From 2cb27d0196a363dbf101e0402a1a3a51f1ca272e Mon Sep 17 00:00:00 2001
From: 4luojing <31199131+4luojing@users.noreply.github.com>
Date: Thu, 21 Jun 2018 11:26:57 +0800
Subject: [PATCH] Add ICNet for image segmentation. (#975)
* Add icnet.
* Fix code.
* Fix utils.py.
---
fluid/icnet/README.md | 110 +++++++++++
fluid/icnet/cityscape.py | 236 +++++++++++++++++++++++
fluid/icnet/eval.py | 96 ++++++++++
fluid/icnet/icnet.py | 301 ++++++++++++++++++++++++++++++
fluid/icnet/images/icnet.png | Bin 0 -> 2090706 bytes
fluid/icnet/images/result.png | Bin 0 -> 212718 bytes
fluid/icnet/images/train_loss.png | Bin 0 -> 19905 bytes
fluid/icnet/infer.py | 133 +++++++++++++
fluid/icnet/train.py | 137 ++++++++++++++
fluid/icnet/utils.py | 114 +++++++++++
10 files changed, 1127 insertions(+)
create mode 100644 fluid/icnet/README.md
create mode 100644 fluid/icnet/cityscape.py
create mode 100644 fluid/icnet/eval.py
create mode 100644 fluid/icnet/icnet.py
create mode 100644 fluid/icnet/images/icnet.png
create mode 100644 fluid/icnet/images/result.png
create mode 100644 fluid/icnet/images/train_loss.png
create mode 100644 fluid/icnet/infer.py
create mode 100644 fluid/icnet/train.py
create mode 100644 fluid/icnet/utils.py
diff --git a/fluid/icnet/README.md b/fluid/icnet/README.md
new file mode 100644
index 00000000..56954720
--- /dev/null
+++ b/fluid/icnet/README.md
@@ -0,0 +1,110 @@
+运行本目录下的程序示例需要使用PaddlePaddle develop最新版本。如果您的PaddlePaddle安装版本低于此要求，请按照[安装文档](http://www.paddlepaddle.org/docs/develop/documentation/zh/build_and_install/pip_install_cn.html)中的说明更新PaddlePaddle安装版本。
+
+
+## 代码结构
+```
+├── network.py # 网络结构定义脚本
+├── train.py # 训练任务脚本
+├── eval.py # 评估脚本
+├── infer.py # 预测脚本
+├── cityscape.py # 数据预处理脚本
+└── utils.py # 定义通用的函数
+```
+
+## 简介
+
+Image Cascade Network（ICNet)主要用于图像实时语义分割。相较于其它压缩计算的方法，ICNet即考虑了速度，也考虑了准确性。
+ICNet的主要思想是将输入图像变换为不同的分辨率，然后用不同计算复杂度的子网络计算不同分辨率的输入，然后将结果合并。ICNet由三个子网络组成，计算复杂度高的网络处理低分辨率输入，计算复杂度低的网络处理分辨率高的网络，通过这种方式在高分辨率图像的准确性和低复杂度网络的效率之间获得平衡。
+
+整个网络结构如下：
+
+
+
+图 1
+
+
+
+## 数据准备
+
+
+
+本文采用Cityscape数据集，请前往[Cityscape官网](https://www.cityscapes-dataset.com)注册下载。下载数据之后，按照[这里](https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/preparation/createTrainIdLabelImgs.py#L3)的说明和工具处理数据。
+处理之后的数据
+```
+data/cityscape/
+|-- gtFine
+| |-- test
+| |-- train
+| `-- val
+|-- leftImg8bit
+| |-- test
+| |-- train
+| `-- val
+|-- train.list
+`-- val.list
+```
+其中，train.list和val.list分别是用于训练和测试的列表文件，第一列为输入图像数据，第二列为标注数据，两列用空格分开。示例如下：
+```
+leftImg8bit/train/stuttgart/stuttgart_000021_000019_leftImg8bit.png gtFine/train/stuttgart/stuttgart_000021_000019_gtFine_labelTrainIds.png
+leftImg8bit/train/stuttgart/stuttgart_000072_000019_leftImg8bit.png gtFine/train/stuttgart/stuttgart_000072_000019_gtFine_labelTrainIds.png
+```
+完成数据下载和准备后，需要修改`cityscape.py`脚本中对应的数据地址。
+
+## 模型训练与预测
+
+### 训练
+执行以下命令进行训练:
+```
+python train.py --batch_size=16 --use_gpu=True
+```
+使用以下命令获得更多使用说明：
+```
+python train.py --help
+```
+训练过程中会根据用户的设置，输出训练集上每个网络分支的`loss`， 示例如下：
+```
+Iter[0]; train loss: 2.338; sub4_loss: 3.367; sub24_loss: 4.120; sub124_loss: 0.151
+```
+### 测试
+执行以下命令在`Cityscape`测试数据集上进行测试：
+```
+python eval.py --model_path="./model/" --use_gpu=True
+```
+需要通过选项`--model_path`指定模型文件。
+测试脚本的输出的评估指标为[mean IoU]()。
+
+### 预测
+执行以下命令对指定的数据进行预测：
+```
+python infer.py \
+--model_path="./model" \
+--images_path="./data/cityscape/" \
+--images_list="./data/cityscape/infer.list"
+```
+通过选项`--images_list`指定列表文件，列表文件中每一行为一个要预测的图片的路径。
+预测结果默认保存到当前路径下的`output`文件夹下。
+
+## 实验结果
+图2为在`CityScape`训练集上的训练的Loss曲线：
+
+
+
+图 2
+
+
+在训练集上训练，在validation数据集上验证的结果为：mean_IoU=67.0%(论文67.7%)
+
+图3是使用`infer.py`脚本预测产生的结果示例，其中，第一行为输入的原始图片，第二行为人工的标注，第三行为我们模型计算的结果。
+
+
+图 3
+
+
+## 其他信息
+|数据集 | pretrained model |
+|---|---|
+|CityScape | [Model]()[md: ] |
+
+## 参考
+
+- [ICNet for Real-Time Semantic Segmentation on High-Resolution Images](https://arxiv.org/abs/1704.08545)
diff --git a/fluid/icnet/cityscape.py b/fluid/icnet/cityscape.py
new file mode 100644
index 00000000..3288b7f1
--- /dev/null
+++ b/fluid/icnet/cityscape.py
@@ -0,0 +1,236 @@
+"""Reader for Cityscape dataset.
+"""
+import os
+import cv2
+import numpy as np
+import paddle.v2 as paddle
+
+DATA_PATH = "./data/cityscape"
+TRAIN_LIST = DATA_PATH + "/train.list"
+TEST_LIST = DATA_PATH + "/val.list"
+IGNORE_LABEL = 255
+NUM_CLASSES = 19
+TRAIN_DATA_SHAPE = (3, 720, 720)
+TEST_DATA_SHAPE = (3, 1024, 2048)
+IMG_MEAN = np.array((103.939, 116.779, 123.68), dtype=np.float32)
+
+
+def train_data_shape():
+ return TRAIN_DATA_SHAPE
+
+
+def test_data_shape():
+ return TEST_DATA_SHAPE
+
+
+def num_classes():
+ return NUM_CLASSES
+
+
+class DataGenerater:
+ def __init__(self, data_list, mode="train", flip=True, scaling=True):
+ self.flip = flip
+ self.scaling = scaling
+ self.image_label = []
+ with open(data_list, 'r') as f:
+ for line in f:
+ image_file, label_file = line.strip().split(' ')
+ self.image_label.append((image_file, label_file))
+
+ def create_train_reader(self, batch_size):
+ """
+ Create a reader for train dataset.
+ """
+
+ def reader():
+ np.random.shuffle(self.image_label)
+ images = []
+ labels_sub1 = []
+ labels_sub2 = []
+ labels_sub4 = []
+ count = 0
+ for image, label in self.image_label:
+ image, label_sub1, label_sub2, label_sub4 = self.process_train_data(
+ image, label)
+ count += 1
+ images.append(image)
+ labels_sub1.append(label_sub1)
+ labels_sub2.append(label_sub2)
+ labels_sub4.append(label_sub4)
+ if count == batch_size:
+ yield self.mask(
+ np.array(images),
+ np.array(labels_sub1),
+ np.array(labels_sub2), np.array(labels_sub4))
+ images = []
+ labels_sub1 = []
+ labels_sub2 = []
+ labels_sub4 = []
+ count = 0
+ if images:
+ yield self.mask(
+ np.array(images),
+ np.array(labels_sub1),
+ np.array(labels_sub2), np.array(labels_sub4))
+
+ return reader
+
+ def create_test_reader(self):
+ """
+ Create a reader for test dataset.
+ """
+
+ def reader():
+ for image, label in self.image_label:
+ image, label = self.load(image, label)
+ image = paddle.image.to_chw(image)[np.newaxis, :]
+ label = label[np.newaxis, :, :, np.newaxis].astype("float32")
+ label_mask = np.where((label != IGNORE_LABEL).flatten())[
+ 0].astype("int32")
+ yield image, label, label_mask
+
+ return reader
+
+ def process_train_data(self, image, label):
+ """
+ Process training data.
+ """
+ image, label = self.load(image, label)
+ if self.flip:
+ image, label = self.random_flip(image, label)
+ if self.scaling:
+ image, label = self.random_scaling(image, label)
+ image, label = self.resize(image, label, out_size=TRAIN_DATA_SHAPE[1:])
+ label = label.astype("float32")
+ label_sub1 = paddle.image.to_chw(self.scale_label(label, factor=4))
+ label_sub2 = paddle.image.to_chw(self.scale_label(label, factor=8))
+ label_sub4 = paddle.image.to_chw(self.scale_label(label, factor=16))
+ image = paddle.image.to_chw(image)
+ return image, label_sub1, label_sub2, label_sub4
+
+ def load(self, image, label):
+ """
+ Load image from file.
+ """
+ image = paddle.image.load_image(
+ DATA_PATH + "/" + image, is_color=True).astype("float32")
+ image -= IMG_MEAN
+ label = paddle.image.load_image(
+ DATA_PATH + "/" + label, is_color=False).astype("float32")
+ return image, label
+
+ def random_flip(self, image, label):
+ """
+ Flip image and label randomly.
+ """
+ r = np.random.rand(1)
+ if r > 0.5:
+ image = paddle.image.left_right_flip(image, is_color=True)
+ label = paddle.image.left_right_flip(label, is_color=False)
+ return image, label
+
+ def random_scaling(self, image, label):
+ """
+ Scale image and label randomly.
+ """
+ scale = np.random.uniform(0.5, 2.0, 1)[0]
+ h_new = int(image.shape[0] * scale)
+ w_new = int(image.shape[1] * scale)
+ image = cv2.resize(image, (w_new, h_new))
+ label = cv2.resize(
+ label, (w_new, h_new), interpolation=cv2.INTER_NEAREST)
+ return image, label
+
+ def padding_as(self, image, h, w, is_color):
+ """
+ Padding image.
+ """
+ pad_h = max(image.shape[0], h) - image.shape[0]
+ pad_w = max(image.shape[1], w) - image.shape[1]
+ if is_color:
+ return np.pad(image, ((0, pad_h), (0, pad_w), (0, 0)), 'constant')
+ else:
+ return np.pad(image, ((0, pad_h), (0, pad_w)), 'constant')
+
+ def resize(self, image, label, out_size):
+ """
+ Resize image and label by padding or cropping.
+ """
+ ignore_label = IGNORE_LABEL
+ label = label - ignore_label
+ if len(label.shape) == 2:
+ label = label[:, :, np.newaxis]
+ combined = np.concatenate((image, label), axis=2)
+ combined = self.padding_as(
+ combined, out_size[0], out_size[1], is_color=True)
+ combined = paddle.image.random_crop(
+ combined, out_size[0], is_color=True)
+ image = combined[:, :, 0:3]
+ label = combined[:, :, 3:4] + ignore_label
+ return image, label
+
+ def scale_label(self, label, factor):
+ """
+ Scale label according to factor.
+ """
+ h = label.shape[0] / factor
+ w = label.shape[1] / factor
+ return cv2.resize(
+ label, (h, w), interpolation=cv2.INTER_NEAREST)[:, :, np.newaxis]
+
+ def mask(self, image, label0, label1, label2):
+ """
+ Get mask for valid pixels.
+ """
+ mask_sub1 = np.where(((label0 < (NUM_CLASSES + 1)) & (
+ label0 != IGNORE_LABEL)).flatten())[0].astype("int32")
+ mask_sub2 = np.where(((label1 < (NUM_CLASSES + 1)) & (
+ label1 != IGNORE_LABEL)).flatten())[0].astype("int32")
+ mask_sub4 = np.where(((label2 < (NUM_CLASSES + 1)) & (
+ label2 != IGNORE_LABEL)).flatten())[0].astype("int32")
+ return image.astype(
+ "float32"), label0, mask_sub1, label1, mask_sub2, label2, mask_sub4
+
+
+def train(batch_size=32, flip=True, scaling=True):
+ """
+ Cityscape training set reader.
+ It returns a reader, in which each result is a batch with batch_size samples.
+
+ :param batch_size: The batch size of each result return by the reader.
+ :type batch_size: int
+ :param flip: Whether flip images randomly.
+ :type batch_size: bool
+ :param scaling: Whether scale images randomly.
+ :type batch_size: bool
+ :return: Training reader.
+ :rtype: callable
+ """
+ reader = DataGenerater(
+ TRAIN_LIST, flip=flip, scaling=scaling).create_train_reader(batch_size)
+ return reader
+
+
+def test():
+ """
+ Cityscape validation set reader.
+ It returns a reader, in which each result is a sample.
+
+ :return: Training reader.
+ :rtype: callable
+ """
+ reader = DataGenerater(TEST_LIST).create_test_reader()
+ return reader
+
+
+def infer(image_list=TEST_LIST):
+ """
+ Infer set reader.
+ It returns a reader, in which each result is a sample.
+
+ :param image_list: The image list file in which each line is a path of image to be infered.
+ :type batch_size: str
+ :return: Infer reader.
+ :rtype: callable
+ """
+ reader = DataGenerater(image_list).create_test_reader()
diff --git a/fluid/icnet/eval.py b/fluid/icnet/eval.py
new file mode 100644
index 00000000..d3253c3c
--- /dev/null
+++ b/fluid/icnet/eval.py
@@ -0,0 +1,96 @@
+"""Evaluator for ICNet model."""
+import paddle.fluid as fluid
+import numpy as np
+from utils import add_arguments, print_arguments, get_feeder_data
+from paddle.fluid.layers.learning_rate_scheduler import _decay_step_counter
+from paddle.fluid.initializer import init_on_cpu
+from icnet import icnet
+import cityscape
+import argparse
+import functools
+import sys
+import os
+
+parser = argparse.ArgumentParser(description=__doc__)
+add_arg = functools.partial(add_arguments, argparser=parser)
+# yapf: disable
+add_arg('model_path', str, None, "Model path.")
+add_arg('use_gpu', bool, True, "Whether use GPU to test.")
+# yapf: enable
+
+
+def cal_mean_iou(wrong, correct):
+ sum = wrong + cerroct
+ true_num = (sum != 0).sum()
+ for i in len(sum):
+ if sum[i] == 0:
+ sum[i] = 1
+ return (cerroct.astype("float64") / sum).sum() / true_num
+
+
+def create_iou(predict, label, mask, num_classes, image_shape):
+ predict = fluid.layers.resize_bilinear(predict, out_shape=image_shape[1:3])
+ predict = fluid.layers.transpose(predict, perm=[0, 2, 3, 1])
+ predict = fluid.layers.reshape(predict, shape=[-1, num_classes])
+ label = fluid.layers.reshape(label, shape=[-1, 1])
+ _, predict = fluid.layers.topk(predict, k=1)
+ predict = fluid.layers.cast(predict, dtype="float32")
+ predict = fluid.layers.gather(predict, mask)
+ label = fluid.layers.gather(label, mask)
+ label = fluid.layers.cast(label, dtype="int32")
+ predict = fluid.layers.cast(predict, dtype="int32")
+ iou, out_w, out_r = fluid.layers.mean_iou(predict, label, num_classes)
+ return iou, out_w, out_r
+
+
+def eval(args):
+ data_shape = cityscape.test_data_shape()
+ num_classes = cityscape.num_classes()
+ # define network
+ images = fluid.layers.data(name='image', shape=data_shape, dtype='float32')
+ label = fluid.layers.data(name='label', shape=[1], dtype='int32')
+ mask = fluid.layers.data(name='mask', shape=[-1], dtype='int32')
+
+ _, _, sub124_out = icnet(images, num_classes,
+ np.array(data_shape[1:]).astype("float32"))
+ iou, out_w, out_r = create_iou(sub124_out, label, mask, num_classes,
+ data_shape)
+ inference_program = fluid.default_main_program().clone(for_test=True)
+ # prepare environment
+ place = fluid.CPUPlace()
+ if args.use_gpu:
+ place = fluid.CUDAPlace(0)
+ exe = fluid.Executor(place)
+ exe.run(fluid.default_startup_program())
+ assert os.path.exists(args.model_path)
+ fluid.io.load_params(exe, args.model_path)
+ print "loaded model from: %s" % args.model_path
+ sys.stdout.flush()
+
+ fetch_vars = [iou, out_w, out_r]
+ out_wrong = np.zeros([num_classes]).astype("int64")
+ out_right = np.zeros([num_classes]).astype("int64")
+ count = 0
+ test_reader = cityscape.test()
+ for data in test_reader():
+ count += 1
+ result = exe.run(inference_program,
+ feed=get_feeder_data(
+ data, place, for_test=True),
+ fetch_list=fetch_vars)
+ out_wrong += result[1]
+ out_right += result[2]
+ print "count: %s; current iou: %.3f;\r" % (count, result[0]),
+ sys.stdout.flush()
+ iou = cal_mean_iou(out_wrong, out_right)
+ print "\nmean iou: %.3f" % iou
+
+
+def main():
+ args = parser.parse_args()
+ print_arguments(args)
+ eval(args)
+
+
+if __name__ == "__main__":
+ main()
diff --git a/fluid/icnet/icnet.py b/fluid/icnet/icnet.py
new file mode 100644
index 00000000..14eaa5fa
--- /dev/null
+++ b/fluid/icnet/icnet.py
@@ -0,0 +1,301 @@
+import paddle.fluid as fluid
+import numpy as np
+import sys
+
+
+def conv(input,
+ k_h,
+ k_w,
+ c_o,
+ s_h,
+ s_w,
+ relu=False,
+ padding="VALID",
+ biased=False,
+ name=None):
+ act = None
+ tmp = input
+ if relu:
+ act = "relu"
+ if padding == "SAME":
+ padding_h = max(k_h - s_h, 0)
+ padding_w = max(k_w - s_w, 0)
+ padding_top = padding_h / 2
+ padding_left = padding_w / 2
+ padding_bottom = padding_h - padding_top
+ padding_right = padding_w - padding_left
+ padding = [
+ 0, 0, 0, 0, padding_top, padding_bottom, padding_left, padding_right
+ ]
+ tmp = fluid.layers.pad(tmp, padding)
+ tmp = fluid.layers.conv2d(
+ tmp,
+ num_filters=c_o,
+ filter_size=[k_h, k_w],
+ stride=[s_h, s_w],
+ groups=1,
+ act=act,
+ bias_attr=biased,
+ use_cudnn=False,
+ name=name)
+ return tmp
+
+
+def atrous_conv(input,
+ k_h,
+ k_w,
+ c_o,
+ dilation,
+ relu=False,
+ padding="VALID",
+ biased=False,
+ name=None):
+ act = None
+ if relu:
+ act = "relu"
+ tmp = input
+ if padding == "SAME":
+ padding_h = max(k_h - s_h, 0)
+ padding_w = max(k_w - s_w, 0)
+ padding_top = padding_h / 2
+ padding_left = padding_w / 2
+ padding_bottom = padding_h - padding_top
+ padding_right = padding_w - padding_left
+ padding = [
+ 0, 0, 0, 0, padding_top, padding_bottom, padding_left, padding_right
+ ]
+ tmp = fluid.layers.pad(tmp, padding)
+
+ tmp = fluid.layers.conv2d(
+ input,
+ num_filters=c_o,
+ filter_size=[k_h, k_w],
+ dilation=dilation,
+ groups=1,
+ act=act,
+ bias_attr=biased,
+ use_cudnn=False,
+ name=name)
+ return tmp
+
+
+def zero_padding(input, padding):
+ return fluid.layers.pad(input,
+ [0, 0, 0, 0, padding, padding, padding, padding])
+
+
+def bn(input, relu=False, name=None, is_test=False):
+ act = None
+ if relu:
+ act = 'relu'
+ name = input.name.split(".")[0] + "_bn"
+ tmp = fluid.layers.batch_norm(
+ input, act=act, momentum=0.95, epsilon=1e-5, name=name)
+ return tmp
+
+
+def avg_pool(input, k_h, k_w, s_h, s_w, name=None, padding=0):
+ temp = fluid.layers.pool2d(
+ input,
+ pool_size=[k_h, k_w],
+ pool_type="avg",
+ pool_stride=[s_h, s_w],
+ pool_padding=padding,
+ name=name)
+ return temp
+
+
+def max_pool(input, k_h, k_w, s_h, s_w, name=None, padding=0):
+ temp = fluid.layers.pool2d(
+ input,
+ pool_size=[k_h, k_w],
+ pool_type="max",
+ pool_stride=[s_h, s_w],
+ pool_padding=padding,
+ name=name)
+ return temp
+
+
+def interp(input, out_shape):
+ out_shape = list(out_shape.astype("int32"))
+ return fluid.layers.resize_bilinear(input, out_shape=out_shape)
+
+
+def dilation_convs(input):
+ tmp = res_block(input, filter_num=256, padding=1, name="conv3_2")
+ tmp = res_block(tmp, filter_num=256, padding=1, name="conv3_3")
+ tmp = res_block(tmp, filter_num=256, padding=1, name="conv3_4")
+
+ tmp = proj_block(tmp, filter_num=512, padding=2, dilation=2, name="conv4_1")
+ tmp = res_block(tmp, filter_num=512, padding=2, dilation=2, name="conv4_2")
+ tmp = res_block(tmp, filter_num=512, padding=2, dilation=2, name="conv4_3")
+ tmp = res_block(tmp, filter_num=512, padding=2, dilation=2, name="conv4_4")
+ tmp = res_block(tmp, filter_num=512, padding=2, dilation=2, name="conv4_5")
+ tmp = res_block(tmp, filter_num=512, padding=2, dilation=2, name="conv4_6")
+
+ tmp = proj_block(
+ tmp, filter_num=1024, padding=4, dilation=4, name="conv5_1")
+ tmp = res_block(tmp, filter_num=1024, padding=4, dilation=4, name="conv5_2")
+ tmp = res_block(tmp, filter_num=1024, padding=4, dilation=4, name="conv5_3")
+ return tmp
+
+
+def pyramis_pooling(input, input_shape):
+ shape = np.ceil(input_shape / 32).astype("int32")
+ h, w = shape
+ pool1 = avg_pool(input, h, w, h, w)
+ pool1_interp = interp(pool1, shape)
+ pool2 = avg_pool(input, h / 2, w / 2, h / 2, w / 2)
+ pool2_interp = interp(pool2, shape)
+ pool3 = avg_pool(input, h / 3, w / 3, h / 3, w / 3)
+ pool3_interp = interp(pool3, shape)
+ pool4 = avg_pool(input, h / 4, w / 4, h / 4, w / 4)
+ pool4_interp = interp(pool4, shape)
+ conv5_3_sum = input + pool4_interp + pool3_interp + pool2_interp + pool1_interp
+ return conv5_3_sum
+
+
+def shared_convs(image):
+ tmp = conv(image, 3, 3, 32, 2, 2, padding='SAME', name="conv1_1_3_3_s2")
+ tmp = bn(tmp, relu=True)
+ tmp = conv(tmp, 3, 3, 32, 1, 1, padding='SAME', name="conv1_2_3_3")
+ tmp = bn(tmp, relu=True)
+ tmp = conv(tmp, 3, 3, 64, 1, 1, padding='SAME', name="conv1_3_3_3")
+ tmp = bn(tmp, relu=True)
+ tmp = max_pool(tmp, 3, 3, 2, 2, padding=[1, 1])
+
+ tmp = proj_block(tmp, filter_num=128, padding=0, name="conv2_1")
+ tmp = res_block(tmp, filter_num=128, padding=1, name="conv2_2")
+ tmp = res_block(tmp, filter_num=128, padding=1, name="conv2_3")
+ tmp = proj_block(tmp, filter_num=256, padding=1, stride=2, name="conv3_1")
+ return tmp
+
+
+def res_block(input, filter_num, padding=0, dilation=None, name=None):
+ tmp = conv(input, 1, 1, filter_num / 4, 1, 1, name=name + "_1_1_reduce")
+ tmp = bn(tmp, relu=True)
+ tmp = zero_padding(tmp, padding=padding)
+ if dilation is None:
+ tmp = conv(tmp, 3, 3, filter_num / 4, 1, 1, name=name + "_3_3")
+ else:
+ tmp = atrous_conv(
+ tmp, 3, 3, filter_num / 4, dilation, name=name + "_3_3")
+ tmp = bn(tmp, relu=True)
+ tmp = conv(tmp, 1, 1, filter_num, 1, 1, name=name + "_1_1_increase")
+ tmp = bn(tmp, relu=False)
+ tmp = input + tmp
+ tmp = fluid.layers.relu(tmp, name=name + "_relu")
+ return tmp
+
+
+def proj_block(input, filter_num, padding=0, dilation=None, stride=1,
+ name=None):
+ proj = conv(
+ input, 1, 1, filter_num, stride, stride, name=name + "_1_1_proj")
+ proj_bn = bn(proj, relu=False)
+
+ tmp = conv(
+ input, 1, 1, filter_num / 4, stride, stride, name=name + "_1_1_reduce")
+ tmp = bn(tmp, relu=True)
+
+ tmp = zero_padding(tmp, padding=padding)
+ if padding == 0:
+ padding = 'SAME'
+ else:
+ padding = 'VALID'
+ if dilation is None:
+ tmp = conv(
+ tmp,
+ 3,
+ 3,
+ filter_num / 4,
+ 1,
+ 1,
+ padding=padding,
+ name=name + "_3_3")
+ else:
+ tmp = atrous_conv(
+ tmp,
+ 3,
+ 3,
+ filter_num / 4,
+ dilation,
+ padding=padding,
+ name=name + "_3_3")
+
+ tmp = bn(tmp, relu=True)
+ tmp = conv(tmp, 1, 1, filter_num, 1, 1, name=name + "_1_1_increase")
+ tmp = bn(tmp, relu=False)
+ tmp = proj_bn + tmp
+ tmp = fluid.layers.relu(tmp, name=name + "_relu")
+ return tmp
+
+
+def sub_net_4(input, input_shape):
+ tmp = interp(input, out_shape=np.ceil(input_shape / 32))
+ tmp = dilation_convs(tmp)
+ tmp = pyramis_pooling(tmp, input_shape)
+ tmp = conv(tmp, 1, 1, 256, 1, 1, name="conv5_4_k1")
+ tmp = bn(tmp, relu=True)
+ tmp = interp(tmp, input_shape / 16)
+ return tmp
+
+
+def sub_net_2(input):
+ tmp = conv(input, 1, 1, 128, 1, 1, name="conv3_1_sub2_proj")
+ tmp = bn(tmp, relu=False)
+ return tmp
+
+
+def sub_net_1(input):
+ tmp = conv(input, 3, 3, 32, 2, 2, padding='SAME', name="conv1_sub1")
+ tmp = bn(tmp, relu=True)
+ tmp = conv(tmp, 3, 3, 32, 2, 2, padding='SAME', name="conv2_sub1")
+ tmp = bn(tmp, relu=True)
+ tmp = conv(tmp, 3, 3, 64, 2, 2, padding='SAME', name="conv3_sub1")
+ tmp = bn(tmp, relu=True)
+ tmp = conv(tmp, 1, 1, 128, 1, 1, name="conv3_sub1_proj")
+ tmp = bn(tmp, relu=False)
+ return tmp
+
+
+def CCF24(sub2_out, sub4_out, input_shape):
+ tmp = zero_padding(sub4_out, padding=2)
+ tmp = atrous_conv(tmp, 3, 3, 128, 2, name="conv_sub4")
+ tmp = bn(tmp, relu=False)
+ tmp = tmp + sub2_out
+ tmp = fluid.layers.relu(tmp)
+ tmp = interp(tmp, input_shape / 8)
+ return tmp
+
+
+def CCF124(sub1_out, sub24_out, input_shape):
+ tmp = zero_padding(sub24_out, padding=2)
+ tmp = atrous_conv(tmp, 3, 3, 128, 2, name="conv_sub2")
+ tmp = bn(tmp, relu=False)
+ tmp = tmp + sub1_out
+ tmp = fluid.layers.relu(tmp)
+ tmp = interp(tmp, input_shape / 4)
+ return tmp
+
+
+def icnet(data, num_classes, input_shape):
+ image_sub1 = data
+ image_sub2 = interp(data, out_shape=input_shape * 0.5)
+
+ s_convs = shared_convs(image_sub2)
+ sub4_out = sub_net_4(s_convs, input_shape)
+ sub2_out = sub_net_2(s_convs)
+ sub1_out = sub_net_1(image_sub1)
+
+ sub24_out = CCF24(sub2_out, sub4_out, input_shape)
+ sub124_out = CCF124(sub1_out, sub24_out, input_shape)
+
+ conv6_cls = conv(
+ sub124_out, 1, 1, num_classes, 1, 1, biased=True, name="conv6_cls")
+ sub4_out = conv(
+ sub4_out, 1, 1, num_classes, 1, 1, biased=True, name="sub4_out")
+ sub24_out = conv(
+ sub24_out, 1, 1, num_classes, 1, 1, biased=True, name="sub24_out")
+
+ return sub4_out, sub24_out, conv6_cls
diff --git a/fluid/icnet/images/icnet.png b/fluid/icnet/images/icnet.png
new file mode 100644
index 0000000000000000000000000000000000000000..f261bb14a85eceac7cd5df282ebc43021b7760d9
GIT binary patch
literal 2090706
zcmZsCbyU;;_df=pgn*O?QYt9j4U-at0g|Ibx{;6^QxT8`>6UH~Ml(QZ7(IG~bTb+@
z>NmgN^LhJu|IYVy&d%;@f9<^Pectiyz4d!|
z*k1}79^AKBQGn&D=H+!$;ZWugX5K!p5UpaIZYKj)GXZm_8Vzv
z+_1fk#%)F7X2lpTHs1TBNy
z6toF%EXrsWVBV;K)CGn!fBVelp*d?5c%l&}`Z55w4XAtnBT<$v#o~JB*N8A{Q~X;*
z`#5(TaDKZ$p)yqGuWu*c_L~ZKNG3n0Zf<U~Y)K+|#P>yl2JSYvWVydG%eovP
z`miND#*~J7{oZ*4Uki(Qfcu-ES6$94l{9{;LO3tJ&2p`|X<9gNntyx2_nC9q8|SxT
zugU%>w{HIsIh>?(uwS3+;j@CTG%SXiK@qX!VT(gUD(hjL{X7qZBc3ktP9@Xm(Ji{-
zSPfn3QFuGjR56QBiHM5uIC46+ycmNRU8a6J`;cJ#1+qsY$8g?$3i3}T#`%$NSu>3Z
zcC@9$G1|WMw&^N(z?2&&w5rA90r)+Aqo1)x#}bD%d&%sqq4KRv`$DQ?l^flcs`I3
zm{4QTL;e*vStjV!d1lHvN&9u~bVEYO%sJU`g6{1+rpK@R>2Y4Lkul!)wO}7cfeE`5
z?w|+4Mv{}#FNQu}gF*Gc>mI6SkVta{IllUb>VrEem-@E_7;)rZ-Pw9YyK#s7=iS<$
zWc4^Czwg|Ze)0A0dsE)upDChm$uUw>e60EYDDHEaDQh8~$9FP;03nx0?w`zm3y9!z
zeS(|boxP>O^C08n%YB0WSEOG8dq^oYg54Qqd+4eHb)#ul?%>Ljy`_u0_gq#+@iyhF
z?e7|ebj3kiSymec8!y}+H{uY>)_qsEd5EF(!GFqgKmV8h4>AW_sHvLBy~r=rzk2ui
zL>Ra}*Zqulx#LH=_eHHe@i&<+0pL>s*+)_`vFtD&GOD)MGBIAy?vpW1
zBZC__?i^w?>0w;2LElq{pKL%k$c{xYlz0JI=Ha}Nk03xg7E0)CC=HYr5h)r-(-m%(
zU#VLvHbJ)?O5ExC-)8-6XKT(9gU{p{h%SJe)e
z?2q+9x^E=pB-H4hw>(!@w0b7=^3yXWN2XoYbmlqcX2#gw*A^tM5_-US87G!ZrUlkk
z_3v!DYMqURQJmN3KGyd)-o(js?g2yl6$XsQ1sBaS23ilLtQ^lE5}`H
zGibD$PkZ{csFtX#H`F!k5O-C9f}`{6lI)^-=!=UN)-TMpz7(rjYBr2{jJ@8{+Y%v`
z<l%W*S_nz|QlPubiI~$wFd>Au~2S!|*Q(H~DWs7B8
zXq>&cVtlR8yYOIKX>4Zv_1Jt7{}^}-F;-XPS+sz7Wt(LC#eT(ZXGEmr!2N+^B%E$9
z?VD)u`s~CP!2yA%0w9|`1l@4j#OTB!qHnZ#R4nI2E46$#E4{c`_TtNF?gQ;^6@~9m
z+4bB(^B{uu*qFgWi(}q=cR`OUkLQgijRB4PK3zWQK7l^FK4-`GF&tOVuAX3;F`9T{
zcuaV>cx8BX#0Nnf?G)`sL03T)5#bT@IT-g6m}A9v6m5gFRW)ZYU(;b
zVcy55!awKeGfgA}6;|dg;icsZ9(dA^yVf)i%sEV@N+lD97e)(wh=@&Ycd(4qALR_-
zB;;)3$ke0J?KE(NZq;FI9G%mh=4`;-LGQ?IG;I*(MmP=|
zPS#G{bx3nhM1~#1yVhSVlh*+OIRQOc17@QmKY2$gCl2t!x(Li2tK3TxEgL7)kE?$=&vNoPDX0CWpG3j7^s&x9&xdS=W)ZsfEyhOgLx$Cw78WNta8~5xR
zns*LdcnPV7MBfPA(B9I&HT!YxW6`VbpYfl-(z)MprN^a7J`a4B|0T)fqG;N48r^h6
zyiU088TfhlVRxn#LF
z**-axIpo=?IVRm+-ErND=4J0dUB;n^ujCyb-R8?V%c)BR)EJK01{aB6QduI480KYp
z?=UWo3Ei}o^Ye$D&(oeOJRAw(4dDzyiDsqY4IYnv%&So@sCU@o+9S`vcgA=AhU(*9
zS9{jL63;5B^g(4lLn3)8LnG!;>nD{KVMYO_U?Q(Yl~4tldHPb4rCIV$VocI|Ek&~Y
z`xksQ{6~Dj(0*trG#8$hxN9i%oZz9LrR;QARa2+H^w8MH4!J60zLezz&DbLrVX-sC
z4A%@EXVZr1LlrpdA^51(_>=xGqY;De20v;|-SgJ_Fxtb#c_lue?Fue^$(AnXfloDG
zYHaNxFRHmXxcwTqtc=uRbVgFLD^L8l4ZKzJ^UuRxSiPYQfo{TmI&@Y;$}e7S2ergn=AB2
zCWROQgRY{4AIrl6t;-(k>nB8}%cy8f<-Tin
zR`x-qg-E)Ual@SV9fqfm*8PUkH_vte-t#Aym4t)Tk;WEAJB=$YMQ%U+zc1roFo@wj
z!4nUr3vT2VH`=4g5cXRNy@*GEnkzDzs9bhOdoE)Zx78<|=N0RjntPkj8#^*T4F
zp;=E13`?!%h5%YixnFd
zN#1u+)N{kZ;d=V#bIU6a@Bs&h5$BDZw3b)e&fI-dI?F5J{RLF=oWp2b!l;m=HW7tV
z<#QFN+fMP-B-S!8Fc1VEu1t5p0<6sOp!AuJTc+8Qv&Q~h(K7O~PArkm3@|OU*P|4okq|zVS?U89`YwN8uSTOI0MAk~S5Bsh3X``;Ho12^M
zbFcb5jyyy?*TZX#%lOrV^EA#PKZt)Eh__9FwmHzf?WrBuTok-sPh$lf!8T&RfgN%d^EaJgk^btaKNi3trVlJN|v{|9Sp6
zIIo!nCUF1WI=rDkyy2tq7><|rD+ek)46@dnfRPq~Lu{jH1QCx{a(OC?pEq7{+)B?*!ZX0WW45em@E9~|v
z{CBlx-t^wYWX;y^a!jOL&y?_c-lZI)4tu>0B>-D?(Ef0820rtrmRq=@Lx^g%tSb=C
zB%LU043*shJU;R4Yabt*NVD_*5$))Rye`Z_505y3fVu(44JFdEO=yI;&&0z4=ieHk^-ZxjG*c-uOT@`UxP3;k`+mE}pJb}zjyPG@RlcUK{|UrE*BP4!x{
z1znA=wtTL0>C8>gulaKrL{BH0_fe;VbuwLLYfmNGF*|QHvwZsj=4NrkR7$A}y%1}^
ztt9%PL|MH_BQ$5k28fx75`-A%Hz|G!tr8?Te+Y)q*eA{>9xRgc-S)Tf08JV>Iy$26
zs)rd02lskFcfShv9M8^8jUGCXf_+y!XJ2PP2nps|6s)hKFXjceP;m;}4Rx%sy{mx^
zwEQZf+!5Y6)32e%<8g`{9Jv5)FaNGM+u+_!*Ca#22bL!47sU0CSI$L-Y(^nF9&^a6
zc|AQEW73r)fzSg8OKb>*n2?4(tL0}NiNE#xPxON3%i`a9clZ?N0z?p+i*k0bC0N&l
zUwSM1etIHD$p_H8){3DP={}}pzE%n(FrukT7w5}`%rHo;;VMk2Z@)tUmM4|e*yOJg
zER!|PHVySoi0Lm|J?y#T6zncPPSiSbLb|Mxu<%dlegQy#*XDb;?bu4jLLBhAq40BD
z3fC1_8k5e3&=c6XxgK=tuM*2!08osheoC@QMhT#@8@NebSVg%sgQhJ4FKc&8ZR3H7
zMDt8T(cmXP^-<4`C~@Ofa1&uK@Gn`CsVIU9YUcI%8MqcKefw0b`cZ=My_bQnX(ooL
zVG4ffm(y-?v{?3vyUWdgp_an)UcSE;Qs#>W^-iq3LkL*YgbvZE+~E9pQy5K)WKFL5
zc2Ool?qT%ltlQ((+immssydOUa^%I`zDy$QJx#*)c1NE(C)_=6%4%v(A{blb0Fxf<
z)iGC6a{Re?+@+I`I=RGhTk`n8Y8hwTs|IiEYm+e75sB5xsQeHv`2oE{G=`eBkdr$F5G>ByL6`gc}4$1w2YlM%Hl_j8j=`o$Zxr3K&F<7V
z{71sk=O|W7I+R@5x@EXWrvbapEX-lLO4T{XMxfCskW6eob;yD*w(8=3c=-MHQnW#>
zZ8}@c&gj`<(1~YnvvZ)|M0^?2S7D3cA}SHp80;4zS~Cj0XR1XyO{S?@F5?b-bP>Rz48>Z}(qKj%SvSZ~u=c#17(
zw5Js{2!y%m(iC($igR|~Mg;8A^M5%9atI*H*Dv<3dWPBEOTRP%O}6A%+bimxMvRLeI#W_
z^T0s=C<9n+A|&!#H{iT5X;q@9G&2qik(RpoR!40AK!Aw*afH7}sSI_nE;v>wJkLwE
zxHX~&Txfo>m5`^L=tTNW)5HXzvUU9oI%7c9O#1UZm>Bq;<>bfMwQ%A|<<%$uoOoiy
zg1G$l7k2+~)mux0g3(dP9VHjAuKZpfaC&?A*`!;SAN`
zOE%<)eF9skD9&E`=s=8I**s%HUvWFA7SOKxZjKz2*X?rCq(Gq3q(KvNDd+y81vkG>
z4>cxTKs4pz^Mc_d@WzDF?R4cMsJ)uiQmlv^t7!hXYYg@LVMZis4}+)2Fx|o;c0QWdcE4%*my7=Q;b;gcX=@K@k_~wRvcgh
zK?s^r(f5qiRxK?4TK^ttJ34cv?a6Yqosf_~La^a{r>v}u-S+31o^M5jAIM5ACrkEV
zK1PhHzW?m`w~&lQjIELoqy%D|p~l2+_arx|Jj+b>;21DVbn~L&1^UbEq;K%HwYjC9
zPFc9YX&+b6i-)4G^8rNF
z*Uj7NnsQdfghlTei_WI5y+t7rHD;rJoMp|g73>S`jl1%9a8dG>fwHpw#;@ifmSHBI
zrymx6*o!r$H}5>Uwf-TV6)?2HKQ582w;e8XW3JCFZlF4Rz<
zqZL^EVJUiEWuECY{^^Mx8k!Sc>2J*M-Pgd2Y%!8DNhWUFZi0Mq@9~}VdR~Yz$KoSd
znndMR@pXYK1-sVv=a{DOxlCG-pgqE&?bT_vTu;^kd~X!>Gz<6sh_4!^cFccqYj7|O
zzr`RTXjVaOGc!sd$F0VUj;VWT%5S}%Zh87}&^<8?YC>ZN188iz&xm(zXHdJX^-tTC
zY;9?J?N0dxtOriGn?xafQyh+#-d;EWpTQOmyioMEJPa={4(Bb4gDPGxQPzd{m^KUsnAdwiH59-BDD*kJZEd=LHo@?eSS(THQE
zg$wrWZr_^g%S?woTm-MQ<7uT&iJ5AaBjCsiaSP8
zBG&oBLAMh=4It^cR*iUM#?81AZ2a{RKiKHQ#R>6gzt|5RSiK|4#OzCX4srYNM0Ebq
zg5UaXAUuKdh?x7Zy0*53DfR*e$hXa)(%5ofIlX-!(?*+nIz-bE_5?0kDLU=v#_N%f
zu;s4b_sJWJP#wc$&_yxCLiT&V@X(S>OiqeVzsmk;0V!$i^SpYeOQUQ9HkHZm9@ybc
z!I2VW)q$_(p?!}qc43U;pjLfAs33D1>YLKmxe7m|TuX4N~L#L;_6gHF}
z{M2P}N5esb>7PlbVuVE*b&!@$}}=UfJm;L|)WOv;*jQ$JBC%
z2p<-W4FwQ|z~}8-sfE)M^Xf_t%L?ZDCsQ?wx(PR*8lZN0;XcYO@OJNu;p1&NF^|Gh
z6xHZZ*Ru}0Tg3OmzPYV#U>gbpzSya`k)%yw*30MXmw?<|+}%T^JogIrfaY11`0UKl
z-@gr=@#}gqkp11rIVH4xlYjK1pzoi06b>ux>#9@mIV
ztAf>xwem&0!}1tn;j84!@@}F_L$0I(!%@~=uX6@;ZG3IUEf3%2aU;ruRgL0wqvuPK
zPqD+ih0F1R>T%0l1#kLk*J;5(Cg3Kt&*hv;G}ZIOV2C!euc=3#Pu!x%m8t62G9c1h
z`eVm-yQJV1dyMLyaEfo;?41Po3SEQHUc6_H!m4nFSjSc?mgYR*f*t;s7?Fah44=T1bA`wxbavxOo>FFs~=D@?lb)g1Gdi{xx
z!4sAr(+ztZU?@
z@fV3#0tNfd4nA)`Fu5seDd9~PsffNx&evnm=`yM|E9^i4HVmN)iZY?O87#?uqE5(*
z-YD;Yel?EV2RKt59v*WtF-kW)5-K18meeEMRd2B>2+Sq7wJ$Slduywf2Be2BYx&KZ
z?TGLAOF(<3+#R@$bY-}($~rsoYn;tUeahx1ioSuaSQ8Sr(Y}aLy^-I7s90Ue47yhg
zrZtG(T^yg7&>1L9{$MfKjVNX-K8@crl0L}^=_cvn^wG3G`;l*PIP`Q{$VnWqzrb1}
zSdfCEaq^|14d)ktfCk}iMK@_90>$4&oq;saGzCiz^uUONaGJw#9QE-fP*%3Rf<1iS
zXtW!2p+sjHai>>sKbNk~+VVreud>;(B%681uLB_K8yGVrBO1^eXW#P9@Yx|Hi$@S$
zcm0qMS&{E2dQfv~xM-v;@Yx=B^{Gb4^l6>;*NVjH9z5a{`mMfWx|gj7sX24jLb(^O
z(1V^m@2}$f)dr2ahdc3|;o{Pc9+r7$8CrGH0r>{=KI^-glL$1a3%ER-n|l^Pj^0ukWAvx6Qo%-lvMf
ztJVT#%(!>TSqe7w=El+JHRg8s!;lVpr6S3+!V@^LY+nyy%nP>fGN-hzKsaF-x+|zD
z&ys7?hnt$oQVk
z><67bhmkBGeFJ{;)7U@{50A-+Fgex4^!@Avdv
zGaDJt+*B~d*uiC2`742M&_}ULk0J45?rBy1+)p8$;=+sa9nZ*{a}|@hG}Uy&4c@^Z
zLx{;<9;%1Lv-e~Eo6_*$l_c*^HfNNEUK59_@W~$UD`L@#8?La(1+hcqENR$lrlVR=(BwM)ywOGehi*cUz9q=lbG4u2$JH&@l0Tzysk6~
zUI}muAQ8N>cqZ+2htsIhJsudD;I^hWy^t3;cKdIk{S*Co`RdVMkoD7wLkM#1@YHY-
zR|42;>yrm9xNr-j;`SH)>JR&gpEsr+X&cG9bcF#-8hY_W?+m#Bw?FjVZzD%(s-0zo
z$cGU;XxmOye3tP(dr+$Hd$-Nc8>>Me{-eX1+#qK>GJ(-JpfOF|ZnMQu<^Z&>Xmoxi
zNpV)+!K6X=sARj!Qnyuo=UQrG2UgNpwZ+PJ5%Vj3WCw%6y!pcv!gg5!#tg?IBW6YmK5Yje=j(nt>Y&D-8O$IH|(iFwSTmnE&+0QbJC+&kMON
z4Y
znj`oL@ZNJdeJl`0FfFm@lSu$pdpo+)m}&`z_BdF;cZXN$Ydm#gJ6;qOq6K
z%sV@3zCK630uGa`Jcwc18oUsg@XZAH{
zJzXE&44l0CPsRKVejLdBt9JSbpO=(=y$r#}U$o@QB{8hFQD>C)x-wJ;fjs3rEVY|C
z_Ex}M@^|lus>2BhMO<9Mq-tzOva$M0YKiFFB`V3$q@x2c`y-I#+hgOnWa#G&1$%#h
z+J+hg()xp=L^2sDgg3H)ncEVqmMXRMsBAiFOMaK&K?!XfE6|gdiK*y2u6=iL@U+{y
zo3tby8isPW<`br9bh9(xR@=`=zq7F`1{n>(;n_LA~A
zUOja)FN&W@Yz*8RpqD4wG{MHK!5wFGF&Jva3j;o1E<-9%u;Zc8Ptm*zsZ!S4Q^J&(;
zwZ(^tk5QJ2_Lma^18TTrq0_VkWdtP6eZj;7X9F3M7yP@syGEU-b+=UPG2C+*85!A`
zWCKNG46C#L`lo2#%r))_EPUWJX_4r$9ue#B0%6t#Hw{u}W{kNl0JV;jH2s;6&y#mh
zpDf`C9#oo{a12L7x4)_tZRj+~-!l0p`q_2wU&F4bIJAe8FL@K;uFV|dl}!Pzt*L2t
zMxMst9RPaa1;f=IOcx=@Q~wS>%x|i>u{5PQEsNv;*!5Z}^&d!6-|ue`fR~Tz=Ge|u
z>0QYT*h45h>)2|s5wI&jY=UmjIuA99miY!unOUn1MQAVfG`RW|yCLJUWeerZyzoY@
zx2Xi?#zsdErI{{@Dk?HrCWF(D4j`+I8wYcy6ozY=?AhegH#2uW4#XKA&E?HPmcKn`
zUqNHMW#LEhwi!cgW&>NoN9#k*_n3DQ1T8OJdj&RswycwOmu~nTy#6#EA}d7xgk$KE
ze>M5W38y(?GG))jNSOR*3jxc}$l1GuBVn|Y|N2IBJ+QTj+KI9*y;JLz>B6hhhNie_
zu<@f*jV%$R03|uYs>$8-r{-%G=FA6gH`xxTODZyZDPvN<;ynqRrx$kf`u)jX4?8>C
z4==n?A_~XgpFU$|Uap4m=K7vJI}aKhxkgB^j@818#^GDAgJN?r>z+sGPigsOZGH7k
z<>iCN^I4yF5Y(D}qdMdjIoIaYv;LEFaQ~>XWziDfLw9b1YBHjUrZaJg%~Y
zi{07mwcLHqcWm}xWvZsOwmf|L)bWt~2?}xPjV89v(UKc0M6rs2!#vDURHMOj;q%1Y;bqyQ>TlsOZAxH+wQ`Pp
zZf~g|Iz(ts&H{8SRK3E&Zy6|0x9%Ap-HFtC^4myO|EREizqGLC6^rkK
zd0xp(AgeP`c9-Em(Hj18ErJ88s)|q@kBVP?rW-%lbGJvV|<%-1=jH7TKF(wMC0Br$<)o?Cf<3o3u~ogi6CR
zdoAoU^C!CGmQd2;NpQ=ScE(0sD>qred=N+{gB1~bdmD(*>sY~+ILFl%(X2~$4FI-M
z>O!7Jk6^Y)q13NqQm;FhgjAz9UrL{1(W$jXmHEQwde&m0^V))z{TFmG&3HP-^d<}T>};WCyGs3`!l5~(Z@z5TNUn8
z+}Mlg$NpP#>CXo?*A(t-al5$AS41KssexYkYDZ3o#%~$CHkeCt@HmW2#>#aeFU}}K
zl6T*oHryI$O_(#Go4WoddH)p%-M-iUm_KkBgV$?d{!=BTJmdaM>^@unL;gr>Ff
zAm}M6DUB67zqai=AS0Eo@XzbEQtxoQC*^jRwrf$gGv=v}j}lfn+R|b>F=R|{Xofta
zJmh=Gbzqt6HUd+X>^y`=N3-%c%0H4ZQU-QZ(#rC
za)uHScqq|4NjSph>G(teVrEIQNw?*cwlp^r}oy!W!|V}%M(j4
zVXv)N*93}4&v27b^guQJCO)b}&tko*;ondhMVez;Z&eXjCrUalBRiT=Kq6}2=n#FD
z($=>x(-%clT!HNlv9^XM{?nb4fQ=o$?@3p^-bHw=cKiQhZVk*sLPwSN+L<)|JCpyU
z_~QUU+x|ayI6_Zr9n}QBsvSFzM`4xF$QN;3kM@Y=>_}v{$%3^W$t|#E|C+v~;fEcv
z;9uiwYE8#Dv{bST|WqlMlWA9Yd5hXejSM
zM9@x7&Z3gE@0e^%DJOh#a)Kysqx7{_i%3k>ED8>t&U|+jxnRrEedz$Nx=sSJP*X#1
zq*CY;pD+`V;3&yh^I1d5W@0ZhlaWw3e(!NDUn1_iEg>Y`Ge94OvoX2ko}{LQ-)VLD
zd*i2>!(udb-sWMS+xJc3`)lIPCLn(AP->!6vP`Q0qGez*z0Q_cF*`d+e6PaQO!dTb
z;#k@qLW(ftpouQx^yyrFR#q0vLxKyuLR%F`>{)L7FKrQBUtizEv7JD9dXf5g0Ije}
z32(d--bf<#A*cBaSf(_k7w`;B+i)f!Vm+7^#f{};+N5B-GVZ&hwWvgMXr|iHFRV2L
zh*xT+lr!;}t#4m#F>3hm$;rvyew)h*;%eNoeB>VaR(_T9P~zOR(yV6F9rLogT89Gq
ze_^bT5ybyap3-)Eh__6}8h)Q1ys1Rz>iy`K5LAPH#0#RK+TE9U2lLlum?HTws%AP1
zu>ZtysW)t5rykOc?!?Tn2@w*64&LO;U
z>*$A^oRC_TlJpQO61y?KY>lnQAPS_KrHEhdB@U(inX+&nmv7j#NyUU(;1o9?v@1|o
zl6ie8BSenb2*}Xy69aioiSz0pWpq>iJ<`9Q@a>^JNXhqC`Mq}9JN
z=hUyxGDZcu`3a!`@G865-jzew@%p@&gRkH2!_S%cCu%6{rmHN|T^|@>g&s9O(6IaW
za1+&7x82=`iLM$MLzDIWmJeJeJo|7xrxhn7A1%rj0`?t9&$w?<3O=Ku1aMgAs6C?h
zX5I2!K{?}(Cwi*I_`-!{Iv_^Ek@Lf|WnB?BHHP%Q_`Js4xDBQid|xly_SxDU=$;}7
zLw99c0OI)`KuF@86NGX=Fd%Yhoh_M
z&jS7be>tfBx3lqs#!Mr4-F|z};#z-De+G$cTB-$C=^T$LPJrYt;Lp+u&|ZzG^E^pWv70ySpmc-BtaeubnnW-2Z))Kl
zoShpZ$p&o>mtQeK1ENXn6t>PBV8LjUB^u)R)pZ)H)Yd5iU|H`B*Z^LYb70`7HgdHVm
zzEutB+}b{KwsSPnaba!b_H*6-T=&}
zM{QZ8)7@aU_AL62zblq)`BwX?VOJ2~E6bhZ_i&;YQug)$+S+nHFBbqu#KEm>kf%fy
z;c;u^&9_t-R3;zzIr$D%)bC0#LKk}GA_W;lvB>-g#<=J@-Y9Q2^PyDvo(x;OkM>|z
zZoioSt}Fen{BxpqbIKWRNREG%*45h6N#$UmzrZ(-^`Xq0m_C<4lkX`7tT|ftF7d0A
zSPBL+*GKj22uTf)cgr(3!H*g+=Q=}tz(OBU0hwth`LfB9ioqxE`iA>->_z%6x_Rb
zLfp%)W6p=pOtO43RO&Pc?aQdOHD*Mh60>jKRb#0v19Uev>AZSYNHihS?Qxn>ia|;y
zoBK`B+A<~}uv&SN+ITZ)es|*>oqL~LEpD0yJ-HM>IYPa9kshO%=fwG}@HWnOEqwz6
zP!}ib%d?((lcw230DCOF&v1f@9lNGb-*#^|*2qxSUaFqgGo5JDo_I;x%fCJK6*tt0#WyDALOkh-3vD^VJ#vwQ2Z
zZ^?XJlnga4rR$n{rAd2dXVh0@?N}#CUv^&|J-N!kZ5amy)?;@|xjZfU
zZijvG>7%FGA#pnbeFyse?uw`v4A^F$c$^-goKV_h#2T+8xt27ck(`dDOC;I2UwB|Y
zv0cL2Ro)__UntUp^Ve{Z-zWFO_Z}aB%`Y~Lx3Qc9d*nX!PR=hk(w;O0(WSw%kibR<
z6pj6+_0f1v&CL}|*LW(^5qUw=F;G?CJjAxVamiwTy
z29In%SoqN=9(YSndQR~~=I*o#=hgTIY&Rw|E6u&aHf-TTr@BlS5)3XpS+Tno2xQk^
zu7&;(NBF8^JDhOgA99l;sRnXcOU2%3g})kePpm0faM
zzXr6v{l54Gi=TbKfMBaD!buMgoYitflA;6!bQhX@-nKY#@|g{bqE>z9-^`_33;wah
zqie9#OeVG|>1f5<$1H=Gm{^e3;K#_C&02qQydV-Nnh(e+?e?fRH5l@Z?Bv+ixAGXk
z&bWGs<$;J@n%0+^N|oUyTzRm~_o*2~BTHFbMJcijIS;=qY6r=%5WlBe)op#`p8081
zfE{vj`C3#thG8=v<{bnSk1s~bF(1@AB2DRd?jlsC@&DDMP@WcXT6?y8E~lMGTWguBM@y&%AC0Nl*S|S$|XVgID%nn(KEgA{Bc`;G9~K?y*+mI>E1>
zViOEQQ>5tBm#4lo7j-5X@9CRs(RuRA?c;w>gg$D^!>b(S-O0ZX9d8rt4(^PK>XhFk0LMr89NZ7lC&i*Y{0YBZy
z!?vpz8V$~~uI0K7Tg}Q?h$60}PL1>P-j#D<7f*0qELJd3Dz!RXxnsdMo|?!zuw0~7
zEV$5EqEpf0D06ZiX7Qn9R>-@HJGwXKIlDXIJLIXt6U1T0?gjR=NN;UyczI*dsfrtcY@6P;)3oSy2kqygs)P(vzw>NC+X7-rha%}Jo~}NLTJY%vd_dG=YQIK&Pa3b
zJ;|4U^Z@idZJjxB?3A5DVu7EM9^6JCPa_Kc-ryPZ$PJm~^+{ukk9)^C_oAe{bk=|c
z(d=EenUVioS?4ZjHz)=U`8QiK@Y`l2-a({*U(YT}H|+&>cQ=qaWp#2eMcH+LJv(a69sx{kLvQ)Q
zBwYLZFy&cq${uai8tLhY>g7Kty0%J+DNqOYa<>c?>rP}6t-kADwEkO!q+3I1&
zL&A1KP+_zu;ObZ5heCAywfL70MHUM+>horH_Vxhroc!0A-|esKr&X)=t`5bol$)9h
z1oIFf8sq;L5c^U2Z=Z`%Kr1PY$Iw%|!S3Z$+C^A
z-~xF`g_@guM+u~k%&|j2+dr+u5g{hzgk%+Fc?jWljK{&DeMc|W@y0&=3{CLstHDHf
zNPM!2gFpA42kYcI30M@4U1exhoyE%{d2scvr6(e*_T%xF%!;37;T
zoVb=^&8>?7w5dqV6)x~&cp`=-b9*_rn}XiWiE62ui|!$@oBvUCh~n63@YUq+KwgC5
zR-jR0V?vD5vxm0abgB>L3S5x_2wp5b0qC{(p;KaZu-@?#qBU*{>dZG>N(sllR@<0b
zb1{f|DD3B0@6(yLE$p;eUG=+{N}lCtX-xy!V%w(kpu+X7i(Odmm?^q|q@(>iClM&-(
z-x<0v2F!2`>7O97v+#Hr!Sjc^rV}p>9dp{~&&DZ9(wHWD3)a+fX)1vycn8e$Oi}e}
zk9$A3M(!~~ZaA$(Tl2~V3mCC@sxPmuu8yCgyazehf4=@JugmY$e(f~G$
z#6nyaXeLdy_|o!_Zl)}B5hYCxqxh%gFc3dP2;521+dB0)*^hJfRaDPbJ&+Q@xFK1
zQEq0n?3rLW-D~0MU9?T}5`@t1Vym-lPXV>x)DjaD6ILvrYWQihkU1*sXr~jq1
zRUta=2Lgky7E0KdL48Y~p3e4{9NjA~otdmmgT`*I)Obn#O~F6W&$W9u2Y+YDPpgmd
zd_i`?e8~})fTZ<7m`Y0DL#aeWK`LwN5H#P=^4aKS7`A9~U}pC(q4kb{@g&8|W=tS6;-_9V&IMdt*mOdn2FP$|^JTAYKH~y?~vJ~5J
z><6;^ej!VXO3hqzR&lku9J>jNj<3$lSCZl-{G1r`+^q_;ET=r+hvg%GJRew=8N(9-
z;iDPSUSeJq;uxPFSy_v*^Np*8wj#X=D|b}K^to!rF#?&nagDb1rcKjOYXufZ!GrxB
zwI7Ypri)sQu@p6|L_^Cb+7lwyY8p^(Oj|6;OiqayueE(bQBrY}K)qztC-#5rIpOpX
z|H}uv#k~_0+|#~oHa~T9_Iuphz>BDDKPiZy^M%2qFVt(M|Ko@`=sE1`$@_y^3?@RJ
zKMc7^C@IgXmz$)~heRdeGssek>0L`%6zY8_r{7#~ow%ARm{oA@q1}||OdVS)GAF~lh#rw26bUJ*(zB;75
z{kivl!R@2XTAFNpB9)3hw`JJ|N2Sd_;P!KG(uQR6EYorhpy$!n`{W9S?Uc_Qa@yWV
zfy9-R^_J*n*exg7AXlenFwcnl{>1+?1DH^bFWnNrCo;W0B;Fw)`>uVc1^67KDRG?8$Jow!9BrkEUk6)bq=W4qbfj`^68NQ
zc(9V58pj_=Pt^YC=}8L}G_7fx^*HhkigjU?$T{5)sg@~r@+ukF$93KKXK7ojCQ?pQ
zEd8z0?po5C{~<9~OjIrHND#1lO5NU0NaC61DDy4TFiI1WpwrBO`ea
z%kj!9xL9n1l>ibB{lV%a+8Q@~gjaW%Xnu0SpQ>+=NNKR1&`;PImb}(m8s6R2nDD))
zUPe^ijj!^iUN6=l#oqX+*iY7Dpjy4v+M-$ByKTR)AN4pTJQU_EDJO{?@6L%8DK%LR
z*w=yd_QWUG^ejz~0z(O#QE)wwRJx?3uukC9^**$&WPN<|?_H+${f>hx{&M9>e1o|S
z3t4DI$m<+Wtw&jIireIVcRKM?8P94K;HX*%R^WgXZU0~E126FZ#kcXaK$9gDMg7RZ
zv}D6B*Vhpt2_z#hZR|fA|Fa_#my>{9k=2<1W_-IbZQLVVEj0g*#o_>DZL66LCe`HC
zXS!A+7A??1e%#C?TlHdaST}29E0ulHZ~q@p$X(4QHSiPZOhu-
zOvkBmKi-w%D}~sS-n_DE7gY&uCqsFAS;bg79bm=K8&{zHs_rBih40>MN>*D{61r34
zxJ0j&PtN7!^;O1*0=x+i@u0=A)V@!TfL$`=kWQJ#rA2^>sV^d=!{a?NgVp27`XpED
zmG|g8%eLg(y&2a(j%A#_(~#yiXeHei-`f*Yu~4fT01YEJD918b4XnSjrZFLg-S5U2
zCHqYZ3~aZRuFQQpqyJ^aq=!;Pw`9hlQn@@L{?_~7C@<(=hU}}Qq5@+#V2P&HmF#RB
zM$w!7aXQfmB!X$Cz`6N^D4By@fm5?PDpPEQU~X=AweTqjt0pm9As5WDO)wR|e3ALy
z^u%NWATE_ncV;VbHKnr59!K*fbL>g+Ui{<4x*@T4(Blr(4+nt2yM*x?wQFJK9v-br
zeZ_a~hz(Gx9i{hPY={Da^j-u3=|X_e
zqjCg60a1!{MIq8cFM-gccd4NzK}rGy5+H;Yl6;$EJ>UJ^|L=YteUgy9XJ*ZsHEX@=
zU6Xo@YNzf623J$5{w2h!!7b6U%1UrtVX`oDse&iL3eLQB@XkxR%^(xmI`Ovl_2n8;Vs
znj@HgnRZ{E`9Y69Y3^`mTVGsh!c%~Sqy$&clQxdriqk251>Tl{Hg4YnX_EBqMMcyL
zM~t7pi^*!tdDi)LmnSKGu`*5K5j8lOeEWE{@_hga)4oh;)*_MeUkTcW5X4dp1f~@M
zJ|`A%w0sgj012)C<9hPu`hViO1|JlSuohJkc(72{6By}9>XOO+kj%{u=1>&(=mPwkV95vCn
z`UB6uMI`4;xPo45d33nlPmaOIls&c_pZ`l>$G-3B1>vshbIprpNgk(dePmvUyv?DP
zA7Z%wH{AXovc;sn9$9*Y|-X{R8gx?317xHc(L0McL3<+lVM1
zp6fO;YgVLB3kBIDBX3?MT6&oq(sr$3IXPpx*=||$iU*T@^jbUGwq$qy(M>2}&t%0h
ziw}R|>N4%RX7w)FDVY-1ai^WEhcE3k61c_MjO&e(2AJ2GOZ$($*1DXL08Uq*8=_1D
z5q%S`I|5HIL4ch+H`cOQ9S}$qP%n7?NF$4u8xI&@`VHT;0%Y$J(cgWw_OF|orMEE8Kt0loATM@
zH{I?j6Rr8$nW8z<6cVlL8Ii3^j&h6T9cLtZs!WWd?I4AQ}-nz
zaV{A7TA^*-@+O)O{but;;q#8IO72AG3`5`dxr;S(+oVgPuWEs5n5stHzskuSrY_}q
z!-xD$qSkpT9EetG8wMiAS|-fR%v7%UM76Z@V!CEnST^;q>~eLI(wAbvd2bK>>G}MU
zeo@4KQH^Jmzr_5mr#s~Bl?emnznXltsIY9+MB-Z
z&)g;soZdsosPEa{o#aSRF%;_$Ny$C#8pXntf$IZ)7Z+b|Ic(-*z}}kL+FWst|qEAp_V)tv0{~I&bHm
z!#Yc|bKX%s7G}5n&$)5A-JiJAlPX_Y;;6VB!a)WkNhHiOirqKoTS^WKiM3Vx-UxR987XppH$kc@zW!re^=-8Jq))~%o=TYfU|{f7mzTrHzS+n~
z9qG#_D%ooaTN4sE_QGn(;Kee^?9Rl6KIhU|RAESd)7Ssy!A3s)&!%z9*n{{n_Wo9I
z62PcYsuQr7dv2g)gwDMUOPr+K{M6euh;V@~+D!YUXB)!wSQGQ2KU_Q!Q3(`DCtu&a
zxJ^ZCm7kLd;l~FxwsR~QGa&-VEK;zYa0x~C#~;7#f233wVoE+>ptiEr`jZRrh_Fmm
z%W=`yN)#*zxmcQbptwH!*0iyBS8`{*2-)&FiT~OB0fzT{g5-j|wV7dNW`-7WR5m{h
z5Mu;mMh<&ey{#)9J*%YCc7ShE{6oNQi-h92%AT`p&!Ufq9gwmghzo9qq!LY6$5T9L
zAIjLOHoFdX${x`<2`MfM=4kV=zsO2
zv78aLu>MAGB>QIU=Tw){+>E7TF}!FeZPCIP1gUZY1R@QWtc}ulX{Q|_BeCOOUSUN
z%uP`MX;2*}ed;FRrR>VsB2`icZ$}6#=HyI{f|HkoK%-r9T3YGT?sRfiTD<=J`)QlPtOY7`4xo+2v^}_2*
z-8FS~{qx6N|IwQz5ug#i2Q&W{8|q2%pM~L1dhQhZa)g-x1;)n*uLE9!W1Pa33wxn{
z8Nk4lyZ?^FHE1aol+kaSQ6eX0En}a5+$v$I$2`FXgY82x;lQ^$=U4jOR&a@!Bwv$M&iKkq5RzZi7=48kQH
z8x2qvO>PVe$r
z;S=Rv_e4dnCoxjClQB368r581q4slJa7=iM4ce@LX=p*p|@#Q(?H(-q@XhRJ{NyXi%1mcwQ(KSScY%cP`V+b30+2x2EVwnW4#nU4w34TI;QhJ~z7_^;Ry<5bX~uwJMf(Z~sMF|cc7`XhV)sBsCTIBPoO}bzRX*nLcf0YB`vXu(V%ak;
zTITHsfW92diw>}(Fo{e6^Ndgu%mg+A`~$`7h}5@NyocQ1adXsE0G&>
zj9-|8IlU$jGW6AEA1*5L$TWPq!EY!)F6%b-=WDN}V}W`zSjlkfy4p77z;(jFrE2T`
zDKVvjGsORu%PPA2UmqhoaZ>W;C&X0oUUh<$MA_^&s}X
zpSo#2A3k7LY~lIZ=6f+Er%|);FZGERcvsn_@hShNSY4Q}`kdW=8Ys7MZ5GDu-*)Gl
zC{n9GX28c#G^e+hX|8@=(x{@se)(-jtJg@`RZlI_*`%QEO9~EU6RnxLDN*;8YPR$;
zvo5Wy82MXS2|`$pHeDF+jlR7Zx4`#|0Mf)t$*~CB3hy)_JEHxZ^YI-};II`~DmQ7@
z+{dgJml2q8JJxF`{Mj_f^`sUo}CCLG|S=ttm_Mg;L
zIX6&JXy@=VhGP^ivo<8_&Vq>}-gz94tlp?Fp0y*~w|
z%>DV6f#H7YLkYN?!-TQKsuRFQ;Wse078sn-G`+b86RamY)xRCnK%w9^QWG-JWcsPf
zW%G!wqDh&od8Mo&Dh%)Nm!OPb{wvT!J@G0vzW6&A(MPN6&z6^!1=*O6Aa8s+(JbWy
z2x~Z1gW3^c)~D&Fh^-Ls7a1En0#xeV#}2%Z=2HMTQ6xqdl?R-0pYJ5oPma<6z{Vbk
zH6*q_LGrZgH?S$H2R~g{4&-MxlHWP~Z+_19AJNw;m^)c|I?=v?GOzUaX!Ljl1I9*h
zo%bK_`s=g%W$*C~spg#oyyK^_tDhdEByhaS0sscd5%TQG()_|M4?0z;ho{zz@4Owo
zgy~TIM87BPQ^92de!n5B``mudu4Tz6}~gtFvk9%r{B(e9fn#FqcfWE%f3lPLhe@JpfE
z>1$d0uG=xLcmhl>V8`0>xw^CQMU8uK?S`7l^-ru$W{9nT*xSD|pypT0lXEkk8
zSv8l+-}feHUuhv82CD@kgSyKv>dKT#ck7P27+qgdt&t4fWi`7?JL9W%#ruPn^W8zw
zTHL^+ig?d2>h`Hc)_eSJR1!KxL)Y$n>^xA-SfI75oPq7rV@xHdbD1%R^XNtY!vNO{lqeE7gPQkLwU)Rr`S$E{Sc_5Nm#`UX2C8P
z13bgZhgMfZYqA4@|Dkz~Yf5_5w1B=G5IUD+IGpotce+3i;e>=)HXUH)-ECcvqct=%
z45P}&7425B3uUv8({DET6xPYgp`m?UIiF#x<`lY4F&!Bb69e(Nm+Tf9y6f9r$!yjM
zZj%pk=&qm8_Muz0MGxy?vaz*;Bs5A^Wleo`d;RrRPylUVl%8`s>Gbvr`op)-;wqPp
zsLloBd~0>)T?<4a-8dCjSkTfNnaW#sXPabqpbcjJlKMN*;T`!sMtM^Dl|^>)@x#7y
zB<;>3VH;N!NIS$8x6rqFlm%7c;k|C@;y;;m7V%UyrUlx3-3mV5Pnw9le>
z9<{d;6100E@IV;k99xe516Ivy^2UU9AZ^4=^kqxI%K}mSV4(#r0UK!A9lDv{l3M+U
zvhj|RQ<)Wn+S=L**YUwLFeMB+ZCeyJgKs16)2E0n;C4}w>g!2$^6`rp)rzpEney?$
zFZX|nhr3dP%*#us(orM}*)^qQm@?LxWP3L`B2v}l{wS66MK%Hyeq
zdoGjHo`ZuP8s)H-FDqWUse6OQ0Yk9xTM%hk;M=#5JdMd
z>ru(wyrB|x=v@?Ir!-Iu`^gY5{3RNgJRkDwYOPyQ6oR4kHv1+xAR5f`u%TAp0^dp}9r
z;}8Salu$Xn8V|z&V=l88LZ1Z73Y#Q>;uBXYzBPX1hdFq#(B@fI~#ta^A?jSv50kr;p{cdv~?bx-6O)F4O({q`{mdIk|PDu}PsI|&aiKq4@f~sxMDCUI8l~c=HeExIj4<9T&n<0VK(Z!sG4d2om
zzafQ=W7nVeGaf)!67KbMI)MthiOWPHWT?5hnZvjyMX(T2Ld;y|K$K;159aLJ?_+yB
zQnTW{Bvp5SPfSV7j9IJ1|z)TuiL=3-*~-T!k0+
zsos-ZxL%3!SsFcHytbuW<7=5z6Byzk67)G8DRrl6m=+O%-#E$wQDzr)vgE*gP~`@Ga@PB8hUU(7U?nl_m^!aWZKp7PWp%Qn-f
zlzJbMwWW6kofA%~F7XfX4ft)-9~-8sf=YOBxm4>eCH%59Cty=)V`lCUsX
z8G@@&m7lx(8!J)J`i|!|n7Y!pw$wL{9#xT_n|AZuoJ<(DRMUo(fgjBunk%-eFQQL1
zav24t3Z!OMj!?(r21E6eRHV|npfS;@i#0>OjnU9%I24QQlAzLP#qmT|IgmkP?&_HD0w%vZ$5N?cojM+9^LS{$KktlV07K<5u=->=)U#lFVR%_s_0hK5I0tc?
zV5ZTBs`l^F9MC&wXU{1LjaCkv1(v^cmt96=>}oM^_(FzP#4&81UV)1@acTupLyn%H
zomI(;hG9?lxa32uht?o5H7K$Kb+UbQ6t7c2Izwr_O|x4E2X;qbv`7P|L7nxSoIZ*Dv3T#P$lQ?1x=1f;pqJ?&vUWF$5l8&Q|eSAReh6WAl&XL|2_AOgEtIG5>z
zp5&>of!LN4g@4(%CsDV4YN0!3)P#+5^93w(Vo|RX+)^|KGo)aYGnrb_ryqMSf`g+KF=m;5GMfvKV=j+2!
z9`WtzH;z}YVh|ruR5rDKX++@4c5!`T8>)b0G_0RwB_koek>bMBEu$p1Qgynb8y!Ph
zVHDqOG~>0a6L=xkyDeD_aw|LRuL!KXoOwJcFgy&y_EIOwSS7$aU`;x0F{ee%(^sWP
z7R~JByqHBf%Dr_%ku?&z^W__@5!kT#EUL1DNb}8Gg=t|bPgyS3?WU)DiLjd+XRp5i
zTo9d%d?Y`thEOeIUd|6a*3K50{
z`ZgDNT%wppRB3nIWDHH5%aUndIKG-j;woO}f(GQL!(e<09gE=^z1Ir1r{!vKZm11L
zO=a99a+7PddN-*1=|ViA=5SKipuZe;h_|@Zo0CS9vOu>R@l+AyhMoaYa
z90iYbt)KKn(JNiHxi_Msc8%5{ayt;WWctKeC|znuwMr5Dl{ph9T3%DXh4S*(DcDd{
z>s3XUEDGl2tJ(pUT($5Xj#7Xd3_J%XJb`
zQtP$l5_IRzh5B1Lh497c%sUQTR+!yKK5IOB^XQQy*_j)T%z4|Zfd|4N%F)|vZ=ZikjTHF&$i`+Wyjp+x>01lT8BRCSzaMNny7Cj1ZyZwI3^m41h#8NK9
z50O+VwPsp*g^j5F&fA4lBeqG#RQc(ex&Cs4U5I8upS-e+ia76tzMx(Z
zeX<0Ai1*c!Wcf?+<7@cG|p
z-@-Y{*V{DRk74*xa6jVI)Raz4R8+8+blYf6;3i{@Gs@$xrYf?tMV^BWHJeY5GL3B2
z{#toeIv1tfvoX-&8#vp0K^;A)9ph=$7vU1dp~g!8=wq3*MqEB~y2@$2;pt{C-O7YH
zssT-F#Jpd{bdT}#LaP-Sfht~X89Xa4I4
zj{@xJs&0=xHcX)Vtv5b=BV3b&zWY?gdQ;aT
z)2**Bc$E1i9HM5d;~@=6+yM&{^k>%{d&cD0kb%~(RyF${voU~;5u1Uz`6VWGL~Px?
zVUSgRr9=3_pl4u9S{otcUTP>YzYIw(1HzC^JMCb^oo$_&B*^9=u32xjU8Zf!Y6!Kq
z%AF3Ci06yyYiP0;sZWKjRpQHA$paOcLesU@^_S5sp@~-Bo@|fCjz9%xz#Z#4(OO#E
zD)*@M!YU~+&2obAtwM|{(pOvOg}irgYMZL&+`5bUQSNspOQgyG;!!KbD(L5U)qZ{a
zPDdR3KS@R;vNV_%PB9+YsXf(Jz`}B5x6RKjL%cqHCeR6(xWtXL^r_mg*y32(7W~?4
z0jU*l9=ojEuQgk#ejvnOZJ`Idb_5kh2W%_aws&S;R#>lC&s4VJy8?D769j}?CRx@M
zkx|h?nG-MGLd3lkxDy;QKjmg`3c1y(LW=5k2ShoIDtmJ#
z#07}oq5yXiB&5pkj_MhXdDYA74}@cHfelCHWS8tM>&mRfjL<~*;WGh`m9Q#?nQAXZ
zkzMO+H8E%^fA~=z1>am{nVYU51>N@CS2OsszWFWlCbMf|I~MHw2m1j3lXlIO9`D!Q
zAGgSdl9k!$*I%|XVtZEQXfp+uN1{{N<&kM(5-PNqCPv2a%(S$q8>Z>fOa3OE{qd!c==&RYbDs>ILleVg}*CX
zV`*2T$+Q3skC((o(KHQufiDqV3kUqp>8+yxN8$5M?>LwBX5c^xh6Apua+Y<8hm+=K
znjJg3RpEEt9bI|rtb33s{YFrodsMB_2kRPruEM>|TeQNh8x`0xNrju_w`|uLD}@sy>^4LLU!|ra
zfFO`#YD@TR)#Ry`oqXvRG#`1T-2E1n+aj}vaPvN>(9Yo`~
z%J(PsSV!PP0XccMfe4)B1V~=5Q@or|Ei)iHLe5EZFn%OVf
z=GYmV+SOCM3|6Bij}j(FNL&*Lu)?slf{aXshj0KlaLsG9oSX@^*q>z0-d2E}0jCu`
z&`xZ@Z$-?LPxfO>Tp=z92)+{SIl(8w#$CsG;iO%?%8=?7YS^*GY?)**3?=b3gU4J?
zqXAaqWtE{Et7+?IH
zW)%(3YHN^@7@zg|nvwH=dyF(RT#F7?Rf>ln7Wh1Qyb@x9eUR9?ktkD1pEe^)xByX?
zIS_T#3t5FH=9BA;KLR@*u-Iv0?AAP$O)h6>m3p2qt8$3zI!jND;k&Q&+OzJCaq>-0ROcqGn
zgjTzkUj4@^G)%D{4EDG4KrZuHc$T_+NTm(V$wZJmiCFw<@ldazwag7&J_9U)WHe!v
z8ZN8`Q~u-+F8}7pI!_>=ONKg4DxKU3y-M{dwhOg_qx2EK$?~4W*`dEWI)i+c9Esgt
zWPZr!0%Sl`3nIu;af0nCSv_zw2{{w4Ggl=KU92X-1nvmWt(n8Sts_>u=GNC8iOH$E
zLuOWqzBeR`2=QvVhyXvEsfg8A89arklhe~qmx!^8lB+ijr#1>IF|-7HVLff$$r-Fm
zxvh@jrPf~vaS759#$~($t>5gYKkpgdjKr-g4}Ysr&L~&N)G(Gv|(Q1>OX9@aS{ilQrn_4=$Kn?k;gj*dc8z6~J$dBNkTMok$^5q|p5U(@c2WiRI1cA8^L*g)&r<&~9
z622bXZJz*X;67Ox9byGT4ba@>qKR|sj?vmWZ8rtUKAh^?#kN$AlW-sw@keB2sO0h!Ca>ce51ZOWq2Ki^O_
zl&;gZM^U2Kp*eA=tyO&a&_S5<-=o~ocQ;n$HZx2&`S<1LZ4tJN#-C)Y=J~8lHJT49
zC}mLXF7(Nx+Drr~>go{DFpPPhT{IB&h@@djUl9)D$3S4rXuGB&s!8Kkz399*t7B$B
zEpMXwg5|KV2-G>WhKcWu3)w}#*SNLsW!KEk!fsV|?w5=Fj)lV?H=gf!(`Fya%a1SE
zd~N&ut=Lqf3Hk~Cn!P!NhoxWPn-nwYTwMlq3qbpz)B~*O
z$JIOIvI^sVz|j&kQ(9cd4t%rgvI^?P^3uEv7fRtVZW5O44G8S$VX
z_;T*q#-StMRF?1kMvhz`|4HBY)hQG3|8pPPFfI~_^@uBx%a&jaH9k0;6*I69@~g*t
zFKh}9eLpQ^z6cZ`TkA>u`j|Rv9uEO(N7sAsgU7k9g9|oD@^jXjDKhm^1X`u`X4B57
zqTH_LRcLs+Yl@jbuXSO_l+4;O@0*$!rLSngKO
zRdmTOTgCj*0qn`|XNG@$AP5IGswZh9#(aHcON)%z{N
zs;9y9GW0+*=73P;*fL<&-r~QqFf>$;A^bsP;)}F(@b8-aYtSFm8xwVV4NUYmtfA66
zteo&Xsz7KkL10}d(eyN0S)=ZwMT_GtkwG5lM=m-2xzv~fY^?#HxN~Nkj0|7z4wHMF
z3!|nT7iRa~|GdO9exxmY${^me-_i(AC2#gRX-X&Lh#VpEKE%wolN3;lsVo36S0@_A
ztcW`BctbGqJQDA7F1aDh7l>wv7m>GrK8>M)>C%r%9GUkeTIOEbh63q815a3-iy$j)
zzy?WQs4I>YP9yyBGaFM{7k}N8+9kXyJ13F2_k|}dZ1ANhmkPA@=5KXnc$|IW$I0wm
zdT2eWFxNgo1_sw#WQk2S`9q?=WyWA~>Bp)n5>K1pmh~wEJcfig&{(3zkI%n%I`3D9
zvxTGT?b(5c7hRu|*vuCC(W!sGW>fxyJ>GEJAY1q3&nL$3d;IaZ~?8q>6%RR%Jr%KU8B$NL^A#dw{6r0R=t=jEIx}x6{IZz
z!++bJBTDn-P>aQr1I8AB->CU@^-BF%oHn6jWln<@3|2{NYX9Do6`!-zb=CMmtFq_j
zBPit$xycRw&>^5(d3pHBvANWN=il48LdsrzRse-U6@T>ndqXDTKLyx$iab4OkWv$<
zd%zrE>|FHH4(@45kCyq3Z!kO#`!y|Ey@Te1t5OIzMc<|QTG|di#jh10^
z;FpLN%tOq74TXWh?5DV*vab-7>4cF2bJ6H;nmmh43{SG+{PjDB=gfhyC$nBH~;`LN|mfFTOjZ}vWPRg!meJ~HU<0tGtz>ykV9s~ucpkEeorc`8(N;h`ij
zhdG7v?)XO>zrM=A;CSWd(EDeT#1Il$5vc?BlUx#gt1s;h#64F8?UjxmU(o*Jp59wO
zCxvmW*iN0eZC;5IWrqUr`Wr7!0lhu(6rG!M8^%Z=D}1W|C=+=i8T(TOuJuloE!2sh
z@}$@%4b;yjtpzKLo?u{jdF2a$d0i-z6G8Xly;c8c)bae!j>Zl-8_Tpxpk8+^h-O*j
z5iI-r(-|1#8ox>-El%A8ms?=<{;Q?G)ZtTiIR)2hf)+E(=Hglq*i+_)+iwE{2q=>)
zgT^53|7-+CKZ)b6XXUe$ql&vTL?WkO7R#Nfr0k_Oc0+-Ebi5zS!0=8#+uM}QmYd^o
z_)N@?@%`KJCx+jX^$l;rz^uex%LPOE`0_X}Weu`Hc0iK8vNL(ajQlRwDgmBO(o&XQ
zY;&<6T6;E*@X_rmPG!$~F0q)CWi_0P?;4wvyR~#d)Wmw&YiO~@t`*~4tedx?Z0?cY
z(N-y(X_R5?GAS4dHzPye72Ip{vU``)^J*31U+wv7!NgQc2kF>)zI9==?lUqwdr`mK
ztX4+eG$X&Y>Rzjtv@egMk?P2>Au)0SaF`ZRC=^+dh$ux54(b!}_&khPmvB_D2O0qn
z*;@J5y8YSPwvNH&g31i9Wt4snEw;G7xCY2w@x7>NN(l(y<(;WfnBZYx07XtuZ=op5
z5?O`K(vvM|?they{xpJ5)PVYOOS8PWn3id*z_hz|dTphN8&3iyG@njlNnfBq*kqq
zdzT3zJ%_xWo$jDoPQU^g@S%x(x9GW62lL7va&Z3!Gd3ao029{?ko|Uk#Ke{mWzzts
zEco4dkd%AB>3R8>8z-K9WFUfklJPhP^IZ#(v&G}9&Noj?u-tLwwtIdqVwvfp%tmeJ
z>*l$?PH=%9#Gi}!>qMVcAQKZWdxm79@dtsECr=^`%im4#z0`qzd-ql6?w)4JyxO|w
z>A716zbZ279*xmJdRPC!VEnZ<*bOxo?5`SM_glbUjpxg?o$8celZRqWj6A*61RV9J
zqmf;oak(mQWjHEZTuNg^K^MR1|9Q=a~f
zMeS*DtuHNn$sCg@<`xPUS7-M{L*3-e?{`;5&fDbHW&K7$eJ{RG}bg?(z
z-15%!UB(l6JcTkBi&0OSTg8peys5IdW;-(plep*!f6|<^?PYt%7B(CLcF>C_=66Qn
zWYaM5E(?h>+q**m7->CnnZMpE5@-!1a`8CB)Tz0(%&kEEUFRzKR_25W?%yl?$q%zN
z_4m!2Hz%C0s#co8bfAtvU=mW*B%Jmp^9uIHDh!TIp6*p73IXw?6}Pzb
z?M(9|3L?!+TQRl{sis#tor8z(4^2acZ+uMP%1Rp7>RWv$k!x6GQ<|RVh8wp1EWqqC
zd;wH0BHKEYZ%9ckfC`j6K+C|-c)w%Mk-ovokrOXTxm)_-g(lM_nB2wU{)M}tCybB3
zc^B;w7;{v^Lg8$237|7?_h01YGva+S>V9o{*UgyoO;yk}ThN@qjT4EicPva)fxkR{
zm@8y05$FgcPnrOfT=iJEzlE`gN;Uhwh
zVzGE}0;|TUHwPI6b+MF|@;2nO*|+Q}PDK9Buq@Z3BfqETTHyC>=tH$%Xn?S7`a-eQLn8eBFzVN!)km+K&
ziM7%Z&*JfjeG$y~<*EytkUGj2i{;WIs?J_49^tNbaJD2>EG?$z
z2%mgo;>3T&^U}G99^;u=paVzmx-OCTUI0qJz>f#GmH?@cN3_Kg*xDD$yLtOC2k>?VwgNVHvn9#BF)u%=B<;nZ{2WSz`Ha@
z^Gf%$NPk)T5Kp7_g=9bL+PZv>j3}3KG>t~1sFAG|P=B1|(vb8c>BxUYKfpWv483@;
zwqfDomDm;`wl^jP<|0=-v(7~ftLPlmKu};hl^arCHnViV$SRj546s2t-^iKgZPXZ@
zi)dqHZQ^`>v6uO#YidGZz@2FYbJIK1=byI#6L;#(sGh3$70>ntx2NZ=v%j-%ufc;5|9+n+c?qAO
zkna2xS%|~NL@M*JS9+j@VV^z-R}Y7s$(E1D732|&AB2J)yvupVd@ld5!(1_LVYaQt
znjgc*rn(l3OP?l2*)$rYB(v1#8{(526YVc}7s*I-`I*YxJm?rIGNWMQEHhPdm^0tK
zv@yXmO~UXwOD}U->IoUYoux@Q$x1x2@CC`gN-hBDI3S;r2dBUO>FWQxglyD)uAWq{
z3rXj3YZm~&-=@}W=yh{{fB#K!?nQ6DO#>d*yZNn2&od9-%9r&z7IWad2nVo^N>T5d
zeQ%x^?7k6rJ;E*QV(7VNm#
z1Bg6jdFsuIWrXHryaPVeUDriCi#gSDKkOgTu7)aZRSZCFyw2{qC7s
zUdRIOnHbJjmg9juM^F-OM4C~QgvhS0=6VU*mpB}n{NTA^!;~w)0>9P;7walXbq-_c
zwZdujJU0f;dX*WGcF_mnxxR(Y4Y{7Rzp&!r5G?tkhQ#4+OK2n}PVz@7Kn7Zh>$}((XEVaQClQ&}A@yt-{
zLB5Mn3w7uSdunu>k#&i^@W?$rR-Ppo5}KX+zOkbtpcd+Q`O(T~@lG)eW+E=lhQX1m
zm_8dn4fSLZ^U=1(NOJMJIWJc4#o#9j$;iw?zKC)2KYUsSx9_oVgDotdrUhMJ(3bs+
zfvfR0%iT|;gv0-u#PY)bpGh>LSbSfsQNcj0ytlGiLDrYAcqYAW2F>_}q@pk{Ioz)F
z-i7z-yJ&mw4xM(dT}VhNtlnduk_dgK9dVaF#C
z&VN(mE!%LG`5wI)&X%c2;PEHzE*yO94wH%&m&H^kPxN&sNs&hVLC!K#kNb-novleX
zB>FxTHEIky#(zLc)>qwPn)@=B@b{OKrGIm#K+-yWMLZ?mq~13Mt#+5tP9BNR9?
zNYllhX3^#HzF4yKwSt5k;%o=mNGzx<*oeZa!T2H}?c^q;6F4bE^lthf+Z
ze#wV;QvCsIT1*7qfGo#=+52uyB_~=GCD}md{)w__a;5~Z-7$H7h+W_C^@IEZVEqi+
z%^W8SgcluV&;P=dHVG=qZ=Fx;nq&LhS3jR%6UHaU
zTyYrb;jQsZNZI6+40Z?|)5z_qJ1BWKG>YahGVCxi=YP+~N0vQBJT%8h`STegll@1G
zDosnDh64GOyqYM6I)~AzC7dZ6+Iatgk(3Y5qI`PBhbDJ!Z`rV6kp~%54>4OD(=&Na
zHarLE2~Ow{YcxqDX&s!J+-da$=CsGh30*yA(WrB;Uc^(!!$bv`REXcPs
zhr_4QU?5qMf~1TxJih!Wj78*(L>2h)
zRv}2czARe!4?Vc?9T+h=i}UuEP9X2bYi|OokQS%!Z?_w4qAOT9f1jD7c{#!1D=S0M
z4YCB!6|ZU}Kn9-c0`;4S)lK7@X#1Sm3do)h44im~GX^aC@$hdO<|
zO9rwfv5^9Cgr_{t4+1r|&3P1L7*bak$veQg{@=`JCU&3wD}*1uXG3s&|??u*UE`=(rVv%Nt<(8GTo{QdsN
zlc0)O%#I129^Mszw=s_Z)Fi}?_;qEuxXK6MTs#BZ#uOy{ISHhVEF_8Vo63KqUo8&=
z*MCCDzj_6=%5%$+PFP#!aR=aNnMjWoHy6*P(Gi2^3;9rqG-K@Ah>!ARhScCX-bzgc
z)E}pcy!_$aq-I*F%CeOUc+V*Q3u()|P|Q=c?+@&)=97n#x+Iu^B5r+Bd6nhF!sXvG
zHUVn!_ugS=Ge;##|2?u*z+MA9=X?%IW^K}3K%_2W2o0Y5rj~TG3)?uhFQvKv2JD}-
zd^bNf+NKHSUcKVkiV7$I*>*h!mTvQ5X#&zu|3Z)C7BU=YMQ%SiuFh!oPp{i1^gB}9
zkZQqht`{v1C3;r_*yrfMHP_QtRv0lvPl~LEvk2N}fMws4x1DqAZdZ{qKT+sEvK=Nr
z*fMQ%Lnw>nrANS=7G00CS9p`xFfUc>v|WJI-kg1Co$HfTVx?M=-nRcq={=FikHe39
zq>%>nNxaFQ@U+dj@4DN46=XUar=Qdr01X6`H|i{IjUaGXlzl|{h(}3HO(B14fGz-B
zKG7FT8z>&{wj~b>`~lP3gaXynfM9{4VKZY%x(<5}6`LWnVtkG^LCYsOj5P~?>))W-LsGeok5?DpSJ7^#`}-j22_gG4*MEr1gdx>!
zT9Ir5S|a?p5L17M7vyCcBsjM|W`g+FI5&j71`gJ~h>%F!>r8;DVC3l-b~X5S@uqjS
z2k2<3~Et^(4%}jZu_YM-Aikn>eyp?&e4T*ZXr}i2<#oe`S693HDH~xQFG2lkcc6-lHky%~xxH`SKm094`Uy$r6`e&75U+5eG9SYAdMDh8})la<^7Doyma4YCqo45Hk@
zuQz^j0e+1Uc#Ny-9}n1L`L(PR?3tbX-l3tPhaSck4gd#!|7xy*>6cUilwJ_1-0<`A
zvk@Eq&lB(bdLqM
z06uCkbpc4PE_uqe*vfm&4`>LV(yM>#y12~O4Ij4E+Al~baSR~K^W!W(GB`%)e
z5JfV;#?BCmT(1Rq80p{}%24%XF0O4^=0$*(byY9G+as+QlOuIfVEHK}cOTeOjY_3iw5dPF~6r^1krJH$bI;ckB)h&_jHGX%-ZqVv6^
zWlWmIgoH?i@kNHDqa3Jb#Fc~>RYPo*Rh4&wj5O)0C#dwtKw}LJtWb-<+v}w6iba5o
z9Ycc%Evk-aEARN59qEfyfs|#@hUQ6Q{e6<-IqG!PQ&d&{Mb!b|?+}f3`s6$=+i-sb$o2
z2dfmP?VuYQ+gkB-LEvHIxvVV8m&2^1X!Z2w09yGxZdI77s<1=XDIm|&b>RuLE?R#q
z(FMv&g?Oyftv}M+%11!3+KW3uuA20iS2W1SauqqbtrxIVYfh?ac3QwXUAdp$l}T5x
z9C;ZB5!?wf(4@1*(CC3KfZ$9-oa=cg+vvS+-uy^URHw%%eMlz}Z)L(kPj4RgbUAuY;44Ok$h@&R=2$`-J(Odsl>gP
zo9gt!7c&IYuQ-6X1{bUq*x_~uKDOqf)+=mC0wS@ou`ve2sHW?xqXW2`nSIDx8H~$A
zAL(iRbn9jsWdfKajHP#0{i_|qHo2BcMSe0va%G0N0h2>toTo?i(_wp^5B*55kEZbr
zm=(@VdrXrj#|G%3)4{9Qk)3pxLR4%`Ao>I~2&_o1tpTfw0cSr6FCtU#7y8sI?;xtP
z=y{oRaM{Smz60KK(?nZU*1axbRCnlNuS+k#3F1f+-p8X+hpNS7{1?=92-
zO7BQQ3sq5+ju3k9B{Zdn5;h7_L+GKS5PB!{`o(+py=R~2o_nABeEE}=xmH$YX2u+O
z$9Uf{C-wPkUAj+NMNdQcgmgzhx!v!vVyd{jD4MmE$rSf8i#-*NR5)*(DBdCxZFB09
z&GXtV#IH5o=w&`^vLVUsbN`igt;0V8dyj#
z)#yfFU_&oj==Yt&hg(?o%Lr=*5V!UMRo;f9`(^w8SxDw5u)0*M8;#38JcHFu)114C
zW}`hO>nu(&JlHypzO6E%8CHp(H1Vu1RqOJG#G|*P(F1
z|FK0PJM-&TT*vl_=h^`FzJ>R9veiJk@8&#%OZP3kEhqPX3y`-9h$k=2JGOD*djt`_
zGYr|PYt%2y#wgJQBY(MSF5CK@@$fa}Z=8&cnOJ%|HWJ;4GhdT)%*mvsw=eP4B{Ltf
zc7-@G`xs|XN|}KS-)Cz~q622=zBf*k5lgTx{3!xklrp!Xeog0#S;M(vuV3@FP{V$t+O@+{!35=IQUg_ma=A==Y~Y
z^PSBD&I6x+q^dI^ni$P@#e^d1+8Uw*asNu};#h};x7zPs&GSk4gr4NmuUyiGkEFii
zo1OgqF-wdrD7kZL;`j&7S`d9fAjRwuv7I&a*RPGyLrt}aAKPC>PtZF%Ul~~@exHbb
z?Ml2`AmF29**-C9zG=beQb#mmhVNR9`#U>3@A)Wdi8D>}$FqLo)zWug5-XjSFE87_
zG!lOmi@7%xxZ2;nr2Ag}dz1?;y*p|4(|`Luf$)(-bFb-rK0mLyULBeL^0Dms+px41$|
zS^NW@emt!+nZ4}Nyl|v20BN_FJDwA3vQBr2UzEVbH&r)HoQf4c-pK#%q9fl_A@QF5
z*%ckzrW>PFED}eNjGNOg=YPJtCer=}PMqy$HWZ1(Uy@kY@@vp@?Ua%g&Hhor{f$iH{%ngNl!qe4p>a+?fTLu_
zmKS|V$8m$eN5Vvz;-y#n7w6EE(b1Dgks0qpi{DDTUrH29#NopIvZF9bC~|cv>mW$h
z62~Gf;#Zm{Q8D7r`B7bU_>{g_lUKW9Yh{bI6z5>>V*1aZlc;id{9h6B7a758mVcW#
zKUx04^S>G83c<79#!hP-}A`r*2gO|o4Q#Z{}=
z2*7;>Gs>K{FAua$owO?fVP(K^C*Tmn!G9=YyoZ_UK?T%*)hpFRrSUn`DXo~#b%Szan`EP1>
z>`n_hSf9F_R7f4KO3zBI8ve(i5P@2D(O%71dt~^Yj(ao%%wBsIFTHfFMg*J%g3UNV9W_jRF=(&3iNfRf0+k(9YaB0nebum5!{f)
zVb@E(#eu&?;}J}OITi1TPIBgNiJ7;T(vDvy@6u~fKXx{2saz*P^Ii|Hfk$zhbj@Ny
zaUu!o=s#1#4qm?3Y<19^hXAIt76F}<-e`U9AF}AWQEBYX&0o?IdWtisT*d<&tE#tOm@%r466fD2(3Ml^#ohy57=xiEi=PY`%#P
znC88tNmazHJ9T2mO1*lj8C3H-=1sL0(8yDE>Bl-8@LA&m%(!GjpkL-eBXjdXH*UGi
zwdEy`2-Tx8zPTxo0mK~5K@~|^zE^F=(GVyGr5LTMJgw{zj8b1K>G!hfM47y)zEbHYI-DYonw%PR#9%x~5k0K59@~@e?F^`X26e>BM*b=!h65JxNxx
zZmXT$m$HOnY@~IZo>4EfStwFhchCqsDr?5NiO^;#WO*Kasd+gcxJ`9&Fni$@5KXi?
z37r%E3-ASIaZ9%=S}ZP*^a8doerB5_es*BX;=w%|d3NEo?R{Q@lR6KVzVhg=$mVBm
z8@W0B7gw~{-;{WNv+2JMe}2FE57Q0}Oiv;0)6=03BKuSCVi6~I11QI*nV&ankeEEc
zcBu380p{>?3xj=pgk4#^Q(ujjkICbzk0f-mHNu!zKDGTR#4Te}3bwfDE}nh|R1{{2
z$LA*FRX>;l@{xXJ!396~)3sgiq$VhpBD*VfBB~g9wJJl}-?UpgR!KHgs`I8+cQ5Kx
zB=?$6a_fh4u7ss9(4HBg<>d6&tTd{;Xa0C
zK$T5X_dN8a$$(n4a!-=F*x$B_m(j;lYYZd__stDz3`m3&^{z_XyQYpT=oUBWu@1JN
zSFPMC2*p3JHyC52M!EH`_9B?=<|VSsDG~_FI~=ByG+OKx4LCb#6Y*K|s%E1}mCWr!
z&iK^3AMlc*`+m7P0@&FVbY4K2@eQ!7f$3hSqf_7JAlGa_=uE(=dd{Pv%_BMA$INam
ziO4y@m|_yjhX|z&B%1YjX2(10SE-5GjnBAa2Gr$3+Bn}@JzHYFWy+0eDb%M$!_w(Kwp%pKdqx{@CL
zc_YYN#H2~LZ{nrfS*3owqFQECyy>4ysJ(*G2?aQ%(5cyKe+8v|myB~mvyHi2;DmJ7
zTxSzH_aUD$g%1edAb~h~6PNAalFvG9z_{vh^^dpF&tn1w
zT6Wf}T2Atbbom<5d86-q6Pa{4SKV;ZZMpnOmGS+HbV)wm8Db6}{^}=k<~vhCa2>wv
zYMtM|e`oG*f1Ur`+W%^K!}soYuIkU=)*4TTJsLiQmZ_HT8V2v9#Mwy>63FF`wOcMq
z`l69dJy;{mBL{J~tniR1laks8s2wO*X5M-VzXIhiQ$RnS^;|#h=Xj?utHhs7ZBdMCe
zSnyK+Ir0<5PTND57$NU(3y+{eQkYcvd8?si=(>XKYWP~sq41P;-0I$_CY9OX5*lN-v*nMJ|&mLG;@i_C*W^)E^BkvqOmC>@c@y6Lxer^
zNshw6lyZ4cHCgckOf~A)#C*CGx!8pZW96vvL}Oerl~bhH8>y3`!Gbtiu
z^5N?87MLkF&@x$7oW>#r$R1{t;&6%2%%cb0#vYavz|=GBv>!ryT?g|*pJCg<>K428
zU|Zp?Kf12NnT(t`_jGwuRmUOI!dr?Kx}q69dvX)NMJc*q`RGEUa``T6$_o39oWnKr
zB%CrAj`fbpi77M^EgTp>*?|BVpWhS`q4PG{&l!0r7-=-h)ucfd{Qa(d;(}S8W32H?
zu(F|Iyo6%C3Y3sxGHvuK8Y<5RKsM|6x~Gn<(F?6_#czE$9<|D&D5k@qih)xIdX4=3+I
zr@#$?ExSobNmkt?W>aJ+0T`Gca%sgJ
zslCa}BVAc=OWdo;w##G&j8HxKrJNpeW5=AxRIvIZrKXl~tR4@xZk{b=Tqdv%%>JA$
z4BU`D>bN{*y=0JMPzmv(Mjgy;T`YNJo|6RK`J3SWf3oa9-iDK||JId(C=stXi*zC_
zZk4MED1(dCt?hu)u;Y}@KhMnQ>Wh0ZZBR}@@>U=PBBsyOXW)aQCv5nuk5>p#!x7X#
z2K-DT8h2)C!JpcE>RvB`Pl58nT@+|ZjYv6Hbd=4u&+Cj}9k}q&
zF$clINvm?*`-$S3K?}AzWh1Mr1xf9D)F=b&
zL@P5x7eDv{kc$QNc4csGZ5){g%^%(S7T@~EOE4rtbwkZi(dCcSg9^~I
z46gcDeO3-S(z<&_8t*#{kzK!XRIXh^wMMtO+UeBW@rAoTn;2W9Z&Y5zIaHGBdi+Y?
z&|?c2xQ4hz+o8qQ;@r)*k}?qa7PB5^{OS2pGUW4;#dn9}
zkg=m*!Iq5y?+F58o2g`5>r*~Up6~QOJzxlAsJutAW>_q~YX6=s+evE7DwWjsbuF`BR^0s`+5{eQt^vI21{OM-Q?kkIC
zVZCskNuQ;2Hky{3rCGnNiUMbLHaOdOy3Rl<+r;|d*3Mss$SUL_;XA)Y&uJd8)6QTX
zF#Pqe6aW5+Rq*p~(xU!Z$Y*7Dof;rwEOY{6s@B?W^R)g*J=kG{*$eTdi|uu5uo?0k
zN}Zw9C-*r&RtzeWQ97#EW3$T*%}!F?FWBNxPYNAYgA1Sjw0fFoOk7z$*aHwJ@Gb*m
zne&{sR$Fi!?BwF7yI_*1U#P5HX2N-Me&?y)q@r$kra0Qk&DXJudL9}&FT
z^{MX{RD$T}Nw|?({iSaz1BttO3*Ni0oc{(GAh3|?)4>*y(gp`b5Lb>MUqaq$hL9-{
z&b#n-u=RwYiMh1-p8yHfYSpVFKiPw79Z(2Aq3HH=Z_FCUn!z+Izl+!IUP#r?J_kdv
z0tewRK>a<%-V1k(FW2hCqonJuj7wbi~<6l%Ho(*u(-
z0HNpPtS5IP-)$Z4icL5^mc@Tl?@5&6_oia5c$)LVNOzO2E=wxcyEOD9g_OXyMD82a
zMaY(3wugVdMZ~P_*JPVM-IVJ)X@<{=<2&Zafkc1(VT-$(SCMpWEMUzy7B6aB@*ww@GlzKEZm0
zKP^>%U2>`2${g?jkrmcK4W=<1Iik0%fFMpSq*bpj*cR88Uohp8UE{H-Y7@c+cTHNQ
zBZEp8Gs88UVq1U0M^s;vjaekgnJ%^JrMSNbTL)qWCEIOE>_KXC9F1YM((&2pM-_
zm{EH~)8Gdu$kpaFh#UuvcyZ&cMfWJVr8%)g&MY@tQllrog0bCV#Ty
zaAj8jx7;-X_BqAi>``+E$H_W07N|cuZ9`SfZqLPz~b@0X4d}vMDibbJP-xoBMo0X
zDF!Lp;vkfqs8kA(0+$%A=8zM4bk^LS0lcxUb$X3x;EwDC`HgWND=HORq_$5_MS7O2Bf(BTB~wrEsfaH=~Z))(^`bwy}o
z7p)qb-e=_sXmA8=k^`CJegeA;tASrhW6$Fr5utS#E9w?+uZ2vr=I)dI${+Js;Hxyf
zK#bnoQADecj!VSQYX{J=<49jnW@=hlAC0OVVy}M=DJGm`7s(qr$-_L0Qg=n4J2!Y;
zvTXGA*#%bWzI&C#9$M=f(^mWiO3!GlRvBH4;O52P9ov7L45Y6~#(yh-SY=u
zq^B91=UOPxOd2tRqR_DD4qlW>!iTL=?g^;z^)5+om3^0`5Mu7@
z+|84;Bxt$aW8+QA5vMyoEw6Iq2d3~0o)p%(0oMeLvPyX^4Q?t&^X2FH_E;8(%rq-@
zBr-9xm=$nPQLRkW6c)K?vT-{AR_`ObFO@P)66*D_1UpfuiFv^M(1T7)`-g|56bb*S
z1@MOX-t`6b?#fhzlU`3fRLR;He_?z2ju?;SQeH|ixk#L@R|KpqVrG4Yzc$Uj2szyL
zE4)+TcV44;Ij1T6e~Z3yul-I(@Y#2>n>hAhwuaUCv?z?+WZu#rsGH
zuclrx+^WW`_{o87sRsp6(DW`UOM~0|QMD1bVzzEEWYITP2?Zo&Qg_1{
z&FH_k?ImP6Dt`c5y)`N|1ATO?DsqCtySUUn=7gp@@-JHOb4{<5tB&M5JiqiF6~_+y8ym{GqO#|R*q`&Toqcj=@nL0)B{-KCU>qm8Eb
z+*CCw!`i-~khdMQbcy7(%IoxUah<-N$}`m(j#7ySenUQ^4ZkFnX4y;LU)6A8bf
zN(jJRR3BL_g_8`_^xUk*^HoeYW7(st4S|UD(B~Q|O+Wqc3iGe_mrsgLpNx|P0SZ+?
z@p~6q)Rtdnc;~PA1K}H2nN9iNxt;56CdrCx1hzSp_3%Vj3bbc6Wd0r}bF@5GMNO9q
zrN@`1x~n!f22ry6rYw0n$k4IT?TyLk2n}(mK1$%4%uR!?xQW7i
zD%Nec>kV`>Nmdi-U+g&_r1x5~{(!7QcBibL7x;Q2O{p2wI%}bCf$~
zI=Qz@n*s7PftC&$YMiK6I|3wDFz!HZSHb!=%=Owl>4Bjs%3ZQ~M$0xjH4cY6eye5S
z!e(ga{D!wk%r2J~Ci#UreF)Hy_A$iKWuNwnzQFC2-aF1O%N4^pmt#G`C=A|%wfUGZ
zbI$?yZGt4b