + vs
        for p, g, m, v in zip(params, grads, ms, vs):
            m_t = beta_1 * m + (1. - beta_1) * g
            v_t = beta_2 * v + (1. - beta_2) * K.square(g)

            m_hat_t = m_t / (1. - beta_1_t)
            v_hat_t = K.sqrt(v_t / (1. - beta_2_t))
            new_p = p - lr * (r_t / (v_hat_t + self.epsilon) + flag - 1.)* m_hat_t

            if getattr(p, "constraint", None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, new_p))
            self.updates.append(K.update(m, m_t))
            self.updates.append(K.update(v, v_t))
        return self.updates 

def yolo_eval(yolo_outputs,
              anchors,
              num_classes,
              image_shape,
              max_boxes=20,
              score_threshold=.6,
              iou_threshold=.5):
    """Evaluate YOLO model on given input and return filtered boxes."""
    anchor_mask = [[6,7,8], [3,4,5], [0,1,2]]
    input_shape = K.shape(yolo_outputs[0])[1:3] * 32
    boxes = []
    box_scores = []
    for l in range(3):
        _boxes, _box_scores = yolo_boxes_and_scores(yolo_outputs[l],
            anchors[anchor_mask[l]], num_classes, input_shape, image_shape)
        boxes.append(_boxes)
        box_scores.append(_box_scores)
    boxes = K.concatenate(boxes, axis=0)
    box_scores = K.concatenate(box_scores, axis=0)

    mask = box_scores >= score_threshold
    max_boxes_tensor = K.constant(max_boxes, dtype='int32')
    boxes_ = []
    scores_ = []
    classes_ = []
    for c in range(num_classes):
        # TODO: use keras backend instead of tf.
        class_boxes = tf.boolean_mask(boxes, mask[:, c])
        class_box_scores = tf.boolean_mask(box_scores[:, c], mask[:, c])
        nms_index = tf.image.non_max_suppression(
            class_boxes, class_box_scores, max_boxes_tensor, iou_threshold=iou_threshold)
        class_boxes = K.gather(class_boxes, nms_index)
        class_box_scores = K.gather(class_box_scores, nms_index)
        classes = K.ones_like(class_box_scores, 'int32') * c
        boxes_.append(class_boxes)
        scores_.append(class_box_scores)
        classes_.append(classes)
    boxes_ = K.concatenate(boxes_, axis=0)
    scores_ = K.concatenate(scores_, axis=0)
    classes_ = K.concatenate(classes_, axis=0)

    return boxes_, scores_, classes_ 

def yolo_loss(args, anchors, num_classes, ignore_thresh=.5):
    '''Return yolo_loss tensor

    Parameters
    ----------
    yolo_outputs: list of tensor, the output of yolo_body
    y_true: list of array, the output of preprocess_true_boxes
    anchors: array, shape=(T, 2), wh
    num_classes: integer
    ignore_thresh: float, the iou threshold whether to ignore object confidence loss

    Returns
    -------
    loss: tensor, shape=(1,)

    '''
    yolo_outputs = args[:3]
    y_true = args[3:]
    anchor_mask = [[6,7,8], [3,4,5], [0,1,2]]
    input_shape = K.cast(K.shape(yolo_outputs[0])[1:3] * 32, K.dtype(y_true[0]))
    grid_shapes = [K.cast(K.shape(yolo_outputs[l])[1:3], K.dtype(y_true[0])) for l in range(3)]
    loss = 0
    m = K.shape(yolo_outputs[0])[0]

    for l in range(3):
        object_mask = y_true[l][..., 4:5]
        true_class_probs = y_true[l][..., 5:]

        pred_xy, pred_wh, pred_confidence, pred_class_probs = yolo_head(yolo_outputs[l],
             anchors[anchor_mask[l]], num_classes, input_shape)
        pred_box = K.concatenate([pred_xy, pred_wh])

        # Darknet box loss.
        xy_delta = (y_true[l][..., :2]-pred_xy)*grid_shapes[l][::-1]
        wh_delta = K.log(y_true[l][..., 2:4]) - K.log(pred_wh)
        # Avoid log(0)=-inf.
        wh_delta = K.switch(object_mask, wh_delta, K.zeros_like(wh_delta))
        box_delta = K.concatenate([xy_delta, wh_delta], axis=-1)
        box_delta_scale = 2 - y_true[l][...,2:3]*y_true[l][...,3:4]

        # Find ignore mask, iterate over each of batch.
        ignore_mask = tf.TensorArray(K.dtype(y_true[0]), size=1, dynamic_size=True)
        object_mask_bool = K.cast(object_mask, 'bool')
        def loop_body(b, ignore_mask):
            true_box = tf.boolean_mask(y_true[l][b,...,0:4], object_mask_bool[b,...,0])
            iou = box_iou(pred_box[b], true_box)
            best_iou = K.max(iou, axis=-1)
            ignore_mask = ignore_mask.write(b, K.cast(best_iou<ignore_thresh, K.dtype(true_box)))
            return b+1, ignore_mask
        _, ignore_mask = K.control_flow_ops.while_loop(lambda b,*args: b<m, loop_body, [0, ignore_mask])
        ignore_mask = ignore_mask.stack()
        ignore_mask = K.expand_dims(ignore_mask, -1)

        box_loss = object_mask * K.square(box_delta*box_delta_scale)
        confidence_loss = object_mask * K.square(1-pred_confidence) + \
            (1-object_mask) * K.square(0-pred_confidence) * ignore_mask
        class_loss = object_mask * K.square(true_class_probs-pred_class_probs)
        loss += K.sum(box_loss) + K.sum(confidence_loss) + K.sum(class_loss)
    return loss / K.cast(m, K.dtype(loss)) 

def yolo_eval(yolo_outputs,
              anchors,
              num_classes,
              image_shape,
              max_boxes=20,
              score_threshold=.6,
              iou_threshold=.5):
    """Evaluate YOLO model on given input and return filtered boxes."""
    num_layers = len(yolo_outputs)
    anchor_mask = [[6,7,8], [3,4,5], [0,1,2]] if num_layers==3 else [[3,4,5], [1,2,3]] # default setting
    input_shape = K.shape(yolo_outputs[0])[1:3] * 32
    boxes = []
    box_scores = []
    for l in range(num_layers):
        _boxes, _box_scores = yolo_boxes_and_scores(yolo_outputs[l],
            anchors[anchor_mask[l]], num_classes, input_shape, image_shape)
        boxes.append(_boxes)
        box_scores.append(_box_scores)
    boxes = K.concatenate(boxes, axis=0)
    box_scores = K.concatenate(box_scores, axis=0)

    mask = box_scores >= score_threshold
    max_boxes_tensor = K.constant(max_boxes, dtype='int32')
    boxes_ = []
    scores_ = []
    classes_ = []
    for c in range(num_classes):
        # TODO: use keras backend instead of tf.
        class_boxes = tf.boolean_mask(boxes, mask[:, c])
        class_box_scores = tf.boolean_mask(box_scores[:, c], mask[:, c])
        nms_index = tf.image.non_max_suppression(
            class_boxes, class_box_scores, max_boxes_tensor, iou_threshold=iou_threshold)
        class_boxes = K.gather(class_boxes, nms_index)
        class_box_scores = K.gather(class_box_scores, nms_index)
        classes = K.ones_like(class_box_scores, 'int32') * c
        boxes_.append(class_boxes)
        scores_.append(class_box_scores)
        classes_.append(classes)
    boxes_ = K.concatenate(boxes_, axis=0)
    scores_ = K.concatenate(scores_, axis=0)
    classes_ = K.concatenate(classes_, axis=0)

    return boxes_, scores_, classes_ 

def yolo_eval(yolo_outputs,
              anchors,
              num_classes,
              image_shape,
              max_boxes=20,
              score_threshold=.6,
              iou_threshold=.5):
    """Evaluate YOLO model on given input and return filtered boxes."""
    num_layers = len(yolo_outputs)
    anchor_mask = [[6,7,8], [3,4,5], [0,1,2]] if num_layers==3 else [[3,4,5], [1,2,3]] # default setting
    input_shape = K.shape(yolo_outputs[0])[1:3] * 32
    boxes = []
    box_scores = []
    for l in range(num_layers):
        _boxes, _box_scores = yolo_boxes_and_scores(yolo_outputs[l],
            anchors[anchor_mask[l]], num_classes, input_shape, image_shape)
        boxes.append(_boxes)
        box_scores.append(_box_scores)
    boxes = K.concatenate(boxes, axis=0)
    box_scores = K.concatenate(box_scores, axis=0)

    mask = box_scores >= score_threshold
    max_boxes_tensor = K.constant(max_boxes, dtype='int32')
    boxes_ = []
    scores_ = []
    classes_ = []
    for c in range(num_classes):
        # TODO: use keras backend instead of tf.
        class_boxes = tf.boolean_mask(boxes, mask[:, c])
        class_box_scores = tf.boolean_mask(box_scores[:, c], mask[:, c])
        nms_index = tf.image.non_max_suppression(
            class_boxes, class_box_scores, max_boxes_tensor, iou_threshold=iou_threshold)
        class_boxes = K.gather(class_boxes, nms_index)
        class_box_scores = K.gather(class_box_scores, nms_index)
        classes = K.ones_like(class_box_scores, 'int32') * c
        boxes_.append(class_boxes)
        scores_.append(class_box_scores)
        classes_.append(classes)
    boxes_ = K.concatenate(boxes_, axis=0)
    scores_ = K.concatenate(scores_, axis=0)
    classes_ = K.concatenate(classes_, axis=0)

    return boxes_, scores_, classes_ 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]
        self.updates.append(K.update_add(self.t_cur, 1))

        lr = self.learning_rate
        if self.initial_decay > 0:
            lr = lr * (1. / (1. + self.decay * K.cast(self.iterations,
                                                      K.dtype(self.decay))))
        # momentum
        shapes = [K.int_shape(p) for p in params]
        moments = [K.zeros(shape, name='moment_' + str(i))
                   for (i, shape) in enumerate(shapes)]
        self.weights = [self.iterations] + moments

        total_iterations = self.total_iterations
        # Cosine annealing
        if self.use_cosine_annealing and total_iterations != 0:
            self.eta_t = _compute_eta_t(self)
        self.lr_t = lr * self.eta_t  # for external tracking

        for p, g, m in zip(params, grads, moments):
            # Learning rate multipliers
            lr_t = self.learning_rate
            if self.lr_multipliers is not None:
                lr_t = _apply_lr_multiplier(self, lr_t, p)

            v = self.momentum * m - self.eta_t * lr_t * g  # velocity
            self.updates.append(K.update(m, v))

            if self.nesterov:
                p_t = p + self.momentum * v - self.eta_t * lr_t * g
            else:
                p_t = p + v

            # Weight decays
            if p.name in self.weight_decays.keys() and total_iterations != 0:
                p_t = _apply_weight_decays(self, p, p_t)
            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, new_p))

        self._init_notified = True
        return self.updates 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]
        self.updates.append(K.update_add(self.t_cur, 1))

        lr = self.lr
        if self.initial_decay > 0:
            lr = lr * (1. / (1. + self.decay * K.cast(self.iterations,
                                                      K.dtype(self.decay))))
        # momentum
        shapes = [K.int_shape(p) for p in params]
        moments = [K.zeros(shape) for shape in shapes]
        self.weights = [self.iterations] + moments

        total_iterations = self.total_iterations
        # Cosine annealing
        if self.use_cosine_annealing and total_iterations != 0:
            self.eta_t = _compute_eta_t(self)
        self.lr_t = lr * self.eta_t  # for external tracking

        for p, g, m in zip(params, grads, moments):
            # Learning rate multipliers
            lr_t = self.lr
            if self.lr_multipliers is not None:
                lr_t = _apply_lr_multiplier(self, lr_t, p)

            v = self.momentum * m - self.eta_t * lr_t * g  # velocity
            self.updates.append(K.update(m, v))

            if self.nesterov:
                p_t = p + self.momentum * v - self.eta_t * lr_t * g
            else:
                p_t = p + v

            # Weight decays
            if p.name in self.weight_decays.keys() and total_iterations != 0:
                p_t = _apply_weight_decays(self, p, p_t)
            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, new_p))

        self._init_notified = True
        return self.updates 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.lr
        if self.initial_decay > 0:
            lr = lr * (1. / (1. + self.decay * K.cast(self.iterations,
                                                      K.dtype(self.decay))))

        t = K.cast(self.iterations, K.floatx()) + 1
        lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
                     (1. - K.pow(self.beta_1, t)))

        ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        vhats = [K.zeros(1) for _ in params]
        self.weights = [self.iterations] + ms + vs + vhats

        for p, g, m, v, vhat in zip(params, grads, ms, vs, vhats):
            m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
            v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)

            eta_l_t = self.terminal_bound - \
                (self.terminal_bound - self.lower_bound) / \
                ((1. - self.beta_2) * t + 1)
            eta_u_t = self.terminal_bound + \
                (self.upper_bound - self.terminal_bound) / \
                ((1. - self.beta_2) * t)

            clipped_lr_t = K.minimum(
                K.maximum(lr_t / (K.sqrt(v_t) + self.epsilon), eta_l_t), eta_u_t)
            p_t = p - clipped_lr_t * m_t

            self.updates.append(K.update(m, m_t))
            self.updates.append(K.update(v, v_t))
            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, new_p))
        return self.updates 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.lr
        if self.initial_decay > 0:
            lr = lr * (1. / (1. + self.decay * K.cast(self.iterations,
                                                      K.dtype(self.decay))))

        t = K.cast(self.iterations, K.floatx()) + 1
        lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
                     (1. - K.pow(self.beta_1, t)))

        ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        vhats = [K.zeros(1) for _ in params]
        self.weights = [self.iterations] + ms + vs + vhats

        for p, g, m, v, vhat in zip(params, grads, ms, vs, vhats):
            m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
            v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)

            eta_l_t = self.terminal_bound - \
                (self.terminal_bound - self.lower_bound) / \
                ((1. - self.beta_2) * t + 1)
            eta_u_t = self.terminal_bound + \
                (self.upper_bound - self.terminal_bound) / \
                ((1. - self.beta_2) * t)

            clipped_lr_t = K.minimum(
                K.maximum(lr_t / (K.sqrt(v_t) + self.epsilon), eta_l_t), eta_u_t)
            p_t = p - clipped_lr_t * m_t

            self.updates.append(K.update(m, m_t))
            self.updates.append(K.update(v, v_t))
            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, new_p))
        return self.updates 

def yolo_eval(yolo_outputs,
              anchors,
              num_classes,
              image_shape,
              max_boxes=20,
              score_threshold=.6,
              iou_threshold=.5):
    """Evaluate YOLO model on given input and return filtered boxes."""
    num_layers = len(yolo_outputs)
    anchor_mask = [[6,7,8], [3,4,5], [0,1,2]] if num_layers==3 else [[3,4,5], [1,2,3]] # default setting
    input_shape = K.shape(yolo_outputs[0])[1:3] * 32
    boxes = []
    box_scores = []
    for l in range(num_layers):
        _boxes, _box_scores = yolo_boxes_and_scores(yolo_outputs[l],
            anchors[anchor_mask[l]], num_classes, input_shape, image_shape)
        boxes.append(_boxes)
        box_scores.append(_box_scores)
    boxes = K.concatenate(boxes, axis=0)
    box_scores = K.concatenate(box_scores, axis=0)

    mask = box_scores >= score_threshold
    max_boxes_tensor = K.constant(max_boxes, dtype='int32')
    boxes_ = []
    scores_ = []
    classes_ = []
    for c in range(num_classes):
        # TODO: use keras backend instead of tf.
        class_boxes = tf.boolean_mask(boxes, mask[:, c])
        class_box_scores = tf.boolean_mask(box_scores[:, c], mask[:, c])
        nms_index = tf.image.non_max_suppression(
            class_boxes, class_box_scores, max_boxes_tensor, iou_threshold=iou_threshold)
        class_boxes = K.gather(class_boxes, nms_index)
        class_box_scores = K.gather(class_box_scores, nms_index)
        classes = K.ones_like(class_box_scores, 'int32') * c
        boxes_.append(class_boxes)
        scores_.append(class_box_scores)
        classes_.append(classes)
    boxes_ = K.concatenate(boxes_, axis=0)
    scores_ = K.concatenate(scores_, axis=0)
    classes_ = K.concatenate(classes_, axis=0)

    return boxes_, scores_, classes_ 

def yolo_eval(yolo_outputs,
              anchors,
              num_classes,
              image_shape,
              max_boxes=20,
              score_threshold=.6,
              iou_threshold=.5):
    """Evaluate YOLO model on given input and return filtered boxes."""
    anchor_mask = [[6,7,8], [3,4,5], [0,1,2]]
    input_shape = K.shape(yolo_outputs[0])[1:3] * 32
    boxes = []
    box_scores = []
    for l in range(3):
        _boxes, _box_scores = yolo_boxes_and_scores(yolo_outputs[l],
            anchors[anchor_mask[l]], num_classes, input_shape, image_shape)
        boxes.append(_boxes)
        box_scores.append(_box_scores)
    boxes = K.concatenate(boxes, axis=0)
    box_scores = K.concatenate(box_scores, axis=0)

    mask = box_scores >= score_threshold
    max_boxes_tensor = K.constant(max_boxes, dtype='int32')
    boxes_ = []
    scores_ = []
    classes_ = []
    for c in range(num_classes):
        # TODO: use keras backend instead of tf.
        class_boxes = tf.boolean_mask(boxes, mask[:, c])
        class_box_scores = tf.boolean_mask(box_scores[:, c], mask[:, c])
        nms_index = tf.image.non_max_suppression(
            class_boxes, class_box_scores, max_boxes_tensor, iou_threshold=iou_threshold)
        class_boxes = K.gather(class_boxes, nms_index)
        class_box_scores = K.gather(class_box_scores, nms_index)
        classes = K.ones_like(class_box_scores, 'int32') * c
        boxes_.append(class_boxes)
        scores_.append(class_box_scores)
        classes_.append(classes)
    boxes_ = K.concatenate(boxes_, axis=0)
    scores_ = K.concatenate(scores_, axis=0)
    classes_ = K.concatenate(classes_, axis=0)

    return boxes_, scores_, classes_ 

def yolo_loss(args, anchors, num_classes, ignore_thresh=.5):
    '''Return yolo_loss tensor

    Parameters
    ----------
    yolo_outputs: list of tensor, the output of yolo_body
    y_true: list of array, the output of preprocess_true_boxes
    anchors: array, shape=(T, 2), wh
    num_classes: integer
    ignore_thresh: float, the iou threshold whether to ignore object confidence loss

    Returns
    -------
    loss: tensor, shape=(1,)

    '''
    yolo_outputs = args[:3]
    y_true = args[3:]
    anchor_mask = [[6,7,8], [3,4,5], [0,1,2]]
    input_shape = K.cast(K.shape(yolo_outputs[0])[1:3] * 32, K.dtype(y_true[0]))
    grid_shapes = [K.cast(K.shape(yolo_outputs[l])[1:3], K.dtype(y_true[0])) for l in range(3)]
    loss = 0
    m = K.shape(yolo_outputs[0])[0]

    for l in range(3):
        object_mask = y_true[l][..., 4:5]
        true_class_probs = y_true[l][..., 5:]

        pred_xy, pred_wh, pred_confidence, pred_class_probs = yolo_head(yolo_outputs[l],
             anchors[anchor_mask[l]], num_classes, input_shape)
        pred_box = K.concatenate([pred_xy, pred_wh])

        # Darknet box loss.
        xy_delta = (y_true[l][..., :2]-pred_xy)*grid_shapes[l][::-1]
        wh_delta = K.log(y_true[l][..., 2:4]) - K.log(pred_wh)
        # Avoid log(0)=-inf.
        wh_delta = K.switch(object_mask, wh_delta, K.zeros_like(wh_delta))
        box_delta = K.concatenate([xy_delta, wh_delta], axis=-1)
        box_delta_scale = 2 - y_true[l][...,2:3]*y_true[l][...,3:4]

        # Find ignore mask, iterate over each of batch.
        ignore_mask = tf.TensorArray(K.dtype(y_true[0]), size=1, dynamic_size=True)
        object_mask_bool = K.cast(object_mask, 'bool')
        def loop_body(b, ignore_mask):
            true_box = tf.boolean_mask(y_true[l][b,...,0:4], object_mask_bool[b,...,0])
            iou = box_iou(pred_box[b], true_box)
            best_iou = K.max(iou, axis=-1)
            ignore_mask = ignore_mask.write(b, K.cast(best_iou<ignore_thresh, K.dtype(true_box)))
            return b+1, ignore_mask
        _, ignore_mask = K.control_flow_ops.while_loop(lambda b,*args: b<m, loop_body, [0, ignore_mask])
        ignore_mask = ignore_mask.stack()
        ignore_mask = K.expand_dims(ignore_mask, -1)

        box_loss = object_mask * K.square(box_delta*box_delta_scale)
        confidence_loss = object_mask * K.square(1-pred_confidence) + \
            (1-object_mask) * K.square(0-pred_confidence) * ignore_mask
        class_loss = object_mask * K.square(true_class_probs-pred_class_probs)
        loss += K.sum(box_loss) + K.sum(confidence_loss) + K.sum(class_loss)
    return loss / K.cast(m, K.dtype(loss)) 

def yolo_eval(yolo_outputs,
              anchors,
              num_classes,
              image_shape,
              max_boxes=20,
              score_threshold=.6,
              iou_threshold=.5):
    """Evaluate YOLO model on given input and return filtered boxes."""
    anchor_mask = [[6,7,8], [3,4,5], [0,1,2]]
    input_shape = K.shape(yolo_outputs[0])[1:3] * 32
    boxes = []
    box_scores = []
    for l in range(3):
        _boxes, _box_scores = yolo_boxes_and_scores(yolo_outputs[l],
            anchors[anchor_mask[l]], num_classes, input_shape, image_shape)
        boxes.append(_boxes)
        box_scores.append(_box_scores)
    boxes = K.concatenate(boxes, axis=0)
    box_scores = K.concatenate(box_scores, axis=0)

    mask = box_scores >= score_threshold
    max_boxes_tensor = K.constant(max_boxes, dtype='int32')
    boxes_ = []
    scores_ = []
    classes_ = []
    for c in range(num_classes):
        # TODO: use keras backend instead of tf.
        class_boxes = tf.boolean_mask(boxes, mask[:, c])
        class_box_scores = tf.boolean_mask(box_scores[:, c], mask[:, c])
        nms_index = tf.image.non_max_suppression(
            class_boxes, class_box_scores, max_boxes_tensor, iou_threshold=iou_threshold)
        class_boxes = K.gather(class_boxes, nms_index)
        class_box_scores = K.gather(class_box_scores, nms_index)
        classes = K.ones_like(class_box_scores, 'int32') * c
        boxes_.append(class_boxes)
        scores_.append(class_box_scores)
        classes_.append(classes)
    boxes_ = K.concatenate(boxes_, axis=0)
    scores_ = K.concatenate(scores_, axis=0)
    classes_ = K.concatenate(classes_, axis=0)

    return boxes_, scores_, classes_ 

def yolo_loss(args, anchors, num_classes, ignore_thresh=.5):
    '''Return yolo_loss tensor

    Parameters
    ----------
    yolo_outputs: list of tensor, the output of yolo_body
    y_true: list of array, the output of preprocess_true_boxes
    anchors: array, shape=(T, 2), wh
    num_classes: integer
    ignore_thresh: float, the iou threshold whether to ignore object confidence loss

    Returns
    -------
    loss: tensor, shape=(1,)

    '''
    yolo_outputs = args[:3]
    y_true = args[3:]
    anchor_mask = [[6,7,8], [3,4,5], [0,1,2]]
    input_shape = K.cast(K.shape(yolo_outputs[0])[1:3] * 32, K.dtype(y_true[0]))
    grid_shapes = [K.cast(K.shape(yolo_outputs[l])[1:3], K.dtype(y_true[0])) for l in range(3)]
    loss = 0
    m = K.shape(yolo_outputs[0])[0]

    for l in range(3):
        object_mask = y_true[l][..., 4:5]
        true_class_probs = y_true[l][..., 5:]

        pred_xy, pred_wh, pred_confidence, pred_class_probs = yolo_head(yolo_outputs[l],
             anchors[anchor_mask[l]], num_classes, input_shape)
        pred_box = K.concatenate([pred_xy, pred_wh])

        # Darknet box loss.
        xy_delta = (y_true[l][..., :2]-pred_xy)*grid_shapes[l][::-1]
        wh_delta = K.log(y_true[l][..., 2:4]) - K.log(pred_wh)
        # Avoid log(0)=-inf.
        wh_delta = K.switch(object_mask, wh_delta, K.zeros_like(wh_delta))
        box_delta = K.concatenate([xy_delta, wh_delta], axis=-1)
        box_delta_scale = 2 - y_true[l][...,2:3]*y_true[l][...,3:4]

        # Find ignore mask, iterate over each of batch.
        ignore_mask = tf.TensorArray(K.dtype(y_true[0]), size=1, dynamic_size=True)
        object_mask_bool = K.cast(object_mask, 'bool')
        def loop_body(b, ignore_mask):
            true_box = tf.boolean_mask(y_true[l][b,...,0:4], object_mask_bool[b,...,0])
            iou = box_iou(pred_box[b], true_box)
            best_iou = K.max(iou, axis=-1)
            ignore_mask = ignore_mask.write(b, K.cast(best_iou<ignore_thresh, K.dtype(true_box)))
            return b+1, ignore_mask
        _, ignore_mask = K.control_flow_ops.while_loop(lambda b,*args: b<m, loop_body, [0, ignore_mask])
        ignore_mask = ignore_mask.stack()
        ignore_mask = K.expand_dims(ignore_mask, -1)

        box_loss = object_mask * K.square(box_delta*box_delta_scale)
        confidence_loss = object_mask * K.square(1-pred_confidence) + \
            (1-object_mask) * K.square(0-pred_confidence) * ignore_mask
        class_loss = object_mask * K.square(true_class_probs-pred_class_probs)
        loss += K.sum(box_loss) + K.sum(confidence_loss) + K.sum(class_loss)
    return loss / K.cast(m, K.dtype(loss)) 

def yolo_eval(yolo_outputs,
              anchors,
              num_classes,
              image_shape,
              max_boxes=20,
              score_threshold=.6,
              iou_threshold=.5):
    """Evaluate YOLO model on given input and return filtered boxes."""
    num_layers = len(yolo_outputs)
    anchor_mask = [[6,7,8], [3,4,5], [0,1,2]] if num_layers==3 else [[3,4,5], [1,2,3]] # default setting
    input_shape = K.shape(yolo_outputs[0])[1:3] * 32
    boxes = []
    box_scores = []
    for l in range(num_layers):
        _boxes, _box_scores = yolo_boxes_and_scores(yolo_outputs[l],
            anchors[anchor_mask[l]], num_classes, input_shape, image_shape)
        boxes.append(_boxes)
        box_scores.append(_box_scores)
    boxes = K.concatenate(boxes, axis=0)
    box_scores = K.concatenate(box_scores, axis=0)

    mask = box_scores >= score_threshold
    max_boxes_tensor = K.constant(max_boxes, dtype='int32')
    boxes_ = []
    scores_ = []
    classes_ = []
    for c in range(num_classes):
        # TODO: use keras backend instead of tf.
        class_boxes = tf.boolean_mask(boxes, mask[:, c])
        class_box_scores = tf.boolean_mask(box_scores[:, c], mask[:, c])
        nms_index = tf.image.non_max_suppression(
            class_boxes, class_box_scores, max_boxes_tensor, iou_threshold=iou_threshold)
        class_boxes = K.gather(class_boxes, nms_index)
        class_box_scores = K.gather(class_box_scores, nms_index)
        classes = K.ones_like(class_box_scores, 'int32') * c
        boxes_.append(class_boxes)
        scores_.append(class_box_scores)
        classes_.append(classes)
    boxes_ = K.concatenate(boxes_, axis=0)
    scores_ = K.concatenate(scores_, axis=0)
    classes_ = K.concatenate(classes_, axis=0)

    return boxes_, scores_, classes_ 

def yolo_eval(yolo_outputs,
              anchors,
              num_classes,
              image_shape,
              max_boxes=20,
              score_threshold=.6,
              iou_threshold=.5):
    """Evaluate YOLO model on given input and return filtered boxes."""
    num_layers = len(yolo_outputs)
    anchor_mask = [[6,7,8], [3,4,5], [0,1,2]] if num_layers==3 else [[3,4,5], [1,2,3]] # default setting
    input_shape = K.shape(yolo_outputs[0])[1:3] * 32
    boxes = []
    box_scores = []
    for l in range(num_layers):
        _boxes, _box_scores = yolo_boxes_and_scores(yolo_outputs[l],
            anchors[anchor_mask[l]], num_classes, input_shape, image_shape)
        boxes.append(_boxes)
        box_scores.append(_box_scores)
    boxes = K.concatenate(boxes, axis=0)
    box_scores = K.concatenate(box_scores, axis=0)

    mask = box_scores >= score_threshold
    max_boxes_tensor = K.constant(max_boxes, dtype='int32')
    boxes_ = []
    scores_ = []
    classes_ = []
    for c in range(num_classes):
        # TODO: use keras backend instead of tf.
        class_boxes = tf.boolean_mask(boxes, mask[:, c])
        class_box_scores = tf.boolean_mask(box_scores[:, c], mask[:, c])
        nms_index = tf.image.non_max_suppression(
            class_boxes, class_box_scores, max_boxes_tensor, iou_threshold=iou_threshold)
        class_boxes = K.gather(class_boxes, nms_index)
        class_box_scores = K.gather(class_box_scores, nms_index)
        classes = K.ones_like(class_box_scores, 'int32') * c
        boxes_.append(class_boxes)
        scores_.append(class_box_scores)
        classes_.append(classes)
    boxes_ = K.concatenate(boxes_, axis=0)
    scores_ = K.concatenate(scores_, axis=0)
    classes_ = K.concatenate(classes_, axis=0)

    return boxes_, scores_, classes_ 

def get_updates(self, loss, params):
        grads = self.get_gradients(loss, params)
        self.updates = [K.update_add(self.iterations, 1)]

        lr = self.lr

        completed_updates = K.cast(K.tf.floor(self.iterations / self.accum_iters), K.floatx())

        if self.initial_decay > 0:
            lr = lr * (1. / (1. + self.decay * completed_updates))

        t = completed_updates + 1

        lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) / (1. - K.pow(self.beta_1, t)))

        # self.iterations incremented after processing a batch
        # batch:              1 2 3 4 5 6 7 8 9
        # self.iterations:    0 1 2 3 4 5 6 7 8
        # update_switch = 1:        x       x    (if accum_iters=4)  
        update_switch = K.equal((self.iterations + 1) % self.accum_iters, 0)
        update_switch = K.cast(update_switch, K.floatx())

        ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        gs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]

        if self.amsgrad:
            vhats = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
        else:
            vhats = [K.zeros(1) for _ in params]

        self.weights = [self.iterations] + ms + vs + vhats

        for p, g, m, v, vhat, tg in zip(params, grads, ms, vs, vhats, gs):

            sum_grad = tg + g
            avg_grad = sum_grad / self.accum_iters_float

            m_t = (self.beta_1 * m) + (1. - self.beta_1) * avg_grad
            v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(avg_grad)

            if self.amsgrad:
                vhat_t = K.maximum(vhat, v_t)
                p_t = p - lr_t * m_t / (K.sqrt(vhat_t) + self.epsilon)
                self.updates.append(K.update(vhat, (1 - update_switch) * vhat + update_switch * vhat_t))
            else:
                p_t = p - lr_t * m_t / (K.sqrt(v_t) + self.epsilon)

            self.updates.append(K.update(m, (1 - update_switch) * m + update_switch * m_t))
            self.updates.append(K.update(v, (1 - update_switch) * v + update_switch * v_t))
            self.updates.append(K.update(tg, (1 - update_switch) * sum_grad))
            new_p = p_t

            # Apply constraints.
            if getattr(p, 'constraint', None) is not None:
                new_p = p.constraint(new_p)

            self.updates.append(K.update(p, (1 - update_switch) * p + update_switch * new_p))
        return self.updates 

def _eval_pred(
            self,
            boxes,
            box_scores,
            num_classes,
            max_boxes=20,
            score_threshold=.6,
            iou_threshold=.5):
        """ Evaluate logits for boxes and scores to final boxes, class, scores
            results

        """
        import tensorflow as tf

        def process_batch(params):
            boxes, box_scores = params
            mask = box_scores >= score_threshold
            max_boxes_tensor = K.constant(max_boxes, dtype='int32')
            boxes_ = []
            scores_ = []
            classes_ = []
            for c in range(num_classes):
                # TODO: use keras backend instead of tf.
                class_boxes = tf.boolean_mask(boxes, mask[:, c])
                class_box_scores = tf.boolean_mask(
                    box_scores[:, c], mask[:, c])
                nms_index = tf.image.non_max_suppression(
                    class_boxes, class_box_scores,
                    max_boxes_tensor, iou_threshold=iou_threshold)
                class_boxes = K.gather(class_boxes, nms_index)
                class_box_scores = K.gather(class_box_scores, nms_index)
                classes = K.ones_like(class_box_scores, 'int32') * c
                boxes_.append(class_boxes)
                scores_.append(class_box_scores)
                classes_.append(classes)
            boxes_ = K.concatenate(boxes_, axis=0)
            scores_ = K.concatenate(scores_, axis=0)
            classes_ = K.concatenate(classes_, axis=0)

            pad_len = max_boxes - tf.shape(boxes_)[0]
            pad_boxes = tf.zeros([pad_len, 4], dtype=tf.float32)
            pad_scores = tf.zeros(pad_len, dtype=tf.float32)
            pad_classes = tf.zeros(pad_len, dtype=tf.int32)
            boxes_ = tf.concat([boxes_, pad_boxes], axis=0)
            scores_ = tf.concat([scores_, pad_scores], axis=0)
            classes_ = tf.concat([classes_, pad_classes], axis=0)

            return boxes_, scores_, classes_

        boxes_, scores_, classes_ = tf.map_fn(
            process_batch,
            (boxes, box_scores),
            dtype=(tf.float32, tf.float32, tf.int32))

        return boxes_, scores_, classes_ 

def yolo_eval(yolo_outputs,
              anchors,
              num_classes,
              image_shape,
              max_boxes=20,
              score_threshold=.6,
              iou_threshold=.5):

    num_layers = len(yolo_outputs)
    anchor_mask = [[6,7,8], [3,4,5], [0,1,2]] if num_layers==3 else [[3,4,5], [1,2,3]]
    input_shape = K.shape(yolo_outputs[0])[1:3] * 32
    boxes = []
    box_scores = []
    for l in range(num_layers):
        _boxes, _box_scores = yolo_boxes_and_scores(yolo_outputs[l],
            anchors[anchor_mask[l]], num_classes, input_shape, image_shape)
        boxes.append(_boxes)
        box_scores.append(_box_scores)
    boxes = K.concatenate(boxes, axis=0)
    box_scores = K.concatenate(box_scores, axis=0)

    mask = box_scores >= score_threshold
    max_boxes_tensor = K.constant(max_boxes, dtype='int32')
    boxes_ = []
    scores_ = []
    classes_ = []
    for c in range(num_classes):
        class_boxes = tf.boolean_mask(boxes, mask[:, c])
        class_box_scores = tf.boolean_mask(box_scores[:, c], mask[:, c])
        nms_index = tf.image.non_max_suppression(
            class_boxes, class_box_scores, max_boxes_tensor, iou_threshold=iou_threshold)
        class_boxes = K.gather(class_boxes, nms_index)
        class_box_scores = K.gather(class_box_scores, nms_index)
        classes = K.ones_like(class_box_scores, 'int32') * c
        boxes_.append(class_boxes)
        scores_.append(class_box_scores)
        classes_.append(classes)
    boxes_ = K.concatenate(boxes_, axis=0)
    scores_ = K.concatenate(scores_, axis=0)
    classes_ = K.concatenate(classes_, axis=0)

    return boxes_, scores_, classes_