├── bert-base-uncased ├── tokenizer_config.json ├── pytorch_model.bin └── config.json ├── DATASET ├── UWN18RR │ └── relation_id.tsv ├── CN15k │ ├── check.py │ └── relations.txt ├── NL27K │ ├── check.py │ └── relations.txt ├── UFB15K237 │ └── relation_id.tsv └── confidece_analyse.ipynb ├── transformer ├── Modules.py ├── Optim.py ├── Layers.py ├── SubLayers.py ├── dataset.py ├── Models.py └── Translator.py ├── .gitignore ├── README.md ├── translate_train.py ├── translate_decode.py ├── confidence_prediction.py └── LICENSE /bert-base-uncased/tokenizer_config.json: -------------------------------------------------------------------------------- 1 | { 2 | "do_lower_case": true 3 | } 4 | -------------------------------------------------------------------------------- /bert-base-uncased/pytorch_model.bin: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/seucoin/UKRM/HEAD/bert-base-uncased/pytorch_model.bin -------------------------------------------------------------------------------- /DATASET/UWN18RR/relation_id.tsv: -------------------------------------------------------------------------------- 1 | 0 _hypernym 2 | 1 _derivationally_related_form 3 | 2 _instance_hypernym 4 | 3 _also_see 5 | 4 _member_meronym 6 | 5 _synset_domain_topic_of 7 | 6 _has_part 8 | 7 _member_of_domain_usage 9 | 8 _member_of_domain_region 10 | 9 _verb_group 11 | 10 _similar_to 12 | -------------------------------------------------------------------------------- /DATASET/CN15k/check.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | 3 | 4 | data = [] 5 | with open('train.txt','r') as f: 6 | lines = f.readlines() 7 | for line in lines: 8 | data.append(eval(line.strip().split('\t')[-1])) 9 | 10 | # 求均值 11 | mean_value = np.mean(data) 12 | print("均值:", mean_value) 13 | 14 | # 求方差 15 | variance_value = np.var(data) 16 | print("方差:", variance_value) 17 | -------------------------------------------------------------------------------- /DATASET/NL27K/check.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | 3 | 4 | data = [] 5 | with open('train.txt','r') as f: 6 | lines = f.readlines() 7 | for line in lines: 8 | data.append(eval(line.strip().split('\t')[-1])) 9 | 10 | # 求均值 11 | mean_value = np.mean(data) 12 | print("均值:", mean_value) 13 | 14 | # 求方差 15 | variance_value = np.var(data) 16 | print("方差:", variance_value) 17 | -------------------------------------------------------------------------------- /DATASET/CN15k/relations.txt: -------------------------------------------------------------------------------- 1 | relatedto 2 | hascontext 3 | isa 4 | synonym 5 | atlocation 6 | etymologicallyrelatedto 7 | distinctfrom 8 | haslastsubevent 9 | usedfor 10 | similarto 11 | desires 12 | antonym 13 | dbpedia 14 | partof 15 | formof 16 | hasa 17 | capableof 18 | instanceof 19 | hasprerequisite 20 | motivatedbygoal 21 | derivedfrom 22 | hassubevent 23 | causes 24 | receivesaction 25 | hasproperty 26 | entails 27 | hasfirstsubevent 28 | notdesires 29 | causesdesire 30 | madeof 31 | nothasproperty 32 | createdby 33 | locatednear 34 | notcapableof 35 | definedas 36 | mannerof 37 | -------------------------------------------------------------------------------- /bert-base-uncased/config.json: -------------------------------------------------------------------------------- 1 | { 2 | "architectures": [ 3 | "BertForMaskedLM" 4 | ], 5 | "attention_probs_dropout_prob": 0.1, 6 | "gradient_checkpointing": false, 7 | "hidden_act": "gelu", 8 | "hidden_dropout_prob": 0.1, 9 | "hidden_size": 768, 10 | "initializer_range": 0.02, 11 | "intermediate_size": 3072, 12 | "layer_norm_eps": 1e-12, 13 | "max_position_embeddings": 512, 14 | "model_type": "bert", 15 | "num_attention_heads": 12, 16 | "num_hidden_layers": 12, 17 | "pad_token_id": 0, 18 | "position_embedding_type": "absolute", 19 | "transformers_version": "4.6.0.dev0", 20 | "type_vocab_size": 2, 21 | "use_cache": true, 22 | "vocab_size": 30522 23 | } 24 | -------------------------------------------------------------------------------- /transformer/Modules.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | 5 | 6 | class ScaledDotProductAttention(nn.Module): 7 | ''' Scaled Dot-Product Attention ''' 8 | 9 | def __init__(self, temperature, attn_dropout=0.1): 10 | super().__init__() 11 | self.temperature = temperature 12 | self.dropout = nn.Dropout(attn_dropout) 13 | 14 | def forward(self, q, k, v, mask=None): 15 | 16 | attn = torch.matmul(q / self.temperature, k.transpose(2, 3)) 17 | 18 | if mask is not None: 19 | attn = attn.masked_fill(mask == 0, -1e9) 20 | attn = self.dropout(F.softmax(attn, dim=-1)) 21 | 22 | output = torch.matmul(attn, v) 23 | 24 | return output, attn 25 | -------------------------------------------------------------------------------- /transformer/Optim.py: -------------------------------------------------------------------------------- 1 | '''A wrapper class for scheduled optimizer ''' 2 | import numpy as np 3 | 4 | 5 | class ScheduledOptim(): 6 | '''A simple wrapper class for learning rate scheduling''' 7 | 8 | def __init__(self, optimizer, lr_mul, d_model, n_warmup_steps): 9 | self._optimizer = optimizer 10 | self.lr_mul = lr_mul 11 | self.d_model = d_model 12 | self.n_warmup_steps = n_warmup_steps 13 | self.n_steps = 0 14 | self.lr = 0 15 | 16 | 17 | def step_and_update_lr(self): 18 | "Step with the inner optimizer" 19 | self._update_learning_rate() 20 | self._optimizer.step() 21 | 22 | 23 | def zero_grad(self): 24 | "Zero out the gradients with the inner optimizer" 25 | self._optimizer.zero_grad() 26 | 27 | 28 | def _get_lr_scale(self): 29 | d_model = self.d_model 30 | n_steps, n_warmup_steps = self.n_steps, self.n_warmup_steps 31 | return (d_model ** -0.5) * min(n_steps ** (-0.5), n_steps * n_warmup_steps ** (-1.5)) 32 | 33 | 34 | def _update_learning_rate(self): 35 | ''' Learning rate scheduling per step ''' 36 | 37 | self.n_steps += 1 38 | lr = self.lr_mul * self._get_lr_scale() 39 | self.lr = lr 40 | # lr = 1e-4 41 | 42 | for param_group in self._optimizer.param_groups: 43 | param_group['lr'] = lr 44 | 45 | -------------------------------------------------------------------------------- /transformer/Layers.py: -------------------------------------------------------------------------------- 1 | ''' Define the Layers ''' 2 | import torch.nn as nn 3 | import torch 4 | import torch.nn.functional as F 5 | from transformer.SubLayers import MultiHeadAttention, PositionwiseFeedForward 6 | 7 | 8 | class EncoderLayer(nn.Module): 9 | ''' Compose with two layers ''' 10 | 11 | def __init__(self, d_model, d_inner, n_head, d_k, d_v, dropout=0.1): 12 | super(EncoderLayer, self).__init__() 13 | self.slf_attn = MultiHeadAttention(n_head, d_model, d_k, d_v, dropout=dropout) 14 | self.pos_ffn = PositionwiseFeedForward(d_model, d_inner, dropout=dropout) 15 | 16 | def forward(self, enc_input, slf_attn_mask=None): 17 | enc_output, enc_slf_attn = self.slf_attn(enc_input, enc_input, enc_input, mask=slf_attn_mask) 18 | enc_output = self.pos_ffn(enc_output) 19 | return enc_output, enc_slf_attn 20 | 21 | 22 | class DecoderLayer(nn.Module): 23 | ''' Compose with three layers ''' 24 | 25 | def __init__(self, d_model, d_inner, n_head, d_k, d_v, dropout=0.1): 26 | super(DecoderLayer, self).__init__() 27 | self.slf_attn = MultiHeadAttention(n_head, d_model, d_k, d_v, dropout=dropout) 28 | self.enc_attn = MultiHeadAttention(n_head, d_model, d_k, d_v, dropout=dropout) 29 | self.pos_ffn = PositionwiseFeedForward(d_model, d_inner, dropout=dropout) 30 | 31 | def forward(self, dec_input, enc_output, slf_attn_mask=None, dec_enc_attn_mask=None): 32 | dec_enc_attn = None 33 | dec_output, dec_slf_attn = self.slf_attn(dec_input, dec_input, dec_input, mask=slf_attn_mask) 34 | dec_output, dec_enc_attn = self.enc_attn(dec_output, enc_output, enc_output, mask=dec_enc_attn_mask) 35 | dec_output = self.pos_ffn(dec_output) 36 | return dec_output, dec_slf_attn, dec_enc_attn 37 | -------------------------------------------------------------------------------- /transformer/SubLayers.py: -------------------------------------------------------------------------------- 1 | ''' Define the sublayers in encoder/decoder layer ''' 2 | import torch, time 3 | import numpy as np 4 | import torch.nn as nn 5 | import torch.nn.functional as F 6 | from transformer.Modules import ScaledDotProductAttention 7 | 8 | 9 | class MultiHeadAttention(nn.Module): 10 | ''' Multi-Head Attention module ''' 11 | 12 | def __init__(self, n_head, d_model, d_k, d_v, dropout=0.1): 13 | super().__init__() 14 | 15 | self.n_head = n_head 16 | self.d_k = d_k 17 | self.d_v = d_v 18 | 19 | self.w_qs = nn.Linear(d_model, n_head * d_k, bias=False) 20 | self.w_ks = nn.Linear(d_model, n_head * d_k, bias=False) 21 | self.w_vs = nn.Linear(d_model, n_head * d_v, bias=False) 22 | self.fc = nn.Linear(n_head * d_v, d_model, bias=False) 23 | 24 | self.attention = ScaledDotProductAttention(temperature=d_k ** 0.5) 25 | 26 | self.dropout = nn.Dropout(dropout) 27 | self.layer_norm = nn.LayerNorm(d_model, eps=1e-6) 28 | 29 | 30 | def forward(self, q, k, v, mask=None): 31 | d_k, d_v, n_head = self.d_k, self.d_v, self.n_head 32 | sz_b, len_q, len_k, len_v = q.size(0), q.size(1), k.size(1), v.size(1) 33 | 34 | residual = q 35 | 36 | # Pass through the pre-attention projection: b x lq x (n*dv) 37 | # Separate different heads: b x lq x n x dv 38 | q = self.w_qs(q).view(sz_b, len_q, n_head, d_k) 39 | k = self.w_ks(k).view(sz_b, len_k, n_head, d_k) 40 | v = self.w_vs(v).view(sz_b, len_v, n_head, d_v) 41 | 42 | # Transpose for attention dot product: b x n x lq x dv 43 | q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2) 44 | 45 | if mask is not None: 46 | mask = mask.unsqueeze(1) # For head axis broadcasting. 47 | 48 | q, attn = self.attention(q, k, v, mask=mask) 49 | 50 | # Transpose to move the head dimension back: b x lq x n x dv 51 | # Combine the last two dimensions to concatenate all the heads together: b x lq x (n*dv) 52 | q = q.transpose(1, 2).contiguous().view(sz_b, len_q, -1) 53 | q = self.dropout(self.fc(q)) 54 | q += residual 55 | 56 | q = self.layer_norm(q) 57 | 58 | return q, attn 59 | 60 | 61 | class PositionwiseFeedForward(nn.Module): 62 | ''' A two-feed-forward-layer module ''' 63 | 64 | def __init__(self, d_in, d_hid, dropout=0.1): 65 | super().__init__() 66 | self.w_1 = nn.Linear(d_in, d_hid) # position-wise 67 | self.w_2 = nn.Linear(d_hid, d_in) # position-wise 68 | self.layer_norm = nn.LayerNorm(d_in, eps=1e-6) 69 | self.dropout = nn.Dropout(dropout) 70 | 71 | def forward(self, x): 72 | 73 | residual = x 74 | 75 | x = self.w_2(F.relu(self.w_1(x))) 76 | x = self.dropout(x) 77 | x += residual 78 | 79 | x = self.layer_norm(x) 80 | 81 | return x 82 | -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | # Byte-compiled / optimized / DLL files 2 | __pycache__/ 3 | *.py[cod] 4 | *$py.class 5 | 6 | # C extensions 7 | *.so 8 | 9 | # Distribution / packaging 10 | .Python 11 | build/ 12 | develop-eggs/ 13 | dist/ 14 | downloads/ 15 | eggs/ 16 | .eggs/ 17 | lib/ 18 | lib64/ 19 | parts/ 20 | sdist/ 21 | var/ 22 | wheels/ 23 | share/python-wheels/ 24 | *.egg-info/ 25 | .installed.cfg 26 | *.egg 27 | MANIFEST 28 | 29 | # PyInstaller 30 | # Usually these files are written by a python script from a template 31 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 32 | *.manifest 33 | *.spec 34 | 35 | # Installer logs 36 | pip-log.txt 37 | pip-delete-this-directory.txt 38 | 39 | # Unit test / coverage reports 40 | htmlcov/ 41 | .tox/ 42 | .nox/ 43 | .coverage 44 | .coverage.* 45 | .cache 46 | nosetests.xml 47 | coverage.xml 48 | *.cover 49 | *.py,cover 50 | .hypothesis/ 51 | .pytest_cache/ 52 | cover/ 53 | 54 | # Translations 55 | *.mo 56 | *.pot 57 | 58 | # Django stuff: 59 | *.log 60 | local_settings.py 61 | db.sqlite3 62 | db.sqlite3-journal 63 | 64 | # Flask stuff: 65 | instance/ 66 | .webassets-cache 67 | 68 | # Scrapy stuff: 69 | .scrapy 70 | 71 | # Sphinx documentation 72 | docs/_build/ 73 | 74 | # PyBuilder 75 | .pybuilder/ 76 | target/ 77 | 78 | # Jupyter Notebook 79 | .ipynb_checkpoints 80 | 81 | # IPython 82 | profile_default/ 83 | ipython_config.py 84 | 85 | # pyenv 86 | # For a library or package, you might want to ignore these files since the code is 87 | # intended to run in multiple environments; otherwise, check them in: 88 | # .python-version 89 | 90 | # pipenv 91 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 92 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 93 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 94 | # install all needed dependencies. 95 | #Pipfile.lock 96 | 97 | # poetry 98 | # Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control. 99 | # This is especially recommended for binary packages to ensure reproducibility, and is more 100 | # commonly ignored for libraries. 101 | # https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control 102 | #poetry.lock 103 | 104 | # pdm 105 | # Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control. 106 | #pdm.lock 107 | # pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it 108 | # in version control. 109 | # https://pdm.fming.dev/#use-with-ide 110 | .pdm.toml 111 | 112 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm 113 | __pypackages__/ 114 | 115 | # Celery stuff 116 | celerybeat-schedule 117 | celerybeat.pid 118 | 119 | # SageMath parsed files 120 | *.sage.py 121 | 122 | # Environments 123 | .env 124 | .venv 125 | env/ 126 | venv/ 127 | ENV/ 128 | env.bak/ 129 | venv.bak/ 130 | 131 | # Spyder project settings 132 | .spyderproject 133 | .spyproject 134 | 135 | # Rope project settings 136 | .ropeproject 137 | 138 | # mkdocs documentation 139 | /site 140 | 141 | # mypy 142 | .mypy_cache/ 143 | .dmypy.json 144 | dmypy.json 145 | 146 | # Pyre type checker 147 | .pyre/ 148 | 149 | # pytype static type analyzer 150 | .pytype/ 151 | 152 | # Cython debug symbols 153 | cython_debug/ 154 | 155 | # PyCharm 156 | # JetBrains specific template is maintained in a separate JetBrains.gitignore that can 157 | # be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore 158 | # and can be added to the global gitignore or merged into this file. For a more nuclear 159 | # option (not recommended) you can uncomment the following to ignore the entire idea folder. 160 | #.idea/ 161 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Uncertain Knowledge Graph Completion with Rule Mining 2 | Source code for WISA-2024 paper: Uncertain Knowledge Graph Completion with Rule Mining. 3 | 4 | Since KGs usually suffer from the problem of incompleteness, methods of rule mining and reasoning for knowledge graph completion are extensively studied due to their excellent interpretability. However, previous methods are all conducted under deterministic scenarios, neglecting the uncertainty of knowledge and making them unable to be directly applied to UKGs. In this paper, we propose a new framework on uncertain knowledge graph completion with rule mining. Our framework contains the following components: 1)**The Rule Mining Model** applies an encoder-decoder network transformer to take rule mining as a sequence-to-sequence task to generate rules. It models the uncertainty in UKGs and infers new triples by differentiable reasoning based on TensorLog with mined rules. 2)**The Confidence Prediction Model** uses a pre-trained language model to predict the triple confidence given the rules mined. 5 | 6 |

7 | image 8 |

9 | 10 | ## Requirements 11 | **Step1** Create a virtual environment using `Anaconda` and enter it. 12 | 13 | **Step2** Installing the following packages in the virtual environment: 14 | ``` 15 | pytorch == 2.1.1 16 | transformers == 4.38.2 17 | wandb == 0.16.1 18 | ``` 19 | 20 | ## Datasets 21 | 22 | We adopt CN15K and NL27K datasets to evaluate our models, UKRM and BCP. 23 | 24 | | Dataset | #Entities | #Relations | #Train | #Valid | #Test | 25 | | --------- | ---------- | ---------- | -------- | ------ | ------ | 26 | | CN15K | 15,000 | 36 | 204,984 | 16,881 | 19,293 | 27 | | NL27K | 27,221 | 404 | 149,001 | 12,278 | 14,034 | 28 | 29 | ## Files 30 | - `bert-base-uncased` folder contains the BERT model downloaded from huggingface(https://huggingface.co/google-bert/bert-base-uncased) and it will be used in the confidence prediction model. 31 | - `transformer` folder contains source codes for the rule mining model on uncertain knowledge graph (UKRM). 32 | - `confidence_prediction.py` is the source code for confidence prediction model (BCP). 33 | - `DATASET` folder contains datasets we used in our paper. 34 | - `decode_rules` folder contains input preprocessed for the confidence prediction model. GLM-4 is used in the process so it is a little time-consuming and we offer the data can be used directly. 35 | 36 | ## Usage 37 | To train the rule mining model, please run follow instruction: 38 | ```bash 39 | python translate_train.py 40 | ``` 41 | To decode rules from the rule mining model, please run follow instruction: 42 | ```bash 43 | python translate_decode.py 44 | ``` 45 | To run the confidence prediction model, please run follow instruction: 46 | ```bash 47 | python confidence_predcition.py 48 | ``` 49 | 50 | 51 | ## Argument Descriptions 52 | 53 | Here are explanations of some important args, 54 | 55 | ```bash 56 | --data_path: "path of knowledge graph" 57 | --batch_size: "batch size" 58 | --d_word_vec: "dimension of word vector" 59 | --d_model: "dimension of model (usually same with d_word_vec)" 60 | --d_inner: "dimension of feed forward layer" 61 | --n_layers: "num of layers of both encoder and decoder" 62 | --n_head: "num of attention heads (needs to ensure that d_k*n_head == d_model)" 63 | --d_k: "dimension of attention vector k" 64 | --d_v: "dimension of attention vector v (usually same with d_k)" 65 | --dropout: "probability of dropout" 66 | --n_position: "number of positions" 67 | --lr_mul: "learning rate multiplier" 68 | --n_warmup_steps: "num of warmup steps" 69 | --num_epoch: "num of epochs" 70 | --save_step: "steps to save" 71 | --decode_rule: "decode_rule mode" 72 | ``` 73 | 74 | Configs are set in python files and in case you want to modify them. Normally, other args can be set to default values. 75 | 76 | 77 | ## Citation 78 | Please cite our paper if you use UKRM in your work. 79 | ``` 80 | @inproceedings{chen2024uncertain, 81 | title={Uncertain Knowledge Graph Completion with Rule Mining}, 82 | author={Chen, Yilin and Wu, Tianxing and Liu, Yunchang and Wang, Yuxiang and Qi, Guilin}, 83 | booktitle={International Conference on Web Information Systems and Applications}, 84 | pages={100--112}, 85 | year={2024}, 86 | organization={Springer} 87 | } 88 | ``` 89 | -------------------------------------------------------------------------------- /transformer/dataset.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from collections import defaultdict 3 | import pickle 4 | from torch.utils.data import Dataset 5 | from torch.utils.data import DataLoader 6 | from transformers import AutoTokenizer 7 | import torch.nn.functional as F 8 | 9 | 10 | # 需要先将图转为id，然后加载到数据集，之后输入给模型，进行debug 11 | class KnowledgeGraph: 12 | def __init__(self, data_path): 13 | with open(f"{data_path}/entities.txt") as e, open( 14 | f"{data_path}/relations.txt" 15 | ) as r: 16 | # 转id 17 | self.ents = [x.strip() for x in e.readlines()] 18 | self.rels = [x.strip() for x in r.readlines()] 19 | self.pos_rels = len(self.rels) 20 | self.rels += ["inv_" + x for x in self.rels] + [""] 21 | self.e2id = {self.ents[i]: i for i in range(len(self.ents))} 22 | self.r2id = {self.rels[i]: i for i in range(len(self.rels))} 23 | self.id2r = {i: self.rels[i] for i in range(len(self.rels))} 24 | self.id2e = {i: self.ents[i] for i in range(len(self.ents))} 25 | 26 | # id四元组 27 | self.data = {} 28 | with open(f"{data_path}/train.txt") as f: 29 | train = [item.strip().split("\t") for item in f.readlines()] 30 | self.data["train"] = list( { (self.e2id[h], self.r2id[r], self.e2id[t], eval(s)) for h, r, t, s in train } ) 31 | with open(f"{data_path}/test.txt") as f: 32 | test = [item.strip().split("\t") for item in f.readlines()] 33 | self.data["test"] = list( { (self.e2id[h], self.r2id[r], self.e2id[t], eval(s)) for h, r, t, s in test } ) 34 | with open(f"{data_path}/valid.txt") as f: 35 | valid = [item.strip().split("\t") for item in f.readlines()] 36 | self.data["valid"] = list( { (self.e2id[h], self.r2id[r], self.e2id[t], eval(s)) for h, r, t, s in valid } ) 37 | 38 | self.fact = defaultdict(dict) # 二重字典，h,r双重键，值是（尾实体，置信度）元组构成的列表 39 | for h, r, t,s in self.data['train']: 40 | try: 41 | self.fact[h][r].add((t,s)) 42 | except KeyError: 43 | self.fact[h][r] = set([(t,s)]) 44 | try: 45 | self.fact[t][r+self.pos_rels].add((h,s)) 46 | except KeyError: 47 | self.fact[t][r+self.pos_rels] = set([(h,s)]) 48 | for h in self.fact: 49 | self.fact[h] = {r:list(ts) for r,ts in self.fact[h].items()} 50 | 51 | self.neighbors = defaultdict(dict) # 二重字典，h，r双重键，值是尾实体列表呗 52 | for h, r, t,s in self.data['train']: 53 | try: 54 | self.neighbors[h][r].add(t) 55 | except KeyError: 56 | self.neighbors[h][r] = set([t]) 57 | try: 58 | self.neighbors[t][r+self.pos_rels].add(h) 59 | except KeyError: 60 | self.neighbors[t][r+self.pos_rels] = set([h]) 61 | for h in self.neighbors: 62 | self.neighbors[h] = {r:list(ts) for r,ts in self.neighbors[h].items()} 63 | 64 | 65 | # 获取稀疏矩阵 66 | sparse = data_path + "/sparse.pkl" 67 | try: 68 | with open(sparse, "rb") as db: 69 | self.relations = pickle.load(db) 70 | except: 71 | indices = [[] for _ in range(self.pos_rels)] # train graph转换为稀疏矩阵，为什么只用正关系？ 72 | values = [[] for _ in range(self.pos_rels)] 73 | no_repeat = defaultdict(set) 74 | for h, r, t, s in self.data["train"]: 75 | if (h, t) not in no_repeat[r]: 76 | indices[r].append((h, t)) # 表示位置，即行列 77 | values[r].append(s) # 表示值，即行列位置上的元素，这里都是置信度 78 | no_repeat[r].add((h, t)) 79 | 80 | for i in range(self.pos_rels): 81 | if indices[i] == []: 82 | indices[i].append((0,0)) 83 | values[i].append(0) 84 | 85 | indices = [torch.LongTensor(x).T for x in indices] 86 | values = [torch.FloatTensor(x) for x in values] 87 | size = torch.Size([len(self.ents), len(self.ents)]) 88 | self.relations = [ 89 | torch.sparse.FloatTensor(indices[i], values[i], size).coalesce() 90 | for i in range(self.pos_rels) 91 | ] 92 | # with open(f"{sparse}", "wb") as db: 93 | # pickle.dump(self.relations, db) 94 | 95 | # 所有三元组，包括train，test，valid，以及rev三元组，(头实体,关系)的tuple作为键，尾实体集合作为值存储在filtered_dict字典中 96 | self.filtered_dict = defaultdict(set) 97 | triplets = self.data["train"] + self.data["valid"] + self.data["test"] 98 | for triplet in triplets: 99 | self.filtered_dict[(triplet[0], triplet[1])].add(triplet[2]) 100 | self.filtered_dict[(triplet[2], triplet[1] + self.pos_rels)].add(triplet[0]) 101 | 102 | 103 | class MyDataset(Dataset): 104 | def __init__(self, kg, mode): 105 | self.kg = kg 106 | self.triples = kg.data[mode] 107 | 108 | def __len__(self): 109 | return len(self.triples) 110 | 111 | def __getitem__(self, index): 112 | h, r, t, s = self.triples[index] 113 | return h, r, t, s 114 | 115 | @staticmethod 116 | def collate_fn(data): 117 | query = torch.stack([torch.tensor([d[0], d[1]]) for d in data], dim=0) 118 | t = torch.stack([torch.tensor(d[2]) for d in data], dim=0) 119 | s = torch.stack([torch.tensor(d[3]) for d in data], dim=0) 120 | return query, t, s 121 | 122 | 123 | if __name__ == "__main__": 124 | # 测试代码 125 | data_path = "/data/cyl/MyPaper/MyCodes/DATASETS/CN15k" 126 | kg = KnowledgeGraph(data_path) 127 | train_set = MyDataset(kg, "train") 128 | train_dataloader = DataLoader( 129 | train_set, batch_size=5, collate_fn=MyDataset.collate_fn 130 | ) 131 | for batch in train_dataloader: 132 | print(batch) 133 | break 134 | -------------------------------------------------------------------------------- /translate_train.py: -------------------------------------------------------------------------------- 1 | """ Translate input text with trained model. """ 2 | 3 | import torch 4 | import argparse, os, time 5 | 6 | from transformer.Models import Transformer 7 | from transformer.Translator import Translator 8 | from transformer.dataset import MyDataset, KnowledgeGraph 9 | from transformer.Optim import ScheduledOptim 10 | import torch.optim as optim 11 | import numpy as np 12 | import random 13 | import wandb 14 | import os 15 | from datetime import datetime 16 | 17 | # os.environ['WANDB_DISABLED'] = 'true' 18 | 19 | 20 | def hit_mrr(hits, starttime): 21 | return ( 22 | "MRR:{:.5f} @1:{:.5f} @3:{:.5f} @10:{:.5f} LOS:{:.5f} Time:{:.1f}secs".format( 23 | hits[10] / hits[12], 24 | hits[0] / hits[12], 25 | hits[0:3].sum() / hits[12], 26 | hits[0:10].sum() / hits[12], 27 | hits[11] / hits[12], 28 | time.time() - starttime, 29 | ) 30 | ) 31 | 32 | 33 | def run( translator, data_loader, mode, optimizer, device, epoch, logfile, starttime, decode ): 34 | pred_line = [] 35 | hits = np.zeros(13) # [0:10] for hit, [10] for mrr, [11] for loss, [12] for cnt 36 | for i, (query, t, s) in enumerate(data_loader): 37 | pred_seq, loss, indexL = translator( query.to(device), t.to(device), s.to(device), mode ) 38 | if decode == True: 39 | continue 40 | print(f"\r {mode} {epoch}-{i}/{len(data_loader)}", end=" ") 41 | for j, index in enumerate(indexL): 42 | if index < 10: 43 | hits[index] += 1 44 | hits[10] += 1 / (index + 1) 45 | hits[12] += 1 46 | hits[11] += loss.item() 47 | 48 | 49 | if mode == 'train': 50 | wandb.log({f"Average_{mode}_MRR": hits[10]/hits[12]}) 51 | wandb.log({f"{mode}_loss": loss.item()}) 52 | 53 | if mode == "train": 54 | loss.backward() 55 | optimizer.step_and_update_lr() 56 | wandb.log({"lr": optimizer.lr}) 57 | optimizer.zero_grad() 58 | print(hit_mrr(hits, starttime), end=" ") 59 | 60 | if mode != 'train' and not decode: 61 | wandb.log({f"{mode}_MRR": hits[10]/hits[12]}) 62 | wandb.log({f"{mode}_loss":hits[11]/hits[12]}) 63 | if not decode: 64 | print(f"\r {mode}-{epoch} " + hit_mrr(hits, starttime) + " ") 65 | with open(logfile, "a") as log: 66 | log.write(f"{mode}-{epoch} " + hit_mrr(hits, starttime) + "\n") 67 | return pred_line 68 | 69 | 70 | def main(): 71 | seed = 42 72 | torch.manual_seed(seed) 73 | np.random.seed(seed) 74 | random.seed(seed) 75 | torch.cuda.manual_seed_all(seed) 76 | 77 | kg_path = "/data/cyl/MyPaper/No_sub_Ruleformer4UKG/DATASET/NL27K" 78 | kg = KnowledgeGraph(kg_path) 79 | train_data = MyDataset(kg, "train") 80 | valid_data = MyDataset(kg, "valid") 81 | test_data = MyDataset(kg, "test") 82 | 83 | batch_size = 64 84 | 85 | train_loader = torch.utils.data.DataLoader( dataset=train_data, batch_size=batch_size, shuffle=True, collate_fn=MyDataset.collate_fn, ) # , num_workers=16 86 | valid_loader = torch.utils.data.DataLoader( dataset=valid_data, batch_size=batch_size, shuffle=False, collate_fn=MyDataset.collate_fn, ) 87 | test_loader = torch.utils.data.DataLoader( dataset=test_data, batch_size=batch_size, shuffle=False, collate_fn=MyDataset.collate_fn, ) 88 | 89 | device = torch.device("cuda:0") 90 | 91 | d_word_vec = 384 92 | d_model = 384 93 | d_inner = 1024 94 | n_layers = 6 95 | n_head = 6 96 | d_k = 64 97 | d_v = 64 98 | dropout = 0.1 99 | n_position = 200 100 | lr_mul = 0.02 101 | n_warmup_steps = 5000 102 | 103 | model = Transformer( 104 | n_ent_vocab=len(kg.ents), 105 | n_rel_vocab=len(kg.rels), 106 | d_word_vec=d_word_vec, 107 | d_model=d_model, 108 | d_inner=d_inner, 109 | n_layers=n_layers, 110 | n_head=n_head, 111 | d_k=d_k, 112 | d_v=d_v, 113 | dropout=dropout, 114 | n_position=n_position, 115 | ) 116 | 117 | decode_rule = False 118 | translator = Translator(model=model, body_len=3, device=device, kg=kg).to(device) 119 | optimizer = ScheduledOptim( optim.Adam(translator.parameters(), betas=(0.9, 0.98), eps=1e-09), lr_mul, d_model, n_warmup_steps, ) 120 | 121 | starttime = time.time() 122 | 123 | num_epoch = 200 124 | 125 | savestep = 5 126 | 127 | wandb.init( 128 | # set the wandb project where this run will be logged 129 | project='MyModel', 130 | # track hyperparameters and run metadata 131 | config={ 132 | "batch_size": batch_size, 133 | "epochs": num_epoch, 134 | }, 135 | ) 136 | 137 | logfile = datetime.now().strftime("%Y%m%d_%H%M%S") 138 | if not os.path.exists(f'EXPS/{logfile}'): 139 | os.mkdir(f'EXPS/{logfile}') 140 | 141 | 142 | config = { 143 | "data_path":kg_path, 144 | "batch_size":batch_size, 145 | "d_word_vec": d_word_vec, 146 | "d_model": d_model, 147 | "d_inner": d_inner, 148 | "n_layers": n_layers, 149 | "n_head": n_head, 150 | "d_k": d_k, 151 | "d_v": d_v, 152 | "dropout": dropout, 153 | "n_position": n_position, 154 | "lr_mul":lr_mul, 155 | "n_warmup_steps":n_warmup_steps, 156 | "num_epoch":num_epoch, 157 | "save_step":5, 158 | "decode_rule": decode_rule 159 | } 160 | 161 | # 指定要写入的文件路径 162 | file_path = f'EXPS/{logfile}/config.txt' 163 | 164 | # 将参数写入文件 165 | with open(file_path, "w") as file: 166 | for key, value in config.items(): 167 | file.write(f"{key} = {value}\n") 168 | 169 | 170 | for epoch in range(num_epoch): 171 | if not decode_rule: 172 | run( translator, train_loader, "train", optimizer, device, epoch + 1, logfile, starttime, decode_rule ) 173 | if savestep and (epoch + 1) % savestep == 0: 174 | torch.save(translator.state_dict(), f"EXPS/{logfile}/Translator{epoch+1}.ckpt") 175 | if decode_rule or (epoch + 1) % 5 == 0: 176 | with torch.no_grad(): 177 | if decode_rule: 178 | run( translator, train_loader, "train", optimizer, device, epoch + 1, logfile, starttime, decode_rule ) 179 | run( translator, valid_loader, "valid", optimizer, device, epoch + 1, logfile, starttime, decode_rule ) 180 | run( translator, test_loader, "test", optimizer, device, epoch + 1, logfile, starttime, decode_rule ) 181 | 182 | print("[Info] Finished.") 183 | 184 | 185 | if __name__ == "__main__": 186 | main() 187 | -------------------------------------------------------------------------------- /translate_decode.py: -------------------------------------------------------------------------------- 1 | """ Translate input text with trained model. """ 2 | 3 | import torch 4 | import argparse, os, time 5 | 6 | from transformer.Models import Transformer 7 | from transformer.Translator import Translator 8 | from transformer.dataset import MyDataset, KnowledgeGraph 9 | from transformer.Optim import ScheduledOptim 10 | import torch.optim as optim 11 | import numpy as np 12 | import random 13 | import wandb 14 | import os 15 | from datetime import datetime 16 | 17 | 18 | def hit_mrr(hits, starttime): 19 | return ( 20 | "MRR:{:.5f} @1:{:.5f} @3:{:.5f} @10:{:.5f} LOS:{:.5f} Time:{:.1f}secs".format( 21 | hits[10] / hits[12], 22 | hits[0] / hits[12], 23 | hits[0:3].sum() / hits[12], 24 | hits[0:10].sum() / hits[12], 25 | hits[11] / hits[12], 26 | time.time() - starttime, 27 | ) 28 | ) 29 | 30 | 31 | def run( translator, data_loader, mode, optimizer, device, epoch, logfile, starttime, decode ): 32 | pred_line = [] 33 | hits = np.zeros(13) # [0:10] for hit, [10] for mrr, [11] for loss, [12] for cnt 34 | for i, (query, t, s) in enumerate(data_loader): 35 | pred_seq, loss, indexL = translator( query.to(device), t.to(device), s.to(device), mode ) 36 | if decode == True: 37 | continue 38 | print(f"\r {mode} {epoch}-{i}/{len(data_loader)}", end=" ") 39 | for j, index in enumerate(indexL): 40 | if index < 10: 41 | hits[index] += 1 42 | hits[10] += 1 / (index + 1) 43 | hits[12] += 1 44 | hits[11] += loss.item() 45 | 46 | 47 | if mode == 'train': 48 | wandb.log({f"{mode}_MRR": hits[10]/hits[12]}) 49 | wandb.log({f"{mode}_loss": loss.item()}) 50 | 51 | if mode == "train": 52 | loss.backward() 53 | optimizer.step_and_update_lr() 54 | wandb.log({"lr": optimizer.lr}) 55 | optimizer.zero_grad() 56 | print(hit_mrr(hits, starttime), end=" ") 57 | 58 | if mode != 'train' and not decode: 59 | wandb.log({f"{mode}_MRR": hits[10]/hits[12]}) 60 | wandb.log({f"{mode}_loss":hits[11]/hits[12]}) 61 | if not decode: 62 | print(f"\r {mode}-{epoch} " + hit_mrr(hits, starttime) + " ") 63 | with open(logfile, "a") as log: 64 | log.write(f"{mode}-{epoch} " + hit_mrr(hits, starttime) + "\n") 65 | return pred_line 66 | 67 | 68 | def main(): 69 | seed = 42 70 | torch.manual_seed(seed) 71 | np.random.seed(seed) 72 | random.seed(seed) 73 | torch.cuda.manual_seed_all(seed) 74 | 75 | kg_path = "/data/cyl/MyPaper/No_sub_Ruleformer4UKG/DATASET/CN15k" 76 | kg = KnowledgeGraph(kg_path) 77 | train_data = MyDataset(kg, "train") 78 | valid_data = MyDataset(kg, "valid") 79 | test_data = MyDataset(kg, "test") 80 | 81 | batch_size = 512 82 | 83 | train_loader = torch.utils.data.DataLoader( dataset=train_data, batch_size=batch_size, shuffle=False, collate_fn=MyDataset.collate_fn, ) # , num_workers=16 84 | valid_loader = torch.utils.data.DataLoader( dataset=valid_data, batch_size=batch_size, shuffle=False, collate_fn=MyDataset.collate_fn, ) 85 | test_loader = torch.utils.data.DataLoader( dataset=test_data, batch_size=batch_size, shuffle=False, collate_fn=MyDataset.collate_fn, ) 86 | 87 | device = torch.device("cpu") 88 | 89 | d_word_vec = 384 90 | d_model = 384 91 | d_inner = 1024 92 | n_layers = 4 93 | n_head = 6 94 | d_k = 64 95 | d_v = 64 96 | dropout = 0.1 97 | n_position = 200 98 | lr_mul = 0.02 99 | n_warmup_steps = 5000 100 | 101 | model = Transformer( 102 | n_ent_vocab=len(kg.ents), 103 | n_rel_vocab=len(kg.rels), 104 | d_word_vec=d_word_vec, 105 | d_model=d_model, 106 | d_inner=d_inner, 107 | n_layers=n_layers, 108 | n_head=n_head, 109 | d_k=d_k, 110 | d_v=d_v, 111 | dropout=dropout, 112 | n_position=n_position, 113 | ) 114 | 115 | decode_rule = True 116 | translator = Translator(model=model, body_len=3, device=device, kg=kg, decode = decode_rule).to(device) 117 | ckpt = '/data/cyl/MyPaper/No_sub_Ruleformer4UKG/EXPS/20240208_000822_CN15K_un/Translator40.ckpt' 118 | if ckpt and decode_rule: 119 | print("CKPT APPOINTED!!! DECODE MODE!!!") 120 | decode_rule = True 121 | translator.load_state_dict(torch.load(ckpt)) 122 | 123 | optimizer = ScheduledOptim( optim.Adam(translator.parameters(), betas=(0.9, 0.98), eps=1e-09), lr_mul, d_model, n_warmup_steps, ) 124 | 125 | starttime = time.time() 126 | 127 | num_epoch = 1 128 | 129 | savestep = 0 130 | 131 | if not decode_rule: 132 | wandb.init( 133 | # set the wandb project where this run will be logged 134 | project='MyModel', 135 | # track hyperparameters and run metadata 136 | config={ 137 | "batch_size": batch_size, 138 | "epochs": num_epoch, 139 | }, 140 | ) 141 | 142 | logfile = datetime.now().strftime("%Y%m%d_%H%M%S") 143 | if not os.path.exists(f'EXPS/{logfile}_decode'): 144 | os.mkdir(f'EXPS/{logfile}_decode') 145 | 146 | 147 | config = { 148 | "data_path":kg_path, 149 | "batch_size":batch_size, 150 | "d_word_vec": d_word_vec, 151 | "d_model": d_model, 152 | "d_inner": d_inner, 153 | "n_layers": n_layers, 154 | "n_head": n_head, 155 | "d_k": d_k, 156 | "d_v": d_v, 157 | "dropout": dropout, 158 | "n_position": n_position, 159 | "lr_mul":lr_mul, 160 | "n_warmup_steps":n_warmup_steps, 161 | "num_epoch":num_epoch, 162 | "save_step":5, 163 | "decode_rule": decode_rule 164 | } 165 | 166 | # 指定要写入的文件路径 167 | file_path = f'EXPS/{logfile}_decode/config.txt' 168 | 169 | # 将参数写入文件 170 | with open(file_path, "w") as file: 171 | for key, value in config.items(): 172 | file.write(f"{key} = {value}\n") 173 | 174 | 175 | for epoch in range(num_epoch): 176 | if not decode_rule: 177 | run( translator, train_loader, "train", optimizer, device, epoch + 1, logfile, starttime, decode_rule ) 178 | if savestep and (epoch + 1) % savestep == 0: 179 | torch.save(translator.state_dict(), f"EXPS/{logfile}/Translator{epoch+1}.ckpt") 180 | if decode_rule or (epoch + 1) % 5 == 0: 181 | with torch.no_grad(): 182 | if decode_rule: 183 | run( translator, train_loader, "train", optimizer, device, epoch + 1, logfile, starttime, decode_rule ) 184 | run( translator, valid_loader, "valid", optimizer, device, epoch + 1, logfile, starttime, decode_rule ) 185 | run( translator, test_loader, "test", optimizer, device, epoch + 1, logfile, starttime, decode_rule ) 186 | 187 | print("[Info] Finished.") 188 | 189 | 190 | if __name__ == "__main__": 191 | main() 192 | -------------------------------------------------------------------------------- /confidence_prediction.py: -------------------------------------------------------------------------------- 1 | from transformers import AutoModel, AutoTokenizer 2 | import torch 3 | import torch.nn as nn 4 | from torch.utils.data import Dataset, DataLoader 5 | from transformers import AdamW, get_linear_schedule_with_warmup 6 | import numpy as np 7 | import random 8 | import wandb 9 | 10 | 11 | class PathDataset(Dataset): 12 | def __init__(self, data_path, model_path, max_length): 13 | self.max_legnth = max_length 14 | self.tokenizer = AutoTokenizer.from_pretrained(model_path) 15 | self.lines = open(data_path, "r").readlines() 16 | 17 | def __len__(self): 18 | return len(self.lines) 19 | 20 | def __getitem__(self, index): 21 | score, question, context = self.lines[index].strip("\n").split("\t") 22 | input_txt = f"[CLS] {question} [SEP] {context} [SEP]" 23 | encoding = self.tokenizer.encode_plus( 24 | input_txt, 25 | max_length=self.max_legnth, 26 | add_special_tokens=False, # [CLS]和[SEP] 27 | return_token_type_ids=True, 28 | pad_to_max_length=True, 29 | return_attention_mask=True, 30 | return_tensors="pt", # Pytorch tensor张量 31 | ) 32 | 33 | return { 34 | "input_txt": input_txt, 35 | "input_ids": encoding["input_ids"].flatten(), 36 | "attention_mask": encoding["attention_mask"].flatten(), 37 | "toeken_type_ids": encoding["token_type_ids"].flatten(), 38 | "labels": torch.tensor(eval(score), dtype=torch.float), 39 | } 40 | 41 | 42 | class ConfModel(nn.Module): 43 | def __init__(self, model_path, device): 44 | super().__init__() 45 | self.bert = AutoModel.from_pretrained(model_path) 46 | self.d_model = self.bert.pooler.dense.out_features 47 | self.linear1 = nn.Linear(self.d_model, self.d_model*4) 48 | self.relu = nn.ReLU() 49 | self.linear2 = nn.Linear(self.d_model*4,1) 50 | self.sigmoid = nn.Sigmoid() 51 | self.device = device 52 | self.w_q = nn.Linear(self.d_model, self.d_model) 53 | self.w_k = nn.Linear(self.d_model, self.d_model) 54 | self.w_v = nn.Linear(self.d_model, self.d_model) 55 | self.w_o = nn.Linear(self.d_model, self.d_model) 56 | 57 | def self_attn_encode(self, q, k, v, mask): 58 | # q: batch_size,1,d_model 59 | # k: batch_size,seq_len,d_model -> batch_size,d_model,seq_len 60 | # v: batch_size,seq_len,d_model 61 | q = self.w_q(q) 62 | k = self.w_k(k).transpose(1, 2) 63 | v = self.w_v(v) 64 | 65 | # attn_logit: batch_size,1,seq_len 66 | attn_logit = torch.matmul(q, k) 67 | if mask != None: 68 | attn_logit = attn_logit.masked_fill(mask == 0, -1e9) 69 | attn_weight = torch.softmax(attn_logit, dim=-1) 70 | # batch_size,1,seq_len * batch_size,seq_len,d_model -> batch_size,1,d_model 71 | weight_sum = torch.matmul(attn_weight, v) 72 | output = self.w_o(weight_sum) 73 | return output 74 | 75 | def forward(self, x, mode): 76 | input_ids = x["input_ids"].to(self.device) 77 | attn_mask = x["attention_mask"].to(self.device) 78 | 79 | output = self.bert(input_ids=input_ids, attention_mask=attn_mask) 80 | 81 | last_hidden_state = output[ 82 | "last_hidden_state" 83 | ] # batch_size*max_token_length*d_model 84 | pooler_output = output["pooler_output"] # batch_size,d_model 85 | 86 | if mode == "CLS": 87 | score = pooler_output 88 | elif mode == "avg": 89 | extended_attn_mask = ( 90 | attn_mask.unsqueeze(-1).expand_as(last_hidden_state).float() 91 | ) 92 | last_hidden_state_masked = last_hidden_state * extended_attn_mask 93 | sum_embeddings = torch.sum(last_hidden_state_masked, dim=1) 94 | sum_mask = torch.sum(extended_attn_mask, dim=1) 95 | mean_pooled = sum_embeddings / sum_mask 96 | score = mean_pooled 97 | elif mode == "attn": 98 | output = self.self_attn_encode( 99 | last_hidden_state[:, 0, :].unsqueeze(1), 100 | last_hidden_state, 101 | last_hidden_state, 102 | attn_mask.unsqueeze(1), 103 | ) 104 | score = output 105 | score = self.linear2(self.relu(self.linear1(score))).flatten() 106 | return score 107 | 108 | 109 | def train_model(model, dataloader, criterion, optimizer, scheduler, device, strategy): 110 | model = model.to(device).train() 111 | criterion = criterion.to(device) 112 | 113 | for i, batch in enumerate(dataloader): 114 | # if i < 3: 115 | # print(batch) 116 | input_txt = batch["input_txt"] 117 | labels = batch["labels"].to(device) 118 | output = model(batch, strategy) 119 | 120 | loss = criterion(output, labels) 121 | 122 | if i % 5 == 0: 123 | with torch.no_grad(): 124 | diff = output - labels 125 | abs_diff = torch.abs(diff) 126 | mae = torch.mean(abs_diff) 127 | print(f"{i}\t MSE:{loss.item()}\t MAE:{mae}") 128 | loss.backward() 129 | nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0) 130 | optimizer.step() 131 | scheduler.step() 132 | optimizer.zero_grad() 133 | 134 | 135 | def test_model(model, dataloader, criterion, device, mode, strategy): 136 | with torch.no_grad(): 137 | output_list = None 138 | label_list = None 139 | model = model.to(device).eval() 140 | for i, batch in enumerate(dataloader): 141 | input_txt = batch["input_txt"] 142 | labels = batch["labels"].to(device) 143 | output = model(batch, strategy) 144 | if output_list == None: 145 | output_list = output 146 | else: 147 | output_list = torch.cat((output_list, output), dim=0) 148 | if label_list == None: 149 | label_list = labels 150 | else: 151 | label_list = torch.cat((label_list, labels), dim=0) 152 | 153 | diff = output_list - label_list 154 | abs_diff = torch.abs(diff) 155 | mae = torch.mean(abs_diff) 156 | mse = criterion(output_list, label_list).item() 157 | print(f"{mode} \t MSE: {mse}\t MAE: {mae}") 158 | 159 | 160 | def main(): 161 | # wandb.init() 162 | EXP_name = '' 163 | data_path = "/data/cyl/MyPaper/No_sub_Ruleformer4UKG/decode_rules/dataset/NL27K/" 164 | model_path = "/data/cyl/MyPaper/No_sub_Ruleformer4UKG/bert-base-uncased" 165 | strategy = "attn" 166 | 167 | batch_size = 256 168 | max_length = 90 169 | train_dataset = PathDataset(data_path + "nl27k_path_train.txt", model_path, max_length) 170 | valid_dataset = PathDataset(data_path + "nl27k_path_valid.txt", model_path, max_length) 171 | test_dataset = PathDataset(data_path + "nl27k_path_test.txt", model_path, max_length) 172 | 173 | train_dataloader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True) 174 | valid_dataloader = DataLoader(valid_dataset, batch_size=batch_size, shuffle=False) 175 | test_dataloader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False) 176 | 177 | device = torch.device("cuda:2") 178 | 179 | myModel = ConfModel(model_path, device) 180 | 181 | EPOCHS = 100 182 | criterion = torch.nn.MSELoss() 183 | optimizer = AdamW(myModel.parameters(), lr=2e-5, correct_bias=False) 184 | total_steps = len(train_dataloader) * EPOCHS 185 | scheduler = get_linear_schedule_with_warmup( 186 | optimizer, num_warmup_steps=total_steps // 10, num_training_steps=total_steps 187 | ) 188 | for e in range(EPOCHS): 189 | train_model(myModel, train_dataloader, criterion, optimizer, scheduler, device, strategy) 190 | # torch.save(myModel.state_dict(), f"{EXP_name}/model_{e}.pt") 191 | test_model(myModel, valid_dataloader, criterion, device, "valid", strategy) 192 | test_model(myModel, test_dataloader, criterion, device, "test", strategy) 193 | 194 | 195 | if __name__ == "__main__": 196 | seed = 42 197 | torch.manual_seed(seed) 198 | np.random.seed(seed) 199 | random.seed(seed) 200 | torch.cuda.manual_seed_all(seed) 201 | main() 202 | -------------------------------------------------------------------------------- /transformer/Models.py: -------------------------------------------------------------------------------- 1 | ''' Define the Transformer model ''' 2 | import torch, math 3 | import torch.nn as nn 4 | import numpy as np 5 | from transformer.Layers import EncoderLayer, DecoderLayer 6 | 7 | 8 | def get_pad_mask(seq, pad_idx): 9 | return (seq != pad_idx).unsqueeze(-2) 10 | 11 | 12 | def get_subsequent_mask(seq): 13 | ''' For masking out the subsequent info. ''' 14 | sz_b, len_s = seq.size() 15 | subsequent_mask = (1 - torch.triu( 16 | torch.ones((1, len_s, len_s), device=seq.device), diagonal=1)).bool() 17 | return subsequent_mask 18 | 19 | 20 | class PositionalEncoding(nn.Module): 21 | 22 | def __init__(self, d_hid, n_position=200): 23 | super(PositionalEncoding, self).__init__() 24 | # Not a parameter 25 | self.register_buffer('pos_table', self._get_sinusoid_encoding_table(n_position, d_hid)) 26 | 27 | def _get_sinusoid_encoding_table(self, n_position, d_hid): 28 | ''' Sinusoid position encoding table ''' 29 | # TODO: make it with torch instead of numpy 30 | 31 | def get_position_angle_vec(position): 32 | return [position / np.power(10000, 2 * (hid_j // 2) / d_hid) for hid_j in range(d_hid)] 33 | 34 | sinusoid_table = np.array([get_position_angle_vec(pos_i) for pos_i in range(n_position)]) 35 | sinusoid_table[:, 0::2] = np.sin(sinusoid_table[:, 0::2]) # dim 2i 36 | sinusoid_table[:, 1::2] = np.cos(sinusoid_table[:, 1::2]) # dim 2i+1 37 | 38 | return torch.FloatTensor(sinusoid_table).unsqueeze(0) 39 | 40 | def forward(self, x): 41 | return x + self.pos_table[:, :x.size(1)].clone().detach() 42 | 43 | 44 | class Encoder(nn.Module): 45 | ''' A encoder model with self attention mechanism. ''' 46 | 47 | def __init__(self, d_word_vec, n_layers, n_head, d_k, d_v, d_model, d_inner, dropout=0.1, 48 | n_position=200, scale_emb=False, ent_word_emb=None, rel_word_emb=None): 49 | 50 | super().__init__() 51 | self.ent_word_emb = ent_word_emb 52 | self.rel_word_emb = rel_word_emb 53 | self.position_enc = PositionalEncoding(d_word_vec, n_position=n_position) 54 | self.dropout = nn.Dropout(p=dropout) 55 | self.layer_stack = nn.ModuleList([ EncoderLayer(d_model, d_inner, n_head, d_k, d_v, dropout=dropout) for _ in range(n_layers)]) 56 | self.layer_norm = nn.LayerNorm(d_model, eps=1e-6) 57 | self.scale_emb = scale_emb 58 | self.d_model = d_model 59 | 60 | def forward(self, src_seq, src_mask, return_attns=False): 61 | 62 | enc_slf_attn_list = [] 63 | 64 | # -- Forward 65 | # enc_output = self.src_word_emb(src_seq) 66 | # src_seq: [batch_size,2] 67 | head_emb = self.ent_word_emb(src_seq[:,0]) 68 | rel_emb = self.rel_word_emb(src_seq[:,1]) 69 | enc_output = torch.stack((head_emb,rel_emb),dim=1) 70 | 71 | if self.scale_emb: 72 | enc_output *= self.d_model ** 0.5 73 | enc_output = self.dropout(self.position_enc(enc_output)) 74 | enc_output = self.layer_norm(enc_output) 75 | 76 | for enc_layer in self.layer_stack: 77 | enc_output, enc_slf_attn = enc_layer(enc_output, slf_attn_mask=src_mask) 78 | enc_slf_attn_list += [enc_slf_attn] if return_attns else [] 79 | 80 | if return_attns: 81 | return enc_output, enc_slf_attn_list 82 | return enc_output, 83 | 84 | 85 | class Decoder(nn.Module): 86 | ''' A decoder model with self attention mechanism. ''' 87 | 88 | def __init__( 89 | self, d_word_vec, n_layers, n_head, d_k, d_v, 90 | d_model, d_inner, n_position=200, dropout=0.1, scale_emb=False, rel_word_emb=None): 91 | 92 | super().__init__() 93 | 94 | self.rel_word_emb = rel_word_emb 95 | self.bos_embedding = nn.Embedding(1,d_model) 96 | self.position_enc = PositionalEncoding(d_word_vec, n_position=n_position) 97 | self.dropout = nn.Dropout(p=dropout) 98 | self.layer_stack = nn.ModuleList([ DecoderLayer(d_model, d_inner, n_head, d_k, d_v, dropout=dropout) for _ in range(n_layers)]) 99 | self.layer_norm = nn.LayerNorm(d_model, eps=1e-6) 100 | self.scale_emb = scale_emb 101 | self.d_model = d_model 102 | 103 | self.bos_embedding = torch.nn.Embedding(1,d_model) 104 | 105 | def forward(self, trg_seq, trg_mask, enc_output, src_mask, return_attns=False): 106 | 107 | dec_slf_attn_list, dec_enc_attn_list = [], [] 108 | 109 | # -- Forward 110 | if trg_seq.size(1) == 1: 111 | dec_output = self.rel_word_emb(trg_seq) 112 | else: 113 | bos_embedding = self.bos_embedding(trg_seq[:,0]).unsqueeze(1) 114 | rel_embedding = self.rel_word_emb(trg_seq[:,1:]) 115 | dec_output = torch.cat((bos_embedding,rel_embedding),dim=1) 116 | if self.scale_emb: 117 | dec_output *= self.d_model ** 0.5 118 | dec_output = self.dropout(self.position_enc(dec_output)) 119 | dec_output = self.layer_norm(dec_output) 120 | 121 | for dec_layer in self.layer_stack: 122 | dec_output, dec_slf_attn, dec_enc_attn = dec_layer( 123 | dec_output, enc_output, slf_attn_mask=trg_mask, dec_enc_attn_mask=src_mask) 124 | dec_slf_attn_list += [dec_slf_attn] if return_attns else [] 125 | dec_enc_attn_list += [dec_enc_attn] if return_attns else [] 126 | 127 | if return_attns: 128 | return dec_output, dec_slf_attn_list, dec_enc_attn_list 129 | return dec_output, 130 | 131 | 132 | class Transformer(nn.Module): 133 | ''' A sequence to sequence model with attention mechanism. ''' 134 | 135 | def __init__( 136 | self, n_ent_vocab, n_rel_vocab, 137 | d_word_vec=512, d_model=512, d_inner=2048, 138 | n_layers=6, n_head=8, d_k=64, d_v=64, 139 | dropout=0.1, n_position=200): 140 | 141 | super().__init__() 142 | self.relationE = nn.Embedding(n_rel_vocab, d_model) 143 | self.entityE = nn.Embedding(n_ent_vocab, d_model) 144 | 145 | # self.src_pad_idx, self.trg_pad_idx = src_pad_idx, trg_pad_idx 146 | 147 | assert n_head*d_k == d_model 148 | # In section 3.4 of paper "Attention Is All You Need", there is such detail: 149 | # "In our model, we share the same weight matrix between the two 150 | # embedding layers and the pre-softmax linear transformation... 151 | # In the embedding layers, we multiply those weights by \sqrt{d_model}". 152 | # 153 | # Options here: 154 | # 'emb': multiply \sqrt{d_model} to embedding output 155 | # 'prj': multiply (\sqrt{d_model} ^ -1) to linear projection output 156 | # 'none': no multiplication 157 | 158 | # assert scale_emb_or_prj in ['emb', 'prj', 'none'] 159 | # scale_emb = (scale_emb_or_prj == 'emb') if trg_emb_prj_weight_sharing else False 160 | scale_emb = False 161 | # self.scale_prj = (scale_emb_or_prj == 'prj') if trg_emb_prj_weight_sharing else False 162 | self.d_model = d_model 163 | 164 | self.encoder = Encoder( 165 | n_position=n_position, 166 | d_word_vec=d_word_vec, d_model=d_model, d_inner=d_inner, 167 | n_layers=n_layers, n_head=n_head, d_k=d_k, d_v=d_v, 168 | dropout=dropout, scale_emb=scale_emb, 169 | ent_word_emb=self.entityE, rel_word_emb=self.relationE) 170 | 171 | self.decoder = Decoder( 172 | n_position=n_position, 173 | d_word_vec=d_word_vec, d_model=d_model, d_inner=d_inner, 174 | n_layers=n_layers, n_head=n_head, d_k=d_k, d_v=d_v, 175 | dropout=dropout, scale_emb=scale_emb, rel_word_emb=self.relationE) 176 | 177 | self.trg_word_prj = nn.Linear(d_model, n_rel_vocab, bias=False) 178 | 179 | for n,p in self.named_parameters(): 180 | if p.dim() > 1 and 'entityE' not in n and 'relationE' not in n: 181 | nn.init.xavier_uniform_(p) 182 | 183 | assert d_model == d_word_vec, \ 184 | 'To facilitate the residual connections, \ 185 | the dimensions of all module outputs shall be the same.' 186 | 187 | # if trg_emb_prj_weight_sharing: 188 | # # Share the weight between target word embedding & last dense layer 189 | # self.trg_word_prj.weight = self.decoder.trg_word_emb.weight 190 | 191 | # if emb_src_trg_weight_sharing: 192 | # self.encoder.src_word_emb.weight = self.decoder.trg_word_emb.weight 193 | 194 | 195 | def forward(self, src_seq, trg_seq): 196 | raise NotImplementedError 197 | 198 | src_mask = get_pad_mask(src_seq, self.src_pad_idx) 199 | trg_mask = get_pad_mask(trg_seq, self.trg_pad_idx) & get_subsequent_mask(trg_seq) 200 | 201 | enc_output, *_ = self.encoder(src_seq, src_mask) 202 | dec_output, *_ = self.decoder(trg_seq, trg_mask, enc_output, src_mask) 203 | seq_logit = self.trg_word_prj(dec_output) 204 | if self.scale_prj: 205 | seq_logit *= self.d_model ** -0.5 206 | 207 | return seq_logit.view(-1, seq_logit.size(2)) 208 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright [yyyy] [name of copyright owner] 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. 202 | -------------------------------------------------------------------------------- /transformer/Translator.py: -------------------------------------------------------------------------------- 1 | """ 2 | 修改了新的解码方式，用邻接矩阵判断解码的关系是否相连，而不是用循环去判断 3 | """ 4 | """ This module will handle the text generation with beam search. """ 5 | import torch, math 6 | import torch.nn as nn 7 | import torch.nn.functional as F 8 | from transformer.Models import Transformer, get_pad_mask, get_subsequent_mask 9 | from collections import defaultdict 10 | import numpy, time 11 | from itertools import product 12 | from datetime import datetime 13 | 14 | 15 | class Translator(nn.Module): 16 | """Load a trained model and translate in beam search fashion.""" 17 | 18 | def __init__(self, model, body_len, device=None, kg=None, decode=False): 19 | super(Translator, self).__init__() 20 | 21 | self.alpha = 0.7 22 | self.body_len = body_len 23 | 24 | self.n_rel_vocab = len(kg.rels) 25 | self.n_ent_vocab = len(kg.ents) 26 | # self.src_pad_idx = opt.src_pad_idx 27 | # self.n_trg_vocab = opt.trg_vocab_size 28 | # self.n_src_vocab = opt.src_vocab_size 29 | 30 | self.device = device 31 | 32 | self.model = model 33 | self.model.train() 34 | self.database = [x.to(device) for x in kg.relations] 35 | self.filtered_dict = kg.filtered_dict 36 | 37 | self.decode = decode 38 | if self.decode: 39 | self.graph = kg.neighbors 40 | self.decode_rule_num = 0 41 | self.the_rel = 0 42 | self.the_rel_min = 0 43 | self.the_all = 0 44 | self.id2r = kg.id2r 45 | self.id2e = kg.id2e 46 | self.rules = defaultdict(dict) 47 | self.decode_rule_num_filter = 0 48 | self.decode_file = 'cn15k_rules_0221.txt' 49 | 50 | self.register_buffer("thr", torch.FloatTensor([1e-20])) 51 | 52 | def _model_decode(self, trg_seq, enc_output, src_mask): 53 | trg_mask = None 54 | dec_output, *_ = self.model.decoder(trg_seq, trg_mask, enc_output, src_mask) 55 | return F.softmax(self.model.trg_word_prj(dec_output), dim=-1) 56 | 57 | def _get_init_state(self, src_seq, src_mask): 58 | enc_output = None 59 | enc_output, *_ = self.model.encoder(src_seq, src_mask) 60 | dec_output = self._model_decode(self.init_seq, enc_output, src_mask) 61 | 62 | best_k_probs, best_k_idx = dec_output[:, -1, :].topk(1) 63 | 64 | scores = torch.log(best_k_probs).view(self.batch_size) 65 | gen_seq = self.blank_seqs.clone().detach() 66 | gen_seq[:, 1] = best_k_idx.squeeze() 67 | return enc_output, gen_seq, scores, dec_output 68 | 69 | def _get_the_best_score_and_idx(self, gen_seq, dec_output, scores, step): 70 | # Get k candidates for each beam, k^2 candidates in total. 71 | best_k2_probs, best_k2_idx = dec_output[:, -1, :].topk(1) 72 | assert dec_output.size(1) == step 73 | 74 | gen_seq[:, step] = best_k2_idx.squeeze() 75 | 76 | return gen_seq, scores, dec_output 77 | 78 | def forwardAllNLP(self, query, t, s, attention, mode): 79 | num_pos_rel = self.n_rel_vocab // 2 80 | batch_size = query.size(0) 81 | database = [x.clone().detach() for x in self.database] 82 | for (hh, rr), tt in zip(query, t): 83 | if rr >= num_pos_rel: 84 | continue 85 | idxs = torch.where( (hh == database[rr].indices()[0]) & (tt == database[rr].indices()[1]) ) 86 | database[rr].values()[idxs] = 0 87 | 88 | memories = F.one_hot(query[:,0], num_classes=self.n_ent_vocab).float().to_sparse() 89 | for step in range(self.body_len): 90 | added_results = torch.zeros(batch_size, self.n_ent_vocab).to(self.device) 91 | for r in range(num_pos_rel): 92 | for links, atta in zip( [database[r], database[r].transpose(0, 1)], [attention[:, step, r], attention[:, step, r + num_pos_rel]], ): 93 | added_results = added_results + torch.sparse.mm( memories, links ).to_dense() * atta.unsqueeze(1) 94 | added_results = added_results + memories.to_dense() * attention[ :, step, -1 ].unsqueeze(1) 95 | # added_results = added_results / torch.max(self.thr, torch.sum(added_results, dim=1).unsqueeze(1)) 96 | memories = added_results.to_sparse() 97 | memories = memories.to_dense() 98 | memories = memories / torch.max( self.thr, torch.sum(memories, dim=1).unsqueeze(1) ) 99 | targets = F.one_hot(t, num_classes=self.n_ent_vocab).float() 100 | final_loss = -torch.sum( targets * torch.log(torch.max(self.thr, memories)), dim=1 ) 101 | batch_loss = torch.mean(final_loss) 102 | if mode != "train": 103 | for i in range(batch_size): 104 | for idx in self.filtered_dict[(query[i,0].item(), query[i,1].item())]: 105 | if idx != t[i].item(): 106 | memories[i, idx] = 0 107 | idxs = torch.argsort(memories, descending=True) 108 | indexs = torch.where(idxs == t.unsqueeze(-1))[1].tolist() 109 | return batch_loss, indexs 110 | 111 | def forward(self, query, t, s, mode="train"): 112 | batch_size = query.size(0) 113 | self.batch_size = batch_size 114 | # self.init_seq = trg[:, 1].unsqueeze(-1).clone().detach() 115 | self.init_seq = torch.zeros((batch_size, 1),dtype=torch.long).to(self.device) # decoder输入 116 | self.blank_seqs = torch.zeros((batch_size,self.body_len+1),dtype=torch.long).detach().to(self.device) # 存放解码序列，不参与梯度计算 117 | # src_mask = get_pad_mask(src_seq, self.src_pad_idx) 118 | src_mask = None 119 | enc_output, gen_seq, scores, dec_output = self._get_init_state(query, src_mask) 120 | for step in range(2, self.body_len + 1): 121 | dec_output = self._model_decode( gen_seq[:, :step].clone().detach(), enc_output, src_mask ) 122 | gen_seq, scores, dec_output = self._get_the_best_score_and_idx( gen_seq, dec_output, scores, step ) 123 | 124 | gen_seq[:,0] = query[:,1] 125 | if self.decode: 126 | self.decode_rule(dec_output, query, t, s, mode) 127 | loss, index = 0, [0] 128 | else: 129 | loss, index = self.forwardAllNLP(query, t, s, dec_output, mode) 130 | 131 | return gen_seq, loss, index 132 | 133 | 134 | def decode_rule(self, attention, query, t, s, mode): 135 | def dfs(database,memories,rela_attention,step,body_len,tail,path): 136 | if step == body_len: 137 | return tail 138 | to_add = None 139 | step_attention = rela_attention[step] 140 | added_results = torch.zeros(1, self.n_ent_vocab).to(self.device) 141 | for r in range(self.n_rel_vocab): 142 | if r < self.n_rel_vocab//2: 143 | if to_add == None: 144 | to_add = torch.sparse.mm(memories,database[r]).to_dense()*step_attention[r] 145 | else: 146 | to_add = torch.cat([to_add,torch.sparse.mm(memories,database[r]).to_dense()*step_attention[r]]) 147 | elif r < self.n_rel_vocab-1: 148 | if to_add == None: 149 | to_add = torch.sparse.mm(memories,database[r-self.n_rel_vocab//2].transpose(0,1)).to_dense()*step_attention[r] 150 | else: 151 | to_add = torch.cat([to_add,torch.sparse.mm(memories,database[r-self.n_rel_vocab//2].transpose(0,1)).to_dense()*step_attention[r]]) 152 | else: 153 | if to_add == None: 154 | to_add = memories.to_dense()*step_attention[r] 155 | else: 156 | to_add = torch.cat([to_add,memories.to_dense()*step_attention[r]]) 157 | added_results = added_results + torch.sum(to_add,dim=0) 158 | 159 | added_results = added_results.to_dense() / torch.max(self.thr, torch.sum(added_results, dim=1).unsqueeze(1)) 160 | memories = added_results.to_sparse() 161 | 162 | target = dfs(database,memories,rela_attention,step+1,body_len,tail,path) 163 | path.append(target) 164 | 165 | rel_idx = torch.argmax(to_add[:,target]) 166 | path.append(rel_idx) 167 | 168 | if step == 0: 169 | return 170 | 171 | if rel_idx < self.n_rel_vocab//2: 172 | adj_matrix = database[rel_idx].to_dense() 173 | elif rel_idx < self.n_rel_vocab-1: 174 | adj_matrix = database[rel_idx-self.n_rel_vocab//2].transpose(0,1).to_dense() 175 | else: 176 | adj_matrix = torch.eye(self.n_ent_vocab) 177 | 178 | adj_vec = adj_matrix[:,target].to(self.device) 179 | dots = memories*adj_vec 180 | return torch.argmax(dots.to_dense()) 181 | 182 | 183 | 184 | num_pos_rel = self.n_rel_vocab // 2 185 | batch_size = query.size(0) 186 | database = [x.clone().detach() for x in self.database] 187 | 188 | for batch in range(batch_size): 189 | head_id = query[batch,0] 190 | rela_id = query[batch,1] 191 | tail_id = t[batch] 192 | conf = s[batch] 193 | 194 | rela_attention = attention[batch] 195 | 196 | idxs = torch.where((head_id == database[rela_id].indices()[0]) & (tail_id == database[rela_id].indices()[1])) 197 | database[rela_id].values()[idxs] = 0 198 | 199 | path = [] 200 | memories = F.one_hot(torch.tensor([head_id]),num_classes=self.n_ent_vocab).float().to_sparse().to(self.device) 201 | dfs(database,memories,rela_attention,0,self.body_len,tail_id,path) 202 | path_item = list(reversed([x.item() for x in path])) 203 | path_item = [head_id.item()]+path_item 204 | 205 | for i in range(len(path_item)): 206 | if i%2 == 0: 207 | path_item[i] = self.id2e[path_item[i]] 208 | else: 209 | path_item[i] = self.id2r[path_item[i]] 210 | 211 | triple = [str(conf.item()),self.id2e[head_id.item()],self.id2r[rela_id.item()],self.id2e[tail_id.item()]] 212 | with open(self.decode_file,'a') as f: 213 | f.write(mode+'\t'+'\t'.join(triple+path_item)+'\n') 214 | 215 | 216 | 217 | # def decode_rule(self, dec_output, query, mode): 218 | # relation_attention_list = dec_output 219 | # batch_size = query.size(0) 220 | # num_step = self.body_len 221 | # for batch in range(batch_size): 222 | # paths = {t + 1: [] for t in range(num_step)} 223 | # # paths at hop 0, in the format of ([rel1,..],[ent1,..],weight) 224 | # paths[0] = [([-1], [query[batch, 0].item()], 1.0)] 225 | # relation_attentions = relation_attention_list[batch] 226 | # for step in range(num_step): 227 | # if not paths[step]: 228 | # break 229 | # relation_attention_ori = relation_attentions[step] 230 | # for rels, pths, wei in paths[step]: 231 | # if pths[-1] not in self.graph: 232 | # continue 233 | # # select relations(including self-loop) connected to the tail of each path 234 | # sel = torch.LongTensor(list(self.graph[pths[-1]].keys()) + [self.n_rel_vocab - 1]) 235 | # relation_attention = torch.zeros(self.n_rel_vocab).to(self.device) 236 | # relation_attention[sel] = relation_attention_ori[sel].clone() 237 | # rel_att_max = torch.max(relation_attention).item() 238 | # relation_attention /= rel_att_max 239 | 240 | # for rr in torch.nonzero(relation_attention> max(self.the_rel, self.the_rel_min / rel_att_max)): # relations which exceed threshold 241 | # # 归一化之后大于self.the_rel，归一化之前大于self.thr_rel_min 242 | # rr = rr.item() 243 | # if rr == self.n_rel_vocab - 1: # 244 | # paths[step + 1].append((rels + [rr],pths + [pths[-1]],wei * relation_attention[rr].item(),)) 245 | # elif rr in self.graph[pths[-1]].keys(): 246 | # for tail in self.graph[pths[-1]][rr]: 247 | # paths[step + 1].append((rels + [rr],pths + [tail],wei * relation_attention[rr].item(),)) 248 | 249 | # for path in paths[step + 1]: 250 | # rels, pths, wei = path 251 | # if path[2] > self.the_all: 252 | # self.decode_rule_num += 1 253 | # print( "\rWrite {}-{} Rule(s)".format( self.decode_rule_num, self.decode_rule_num_filter ), end="", ) 254 | # head_rule = self.id2r[query[batch, 1].item()] 255 | # rule_body = "^".join([self.id2r[r] for r in rels[1:]]) 256 | # try: 257 | # self.rules[head_rule][rule_body].append(wei) 258 | # except KeyError: 259 | # self.rules[head_rule][rule_body] = [wei] 260 | # self.decode_rule_num_filter += 1 261 | # with open(self.decode_file, "a") as f: 262 | # f.write(mode+'\t'+self.id2e[query[batch,0].item()]+'\t'+head_rule + "<-" + rule_body + "\n") 263 | -------------------------------------------------------------------------------- /DATASET/NL27K/relations.txt: -------------------------------------------------------------------------------- 1 | concept:arthropodthatfeedoninsect 2 | concept:teamplaysincity 3 | concept:organizationhastopmember 4 | concept:citystadiums 5 | concept:createdbyagent 6 | concept:proxyfor 7 | concept:organizationcreatedatdate 8 | concept:cityliesonriver 9 | concept:coachworksincountry 10 | concept:academicprogramatuniversity 11 | concept:countryisthehomeofsportsteam 12 | concept:arthropodlookslikeinsect 13 | concept:cityhotels 14 | concept:subpartoforganization 15 | concept:athleteinjuredhisbodypart 16 | concept:persondiedinlocation 17 | concept:worker 18 | concept:organizationcreatedbyperson 19 | concept:inverseofbankchiefexecutiveceo 20 | concept:countrycities 21 | concept:stateorprovinceresidenceofperson 22 | concept:citynewspaper 23 | concept:personhasresidenceincountry 24 | concept:leagueplayers 25 | concept:musicianplaysinstrument 26 | concept:generalizationof 27 | concept:personleadscity 28 | concept:sportusesstadium 29 | concept:personhasparent 30 | concept:wifeof 31 | concept:topmemberoforganization 32 | concept:istallerthan 33 | concept:publicationjournalist 34 | concept:statehascapital 35 | concept:statecontainsfarm 36 | concept:teamplayssport 37 | concept:countryhascompanyoffice 38 | concept:countryofpersonbirth 39 | concept:locationofitemexistence 40 | concept:inverseofagriculturalproductcookedwithagricultureproduct 41 | concept:ismultipleof 42 | concept:dateof 43 | concept:countrycontainsfarm 44 | concept:inverseofanimalsuchasinsect 45 | concept:equipmentusedbysport 46 | concept:organizationhasagent 47 | concept:companyceo 48 | concept:inverseoflanguegeschoolincity 49 | concept:politicianholdsoffice 50 | concept:geopoliticallocationresidenceofpersion 51 | concept:stadiumhometoathlete 52 | concept:officeheldbypolitician 53 | concept:politicianusendorsespoliticianus 54 | concept:inverseofagriculturalproducttoattractinsect 55 | concept:persondeathdate 56 | concept:winneringame 57 | concept:beveragemadefrombeverage 58 | concept:athleteplaysforteam 59 | concept:instrumentplayedbymusician 60 | concept:agriculturalproductcookedwithagriculturalproduct 61 | concept:inverseofanimalssuchasmammals 62 | concept:arthropodandotherarthropod 63 | concept:geopoliticallocationcontainscity 64 | concept:agentinvolvedwithitem 65 | concept:organizationacronymhasname 66 | concept:stadiumhometeam 67 | concept:marriedinyear 68 | concept:musicalartisthadapetanimal 69 | concept:chemicaltypehaschemical 70 | concept:worksfor 71 | concept:producedby 72 | concept:drugworkedonbyagent 73 | concept:geopoliticalorganizationleadbyperson 74 | concept:inverseofwriterwasbornincity 75 | concept:athletecoach 76 | concept:organizationheadquarteredincountry 77 | concept:automakerproducesmodel 78 | concept:stadiumhometosport 79 | concept:athleteplayssport 80 | concept:inverseofinsecteatsgrain 81 | concept:sportusesequipment 82 | concept:companyalsoknownas 83 | concept:acquired 84 | concept:cityhasairport 85 | concept:iteminvolvedwithagent 86 | concept:animaltypehasanimal 87 | concept:locationofpersondeath 88 | concept:musicgenreartist 89 | concept:inverseofclothingmadefromplant 90 | concept:inverseofarachnidiseatenbybird 91 | concept:stateorprovinceoforganizationheadquarters 92 | concept:agriculturalproductgrowninlandscapefeatures 93 | concept:animaleatfood 94 | concept:inverseofvegetableproductioninstateorprovince 95 | concept:jobpositionheldbyperson 96 | concept:professiontypehasprofession 97 | concept:inverseofagriculturalproductgrowninlandscapefeature 98 | concept:agriculturalproducttoattractinsect 99 | concept:hospitalincity 100 | concept:personbornincountry 101 | concept:dateoforganizationcreation 102 | concept:hasofficeincity 103 | concept:statehaslake 104 | concept:producesproduct 105 | concept:stateorprovinceisborderedbystateorprovince 106 | concept:museumincity 107 | concept:countrycapital 108 | concept:personhasresidenceinstateorprovince 109 | concept:vegetableproductioninstateorprovince 110 | concept:organizationheadquarteredincity 111 | concept:personbelongstoorganization 112 | concept:personhasresidenceincity 113 | concept:inverseofprofessionactsatthesameareaofpoliticissue 114 | concept:fatherofperson 115 | concept:inverseofdiseasemaybecausedbydrug 116 | concept:athletehomestadium 117 | concept:chemicalistypeofchemical 118 | concept:sportnamehasoriginlanguage 119 | concept:televisioncompanyaffiliate 120 | concept:productproducedincountry 121 | concept:diseasemaybecausedbydrug 122 | concept:personborninlocation 123 | concept:sportfansincountry 124 | concept:inverseofmammalsuchasmammal 125 | concept:citylocatedincountry 126 | concept:coachesteam 127 | concept:specializationof 128 | concept:farmlocatedincountry 129 | concept:inverseofanimaleatfood 130 | concept:personalsoknownas 131 | concept:inverseofinvertebratefeedonfood 132 | concept:subjectconcernedbyacademicfield 133 | concept:citytelevisionstation 134 | concept:inverseofcoachcanspeaklanguage 135 | concept:agentcontrols 136 | concept:agentcreated 137 | concept:objectfoundinscene 138 | concept:personattendsschool 139 | concept:organizationnamehasacronym 140 | concept:countrylanguage 141 | concept:inverseofvegetableisproducedatcountry 142 | concept:weaponmadeincountry 143 | concept:bodypartcontainsbodypart 144 | concept:agentcontributedtocreativework 145 | concept:agentparticipatedinevent 146 | concept:economicsectorcompany 147 | concept:schoolgraduatedperson 148 | concept:cityhashospital 149 | concept:players 150 | concept:jobpositionusesacademicfield 151 | concept:inverseofanimalfeedoninsect 152 | concept:athletealsoknownas 153 | concept:arthropodcanbeveryirritatingtomammal 154 | concept:vegetableisproducedatcountry 155 | concept:languageofuniversity 156 | concept:teammember 157 | concept:inverseofarthropodcalledarthropod 158 | concept:countryhascitizen 159 | concept:inverseofagriculturalproductgrowinginstateorprovince 160 | concept:creativeworkcontributedtobyagent 161 | concept:inverseofsportschoolincountry 162 | concept:itemfoundinroom 163 | concept:academicfieldusedinjobposition 164 | concept:inverseofweaponmadeincountry 165 | concept:languageofsportsgame 166 | concept:personbirthdate 167 | concept:inverseofautomobilemakerdealersincountry 168 | concept:academicfieldsuchasacademicfield 169 | concept:countryofpersondeath 170 | concept:animalsuchasfish 171 | concept:cityalsoknownas 172 | concept:inverseofcountriessuchascountries 173 | concept:teamhomestadium 174 | concept:parkincity 175 | concept:superpartof 176 | concept:organizationleadbyperson 177 | concept:companiesheadquarteredhere 178 | concept:academicfieldusedbyeconomicsector 179 | concept:locationlocatedwithinlocation 180 | concept:inverseofbankbankincountry 181 | concept:countryoforganizationheadquarters 182 | concept:motherofperson 183 | concept:bankchiefexecutiveceo 184 | concept:agriculturalproductcamefromcountry 185 | concept:agentholdssharesincompany 186 | concept:clothingmadefromplant 187 | concept:personterminatedbyorganization 188 | concept:cityofpersonbirth 189 | concept:inverseofarthropodlooklikeinsect 190 | concept:academicfieldconcernssubject 191 | concept:politicianusholdsoffice 192 | concept:politicianuswenttoschool 193 | concept:economicsectorusingacademicfield 194 | concept:organizationhasofficialwebsite 195 | concept:automobilemakerdealersincountry 196 | concept:officialwebsiteoforganization 197 | concept:inverseofarthropodfeedoninsect 198 | concept:dateofsportsgame 199 | concept:persondiedincity 200 | concept:organizationterminatedperson 201 | concept:atlocation 202 | concept:stateorprovinceofpersonbirth 203 | concept:inverseofmusicalartisthadapetanimal 204 | concept:locationofpersonbirth 205 | concept:radiostationincity 206 | concept:cityhascompanyoffice 207 | concept:persongraduatedfromuniversity 208 | concept:personhasjobposition 209 | concept:superpartoforganization 210 | concept:geopoliticallocationcontainscountry 211 | concept:countrylocatedingeopoliticallocation 212 | concept:inverseoffooddecreasestheriskofdisease 213 | concept:musicartistmusician 214 | concept:animalsuchasinsect 215 | concept:synonymfor 216 | concept:animalpreyson 217 | concept:inverseofacademicfieldsuchasacademicfield 218 | concept:yearofmarriage 219 | concept:dateofpersondeath 220 | concept:mammalcanbeirritatedbyarthropod 221 | concept:inverseofbeveragemadefrombeverage 222 | concept:inverseofanimaleatvegetable 223 | concept:automobilemakercardealersinstateorprovince 224 | concept:isoneoccurrenceof 225 | concept:itemexistsatlocation 226 | concept:countrycurrency 227 | concept:coachcanspeaklanguage 228 | concept:citylocatedinstate 229 | concept:citysportsteams 230 | concept:attractionofcity 231 | concept:professionistypeofprofession 232 | concept:citylanguage 233 | concept:cityparks 234 | concept:personleadsgeopoliticalorganization 235 | concept:invertebrateturnintoinsect 236 | concept:politicianusendorsedbypoliticianus 237 | concept:stadiumoreventvenuedisclosescompany 238 | concept:animalthatfeedoninsect 239 | concept:sportteam 240 | concept:citycapitalofcountry 241 | concept:televisionstationincity 242 | concept:inverseofagriculturalproductincludingagriculturalproduct 243 | concept:cityofpersondeath 244 | concept:lakeinstate 245 | concept:politicalgroupofpoliticianus 246 | concept:teammate 247 | concept:riverflowsthroughcity 248 | concept:inverseofanimalsuchasfish 249 | concept:organizationalsoknownas 250 | concept:personhascitizenship 251 | concept:inverseofpersongraduatedfromuniversity 252 | concept:arachnidiseatenbybird 253 | concept:personleadsorganization 254 | concept:languageofcity 255 | concept:leaguestadiums 256 | concept:cityattractions 257 | concept:subpartof 258 | concept:animalsuchasinvertebrate 259 | concept:animalssuchasmammals 260 | concept:bodypartwithinbodypart 261 | concept:inverseofanimalsuchasmollusk 262 | concept:transportationincity 263 | concept:automobilemakerchiefexecutiveceo 264 | concept:teamplaysagainstteam 265 | concept:dateoforganzationdissolution 266 | concept:statehasmountain 267 | concept:headquarteredin 268 | concept:leagueteams 269 | concept:mammalsuchasmammal 270 | concept:persongraduatedschool 271 | concept:inverseofathleteinjuredhisbodypart 272 | concept:inverseofstadiumoreventvenuedisclosescompany 273 | concept:competeswith 274 | concept:statelocatedingeopoliticallocation 275 | concept:companyeconomicsector 276 | concept:inverseofautomobilemakerdealersincity 277 | concept:parentofperson 278 | concept:arthropodcalledarthropod 279 | concept:musicianinmusicartist 280 | concept:citycapitalofstate 281 | concept:politicianusmemberofpoliticalgroup 282 | concept:universityoperatesinlanguage 283 | concept:organizationheadquarteredinstateorprovince 284 | concept:countryresidenceofperson 285 | concept:farmlocatedinstate 286 | concept:personhasmother 287 | concept:insecteatsgrain 288 | concept:scenecontainsobject 289 | concept:universityincity 290 | concept:sportsgamedate 291 | concept:locatedat 292 | concept:organizationdissolvedatdate 293 | concept:cityleadbyperson 294 | concept:personwrittenaboutinpublication 295 | concept:agentcompeteswithagent 296 | concept:companyhasshareholder 297 | concept:bankbankincountry 298 | concept:citymuseums 299 | concept:farmproducesagriculturalproduct 300 | concept:athleteledsportsteam 301 | concept:personhasfather 302 | concept:agentworkedondrug 303 | concept:bankboughtbank 304 | concept:cityuniversities 305 | concept:sportsgameteam 306 | concept:agriculturalproductproducedbyfarm 307 | concept:clothingtogowithclothing 308 | concept:personbornincity 309 | concept:inverseofanimalsuchasinvertebrate 310 | concept:inverseofathleteledsportsteam 311 | concept:citycontainsbuilding 312 | concept:agentactsinlocation 313 | concept:airportincity 314 | concept:agentcollaborateswithagent 315 | concept:automodelproducedbymaker 316 | concept:mutualproxyfor 317 | concept:personmovedtostateorprovince 318 | concept:hotelincity 319 | concept:officeheldbypoliticianus 320 | concept:teamcoach 321 | concept:invertebratefeedonfood 322 | concept:roomcancontainitem 323 | concept:dateofpersonbirth 324 | concept:schoolattendedbyperson 325 | concept:atdate 326 | concept:buildinglocatedincity 327 | concept:cityhostsattraction 328 | concept:animaldevelopdisease 329 | concept:hassibling 330 | concept:stadiumlocatedincity 331 | concept:languageofcountry 332 | concept:cityradiostation 333 | concept:leaguecoaches 334 | concept:inverseofarthropodandotherarthropod 335 | concept:locationcontainslocation 336 | concept:countriessuchascountries 337 | concept:inverseofautomobilemakerchiefexecutiveceo 338 | concept:husbandof 339 | concept:universityhasacademicprogram 340 | concept:hasfamilymember 341 | concept:inverseofsportfansincountry 342 | concept:ageofperson 343 | concept:ceoof 344 | concept:inverseofagriculturalproductcamefromcountry 345 | concept:agenthaswebsite 346 | concept:sportschoolincountry 347 | concept:inverseofautomobilemakercardealersinstateorprovince 348 | concept:sportsgamewinner 349 | concept:cityhastransportation 350 | concept:politicianrepresentslocation 351 | concept:personhasage 352 | concept:teamplaysinleague 353 | concept:journalistwritesforpublication 354 | concept:agriculturalproductgrowinginstateorprovince 355 | concept:organizationhasperson 356 | concept:locationactedinbyagent 357 | concept:proxyof 358 | concept:personhiredbyorganization 359 | concept:inverseofpersonmovedtostateorprovince 360 | concept:acquiredby 361 | concept:stadiumhometoleague 362 | concept:animalpredators 363 | concept:professionactsatthesameareaofpoliticissue 364 | concept:cityoforganizationheadquarters 365 | concept:countryalsoknownas 366 | concept:fooddecreasestheriskofdisease 367 | concept:mountaininstate 368 | concept:countrystates 369 | concept:personborninstateorprovince 370 | concept:eventhasparticipantagent 371 | concept:teamingame 372 | concept:inverseofanimaldevelopdisease 373 | concept:languageschoolincity 374 | concept:musicartistgenre 375 | concept:hasofficeincountry 376 | concept:hasspouse 377 | concept:inverseofinvertebrateturnintoinsect 378 | concept:athleteplaysinleague 379 | concept:websiteoperatedbyagent 380 | concept:hashusband 381 | concept:agentcreatedorganization 382 | concept:agriculturalproductincludingagriculturalproduct 383 | concept:currencycountry 384 | concept:agentowns 385 | concept:statecontainscity 386 | concept:automobilemakerdealersincity 387 | concept:agentbelongstoorganization 388 | concept:persondiedincountry 389 | concept:fishservedwithfood 390 | concept:controlledbyagent 391 | concept:coachesinleague 392 | concept:haswife 393 | concept:locationrepresentedbypolitician 394 | concept:ownedbyagent 395 | concept:televisionstationaffiliatedwith 396 | concept:attractionbefallincity 397 | concept:animalistypeofanimal 398 | concept:countryproducesproduct 399 | concept:citylocatedingeopoliticallocation 400 | concept:inverseofbankboughtbank 401 | concept:inverseoffishservedwithfood 402 | concept:statelocatedincountry 403 | concept:teamalsoknownas 404 | concept:inverseofpoliticianuswenttoschool 405 | concept:publicationwritesabout 406 | concept:coachesathlete 407 | concept:organizationhiredperson 408 | concept:isshorterthan 409 | concept:personhasresidenceingeopoliticallocation 410 | concept:inverseofcoachworksincountry 411 | concept:inverseofcountryisthehomeofsportsteam 412 | concept:newspaperincity 413 | concept:animalsuchasmollusk 414 | concept:cityresidenceofperson 415 | concept:geopoliticallocationcontainsstate 416 | concept:writerwasbornincity 417 | concept:animaleatvegetable 418 | -------------------------------------------------------------------------------- /DATASET/UFB15K237/relation_id.tsv: -------------------------------------------------------------------------------- 1 | 0 /location/country/form_of_government 2 | 1 /tv/tv_program/regular_cast./tv/regular_tv_appearance/actor 3 | 2 /media_common/netflix_genre/titles 4 | 3 /award/award_winner/awards_won./award/award_honor/award_winner 5 | 4 /soccer/football_team/current_roster./sports/sports_team_roster/position 6 | 5 /soccer/football_team/current_roster./soccer/football_roster_position/position 7 | 6 /film/actor/film./film/performance/film 8 | 7 /award/award_category/nominees./award/award_nomination/nominated_for 9 | 8 /award/award_nominee/award_nominations./award/award_nomination/award_nominee 10 | 9 /music/performance_role/regular_performances./music/group_membership/role 11 | 10 /award/award_category/winners./award/award_honor/ceremony 12 | 11 /film/film/release_date_s./film/film_regional_release_date/film_release_distribution_medium 13 | 12 /award/award_winning_work/awards_won./award/award_honor/award_winner 14 | 13 /film/film/release_date_s./film/film_regional_release_date/film_release_region 15 | 14 /film/film/language 16 | 15 /location/location/contains 17 | 16 /organization/organization/headquarters./location/mailing_address/country 18 | 17 /people/person/profession 19 | 18 /location/statistical_region/religions./location/religion_percentage/religion 20 | 19 /award/award_nominee/award_nominations./award/award_nomination/award 21 | 20 /music/genre/artists 22 | 21 /influence/influence_node/influenced_by 23 | 22 /education/educational_institution/students_graduates./education/education/student 24 | 23 /organization/organization_member/member_of./organization/organization_membership/organization 25 | 24 /people/person/place_of_birth 26 | 25 /common/topic/webpage./common/webpage/category 27 | 26 /film/film/other_crew./film/film_crew_gig/film_crew_role 28 | 27 /award/award_ceremony/awards_presented./award/award_honor/award_winner 29 | 28 /film/film/country 30 | 29 /olympics/olympic_sport/athletes./olympics/olympic_athlete_affiliation/country 31 | 30 /award/award_winning_work/awards_won./award/award_honor/award 32 | 31 /film/film/genre 33 | 32 /film/film_distributor/films_distributed./film/film_film_distributor_relationship/film 34 | 33 /location/statistical_region/rent50_2./measurement_unit/dated_money_value/currency 35 | 34 /film/film/estimated_budget./measurement_unit/dated_money_value/currency 36 | 35 /government/legislative_session/members./government/government_position_held/district_represented 37 | 36 /sports/sports_team/roster./american_football/football_historical_roster_position/position_s 38 | 37 /government/government_office_category/officeholders./government/government_position_held/jurisdiction_of_office 39 | 38 /award/award_nominee/award_nominations./award/award_nomination/nominated_for 40 | 39 /people/deceased_person/place_of_death 41 | 40 /sports/pro_athlete/teams./sports/sports_team_roster/team 42 | 41 /music/artist/track_contributions./music/track_contribution/role 43 | 42 /music/instrument/instrumentalists 44 | 43 /people/person/gender 45 | 44 /business/job_title/people_with_this_title./business/employment_tenure/company 46 | 45 /base/popstra/celebrity/friendship./base/popstra/friendship/participant 47 | 46 /business/business_operation/industry 48 | 47 /award/award_category/winners./award/award_honor/award_winner 49 | 48 /people/person/nationality 50 | 49 /sports/sports_team/roster./baseball/baseball_roster_position/position 51 | 50 /people/cause_of_death/people 52 | 51 /tv/tv_program/languages 53 | 52 /music/genre/parent_genre 54 | 53 /base/popstra/celebrity/breakup./base/popstra/breakup/participant 55 | 54 /location/country/capital 56 | 55 /film/special_film_performance_type/film_performance_type./film/performance/film 57 | 56 /location/hud_county_place/place 58 | 57 /music/performance_role/guest_performances./music/recording_contribution/performance_role 59 | 58 /location/location/adjoin_s./location/adjoining_relationship/adjoins 60 | 59 /film/film_subject/films 61 | 60 /base/schemastaging/person_extra/net_worth./measurement_unit/dated_money_value/currency 62 | 61 /education/educational_institution/colors 63 | 62 /base/aareas/schema/administrative_area/administrative_parent 64 | 63 /organization/role/leaders./organization/leadership/organization 65 | 64 /government/legislative_session/members./government/government_position_held/legislative_sessions 66 | 65 /olympics/olympic_games/sports 67 | 66 /people/person/places_lived./people/place_lived/location 68 | 67 /award/award_category/disciplines_or_subjects 69 | 68 /sports/professional_sports_team/draft_picks./sports/sports_league_draft_pick/school 70 | 69 /olympics/olympic_participating_country/medals_won./olympics/olympic_medal_honor/olympics 71 | 70 /language/human_language/countries_spoken_in 72 | 71 /people/ethnicity/people 73 | 72 /olympics/olympic_participating_country/athletes./olympics/olympic_athlete_affiliation/olympics 74 | 73 /olympics/olympic_participating_country/medals_won./olympics/olympic_medal_honor/medal 75 | 74 /film/film/other_crew./film/film_crew_gig/crewmember 76 | 75 /music/performance_role/track_performances./music/track_contribution/role 77 | 76 /music/record_label/artist 78 | 77 /base/popstra/celebrity/dated./base/popstra/dated/participant 79 | 78 /base/locations/continents/countries_within 80 | 79 /sports/sports_league_draft/picks./sports/sports_league_draft_pick/school 81 | 80 /sports/sports_team/sport 82 | 81 /education/educational_institution_campus/educational_institution 83 | 82 /people/person/religion 84 | 83 /organization/organization/headquarters./location/mailing_address/citytown 85 | 84 /film/film/produced_by 86 | 85 /people/person/spouse_s./people/marriage/type_of_union 87 | 86 /film/film/production_companies 88 | 87 /location/country/second_level_divisions 89 | 88 /base/culturalevent/event/entity_involved 90 | 89 /education/educational_degree/people_with_this_degree./education/education/institution 91 | 90 /film/film/distributors./film/film_film_distributor_relationship/film_distribution_medium 92 | 91 /tv/tv_program/tv_producer./tv/tv_producer_term/producer_type 93 | 92 /music/performance_role/regular_performances./music/group_membership/group 94 | 93 /government/politician/government_positions_held./government/government_position_held/legislative_sessions 95 | 94 /sports/sports_team_location/teams 96 | 95 /sports/sports_team/colors 97 | 96 /sports/professional_sports_team/draft_picks./sports/sports_league_draft_pick/draft 98 | 97 /award/award_nominated_work/award_nominations./award/award_nomination/nominated_for 99 | 98 /baseball/baseball_team/team_stats./baseball/baseball_team_stats/season 100 | 99 /user/alexander/philosophy/philosopher/interests 101 | 100 /people/person/languages 102 | 101 /education/educational_institution/students_graduates./education/education/major_field_of_study 103 | 102 /film/director/film 104 | 103 /award/award_ceremony/awards_presented./award/award_honor/honored_for 105 | 104 /olympics/olympic_sport/athletes./olympics/olympic_athlete_affiliation/olympics 106 | 105 /film/film/costume_design_by 107 | 106 /film/film/featured_film_locations 108 | 107 /tv/tv_program/genre 109 | 108 /organization/endowed_organization/endowment./measurement_unit/dated_money_value/currency 110 | 109 /film/film/written_by 111 | 110 /tv/tv_producer/programs_produced./tv/tv_producer_term/producer_type 112 | 111 /location/hud_county_place/county 113 | 112 /education/educational_degree/people_with_this_degree./education/education/major_field_of_study 114 | 113 /base/schemastaging/organization_extra/phone_number./base/schemastaging/phone_sandbox/contact_category 115 | 114 /base/saturdaynightlive/snl_cast_member/seasons./base/saturdaynightlive/snl_season_tenure/cast_members 116 | 115 /film/film/film_format 117 | 116 /film/film/music 118 | 117 /sports/sports_team/roster./american_football/football_roster_position/position 119 | 118 /user/tsegaran/random/taxonomy_subject/entry./user/tsegaran/random/taxonomy_entry/taxonomy 120 | 119 /celebrities/celebrity/celebrity_friends./celebrities/friendship/friend 121 | 120 /location/hud_foreclosure_area/estimated_number_of_mortgages./measurement_unit/dated_integer/source 122 | 121 /base/x2010fifaworldcupsouthafrica/world_cup_squad/current_world_cup_squad./base/x2010fifaworldcupsouthafrica/current_world_cup_squad/current_club 123 | 122 /base/popstra/location/vacationers./base/popstra/vacation_choice/vacationer 124 | 123 /base/popstra/celebrity/canoodled./base/popstra/canoodled/participant 125 | 124 /education/educational_institution/school_type 126 | 125 /sports/sport/pro_athletes./sports/pro_sports_played/athlete 127 | 126 /award/hall_of_fame/inductees./award/hall_of_fame_induction/inductee 128 | 127 /base/biblioness/bibs_location/state 129 | 128 /education/university/local_tuition./measurement_unit/dated_money_value/currency 130 | 129 /business/business_operation/revenue./measurement_unit/dated_money_value/currency 131 | 130 /sports/sports_position/players./sports/sports_team_roster/team 132 | 131 /military/military_conflict/combatants./military/military_combatant_group/combatants 133 | 132 /sports/sports_league/teams./sports/sports_league_participation/team 134 | 133 /base/biblioness/bibs_location/country 135 | 134 /music/group_member/membership./music/group_membership/role 136 | 135 /location/administrative_division/country 137 | 136 /film/film/executive_produced_by 138 | 137 /food/food/nutrients./food/nutrition_fact/nutrient 139 | 138 /music/group_member/membership./music/group_membership/group 140 | 139 /location/statistical_region/gdp_nominal_per_capita./measurement_unit/dated_money_value/currency 141 | 140 /education/field_of_study/students_majoring./education/education/student 142 | 141 /award/award_winning_work/awards_won./award/award_honor/honored_for 143 | 142 /sports/sports_team/roster./basketball/basketball_roster_position/position 144 | 143 /organization/organization/headquarters./location/mailing_address/state_province_region 145 | 144 /base/aareas/schema/administrative_area/capital 146 | 145 /tv/tv_producer/programs_produced./tv/tv_producer_term/program 147 | 146 /education/educational_institution/campuses 148 | 147 /film/film/distributors./film/film_film_distributor_relationship/region 149 | 148 /travel/travel_destination/climate./travel/travel_destination_monthly_climate/month 150 | 149 /medicine/disease/notable_people_with_this_condition 151 | 150 /base/schemastaging/organization_extra/phone_number./base/schemastaging/phone_sandbox/service_location 152 | 151 /people/ethnicity/geographic_distribution 153 | 152 /military/military_combatant/military_conflicts./military/military_combatant_group/combatants 154 | 153 /film/film/personal_appearances./film/personal_film_appearance/person 155 | 154 /location/location/time_zones 156 | 155 /film/film/story_by 157 | 156 /base/petbreeds/city_with_dogs/top_breeds./base/petbreeds/dog_city_relationship/dog_breed 158 | 157 /olympics/olympic_games/participating_countries 159 | 158 /tv/tv_network/programs./tv/tv_network_duration/program 160 | 159 /people/person/employment_history./business/employment_tenure/company 161 | 160 /music/artist/origin 162 | 161 /people/marriage_union_type/unions_of_this_type./people/marriage/location_of_ceremony 163 | 162 /tv/tv_writer/tv_programs./tv/tv_program_writer_relationship/tv_program 164 | 163 /time/event/locations 165 | 164 /celebrities/celebrity/sexual_relationships./celebrities/romantic_relationship/celebrity 166 | 165 /government/political_party/politicians_in_this_party./government/political_party_tenure/politician 167 | 166 /education/university/domestic_tuition./measurement_unit/dated_money_value/currency 168 | 167 /organization/non_profit_organization/registered_with./organization/non_profit_registration/registering_agency 169 | 168 /user/ktrueman/default_domain/international_organization/member_states 170 | 169 /american_football/football_team/current_roster./sports/sports_team_roster/position 171 | 170 /soccer/football_player/current_team./sports/sports_team_roster/team 172 | 171 /olympics/olympic_games/medals_awarded./olympics/olympic_medal_honor/medal 173 | 172 /award/ranked_item/appears_in_ranked_lists./award/ranking/list 174 | 173 /film/film_set_designer/film_sets_designed 175 | 174 /user/jg/default_domain/olympic_games/sports 176 | 175 /people/person/sibling_s./people/sibling_relationship/sibling 177 | 176 /base/marchmadness/ncaa_basketball_tournament/seeds./base/marchmadness/ncaa_tournament_seed/team 178 | 177 /music/artist/contribution./music/recording_contribution/performance_role 179 | 178 /sports/sports_position/players./sports/sports_team_roster/position 180 | 179 /government/politician/government_positions_held./government/government_position_held/basic_title 181 | 180 /sports/sports_position/players./american_football/football_historical_roster_position/position_s 182 | 181 /people/profession/specialization_of 183 | 182 /award/award_category/category_of 184 | 183 /film/film/cinematography 185 | 184 /film/film/film_production_design_by 186 | 185 /tv/tv_program/country_of_origin 187 | 186 /film/film/edited_by 188 | 187 /tv/tv_personality/tv_regular_appearances./tv/tv_regular_personal_appearance/program 189 | 188 /education/field_of_study/students_majoring./education/education/major_field_of_study 190 | 189 /education/educational_degree/people_with_this_degree./education/education/student 191 | 190 /location/us_county/county_seat 192 | 191 /film/film/film_festivals 193 | 192 /people/person/spouse_s./people/marriage/spouse 194 | 193 /film/film/runtime./film/film_cut/film_release_region 195 | 194 /location/capital_of_administrative_division/capital_of./location/administrative_division_capital_relationship/administrative_division 196 | 195 /base/aareas/schema/administrative_area/administrative_area_type 197 | 196 /influence/influence_node/peers./influence/peer_relationship/peers 198 | 197 /government/governmental_body/members./government/government_position_held/legislative_sessions 199 | 198 /base/schemastaging/organization_extra/phone_number./base/schemastaging/phone_sandbox/service_language 200 | 199 /organization/organization/place_founded 201 | 200 /base/eating/practicer_of_diet/diet 202 | 201 /tv/tv_program/program_creator 203 | 202 /people/deceased_person/place_of_burial 204 | 203 /film/film/release_date_s./film/film_regional_release_date/film_regional_debut_venue 205 | 204 /ice_hockey/hockey_team/current_roster./sports/sports_team_roster/position 206 | 205 /location/administrative_division/first_level_division_of 207 | 206 /location/statistical_region/gdp_real./measurement_unit/adjusted_money_value/adjustment_currency 208 | 207 /organization/organization/child./organization/organization_relationship/child 209 | 208 /business/business_operation/assets./measurement_unit/dated_money_value/currency 210 | 209 /film/person_or_entity_appearing_in_film/films./film/personal_film_appearance/type_of_appearance 211 | 210 /time/event/instance_of_recurring_event 212 | 211 /tv/non_character_role/tv_regular_personal_appearances./tv/tv_regular_personal_appearance/person 213 | 212 /film/actor/film./film/performance/special_performance_type 214 | 213 /business/business_operation/operating_income./measurement_unit/dated_money_value/currency 215 | 214 /people/ethnicity/languages_spoken 216 | 215 /base/americancomedy/celebrity_impressionist/celebrities_impersonated 217 | 216 /medicine/symptom/symptom_of 218 | 217 /education/university/fraternities_and_sororities 219 | 218 /location/statistical_region/places_exported_to./location/imports_and_exports/exported_to 220 | 219 /location/statistical_region/gdp_nominal./measurement_unit/dated_money_value/currency 221 | 220 /travel/travel_destination/how_to_get_here./travel/transportation/mode_of_transportation 222 | 221 /location/country/official_language 223 | 222 /location/location/partially_contains 224 | 223 /film/film/film_art_direction_by 225 | 224 /people/person/spouse_s./people/marriage/location_of_ceremony 226 | 225 /education/university/international_tuition./measurement_unit/dated_money_value/currency 227 | 226 /organization/organization_founder/organizations_founded 228 | 227 /film/film/dubbing_performances./film/dubbing_performance/actor 229 | 228 /dataworld/gardening_hint/split_to 230 | 229 /medicine/disease/risk_factors 231 | 230 /film/film/prequel 232 | 231 /base/localfood/seasonal_month/produce_available./base/localfood/produce_availability/seasonal_months 233 | 232 /film/actor/dubbing_performances./film/dubbing_performance/language 234 | 233 /broadcast/content/artist 235 | 234 /location/statistical_region/gni_per_capita_in_ppp_dollars./measurement_unit/dated_money_value/currency 236 | 235 /music/instrument/family 237 | 236 /government/politician/government_positions_held./government/government_position_held/jurisdiction_of_office 238 | -------------------------------------------------------------------------------- /DATASET/confidece_analyse.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "cell_type": "code", 5 | "execution_count": 1, 6 | "metadata": {}, 7 | "outputs": [], 8 | "source": [ 9 | "from tqdm import tqdm" 10 | ] 11 | }, 12 | { 13 | "cell_type": "code", 14 | "execution_count": 2, 15 | "metadata": {}, 16 | "outputs": [], 17 | "source": [ 18 | "file1 = 'CN15k/train.txt'\n", 19 | "file2 = 'NL27K/train.txt'\n", 20 | "file3 = 'UWN18RR/train2id.tsv'\n", 21 | "file4 = 'UFB15K237/train2id.tsv'" 22 | ] 23 | }, 24 | { 25 | "cell_type": "code", 26 | "execution_count": 3, 27 | "metadata": {}, 28 | "outputs": [ 29 | { 30 | "name": "stderr", 31 | "output_type": "stream", 32 | "text": [ 33 | "100%|██████████| 204984/204984 [00:00<00:00, 210743.55it/s]\n", 34 | "100%|██████████| 149100/149100 [00:00<00:00, 217059.48it/s]\n", 35 | "100%|██████████| 94863/94863 [00:00<00:00, 221649.82it/s]\n", 36 | "100%|██████████| 316319/316319 [00:01<00:00, 235975.54it/s]\n" 37 | ] 38 | } 39 | ], 40 | "source": [ 41 | "with open(file1,'r') as f1, open(file2,'r') as f2,open(file3,'r') as f3, open(file4,'r') as f4:\n", 42 | " lines1 = f1.readlines()\n", 43 | " lines2 = f2.readlines()\n", 44 | " lines3 = f3.readlines()\n", 45 | " lines4 = f4.readlines()\n", 46 | " data1,data2,data3,data4 = [],[],[],[]\n", 47 | " \n", 48 | " for line in tqdm(lines1):\n", 49 | " data1.append(eval(line.strip().split('\\t')[-1]))\n", 50 | " for line in tqdm(lines2):\n", 51 | " data2.append(eval(line.strip().split('\\t')[-1]))\n", 52 | " for line in tqdm(lines3):\n", 53 | " data3.append(eval(line.strip().split('\\t')[-1]))\n", 54 | " for line in tqdm(lines4):\n", 55 | " data4.append(eval(line.strip().split('\\t')[-1]))" 56 | ] 57 | }, 58 | { 59 | "cell_type": "code", 60 | "execution_count": 18, 61 | "metadata": {}, 62 | "outputs": [ 63 | { 64 | "name": "stderr", 65 | "output_type": "stream", 66 | "text": [ 67 | "The PostScript backend does not support transparency; partially transparent artists will be rendered opaque.\n" 68 | ] 69 | }, 70 | { 71 | "data": { 72 | "image/png": "iVBORw0KGgoAAAANSUhEUgAABKUAAAMVCAYAAACm0EewAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjcuMiwgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy8pXeV/AAAACXBIWXMAAA9hAAAPYQGoP6dpAADD9UlEQVR4nOzde1xVdb7/8feWyxZJtggBkphWSBpaiolohaaAjkjmzFhDUZahhUkUZJldtEnMu42UqVmaaHRmHDulE6GlzHC8MzKTl9FmxrwkqCVu1GiDuH5/9Gj92oLXYHPx9Xw81uNx9vp+1lqftdaZ+PjZ372WxTAMQwAAAAAAAIALNavvBAAAAAAAAHD1oSkFAAAAAAAAl6MpBQAAAAAAAJejKQUAAAAAAACXoykFAAAAAAAAl6MpBQAAAAAAAJejKQUAAAAAAACXoykFAAAAAAAAl6MpBQAAAAAAAJejKQWgSfnnP/+pRx55RB06dFDz5s11zTXXqHv37po2bZqOHz8uSerbt68sFosGDhxYbfuvv/5aFotFM2bMcFr/4osvKj4+Xtddd50sFotGjBhR4/EnTpwoi8VSbWnevHm1WIvFoieffLLa+ldeeUUWi0VPPPGEzp49ewVXAQAAoLrFixebdcn+/furjfft21fh4eHm5/bt2ys+Pv6C+3znnXc0dOhQtW/fXl5eXrrpppv0xBNPqLi42Clu/fr1NdZIPy2PP/54tTy3bdvmtI9vv/1WPXr00DXXXKM1a9ZcySUA0MC413cCAFBbFi5cqJSUFIWFhenZZ59V586dVVlZqW3btuntt9/Wxo0btXLlSjP+s88+0xdffKG77777ovuePXu2unbtqoSEBL377rsXjc/NzZXNZjM/N2t28e8ADMPQU089pblz5+r555/XlClTLroNAADA5XI4HHrxxRe1dOnSX7yvV155Rf369VNmZqauu+467dmzR7///e/1v//7v9q+fbsCAwMlSd27d9fGjRurbT9v3jy9//77uvfeey94nEOHDikmJkZHjhzR2rVr1atXr1+cO4D6R1MKQJOwceNGPfHEE4qJidFHH30kq9VqjsXExCg9PV25ubnmuo4dO+rMmTMaN26ctm7dKovFcsH9nzx50mwsXUoBFxERIX9//0vO/8yZM3r00Ue1dOlSTZ8+XRkZGZe8LQAAwOUYOHCgli9froyMDN16662/aF/bt29XQECA+Tk6Olrdu3fX7bffroULF+rFF1+UJPn4+FRrJBmGoQceeEDXX3+9YmJiznuMr776SgMGDFBlZaXy8/PVpUuXX5QzgIaDn+8BaBIyMzNlsVi0YMECp4bUTzw9PZWQkGB+9vDw0OTJk1VYWKgPP/zwovu/lJlOV+qHH37Qr3/9ay1fvlzvvPMODSkAAFCnxo0bJz8/Pz333HO/eF8/b0j9JCIiQm5ubjp48OAFt123bp3++9//6pFHHjlvrVVUVKQ77rhD7u7uKigooCEFNDE0pQA0elVVVfriiy8UERGhkJCQS97uvvvuU0REhF588UVVVlbWak5dunSRm5ubAgMD9dBDD+nAgQM1xp08eVKDBg1Sbm6uPvzwQ40cObJW8wAAADhXy5Yt9eKLL5qPMqht+fn5qqqq0i233HLBuEWLFqlZs2Z65JFHahwvKChQ3759FRAQoIKCAt1www21niuA+kVTCkCj9+233+r7779Xhw4dLms7i8WiqVOn6j//+Y/mz59fK7nceOONmjx5st59912tXbtWzzzzjFavXq2ePXvqm2++qRb//vvva/369crKytKvf/3rWskBAADgYh5//HHdcMMNeu6552QYRq3t9+TJk0pJSVFISIgeffTR88adOHFCf/7znxUTE6N27drVGPP0009Lkr744gu1adOm1nIE0HDQlAJwVevfv79iY2P16quv6uTJk794f0lJSXrhhRc0aNAg9evXT88995w+/fRTHTt2TNOmTasWf+edd6pVq1aaNGmS/v3vf//i4wMAAFwKT09Pvfbaa9q2bZv+53/+p1b2+cMPP2jYsGHav3+//vjHP+qaa645b+yyZcv0ww8/6LHHHjtvTEJCgux2u9LS0lRVVVUrOQJoWGhKAWj0/P391aJFC+3bt++Ktp86daq+/fZbzZgxo5Yz+1HPnj3VsWNHbdq0qdpY165dtXbtWn3//feKjo7W3r176yQHAACAc91///3q3r27JkyY8IsfZeBwOHTvvfeqoKBAH3/8sSIjIy8Yv2jRIl177bW65557zhvz0ksv6eWXX9by5cv14IMP0pgCmiCaUgAaPTc3N/Xv31+FhYU6dOjQZW9/22236Xe/+51mzZqlI0eO1EGGP75d5nwP8IyIiNDatWv1ww8/qF+/ftqzZ0+d5AAAAPBzP3+UwYIFC654Pw6HQ0OHDtW6dev00UcfqX///heM3759u7Zv366HHnpIHh4eF4ydNGmSXnnlFeXk5CgxMVFnzpy54jwBNDw0pQA0CePHj5dhGEpOTlZFRUW18crKSn3yySfn3f61115TRUWFJk2aVOu5bdq0SV999VW11yD/XPfu3fX555/L4XCoX79++te//lXreQAAAJxrwIABiomJ0auvvqpTp05d9vY/zZD64osvtGLFCsXFxV10m0WLFknSJb/gZeLEiZo0aZL+53/+h8YU0MS413cCAFAboqKiNG/ePKWkpCgiIkJPPPGEbrnlFlVWVmr79u1asGCBwsPDNWTIkBq379Chg5544gm98cYbNY7n5+fr2LFjkn5829/+/fv1pz/9SZIUHR2ta6+9VpJ066236sEHH1SnTp3UvHlzbdmyRdOnT1dQUJDGjRt3wXO47bbb9Pnnn6t///7q16+fvvjiC3Xq1OlKLwkAAMAlmTp1qiIiInT06NFqb8wrKSkxa56fa9++vXr06KHf/OY3+vTTTzVhwgT5+fk5Pa7Ax8dHnTt3dtruhx9+0PLly9W7d+/LqnNefvllNWvWTC+99JIMw9AHH3wgd3f+OQs0dvyvGECTkZycrJ49e2r27NmaOnWqSkpK5OHhoY4dOyoxMVFPPvnkBbd/8cUX9d5776msrKza2CuvvKL8/Hzz8/r167V+/XpJ0rp169S3b19JUufOnbVgwQIVFxeroqJCwcHBuv/++/Xyyy9f0ltjbr31Vn3xxRdOjalzizkAAIDa1K1bN/3ud7/T8uXLq40VFhbqt7/9bbX1Dz/8sBYvXqxVq1ZJkiZPnqzJkyc7xURHR5v10k/+/Oc/q7S09IIPOD+fF198Uc2aNdOECRN09uxZ5eTkXPTnfwAaNotRm+//BAAAAAAAAC4Bz5QCAAAAAACAy9GUAgAAAAAAgMvRlAIAAAAAAIDL0ZQCAAAAAACAy9GUAgAAAAAAgMu513cCTc3Zs2d1+PBhtWzZUhaLpb7TAQAAl8gwDJ08eVLBwcFq1ozv7eoKtRIAAI1TXdRKNKVq2eHDhxUSElLfaQAAgCt08OBBtW3btr7TaLKolQAAaNxqs1aiKVXLWrZsKenHm+Tj41PP2QAAgEtVVlamkJAQ82856ga1EgAAjVNd1Eo0pWrZT9PQfXx8KLQAAGiE+ElZ3aJWAgCgcavNWokHJgAAAAAAAMDlaEoBAAAAAADA5WhKAQAAAAAAwOVoSgEAAAAAAMDlaEoBAAAAAADA5WhKAQAAAAAAwOVoSgEAAAAAAMDlaEoBAAAAAADA5WhKAQAAAAAAwOVoSgEAAAAAAMDlaEoBAAAAAADA5WhKAQAAAAAAwOXc6zsBAEDjNXvN3lrd39MxHWt1fwAAAEBDRS3NTCkAAAAAAADUA5pSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAADVj79u1lsViqLWPGjJEkGYahiRMnKjg4WF5eXurbt6927tzptA+Hw6GxY8fK399f3t7eSkhI0KFDh5xiSktLlZSUJJvNJpvNpqSkJJ04ccIp5sCBAxoyZIi8vb3l7++v1NRUVVRU1On5AwCApoumFAAAQAO2detWFRcXm8uaNWskSb/97W8lSdOmTdOsWbOUlZWlrVu3KigoSDExMTp58qS5j7S0NK1cuVI5OTkqKCjQqVOnFB8fr6qqKjMmMTFRRUVFys3NVW5uroqKipSUlGSOV1VVafDgwTp9+rQKCgqUk5OjFStWKD093UVXAgAANDXu9Z0AAAAAzu/aa691+vz666/rxhtvVHR0tAzD0Jw5czRhwgQNGzZMkrRkyRIFBgZq+fLlGj16tOx2uxYtWqSlS5dqwIABkqTs7GyFhIRo7dq1iouL0+7du5Wbm6tNmzYpMjJSkrRw4UJFRUVpz549CgsLU15ennbt2qWDBw8qODhYkjRz5kyNGDFCkydPlo+PT435OxwOORwO83NZWVmtXyMAANA4MVMKAACgkaioqFB2drYeffRRWSwW7du3TyUlJYqNjTVjrFaroqOjtWHDBklSYWGhKisrnWKCg4MVHh5uxmzcuFE2m81sSElSr169ZLPZnGLCw8PNhpQkxcXFyeFwqLCw8Lw5T5kyxfxJoM1mU0hISO1cDAAA0OjRlAIAAGgkPvroI504cUIjRoyQJJWUlEiSAgMDneICAwPNsZKSEnl6esrX1/eCMQEBAdWOFxAQ4BRz7nF8fX3l6elpxtRk/Pjxstvt5nLw4MHLOGMAANCU8fM9AACARmLRokUaNGiQ02wlSbJYLE6fDcOotu5c58bUFH8lMeeyWq2yWq0XzAUAAFydmCkFAADQCOzfv19r167VY489Zq4LCgqSpGozlY4ePWrOagoKClJFRYVKS0svGHPkyJFqxzx27JhTzLnHKS0tVWVlZbUZVAAAAJeCphQAAEAj8N577ykgIECDBw8213Xo0EFBQUHmG/mkH587lZ+fr969e0uSIiIi5OHh4RRTXFysHTt2mDFRUVGy2+3asmWLGbN582bZ7XanmB07dqi4uNiMycvLk9VqVURERN2cNAAAaNL4+R4AAEADd/bsWb333nt6+OGH5e7+/8s3i8WitLQ0ZWZmKjQ0VKGhocrMzFSLFi2UmJgoSbLZbBo5cqTS09Pl5+en1q1bKyMjQ126dDHfxtepUycNHDhQycnJmj9/viRp1KhRio+PV1hYmCQpNjZWnTt3VlJSkqZPn67jx48rIyNDycnJ533zHgAAwIXQlAIAAGjg1q5dqwMHDujRRx+tNjZu3DiVl5crJSVFpaWlioyMVF5enlq2bGnGzJ49W+7u7ho+fLjKy8vVv39/LV68WG5ubmbMsmXLlJqaar6lLyEhQVlZWea4m5ubVq9erZSUFPXp00deXl5KTEzUjBkz6vDMAQBAU2YxDMOo7ySakrKyMtlsNtntdr41BNDkzV6zt1b393RMx1rdH3A5+BvuGlxnAAB+1Nhq6br4G84zpQAAAAAAAOByNKUAAAAAAADgcjSlAAAAAAAA4HI0pQAAAAAAAOByNKUAAAAAAADgcjSlAAAAAAAA4HI0pQAAAAAAAOByNKUAAAAAAADgcjSlAAAAAAAA4HL13pQ6c+aMXnzxRXXo0EFeXl664YYb9Oqrr+rs2bNmjGEYmjhxooKDg+Xl5aW+fftq586dTvtxOBwaO3as/P395e3trYSEBB06dMgpprS0VElJSbLZbLLZbEpKStKJEyecYg4cOKAhQ4bI29tb/v7+Sk1NVUVFRZ2dPwAAAAAAwNWo3ptSU6dO1dtvv62srCzt3r1b06ZN0/Tp0zV37lwzZtq0aZo1a5aysrK0detWBQUFKSYmRidPnjRj0tLStHLlSuXk5KigoECnTp1SfHy8qqqqzJjExEQVFRUpNzdXubm5KioqUlJSkjleVVWlwYMH6/Tp0yooKFBOTo5WrFih9PR011wMAAAAAACAq4R7fSewceNG3XPPPRo8eLAkqX379vrggw+0bds2ST/OkpozZ44mTJigYcOGSZKWLFmiwMBALV++XKNHj5bdbteiRYu0dOlSDRgwQJKUnZ2tkJAQrV27VnFxcdq9e7dyc3O1adMmRUZGSpIWLlyoqKgo7dmzR2FhYcrLy9OuXbt08OBBBQcHS5JmzpypESNGaPLkyfLx8XH15QEAAAAAAGiS6n2m1B133KHPP/9ce/fulST94x//UEFBgX71q19Jkvbt26eSkhLFxsaa21itVkVHR2vDhg2SpMLCQlVWVjrFBAcHKzw83IzZuHGjbDab2ZCSpF69eslmsznFhIeHmw0pSYqLi5PD4VBhYWGN+TscDpWVlTktAAAAAAAAuLB6nyn13HPPyW636+abb5abm5uqqqo0efJk/e53v5MklZSUSJICAwOdtgsMDNT+/fvNGE9PT/n6+laL+Wn7kpISBQQEVDt+QECAU8y5x/H19ZWnp6cZc64pU6Zo0qRJl3vaAAAAAAAAV7V6nyn14YcfKjs7W8uXL9ff//53LVmyRDNmzNCSJUuc4iwWi9NnwzCqrTvXuTE1xV9JzM+NHz9edrvdXA4ePHjBnAAAAAAAANAAZko9++yzev7553X//fdLkrp06aL9+/drypQpevjhhxUUFCTpx1lMbdq0Mbc7evSoOaspKChIFRUVKi0tdZotdfToUfXu3duMOXLkSLXjHzt2zGk/mzdvdhovLS1VZWVltRlUP7FarbJarVd6+gAAAAAAAFelep8p9f3336tZM+c03NzcdPbsWUlShw4dFBQUpDVr1pjjFRUVys/PNxtOERER8vDwcIopLi7Wjh07zJioqCjZ7XZt2bLFjNm8ebPsdrtTzI4dO1RcXGzG5OXlyWq1KiIiopbPHAAAAAAA4OpV7zOlhgwZosmTJ6tdu3a65ZZbtH37ds2aNUuPPvqopB9/TpeWlqbMzEyFhoYqNDRUmZmZatGihRITEyVJNptNI0eOVHp6uvz8/NS6dWtlZGSoS5cu5tv4OnXqpIEDByo5OVnz58+XJI0aNUrx8fEKCwuTJMXGxqpz585KSkrS9OnTdfz4cWVkZCg5OZk37wEAAAAAANSiem9KzZ07Vy+99JJSUlJ09OhRBQcHa/To0Xr55ZfNmHHjxqm8vFwpKSkqLS1VZGSk8vLy1LJlSzNm9uzZcnd31/Dhw1VeXq7+/ftr8eLFcnNzM2OWLVum1NRU8y19CQkJysrKMsfd3Ny0evVqpaSkqE+fPvLy8lJiYqJmzJjhgisBAAAAAABw9bAYhmHUdxJNSVlZmWw2m+x2O7OrADR5s9fsrdX9PR3TsVb3B1wO/oa7BtcZAIAfNbZaui7+htf7M6UAAAAAAABw9aEpBQAAAAAAAJejKQUAAAAAAACXoykFAAAAAAAAl6MpBQAAAAAAAJejKQUAAAAAAACXoykFAAAAAAAAl6MpBQAAAAAAAJejKQUAAAAAAACXoykFAAAAAAAAl6MpBQAA0MB98803evDBB+Xn56cWLVrotttuU2FhoTluGIYmTpyo4OBgeXl5qW/fvtq5c6fTPhwOh8aOHSt/f395e3srISFBhw4dcoopLS1VUlKSbDabbDabkpKSdOLECaeYAwcOaMiQIfL29pa/v79SU1NVUVFRZ+cOAACaLppSAAAADVhpaan69OkjDw8Pffrpp9q1a5dmzpypVq1amTHTpk3TrFmzlJWVpa1btyooKEgxMTE6efKkGZOWlqaVK1cqJydHBQUFOnXqlOLj41VVVWXGJCYmqqioSLm5ucrNzVVRUZGSkpLM8aqqKg0ePFinT59WQUGBcnJytGLFCqWnp7vkWgAAgKbFvb4TAAAAwPlNnTpVISEheu+998x17du3N/9vwzA0Z84cTZgwQcOGDZMkLVmyRIGBgVq+fLlGjx4tu92uRYsWaenSpRowYIAkKTs7WyEhIVq7dq3i4uK0e/du5ebmatOmTYqMjJQkLVy4UFFRUdqzZ4/CwsKUl5enXbt26eDBgwoODpYkzZw5UyNGjNDkyZPl4+NTLX+HwyGHw2F+Lisrq/VrBAAAGidmSgEAADRgH3/8sXr06KHf/va3CggIULdu3bRw4UJzfN++fSopKVFsbKy5zmq1Kjo6Whs2bJAkFRYWqrKy0ikmODhY4eHhZszGjRtls9nMhpQk9erVSzabzSkmPDzcbEhJUlxcnBwOh9PPCX9uypQp5s8BbTabQkJCauGqAACApoCmFAAAQAP23//+V/PmzVNoaKg+++wzPf7440pNTdX7778vSSopKZEkBQYGOm0XGBhojpWUlMjT01O+vr4XjAkICKh2/ICAAKeYc4/j6+srT09PM+Zc48ePl91uN5eDBw9e7iUAAABNFD/fAwAAaMDOnj2rHj16KDMzU5LUrVs37dy5U/PmzdNDDz1kxlksFqftDMOotu5c58bUFH8lMT9ntVpltVovmAcAALg6MVMKAACgAWvTpo06d+7stK5Tp046cOCAJCkoKEiSqs1UOnr0qDmrKSgoSBUVFSotLb1gzJEjR6od/9ixY04x5x6ntLRUlZWV1WZQAQAAXAxNKQAAgAasT58+2rNnj9O6vXv36vrrr5ckdejQQUFBQVqzZo05XlFRofz8fPXu3VuSFBERIQ8PD6eY4uJi7dixw4yJioqS3W7Xli1bzJjNmzfLbrc7xezYsUPFxcVmTF5enqxWqyIiImr5zAEAQFPHz/cAAAAasKefflq9e/dWZmamhg8fri1btmjBggVasGCBpB9/TpeWlqbMzEyFhoYqNDRUmZmZatGihRITEyVJNptNI0eOVHp6uvz8/NS6dWtlZGSoS5cu5tv4OnXqpIEDByo5OVnz58+XJI0aNUrx8fEKCwuTJMXGxqpz585KSkrS9OnTdfz4cWVkZCg5ObnGN+8BAABcCE0pAACABuz222/XypUrNX78eL366qvq0KGD5syZowceeMCMGTdunMrLy5WSkqLS0lJFRkYqLy9PLVu2NGNmz54td3d3DR8+XOXl5erfv78WL14sNzc3M2bZsmVKTU0139KXkJCgrKwsc9zNzU2rV69WSkqK+vTpIy8vLyUmJmrGjBkuuBIAAKCpsRiGYdR3Ek1JWVmZbDab7HY73xgCaPJmr9lbq/t7OqZjre4PuBz8DXcNrjMAAD9qbLV0XfwN55lSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAADdjEiRNlsViclqCgIHPcMAxNnDhRwcHB8vLyUt++fbVz506nfTgcDo0dO1b+/v7y9vZWQkKCDh065BRTWlqqpKQk2Ww22Ww2JSUl6cSJE04xBw4c0JAhQ+Tt7S1/f3+lpqaqoqKizs4dAAA0bTSlAAAAGrhbbrlFxcXF5vLll1+aY9OmTdOsWbOUlZWlrVu3KigoSDExMTp58qQZk5aWppUrVyonJ0cFBQU6deqU4uPjVVVVZcYkJiaqqKhIubm5ys3NVVFRkZKSkszxqqoqDR48WKdPn1ZBQYFycnK0YsUKpaenu+YiAACAJse9vhMAAADAhbm7uzvNjvqJYRiaM2eOJkyYoGHDhkmSlixZosDAQC1fvlyjR4+W3W7XokWLtHTpUg0YMECSlJ2drZCQEK1du1ZxcXHavXu3cnNztWnTJkVGRkqSFi5cqKioKO3Zs0dhYWHKy8vTrl27dPDgQQUHB0uSZs6cqREjRmjy5Mny8fGpMXeHwyGHw2F+Lisrq9VrAwAAGi9mSgEAADRwX331lYKDg9WhQwfdf//9+u9//ytJ2rdvn0pKShQbG2vGWq1WRUdHa8OGDZKkwsJCVVZWOsUEBwcrPDzcjNm4caNsNpvZkJKkXr16yWazOcWEh4ebDSlJiouLk8PhUGFh4XlznzJlivmTQJvNppCQkFq4IgAAoCmgKQUAANCARUZG6v3339dnn32mhQsXqqSkRL1799Z3332nkpISSVJgYKDTNoGBgeZYSUmJPD095evre8GYgICAascOCAhwijn3OL6+vvL09DRjajJ+/HjZ7XZzOXjw4GVeAQAA0FQ1iKbUN998owcffFB+fn5q0aKFbrvtNqdv3HiAJwAAuFoNGjRIv/71r9WlSxcNGDBAq1evlvTjz/R+YrFYnLYxDKPaunOdG1NT/JXEnMtqtcrHx8dpAQAAkBpAU6q0tFR9+vSRh4eHPv30U+3atUszZ85Uq1atzBge4AkAAPAjb29vdenSRV999ZX5nKlzZyodPXrUnNUUFBSkiooKlZaWXjDmyJEj1Y517Ngxp5hzj1NaWqrKyspqM6gAAAAuRb03paZOnaqQkBC999576tmzp9q3b6/+/fvrxhtvlFT9AZ7h4eFasmSJvv/+ey1fvlySzAd4zpw5UwMGDFC3bt2UnZ2tL7/8UmvXrpUk8wGe77zzjqKiohQVFaWFCxdq1apV2rNnjySZD/DMzs5Wt27dNGDAAM2cOVMLFy4870M5HQ6HysrKnBYAAIC64nA4tHv3brVp00YdOnRQUFCQ1qxZY45XVFQoPz9fvXv3liRFRETIw8PDKaa4uFg7duwwY6KiomS327VlyxYzZvPmzbLb7U4xO3bsUHFxsRmTl5cnq9WqiIiIOj1nAADQNNV7U+rjjz9Wjx499Nvf/lYBAQHq1q2bFi5caI439Ad48vBOAABQlzIyMpSfn699+/Zp8+bN+s1vfqOysjI9/PDDslgsSktLU2ZmplauXKkdO3ZoxIgRatGihRITEyVJNptNI0eOVHp6uj7//HNt375dDz74oPlzQEnq1KmTBg4cqOTkZG3atEmbNm1ScnKy4uPjFRYWJkmKjY1V586dlZSUpO3bt+vzzz9XRkaGkpOT+UkeAAC4IvXelPrvf/+refPmKTQ0VJ999pkef/xxpaam6v3335ekBv8ATx7eCQAA6tKhQ4f0u9/9TmFhYRo2bJg8PT21adMmXX/99ZKkcePGKS0tTSkpKerRo4e++eYb5eXlqWXLluY+Zs+eraFDh2r48OHq06ePWrRooU8++URubm5mzLJly9SlSxfFxsYqNjZWXbt21dKlS81xNzc3rV69Ws2bN1efPn00fPhwDR06VDNmzHDdxQAAAE2Ke30ncPbsWfXo0UOZmZmSpG7dumnnzp2aN2+eHnroITOuoT7A02q1ymq1XjAPAACAK5WTk3PBcYvFookTJ2rixInnjWnevLnmzp2ruXPnnjemdevWys7OvuCx2rVrp1WrVl0wBgAA4FLV+0ypNm3aqHPnzk7rOnXqpAMHDkgSD/AEAAAAAABoguq9KdWnTx/zQeM/2bt3rzklnQd4AgAAAAAAND31/vO9p59+Wr1791ZmZqaGDx+uLVu2aMGCBVqwYIEkOT3AMzQ0VKGhocrMzDzvAzz9/PzUunVrZWRknPcBnvPnz5ckjRo16rwP8Jw+fbqOHz/OAzwBAAAAAADqQL03pW6//XatXLlS48eP16uvvqoOHTpozpw5euCBB8yYcePGqby8XCkpKSotLVVkZGSND/B0d3fX8OHDVV5erv79+2vx4sXVHuCZmppqvqUvISFBWVlZ5vhPD/BMSUlRnz595OXlpcTERB7gCQAAAAAAUMsshmEY9Z1EU1JWViabzSa73c7sKgBN3uw1e2t1f0/HdKzV/QGXg7/hrsF1BgDgR42tlq6Lv+H1/kwpAAAAAAAAXH1oSgEAAAAAAMDlaEoBAAAAAADA5WhKAQAAAAAAwOVoSgEAAAAAAMDlaEoBAAAAAADA5WhKAQAAAAAAwOVoSgEAAAAAAMDlaEoBAAAAAADA5WhKAQAAAAAAwOVoSgEAAAAAAMDlaEoBAAAAAADA5WhKAQAAAAAAwOVoSgEAAAAAAMDlrqgp5ebmpi1bttQ4VlhYKDc3t1+UFAAAQGNGrQQAAHBxV9SUMgzjvGNnz56VxWK54oQAAAAaO2olAACAi7vin++dr5gqLCyUzWa74oQAAACaAmolAACAC3O/1MA33nhDb7zxhqQfi6yhQ4fKarU6xZSXl+vo0aP6zW9+U7tZAgAANHDUSgAAAJfnkptSAQEBuuWWWyRJX3/9tW644Qa1atXKKcZqtapLly566qmnajVJAACAho5aCbg6zV6zt9b3+XRMx1rfJwA0RJfclPrd736n3/3ud5Kkfv36ad68ebr55pvrLDEAAIDGhFoJAADg8lxyU+rn1q1bV9t5AAAANBnUSgAAABd3RU0p6ce3ymzdulX79+9XeXl5tfGHHnroFyUGAADQmFErAQAAXNgVNaX27t2rhIQEffXVVzW+8thisVBoAQCAqxa1EgAAwMVdUVNqzJgx+uGHH/Thhx+qa9eu1d4sAwAAcDWjVgIAALi4K2pKbdmyRQsXLuR1xgAAADWgVgIAALi4Zley0TXXXCMfH5/azgUAAKBJqKtaacqUKbJYLEpLSzPXGYahiRMnKjg4WF5eXurbt6927tzptJ3D4dDYsWPl7+8vb29vJSQk6NChQ04xpaWlSkpKks1mk81mU1JSkk6cOOEUc+DAAQ0ZMkTe3t7y9/dXamqqKioqav08AQDA1eGKmlKPPPKIli9fXtu5AAAANAl1UStt3bpVCxYsUNeuXZ3WT5s2TbNmzVJWVpa2bt2qoKAgxcTE6OTJk2ZMWlqaVq5cqZycHBUUFOjUqVOKj49XVVWVGZOYmKiioiLl5uYqNzdXRUVFSkpKMserqqo0ePBgnT59WgUFBcrJydGKFSuUnp5eq+cJAACuHlf0873w8HB98MEHSkhI0JAhQ+Tn51ctZtiwYb84OQAAgMaotmulU6dO6YEHHtDChQv12muvmesNw9CcOXM0YcIEc39LlixRYGCgli9frtGjR8tut2vRokVaunSpBgwYIEnKzs5WSEiI1q5dq7i4OO3evVu5ubnatGmTIiMjJUkLFy5UVFSU9uzZo7CwMOXl5WnXrl06ePCggoODJUkzZ87UiBEjNHny5PPODHM4HHI4HObnsrKySz5vAADQtF1RUyoxMVGStG/fPq1atarauMVicfrmDQAA4GpS27XSmDFjNHjwYA0YMMCpKbVv3z6VlJQoNjbWXGe1WhUdHa0NGzZo9OjRKiwsVGVlpVNMcHCwwsPDtWHDBsXFxWnjxo2y2WxmQ0qSevXqJZvNpg0bNigsLEwbN25UeHi42ZCSpLi4ODkcDhUWFqpfv3415j5lyhRNmjTpks8VAABcPa6oKbVu3brazgMAAKDJqM1aKScnR3//+9+1devWamMlJSWSpMDAQKf1gYGB2r9/vxnj6ekpX1/fajE/bV9SUqKAgIBq+w8ICHCKOfc4vr6+8vT0NGNqMn78eD3zzDPm57KyMoWEhJw3HgAAXD2uqCkVHR1d23kAAAA0GbVVKx08eFBPPfWU8vLy1Lx58/PGWSwWp8+GYVRbd65zY2qKv5KYc1mtVlmt1gvmAgAArk5X9KBzAAAA1L3CwkIdPXpUERERcnd3l7u7u/Lz8/WHP/xB7u7u5sylc2cqHT161BwLCgpSRUWFSktLLxhz5MiRasc/duyYU8y5xyktLVVlZWW1GVQAAACX4opmSt19990XHLdYLPr888+vKCEAAIDGrrZqpf79++vLL790WvfII4/o5ptv1nPPPacbbrhBQUFBWrNmjbp16yZJqqioUH5+vqZOnSpJioiIkIeHh9asWaPhw4dLkoqLi7Vjxw5NmzZNkhQVFSW73a4tW7aoZ8+ekqTNmzfLbrerd+/eZszkyZNVXFysNm3aSJLy8vJktVoVERFxqZcGAADAdEVNqbNnz1abpv3tt99qz549CggIUMeOHWslOQAAgMaotmqlli1bKjw83Gmdt7e3/Pz8zPVpaWnKzMxUaGioQkNDlZmZqRYtWpgPW7fZbBo5cqTS09Pl5+en1q1bKyMjQ126dDHfxtepUycNHDhQycnJmj9/viRp1KhRio+PV1hYmCQpNjZWnTt3VlJSkqZPn67jx48rIyNDycnJ533zHgAAwIVcUVNq/fr1Na7fu3ev7rnnHr3yyiu/JCcAAIBGzZW10rhx41ReXq6UlBSVlpYqMjJSeXl5atmypRkze/Zsubu7a/jw4SovL1f//v21ePFiubm5mTHLli1Tamqq+Za+hIQEZWVlmeNubm5avXq1UlJS1KdPH3l5eSkxMVEzZsyotXMBAABXF4thGEZt7vDdd9/V/PnztXnz5trcbaNRVlYmm80mu93Ot4YAmrzZa/bW6v6ejmGmLeqPq/6GUytRK6Fpqe2/hRJ/D4GrRWOrpevib3itP+i8ffv22rFjR23vFgAAoEmgVgIAAPhRrTelVqxYoeDg4NreLQAAQJNArQQAAPCjK3qm1KOPPlptncPh0D//+U/t2rXLfJMLAADA1YhaCQAA4OKuqCn1xRdfVHujTPPmzdW+fXuNHz/efNsLAADA1YhaCQAA4OKuqCn19ddf13IaAAAATQe1EgAAwMXV+jOlAAAAAAAAgIu5oplSknT8+HHNnj1bn3/+ub777jv5+/trwIABSktLk6+vb23mCAAA0OhQKwEAAFzYFc2U+uabb9S9e3dNnjxZdrtd7dq104kTJ/T73/9e3bt31+HDh2s7TwAAgEaDWgkAAODirqgp9cILL6i8vFybN2/Wzp07tWbNGu3cuVObN29WeXm5XnjhhdrOEwAAoNGgVgIAALi4K2pK5ebm6rXXXtPtt9/utP7222/Xq6++qk8//bRWkgMAAGiMqJUAAAAu7oqaUna7Xe3bt69xrEOHDrLb7b8kJwAAgEaNWgkAAODirqgp1aFDB61evbrGsU8//VQdOnT4RUkBAAA0ZtRKAAAAF3dFb9975JFH9Pzzz+vs2bN6+OGH1aZNGxUXFys7O1tz587V66+/Xtt5AgAANBrUSgAAABd3RU2pZ599Vv/5z3+UlZWlN99801xvGIZGjRqljIyMWksQAACgsaFWAgAAuLgrakpZLBbNnz9fzzzzjNatW6fvvvtOfn5+uvvuu9WxY8fazhEAAKBRoVYCAAC4uEt+plRpaal+/etfa9WqVea6sLAwPf7445owYYIef/xx7d27V7/+9a/13Xff1UmyAAAADRW1EgAAwOW55KbUO++8o3/84x8aOHDgeWMGDhyoL7/80mmaOgAAwNWAWgkAAODyXHJTKicnR8nJyXJ3P/8v/tzd3ZWcnKyPP/64VpIDAABoLKiVAAAALs8lN6X27t2rHj16XDSue/fu2rt37y9KCgAAoLGhVgIAALg8l9yUOnPmjDw8PC4a5+HhocrKyl+UFAAAQGNDrQQAAHB5Lrkp1aZNG+3ateuicTt37lRQUNAVJTNlyhRZLBalpaWZ6wzD0MSJExUcHCwvLy/17dtXO3fudNrO4XBo7Nix8vf3l7e3txISEnTo0CGnmNLSUiUlJclms8lmsykpKUknTpxwijlw4ICGDBkib29v+fv7KzU1VRUVFVd0LgAA4OriiloJAACgKbnkplR0dLTeeuutC36zV1lZqXnz5qlfv36XncjWrVu1YMECde3a1Wn9tGnTNGvWLGVlZWnr1q0KCgpSTEyMTp48acakpaVp5cqVysnJUUFBgU6dOqX4+HhVVVWZMYmJiSoqKlJubq5yc3NVVFSkpKQkc7yqqkqDBw/W6dOnVVBQoJycHK1YsULp6emXfS4AAODqU9e1EgAAQFNzyU2pp59+Wv/6179077336vDhw9XGDx8+rKFDh2rPnj16+umnLyuJU6dO6YEHHtDChQvl6+trrjcMQ3PmzNGECRM0bNgwhYeHa8mSJfr++++1fPlySZLdbteiRYs0c+ZMDRgwQN26dVN2dra+/PJLrV27VpK0e/du5ebm6p133lFUVJSioqK0cOFCrVq1Snv27JEk5eXladeuXcrOzla3bt00YMAAzZw5UwsXLlRZWdl5c3c4HCorK3NaAADA1acuayUAAICm6JKbUl27dtWbb76pzz77TB06dFDv3r31wAMP6IEHHlDv3r3VoUMH5eXl6c0331SXLl0uK4kxY8Zo8ODBGjBggNP6ffv2qaSkRLGxseY6q9Wq6OhobdiwQZJUWFioyspKp5jg4GCFh4ebMRs3bpTNZlNkZKQZ06tXL9lsNqeY8PBwBQcHmzFxcXFyOBwqLCw8b+5TpkwxfxJos9kUEhJyWecOAACahrqslQAAAJqi87+zuAbJyckKDw9XZmam1q1bp02bNkmSWrRooYEDB2r8+PHq1avXZSWQk5Ojv//979q6dWu1sZKSEklSYGCg0/rAwEDt37/fjPH09HSaYfVTzE/bl5SUKCAgoNr+AwICnGLOPY6vr688PT3NmJqMHz9ezzzzjPm5rKyMxhQAAFepuqiVAAAAmqrLakpJUlRUlD755BOdPXtW3377rSTJ399fzZpd8qQr08GDB/XUU08pLy9PzZs3P2+cxWJx+mwYRrV15zo3pqb4K4k5l9VqldVqvWAuAADg6lGbtRIAAEBTdsXVUbNmzRQQEKCAgIArLrIKCwt19OhRRUREyN3dXe7u7srPz9cf/vAHubu7mzOXzp2pdPToUXMsKChIFRUVKi0tvWDMkSNHqh3/2LFjTjHnHqe0tFSVlZXVZlABAABcTG3USgAAAE1ZvVZI/fv315dffqmioiJz6dGjhx544AEVFRXphhtuUFBQkNasWWNuU1FRofz8fPXu3VuSFBERIQ8PD6eY4uJi7dixw4yJioqS3W7Xli1bzJjNmzfLbrc7xezYsUPFxcVmTF5enqxWqyIiIur0OgAAAAAAAFxtLvvne7WpZcuWCg8Pd1rn7e0tPz8/c31aWpoyMzMVGhqq0NBQZWZmqkWLFkpMTJQk2Ww2jRw5Uunp6fLz81Pr1q2VkZGhLl26mA9O79SpkwYOHKjk5GTNnz9fkjRq1CjFx8crLCxMkhQbG6vOnTsrKSlJ06dP1/Hjx5WRkaHk5GT5+Pi46pIAAAAAAABcFeq1KXUpxo0bp/LycqWkpKi0tFSRkZHKy8tTy5YtzZjZs2fL3d1dw4cPV3l5ufr376/FixfLzc3NjFm2bJlSU1PNt/QlJCQoKyvLHHdzc9Pq1auVkpKiPn36yMvLS4mJiZoxY4brThYAAAAAAOAq0eAecLB+/XrNmTPH/GyxWDRx4kQVFxfrhx9+UH5+frXZVc2bN9fcuXP13Xff6fvvv9cnn3xS7Q14rVu3VnZ2tsrKylRWVqbs7Gy1atXKKaZdu3ZatWqVvv/+e3333XeaO3cuDzEHAAD1at68eeratat8fHzk4+OjqKgoffrpp+a4YRiaOHGigoOD5eXlpb59+2rnzp1O+3A4HBo7dqz8/f3l7e2thIQEHTp0yCmmtLRUSUlJstlsstlsSkpK0okTJ5xiDhw4oCFDhsjb21v+/v5KTU1VRUVFnZ07AABo2hpcUwoAAAD/X9u2bfX6669r27Zt2rZtm+6++27dc889ZuNp2rRpmjVrlrKysrR161YFBQUpJiZGJ0+eNPeRlpamlStXKicnRwUFBTp16pTi4+NVVVVlxiQmJqqoqEi5ubnKzc1VUVGRkpKSzPGqqioNHjxYp0+fVkFBgXJycrRixQqlp6e77mIAAIAmpcH/fA8AAOBqNmTIEKfPkydP1rx587Rp0yZ17txZc+bM0YQJEzRs2DBJ0pIlSxQYGKjly5dr9OjRstvtWrRokZYuXWo+bzM7O1shISFau3at4uLitHv3buXm5mrTpk2KjIyUJC1cuFBRUVHas2ePwsLClJeXp127dungwYMKDg6WJM2cOVMjRozQ5MmTz/sMTofDIYfDYX4uKyur9WsEAAAaJ2ZKAQAANBJVVVXKycnR6dOnFRUVpX379qmkpMR8ZqYkWa1WRUdHa8OGDZKkwsJCVVZWOsUEBwcrPDzcjNm4caNsNpvZkJKkXr16yWazOcWEh4ebDSlJiouLk8PhUGFh4XlznjJlivmTQJvNVu0RCwAA4OpFUwoAAKCB+/LLL3XNNdfIarXq8ccf18qVK9W5c2eVlJRIkgIDA53iAwMDzbGSkhJ5enrK19f3gjEBAQHVjhsQEOAUc+5xfH195enpacbUZPz48bLb7eZy8ODByzx7AADQVPHzPQAAgAYuLCxMRUVFOnHihFasWKGHH35Y+fn55rjFYnGKNwyj2rpznRtTU/yVxJzLarXy4hgAAFAjZkoBAAA0cJ6enrrpppvUo0cPTZkyRbfeeqveeOMNBQUFSVK1mUpHjx41ZzUFBQWpoqJCpaWlF4w5cuRIteMeO3bMKebc45SWlqqysrLaDCoAAIBLQVMKAACgkTEMQw6HQx06dFBQUJDWrFljjlVUVCg/P1+9e/eWJEVERMjDw8Mppri4WDt27DBjoqKiZLfbtWXLFjNm8+bNstvtTjE7duxQcXGxGZOXlyer1aqIiIg6PV8AANA08fM9AACABuyFF17QoEGDFBISopMnTyonJ0fr169Xbm6uLBaL0tLSlJmZqdDQUIWGhiozM1MtWrRQYmKiJMlms2nkyJFKT0+Xn5+fWrdurYyMDHXp0sV8G1+nTp00cOBAJScna/78+ZKkUaNGKT4+XmFhYZKk2NhYde7cWUlJSZo+fbqOHz+ujIwMJScnn/fNewAAABdCUwoAAKABO3LkiJKSklRcXCybzaauXbsqNzdXMTExkqRx48apvLxcKSkpKi0tVWRkpPLy8tSyZUtzH7Nnz5a7u7uGDx+u8vJy9e/fX4sXL5abm5sZs2zZMqWmpppv6UtISFBWVpY57ubmptWrVyslJUV9+vSRl5eXEhMTNWPGDBddCQAA0NRYDMMw6juJpqSsrEw2m012u51vDQE0ebPX7K3V/T0d07FW99eUcK3rHn/DXYPrjKamtv/7LPHfaOBq0djqu7r4G84zpQAAAAAAAOByNKUAAAAAAADgcjSlAAAAAAAA4HI0pQAAAAAAAOByNKUAAAAAAADgcjSlAAAAAAAA4HI0pQAAAAAAAOByNKUAAAAAAADgcjSlAAAAAAAA4HI0pQAAAAAAAOByNKUAAAAAAADgcjSlAAAAAAAA4HI0pQAAAAAAAOByNKUAAAAAAADgcjSlAAAAAAAA4HI0pQAAAAAAAOByNKUAAAAAAADgcjSlAAAAAAAA4HI0pQAAAAAAAOByNKUAAAAAAADgcjSlAAAAAAAA4HI0pQAAAAAAAOByNKUAAAAAAADgcjSlAAAAAAAA4HI0pQAAAAAAAOByNKUAAAAAAADgcjSlAAAAAAAA4HI0pQAAAAAAAOByNKUAAAAAAADgcjSlAAAAAAAA4HI0pQAAABqwKVOm6Pbbb1fLli0VEBCgoUOHas+ePU4xhmFo4sSJCg4OlpeXl/r27audO3c6xTgcDo0dO1b+/v7y9vZWQkKCDh065BRTWlqqpKQk2Ww22Ww2JSUl6cSJE04xBw4c0JAhQ+Tt7S1/f3+lpqaqoqKiTs4dAAA0bTSlAAAAGrD8/HyNGTNGmzZt0po1a3TmzBnFxsbq9OnTZsy0adM0a9YsZWVlaevWrQoKClJMTIxOnjxpxqSlpWnlypXKyclRQUGBTp06pfj4eFVVVZkxiYmJKioqUm5urnJzc1VUVKSkpCRzvKqqSoMHD9bp06dVUFCgnJwcrVixQunp6a65GAAAoElxr+8EAAAAcH65ublOn9977z0FBASosLBQd911lwzD0Jw5czRhwgQNGzZMkrRkyRIFBgZq+fLlGj16tOx2uxYtWqSlS5dqwIABkqTs7GyFhIRo7dq1iouL0+7du5Wbm6tNmzYpMjJSkrRw4UJFRUVpz549CgsLU15ennbt2qWDBw8qODhYkjRz5kyNGDFCkydPlo+PT7X8HQ6HHA6H+bmsrKxOrhMAAGh8mCkFAADQiNjtdklS69atJUn79u1TSUmJYmNjzRir1aro6Ght2LBBklRYWKjKykqnmODgYIWHh5sxGzdulM1mMxtSktSrVy/ZbDanmPDwcLMhJUlxcXFyOBwqLCysMd8pU6aYPwe02WwKCQmpjcsAAACaAJpSAAAAjYRhGHrmmWd0xx13KDw8XJJUUlIiSQoMDHSKDQwMNMdKSkrk6ekpX1/fC8YEBARUO2ZAQIBTzLnH8fX1laenpxlzrvHjx8tut5vLwYMHL/e0AQBAE8XP9wAAABqJJ598Uv/85z9VUFBQbcxisTh9Ngyj2rpznRtTU/yVxPyc1WqV1Wq9YB4AAODqxEwpAACARmDs2LH6+OOPtW7dOrVt29ZcHxQUJEnVZiodPXrUnNUUFBSkiooKlZaWXjDmyJEj1Y577Ngxp5hzj1NaWqrKyspqM6gAAAAuhqYUAABAA2YYhp588kn9+c9/1hdffKEOHTo4jXfo0EFBQUFas2aNua6iokL5+fnq3bu3JCkiIkIeHh5OMcXFxdqxY4cZExUVJbvdri1btpgxmzdvlt1ud4rZsWOHiouLzZi8vDxZrVZFRETU/skDAIAmjZ/vAQAANGBjxozR8uXL9b//+79q2bKlOVPJZrPJy8tLFotFaWlpyszMVGhoqEJDQ5WZmakWLVooMTHRjB05cqTS09Pl5+en1q1bKyMjQ126dDHfxtepUycNHDhQycnJmj9/viRp1KhRio+PV1hYmCQpNjZWnTt3VlJSkqZPn67jx48rIyNDycnJNb55DwAA4EJoSgEAADRg8+bNkyT17dvXaf17772nESNGSJLGjRun8vJypaSkqLS0VJGRkcrLy1PLli3N+NmzZ8vd3V3Dhw9XeXm5+vfvr8WLF8vNzc2MWbZsmVJTU8239CUkJCgrK8scd3Nz0+rVq5WSkqI+ffrIy8tLiYmJmjFjRh2dPQAAaMpoSgEAADRghmFcNMZisWjixImaOHHieWOaN2+uuXPnau7cueeNad26tbKzsy94rHbt2mnVqlUXzQkAAOBieKYUAAAAAAAAXI6mFAAAAAAAAFyOphQAAAAAAABcjqYUAAAAAAAAXK7em1JTpkzR7bffrpYtWyogIEBDhw7Vnj17nGIMw9DEiRMVHBwsLy8v9e3bVzt37nSKcTgcGjt2rPz9/eXt7a2EhAQdOnTIKaa0tFRJSUmy2Wyy2WxKSkrSiRMnnGIOHDigIUOGyNvbW/7+/kpNTVVFRUWdnDsAAAAAAMDVqt6bUvn5+RozZow2bdqkNWvW6MyZM4qNjdXp06fNmGnTpmnWrFnKysrS1q1bFRQUpJiYGJ08edKMSUtL08qVK5WTk6OCggKdOnVK8fHxqqqqMmMSExNVVFSk3Nxc5ebmqqioSElJSeZ4VVWVBg8erNOnT6ugoEA5OTlasWKF0tPTXXMxAAAAAAAArhLu9Z1Abm6u0+f33ntPAQEBKiws1F133SXDMDRnzhxNmDBBw4YNkyQtWbJEgYGBWr58uUaPHi273a5FixZp6dKlGjBggCQpOztbISEhWrt2reLi4rR7927l5uZq06ZNioyMlCQtXLhQUVFR2rNnj8LCwpSXl6ddu3bp4MGDCg4OliTNnDlTI0aM0OTJk+Xj4+PCKwMAAAAAQNMxe83eWt3f0zEda3V/cL16nyl1LrvdLklq3bq1JGnfvn0qKSlRbGysGWO1WhUdHa0NGzZIkgoLC1VZWekUExwcrPDwcDNm48aNstlsZkNKknr16iWbzeYUEx4ebjakJCkuLk4Oh0OFhYU15utwOFRWVua0AAAAAAAA4MIaVFPKMAw988wzuuOOOxQeHi5JKikpkSQFBgY6xQYGBppjJSUl8vT0lK+v7wVjAgICqh0zICDAKebc4/j6+srT09OMOdeUKVPMZ1TZbDaFhIRc7mkDAAAAAABcdRpUU+rJJ5/UP//5T33wwQfVxiwWi9NnwzCqrTvXuTE1xV9JzM+NHz9edrvdXA4ePHjBnAAAAAAAANCAmlJjx47Vxx9/rHXr1qlt27bm+qCgIEmqNlPp6NGj5qymoKAgVVRUqLS09IIxR44cqXbcY8eOOcWce5zS0lJVVlZWm0H1E6vVKh8fH6cFAAAAAAAAF1bvTSnDMPTkk0/qz3/+s7744gt16NDBabxDhw4KCgrSmjVrzHUVFRXKz89X7969JUkRERHy8PBwiikuLtaOHTvMmKioKNntdm3ZssWM2bx5s+x2u1PMjh07VFxcbMbk5eXJarUqIiKi9k8eAAAAAADgKlXvb98bM2aMli9frv/93/9Vy5YtzZlKNptNXl5eslgsSktLU2ZmpkJDQxUaGqrMzEy1aNFCiYmJZuzIkSOVnp4uPz8/tW7dWhkZGerSpYv5Nr5OnTpp4MCBSk5O1vz58yVJo0aNUnx8vMLCwiRJsbGx6ty5s5KSkjR9+nQdP35cGRkZSk5OZgYUAAAAAABALar3ptS8efMkSX379nVa/95772nEiBGSpHHjxqm8vFwpKSkqLS1VZGSk8vLy1LJlSzN+9uzZcnd31/Dhw1VeXq7+/ftr8eLFcnNzM2OWLVum1NRU8y19CQkJysrKMsfd3Ny0evVqpaSkqE+fPvLy8lJiYqJmzJhRR2cPAAAAAABwdar3ppRhGBeNsVgsmjhxoiZOnHjemObNm2vu3LmaO3fueWNat26t7OzsCx6rXbt2WrVq1UVzAgAAAAAAwJWr96YUANSF2Wv21ur+no7pWKv7AwAAAICrXb0/6BwAAAAAAABXH5pSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAAAAAAcDmaUgAAAA3YX//6Vw0ZMkTBwcGyWCz66KOPnMYNw9DEiRMVHBwsLy8v9e3bVzt37nSKcTgcGjt2rPz9/eXt7a2EhAQdOnTIKaa0tFRJSUmy2Wyy2WxKSkrSiRMnnGIOHDigIUOGyNvbW/7+/kpNTVVFRUVdnDYAALgK0JQCAABowE6fPq1bb71VWVlZNY5PmzZNs2bNUlZWlrZu3aqgoCDFxMTo5MmTZkxaWppWrlypnJwcFRQU6NSpU4qPj1dVVZUZk5iYqKKiIuXm5io3N1dFRUVKSkoyx6uqqjR48GCdPn1aBQUFysnJ0YoVK5Senl53Jw8AAJo09/pOAAAAAOc3aNAgDRo0qMYxwzA0Z84cTZgwQcOGDZMkLVmyRIGBgVq+fLlGjx4tu92uRYsWaenSpRowYIAkKTs7WyEhIVq7dq3i4uK0e/du5ebmatOmTYqMjJQkLVy4UFFRUdqzZ4/CwsKUl5enXbt26eDBgwoODpYkzZw5UyNGjNDkyZPl4+NTY44Oh0MOh8P8XFZWVmvXBgAANG40pQAAABqpffv2qaSkRLGxseY6q9Wq6OhobdiwQaNHj1ZhYaEqKyudYoKDgxUeHq4NGzYoLi5OGzdulM1mMxtSktSrVy/ZbDZt2LBBYWFh2rhxo8LDw82GlCTFxcXJ4XCosLBQ/fr1qzHHKVOmaNKkSXVw9gDQ8M1es7dW9/d0TMda3R9Q3/j5HgAAQCNVUlIiSQoMDHRaHxgYaI6VlJTI09NTvr6+F4wJCAiotv+AgACnmHOP4+vrK09PTzOmJuPHj5fdbjeXgwcPXuZZAgCApoqZUgAAAI2cxWJx+mwYRrV15zo3pqb4K4k5l9VqldVqvWAuAADg6sRMKQAAgEYqKChIkqrNVDp69Kg5qykoKEgVFRUqLS29YMyRI0eq7f/YsWNOMecep7S0VJWVldVmUAEAAFwKmlIAAACNVIcOHRQUFKQ1a9aY6yoqKpSfn6/evXtLkiIiIuTh4eEUU1xcrB07dpgxUVFRstvt2rJlixmzefNm2e12p5gdO3aouLjYjMnLy5PValVERESdnicAAGia+PkeAABAA3bq1Cn9+9//Nj/v27dPRUVFat26tdq1a6e0tDRlZmYqNDRUoaGhyszMVIsWLZSYmChJstlsGjlypNLT0+Xn56fWrVsrIyNDXbp0Md/G16lTJw0cOFDJycmaP3++JGnUqFGKj49XWFiYJCk2NladO3dWUlKSpk+fruPHjysjI0PJycnnffMeAADAhdCUAgAAaMC2bdvm9Ga7Z555RpL08MMPa/HixRo3bpzKy8uVkpKi0tJSRUZGKi8vTy1btjS3mT17ttzd3TV8+HCVl5erf//+Wrx4sdzc3MyYZcuWKTU11XxLX0JCgrKyssxxNzc3rV69WikpKerTp4+8vLyUmJioGTNm1PUlAAAATRRNKQAAgAasb9++MgzjvOMWi0UTJ07UxIkTzxvTvHlzzZ07V3Pnzj1vTOvWrZWdnX3BXNq1a6dVq1ZdNGcAAIBLwTOlAAAAAAAA4HI0pQAAAAAAAOByNKUAAAAAAADgcjSlAAAAAAAA4HI0pQAAAAAAAOByNKUAAAAAAADgcjSlAAAAAAAA4HLu9Z0AmrbZa/bW+j6fjulYJ/v+ab8AAAAAAKDu0ZRqRGjCAAAAAACApoKf7wEAAAAAAMDlaEoBAAAAAADA5fj5HgAAAAAAQA14jE7dYqYUAAAAAAAAXI6mFAAAAAAAAFyOn+8BAIA6wXR3AAAAXAhNKcCF+AcaAAAAAAA/4ud7AAAAAAAAcDlmSkESM3jOxfUAAAAAAKBuMVMKAAAAAAAALkdTCgAAAAAAAC5HUwoAAAAAAAAuxzOlgCaC52ABAAAAABoTZkoBAAAAAADA5ZgpBeCCansGlsQsLAAAAAAAM6UAAAAAAABQD5gpBaDe8BwsAAAAALh6MVMKAAAAAAAALsdMKQC4DHU5u4uZYwAAAACuJjSlAOAqQMMLAAAAQENDUwoA0CDRSAMAAACaNppSNXjrrbc0ffp0FRcX65ZbbtGcOXN055131ndaAIBawE8wgV+OWgkAANQGmlLn+PDDD5WWlqa33npLffr00fz58zVo0CDt2rVL7dq1q+/0AACoVTTScLmolQAAQG2hKXWOWbNmaeTIkXrsscckSXPmzNFnn32mefPmacqUKdXiHQ6HHA6H+dlut0uSysrKaj23H06fqtX9/TzHutp3be+3LvfdGK9HY8y5LvfdGK+HK3Kuy32TM/fwfPut633XhZ/2bxhGnR6nsWvItRJQH+qyVkLj19j+FkrSm1/8u1b3N+bum8z/uy6vR13lTa1Uff+1WisZMDkcDsPNzc3485//7LQ+NTXVuOuuu2rc5pVXXjEksbCwsLCwsDSR5eDBg64oOxolaiUWFhYWFhaW2qyVmCn1M99++62qqqoUGBjotD4wMFAlJSU1bjN+/Hg988wz5uezZ8/q+PHj8vPzk8ViqdN8G6OysjKFhITo4MGD8vHxqe90IO5JQ8P9aHi4Jw1LXd4PwzB08uRJBQcH1+p+mxJqpbrHf3MaHu5Jw8L9aHi4Jw1LY6uVaErV4NwCyTCM8xZNVqtVVqvVaV2rVq3qKrUmw8fHh/9gNTDck4aF+9HwcE8alrq6Hzabrdb32RRRK9U9/pvT8HBPGhbuR8PDPWlYGkut1KxW99bI+fv7y83Nrdo3fUePHq32jSAAAMDVhloJAADUJppSP+Pp6amIiAitWbPGaf2aNWvUu3fvesoKAACgYaBWAgAAtYmf753jmWeeUVJSknr06KGoqCgtWLBABw4c0OOPP17fqTUJVqtVr7zySrVp/Kg/3JOGhfvR8HBPGhbuR/2jVqpb/P94w8M9aVi4Hw0P96RhaWz3w2IYvPf4XG+99ZamTZum4uJihYeHa/bs2brrrrvqOy0AAIAGgVoJAADUBppSAAAAAAAAcDmeKQUAAAAAAACXoykFAAAAAAAAl6MpBQAAAAAAAJejKQUAAAAAAACXoymFWvfWW2+pQ4cOat68uSIiIvS3v/3tvLF//vOfFRMTo2uvvVY+Pj6KiorSZ5995sJsrw6Xc09+7v/+7//k7u6u2267rW4TvMpc7v1wOByaMGGCrr/+elmtVt1444169913XZTt1eFy78myZct06623qkWLFmrTpo0eeeQRfffddy7Ktmn761//qiFDhig4OFgWi0UfffTRRbfJz89XRESEmjdvrhtuuEFvv/123ScK/ALUSg0LdVLDQ63U8FArNRxNrVaiKYVa9eGHHyotLU0TJkzQ9u3bdeedd2rQoEE6cOBAjfF//etfFRMTo7/85S8qLCxUv379NGTIEG3fvt3FmTddl3tPfmK32/XQQw+pf//+Lsr06nAl92P48OH6/PPPtWjRIu3Zs0cffPCBbr75Zhdm3bRd7j0pKCjQQw89pJEjR2rnzp364x//qK1bt+qxxx5zceZN0+nTp3XrrbcqKyvrkuL37dunX/3qV7rzzju1fft2vfDCC0pNTdWKFSvqOFPgylArNSzUSQ0PtVLDQ63UsDS5WskAalHPnj2Nxx9/3GndzTffbDz//POXvI/OnTsbkyZNqu3UrlpXek/uu+8+48UXXzReeeUV49Zbb63DDK8ul3s/Pv30U8NmsxnfffedK9K7Kl3uPZk+fbpxww03OK37wx/+YLRt27bOcrxaSTJWrlx5wZhx48YZN998s9O60aNHG7169arDzIArR63UsFAnNTzUSg0PtVLD1RRqJWZKodZUVFSosLBQsbGxTutjY2O1YcOGS9rH2bNndfLkSbVu3bouUrzqXOk9ee+99/Sf//xHr7zySl2neFW5kvvx8ccfq0ePHpo2bZquu+46dezYURkZGSovL3dFyk3eldyT3r1769ChQ/rLX/4iwzB05MgR/elPf9LgwYNdkTLOsXHjxmr3Ly4uTtu2bVNlZWU9ZQXUjFqpYaFOaniolRoeaqXGr6HXSu71nQCajm+//VZVVVUKDAx0Wh8YGKiSkpJL2sfMmTN1+vRpDR8+vC5SvOpcyT356quv9Pzzz+tvf/ub3N35T0RtupL78d///lcFBQVq3ry5Vq5cqW+//VYpKSk6fvw4z0qoBVdyT3r37q1ly5bpvvvu0w8//KAzZ84oISFBc+fOdUXKOEdJSUmN9+/MmTP69ttv1aZNm3rKDKiOWqlhoU5qeKiVGh5qpcavoddKzJRCrbNYLE6fDcOotq4mH3zwgSZOnKgPP/xQAQEBdZXeVelS70lVVZUSExM1adIkdezY0VXpXXUu538jZ8+elcVi0bJly9SzZ0/96le/0qxZs7R48WK+AaxFl3NPdu3apdTUVL388ssqLCxUbm6u9u3bp8cff9wVqaIGNd2/mtYDDQW1UsNCndTwUCs1PNRKjVtDrpVo76PW+Pv7y83NrVrH/OjRo9U6s+f68MMPNXLkSP3xj3/UgAED6jLNq8rl3pOTJ09q27Zt2r59u5588klJP/6hNwxD7u7uysvL09133+2S3JuiK/nfSJs2bXTdddfJZrOZ6zp16iTDMHTo0CGFhobWac5N3ZXckylTpqhPnz569tlnJUldu3aVt7e37rzzTr322mv1/m3T1SYoKKjG++fu7i4/P796ygqoGbVSw0Kd1PBQKzU81EqNX0OvlZgphVrj6empiIgIrVmzxmn9mjVr1Lt37/Nu98EHH2jEiBFavnw5vzOuZZd7T3x8fPTll1+qqKjIXB5//HGFhYWpqKhIkZGRrkq9SbqS/4306dNHhw8f1qlTp8x1e/fuVbNmzdS2bds6zfdqcCX35Pvvv1ezZs5/Pt3c3CT9/2+d4DpRUVHV7l9eXp569OghDw+PesoKqBm1UsNCndTwUCs1PNRKjV+Dr5Vc/WR1NG05OTmGh4eHsWjRImPXrl1GWlqa4e3tbXz99deGYRjG888/byQlJZnxy5cvN9zd3Y0333zTKC4uNpcTJ07U1yk0OZd7T87FW2Vq1+Xej5MnTxpt27Y1fvOb3xg7d+408vPzjdDQUOOxxx6rr1Noci73nrz33nuGu7u78dZbbxn/+c9/jIKCAqNHjx5Gz5496+sUmpSTJ08a27dvN7Zv325IMmbNmmVs377d2L9/v2EY1e/Hf//7X6NFixbG008/bezatctYtGiR4eHhYfzpT3+qr1MALohaqWGhTmp4qJUaHmqlhqWp1Uo0pVDr3nzzTeP66683PD09je7duxv5+fnm2MMPP2xER0ebn6Ojow1J1ZaHH37Y9Yk3YZdzT85FsVX7Lvd+7N692xgwYIDh5eVltG3b1njmmWeM77//3sVZN22Xe0/+8Ic/GJ07dza8vLyMNm3aGA888IBx6NAhF2fdNK1bt+6Cfxdquh/r1683unXrZnh6ehrt27c35s2b5/rEgctArdSwUCc1PNRKDQ+1UsPR1Goli2Ewfw4AAAAAAACuxTOlAAAAAAAA4HI0pQAAAAAAAOByNKUAAAAAAADgcjSlAAAAAAAA4HI0pQAAAAAAAOByNKUAAAAAAADgcjSlAAAAAAAA4HI0pQDUu3/+85965JFH1KFDBzVv3lzXXHONunfvrmnTpun48eN1dtzt27crOjpaNptNFotFc+bM0fr162WxWLR+/fqLbj9ixAi1b9++zvIDAACQqJUANF3u9Z0AgKvbwoULlZKSorCwMD377LPq3LmzKisrtW3bNr399tvauHGjVq5cWSfHfvTRR3X69Gnl5OTI19dX7du3V4sWLbRx40Z17ty5To4JAABwOaiVADRlFsMwjPpOAsDVaePGjbrzzjsVExOjjz76SFar1Wm8oqJCubm5SkhIqJPje3h4KDk5WW+99dYVbT9ixAitX79eX3/9de0mBgAAIGolAE0fP98DUG8yMzNlsVi0YMGCakWWJHl6eppF1tmzZzVt2jTdfPPNslqtCggI0EMPPaRDhw45bdO3b1+Fh4dr69atuvPOO9WiRQvdcMMNev3113X27FlJ0uLFi2WxWHTmzBnNmzdPFotFFotFks47JX3x4sUKCwuT1WpVp06d9P7779d4ThUVFXrttdfMPK+99lo98sgjOnbsmFNc+/btFR8fr9zcXHXv3l1eXl66+eab9e6771bb5zfffKNRo0YpJCREnp6eCg4O1m9+8xsdOXLEjCkrK1NGRoY6dOggT09PXXfddUpLS9Pp06cvchcAAEBDRa1ErQQ0eQYA1IMzZ84YLVq0MCIjIy8pftSoUYYk48knnzRyc3ONt99+27j22muNkJAQ49ixY2ZcdHS04efnZ4SGhhpvv/22sWbNGiMlJcWQZCxZssQwDMM4evSosXHjRkOS8Zvf/MbYuHGjsXHjRsMwDGPdunWGJGPdunXmPt977z1DknHPPfcYn3zyiZGdnW3cdNNNRkhIiHH99debcVVVVcbAgQMNb29vY9KkScaaNWuMd955x7juuuuMzp07G99//70Ze/311xtt27Y1OnfubLz//vvGZ599Zvz2t781JBn5+flm3KFDh4w2bdoY/v7+xqxZs4y1a9caH374ofHoo48au3fvNgzDME6fPm3cdtttTjFvvPGGYbPZjLvvvts4e/bsZd8fAABQv6iVqJWAqwFNKQD1oqSkxJBk3H///ReN3b17tyHJSElJcVq/efNmQ5LxwgsvmOuio6MNScbmzZudYjt37mzExcU5rZNkjBkzxmnduYVWVVWVERwcbHTv3t2pYPn6668NDw8Pp0Lrgw8+MCQZK1ascNrn1q1bDUnGW2+9Za67/vrrjebNmxv79+8315WXlxutW7c2Ro8eba579NFHDQ8PD2PXrl3nvT5TpkwxmjVrZmzdutVp/Z/+9CdDkvGXv/zlvNsCAICGiVqJWgm4GvDzPQAN3rp16yT9+FyCn+vZs6c6deqkzz//3Gl9UFCQevbs6bSua9eu2r9//2Ufe8+ePTp8+LASExPNaeuSdP3116t3795OsatWrVKrVq00ZMgQnTlzxlxuu+02BQUFVZvmftttt6ldu3bm5+bNm6tjx45OeX766afq16+fOnXqdN4cV61apfDwcN12221Ox42Li7vkt+MAAIDGi1qJWglorHj7HoB64e/vrxYtWmjfvn0Xjf3uu+8kSW3atKk2FhwcXK2A8vPzqxZntVpVXl5+2Xn+dOygoKBqY0FBQU4P7jxy5IhOnDghT0/PGvf17bffXnaex44dU9u2bS+Y45EjR/Tvf/9bHh4el3RcAADQ8FErUSsBVwOaUgDqhZubm/r3769PP/1Uhw4dumAx8VNBUlxcXC3u8OHD8vf3r7M8fzp2SUlJtbFz1/n7+8vPz0+5ubk17qtly5aXffxrr7222gNKz+Xv7y8vL68aH/z50zgAAGhcqJUuDbUS0Ljx8z0A9Wb8+PEyDEPJycmqqKioNl5ZWalPPvlEd999tyQpOzvbaXzr1q3avXu3+vfvX2c5hoWFqU2bNvrggw9kGIa5fv/+/dqwYYNTbHx8vL777jtVVVWpR48e1ZawsLDLPv6gQYO0bt067dmz57wx8fHx+s9//iM/P78aj9u+ffvLPi4AAKh/1EoXR60ENG7MlAJQb6KiojRv3jylpKQoIiJCTzzxhG655RZVVlZq+/btWrBggcLDw7Vy5UqNGjVKc+fOVbNmzTRo0CB9/fXXeumllxQSEqKnn366znJs1qyZfv/73+uxxx7Tvffeq+TkZJ04cUITJ06sNk39/vvv17Jly/SrX/1KTz31lHr27CkPDw8dOnRI69at0z333KN77733so7/6quv6tNPP9Vdd92lF154QV26dNGJEyeUm5urZ555RjfffLPS0tK0YsUK3XXXXXr66afVtWtXnT17VgcOHFBeXp7S09MVGRlZm5cFAAC4ALXSxVErAY0bTSkA9So5OVk9e/bU7NmzNXXqVJWUlMjDw0MdO3ZUYmKinnzySUnSvHnzdOONN2rRokV68803ZbPZNHDgQE2ZMqXG5w3UppEjR0qSpk6dqmHDhql9+/Z64YUXlJ+f7/RgTDc3N3388cd64403tHTpUk2ZMkXu7u5q27atoqOj1aVLl8s+9nXXXactW7bolVde0euvv67vvvtO1157re644w61bt1akuTt7a2//e1vev3117VgwQLt27dPXl5eateunQYMGMC3fwAANGLUShdGrQQ0bhbj53MsAQAAAAAAABfgmVIAAAAAAABwOZpSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAAAAAAcDmaUgAAAAAAAHA5mlIAAAAAAABwOZpSAAAAAAAAcDmaUgAajIkTJ8pisejbb7+tcTw8PFx9+/bVsWPH1KxZMz3xxBPVYp566ilZLBaNHz++2tjIkSPl5uam0tJSSdKIESNksVh0yy23qKqqqlq8xWLRk08+6bRuzpw5GjZsmDp06CCLxaK+ffue93zWrVunmJgYBQQE6JprrlHXrl31hz/8odqx2rdvL4vFYi7e3t7q3r27srKyZBiGU+z69eudYt3c3HTttddqyJAh2rZt23lzAQAAjcel1kSS9PXXXzvVBj9fevToYW7zU93z8xqibdu2Gj58uHbs2FHtGC+++KLi4+N13XXXyWKxaMSIERfM9dylefPm1WJrqq0k6ZVXXpHFYtETTzyhs2fPau/evcrIyFBERIRatWql1q1bq0+fPvrTn/5Ubdu1a9cqJiZGwcHBslqtCggI0N13362//OUvTnEXuk4Wi0UDBw6s8fwA1C33+k4AAC7Xtddeq1tuuUXr1q2rNrZ+/Xp5e3ufd+y2226Tr6+v0/pdu3Zp8eLFGjly5EWP/fbbb8vb21t33323Pvnkk/PGrV27VnFxcbrrrru0cOFCeXt76+OPP9ZTTz2l//znP3rjjTec4vv06aMZM2ZIkg4fPqxZs2Zp7NixKisr0wsvvFBt/5mZmerXr58qKyu1fft2TZo0SdHR0SoqKlJoaOhFzwMAADQtY8eOVWJiotO6a665xumzl5eXvvjiC0nSmTNn9O9//1uvvfaaevfurd27d+u6664zY2fPnq2uXbsqISFB77777kWPn5ubK5vNZn5u1uzi8x8Mw9BTTz2luXPn6vnnn9eUKVMkSXl5eVq9erWSkpJ0++2368yZM/rwww/129/+VpMmTdLLL79s7uO7777TLbfcoscee0xBQUE6fvy43n77bQ0ePFhLly7Vgw8+KElq06aNNm7cWC2Hjz76SFOnTtW999570XwB1D6aUgAapX79+mnu3LkqKSlRUFCQJOn48eP68ssvlZ6erjlz5ujkyZNq2bKlJOnQoUP673//q/T0dKf9/DQr6ZVXXlFiYqK8vLwueNxdu3aZRVZ4ePh54xYvXiwPDw+tWrVK3t7ekqQBAwZoz549Wrx4cbWmVKtWrdSrVy/z84ABA9SuXTvNnz+/xqZUaGioGX/nnXeqVatWevjhh5Wdna1JkyZd8BwAAEDT065dO6daoibNmjVzirnjjjvUrl079e/fX6tXr9aoUaPMsZMnT5o1z9KlSy96/IiICPn7+19yvmfOnNGjjz6qpUuXavr06crIyDDH7r//fo0ZM0YWi8VcN2jQIH377beaOnWqnnvuOVmtVknSfffdp/vuu89p3/Hx8erQoYMWLFhgNqWsVmuN12f8+PFq0aKFfve7311y7gBqDz/fA9Ao9evXT9KPs59+kp+fL3d3d7Oo+dvf/maO/TRz6qftfm7q1Kn65ptvqjWKanIp3/pJkoeHhzw9Pas1uVq1alXjdPZz+fj4qGPHjjpy5MglHe+n6fmXGg8AACDJnN3k4eHhtP5Sa54r8cMPP+jXv/61li9frnfeecepISVJ/v7+Tg2pn/Ts2VPff/+9jh8/fsH9e3h4qFWrVnJ3v/AcjP/85z/Kz8/X8OHD5ePjc/knAuAXoykFoFGKjo5Ws2bNnH6mt27dOvXo0UOBgYGKiIhwalitW7dObm5uuvPOO6vtKyoqSvfee6+mTp160SLnUj3++OOqqKhQamqqDh8+rBMnTmjp0qVauXKlxo0bd9Htz5w5o4MHD6pjx46XdLx9+/ZJ0iXHAwCApuXs2bM6c+aM03LusyklmWM//PCDduzYoWeffVa+vr4aPHjwLzp+ly5d5ObmpsDAQD300EM6cOBAjXEnT57UoEGDlJubqw8//PCSHp/wk3Xr1unaa69VQEBAtbGfzv/w4cN65ZVXtHfv3moz5M/17rvvyjAMPfbYY5ecA4DaRVMKQKPUunVrde3a1anxtH79ekVHR0v6sWn184bV+vXrFRERcd5vwaZMmaKTJ08qMzOzVvKLjIzUF198oZUrV+q6666Tr6+vHnnkEU2ePLnGAskwDLNIPHDggFJSUvTdd9+Zz1Y410+FV3l5uTZs2KD09HR17txZjz76aK3kDwAAGpfnnntOHh4eTsvnn3/uFHP69GlzzMvLS126dNG//vUvffLJJzU2ei7FjTfeqMmTJ+vdd9/V2rVr9cwzz2j16tXq2bOnvvnmm2rx77//vtavX6+srCz9+te/vuTjvPPOO1q/fr1efPFFubm5VRv/1a9+JQ8PD1133XWaM2eOPvzwwws22qqqqrRkyRLdfPPN6tOnzyXnAaB28UwpAI1Wv379NHv2bB0+fFhWq1U7duzQ9OnTJf3YlJo5c6bsdrvsdrv27dun4cOHn3dfYWFhGjlypLKyspSamqp27dr9otwKCwt17733KjIyUvPnz5e3t7e++OILvfjii/rhhx/00ksvOcX/5S9/qTZt/qeHdNbk3GcntGnTRhs2bFCrVq1+Ud4AAKBxeuqpp8znJ/0kLCzM6bOXl5f++te/SvrxC66fHl/wq1/9Srm5uYqKirrs4yYlJTl97tevn/r166eoqChNmzat2uMR7rzzTn355ZeaNGmS+vXrp5tuuumix/j00081ZswY/eY3v9HYsWNrjJk7d65OnDih4uJiZWdn67777tOSJUvO+6yo3NxcffPNN2btCKB+0JQC0GD89Lv/qqqqGsfPnDnj1Lj5qSm1fv16Wa1Wubm5md903XHHHZJ+fK7Ud999Z8ZfyMSJE5Wdna2XXnpJS5Ys+UXnMmbMGAUGBmrlypXmt3n9+vVTs2bNNHHiRD3wwAO64YYbzPg77rhDs2fPVlVVlb766iu99NJLevLJJ3XLLbeY5/JzU6dO1d13363vv/9eeXl5mjJlioYOHarNmzebD/4EAACN0+XWRJLUtm1b8xmT59OsWbNqMXFxcQoJCdEzzzxT49vprkTPnj3VsWNHbdq0qdpY165dNXv2bMXExJgz2y/0+IHPPvtMw4YNU0xMjJYtW1bjs6YkOb19OCEhQYMGDdKYMWN033331fh8rEWLFsnDw0MPPfTQFZwhgNrCz/cANBiBgYGSVONUb8MwVFxcbMZI0l133SU3NzetX79e69evV/fu3c1XH/v4+Oi2227TunXrtH79erm7u190anabNm2Ulpam7Oxs/fOf//xF51JUVKSIiIhq08tvv/12nT17Vrt373Zab7PZ1KNHD0VGRurBBx9UXl6ePDw8lJKSorNnz1bb/w033KAePXrorrvu0muvvaZXX31V//jHPzR37txflDcAAKh/l1sT/RItWrTQjTfeqH/84x+1sr+fGIZx3oelR0REaO3atfrhhx/Ur18/7dmzp8a4zz77TEOHDlV0dLRWrFghT0/PSz5+z549VVpaqmPHjlUbO3r0qFatWqWEhIQr/tkigNpBUwpAg3H33XfLYrHoww8/rDaWm5ursrIyDRgwwFxns9nUrVs3synVt29fp21++vZt/fr16tmzp9mwupDnnntOrVu31vPPP/+LziU4OFjbtm2r9g3nT99Atm3b9oLbh4aGaty4cfryyy9rvB7nGjdunG666Sa9/vrrOnny5JUnDgAA6t3l1kS/xKlTp/Tvf/+7VpszmzZt0ldffaVevXqdN6Z79+76/PPP5XA41K9fP/3rX/9yGs/Ly9PQoUN1xx136KOPPrqsmeCGYSg/P1+tWrWSn59ftfH3339flZWVl/WQdQB1g5/vAWgwbrzxRj355JOaPn26Tpw4oV/96lfy8vLS1q1b9frrr6tHjx5KTEx02qZfv36aPn26LBaLpk6d6jQWHR2t2bNnyzAMPfDAA5eUg4+PjyZMmKCnn366xvFt27bp66+/liSVlZXJMAz96U9/kvTjLKjrr79ekvT0008rNTVVQ4YM0ejRo9WiRQt9/vnnmjlzpgYMGKBbb731orlkZGTo7bff1qRJkzR8+PAaH+r5Ew8PD2VmZmr48OF644039OKLL17S+QIAgIbnSmqiS3H27FnzJ3U/PVPqD3/4g0pLSzVx4kSn2Pz8fHOWUVVVlfbv32/WPNHR0br22mslSbfeeqsefPBBderUSc2bN9eWLVs0ffp0BQUFXfSNw7fddps+//xz9e/fX/369dMXX3yhTp06qaCgQEOHDlVQUJBeeOEFFRUVOW3XuXNn8+U199xzj2699Vbddttt8vPz0+HDh7V48WLl5+frzTffNH8K+XOLFi1SSEiI4uLiLvsaAqhlBgA0IGfPnjXmzZtn9OjRw2jRooXh6elphIaGGs8995xx8uTJavF/+ctfDEmGm5ubYbfbncaOHz9uNGvWzJBkrFmzptq2Dz/8sOHt7V1tvcPhMDp06GBIMsaMGVNtG0k1Lu+9955T7IoVK4w77rjD8Pf3N7y9vY1bbrnF+P3vf2+cOnXKKe766683Bg8eXOP1ePPNNw1JxpIlSwzDMIx169YZkow//vGPNcZHRkYavr6+xokTJ2ocBwAAjcOl1kT79u0zJBnTp0+/4P5qqmECAgKM6OhoY+XKldXio6Ojz1vzrFu3zoy7//77jZtuusnw9vY2PDw8jOuvv954/PHHjcOHD1fbZ021lWEYxj/+8Q/D39/fCAwMNHbu3Gm88sor5z32ucefOnWqcfvttxu+vr6Gm5ub4efnZ8TFxRmrVq2q8Tr83//9nyHJePnlly94vQC4hsUwDMMFvS8AAAAAAADAxDOlAAAAAAAA4HI0pQAAAAAAAOByNKUAAAAAAADgcjSlAAAAAAAA4HI0pQAAAAAAAOByNKUAAAAAAADgcu71nUBTc/bsWR0+fFgtW7aUxWKp73QAAMAlMgxDJ0+eVHBwsJo143u7ukKtBABA41QXtRJNqVp2+PBhhYSE1HcaAADgCh08eFBt27at7zSaLGolAAAat9qslWhK1bKWLVtK+vEm+fj41HM2AADgUpWVlSkkJMT8W466Qa0EAEDjVBe1Ur03pf76179q+vTpKiwsVHFxsVauXKmhQ4fWGDt69GgtWLBAs2fPVlpamrne4XAoIyNDH3zwgcrLy9W/f3+99dZbTp270tJSpaam6uOPP5YkJSQkaO7cuWrVqpUZc+DAAY0ZM0ZffPGFvLy8lJiYqBkzZsjT0/OSz+enaeg+Pj4UWgAANEL8pKxuUSsBANC41WatVO8PTDh9+rRuvfVWZWVlXTDuo48+0ubNmxUcHFxtLC0tTStXrlROTo4KCgp06tQpxcfHq6qqyoxJTExUUVGRcnNzlZubq6KiIiUlJZnjVVVVGjx4sE6fPq2CggLl5ORoxYoVSk9Pr72TBQAAAAAAgKQGMFNq0KBBGjRo0AVjvvnmGz355JP67LPPNHjwYKcxu92uRYsWaenSpRowYIAkKTs7WyEhIVq7dq3i4uK0e/du5ebmatOmTYqMjJQkLVy4UFFRUdqzZ4/CwsKUl5enXbt26eDBg2bja+bMmRoxYoQmT57MN3kAAAAAAAC1qN5nSl3M2bNnlZSUpGeffVa33HJLtfHCwkJVVlYqNjbWXBccHKzw8HBt2LBBkrRx40bZbDazISVJvXr1ks1mc4oJDw93mokVFxcnh8OhwsLC8+bncDhUVlbmtAAAAAAAAODCGnxTaurUqXJ3d1dqamqN4yUlJfL09JSvr6/T+sDAQJWUlJgxAQEB1bYNCAhwigkMDHQa9/X1laenpxlTkylTpshms5kLb5MBAAAAAAC4uAbdlCosLNQbb7yhxYsXX/aDtAzDcNqmpu2vJOZc48ePl91uN5eDBw9eVp4AAAAAAABXowbdlPrb3/6mo0ePql27dnJ3d5e7u7v279+v9PR0tW/fXpIUFBSkiooKlZaWOm179OhRc+ZTUFCQjhw5Um3/x44dc4o5d0ZUaWmpKisrq82g+jmr1Wq+PYa3yAAAAAAAAFyaBt2USkpK0j//+U8VFRWZS3BwsJ599ll99tlnkqSIiAh5eHhozZo15nbFxcXasWOHevfuLUmKioqS3W7Xli1bzJjNmzfLbrc7xezYsUPFxcVmTF5enqxWqyIiIlxxugAAAAAAAFeNen/73qlTp/Tvf//b/Lxv3z4VFRWpdevWateunfz8/JziPTw8FBQUpLCwMEmSzWbTyJEjlZ6eLj8/P7Vu3VoZGRnq0qWL+Ta+Tp06aeDAgUpOTtb8+fMlSaNGjVJ8fLy5n9jYWHXu3FlJSUmaPn26jh8/royMDCUnJzP7CQAAAAAAoJbV+0ypbdu2qVu3burWrZsk6ZlnnlG3bt308ssvX/I+Zs+eraFDh2r48OHq06ePWrRooU8++URubm5mzLJly9SlSxfFxsYqNjZWXbt21dKlS81xNzc3rV69Ws2bN1efPn00fPhwDR06VDNmzKi9kwUAAAAAAIAkyWIYhlHfSTQlZWVlstlsstvtzLACAKAR4W+4a3CdAQBonOrib3i9z5QCAAAAAADA1YemFAAAAAAAAFyu3h90jks3e83eWtvX0zEda21fAAAAAADUhtr8d6/Ev30bOmZKAQAAAAAAwOVoSgEAAAAAAMDlaEoBAAAAAADA5WhKAQAAAAAAwOVoSgEAAAAAAMDlaEoBAAAAAADA5WhKAQAAAAAAwOVoSgEAAAAAAMDlaEoBAAAAAADA5WhKAQAAAAAAwOVoSgEAAAAAAMDlaEoBAAAAAADA5WhKAQAAAAAAwOVoSgEAAAAAAMDlaEoBAAA0YPPmzVPXrl3l4+MjHx8fRUVF6dNPPzXHDcPQxIkTFRwcLC8vL/Xt21c7d+502ofD4dDYsWPl7+8vb29vJSQk6NChQ04xpaWlSkpKks1mk81mU1JSkk6cOOEUc+DAAQ0ZMkTe3t7y9/dXamqqKioq6uzcAQBA00ZTCgAAoAFr27atXn/9dW3btk3btm3T3XffrXvuucdsPE2bNk2zZs1SVlaWtm7dqqCgIMXExOjkyZPmPtLS0rRy5Url5OSooKBAp06dUnx8vKqqqsyYxMREFRUVKTc3V7m5uSoqKlJSUpI5XlVVpcGDB+v06dMqKChQTk6OVqxYofT0dNddDAAA0KRYDMMw6juJpqSsrEw2m012u10+Pj61uu/Za/bW2r6ejulYa/sCAKApqMu/4bWtdevWmj59uh599FEFBwcrLS1Nzz33nKQfZ0UFBgZq6tSpGj16tOx2u6699lotXbpU9913nyTp8OHDCgkJ0V/+8hfFxcVp9+7d6ty5szZt2qTIyEhJ0qZNmxQVFaV//etfCgsL06effqr4+HgdPHhQwcHBkqScnByNGDFCR48ePe81czgccjgc5ueysjKFhIQ0iusMAHC92vx3r8S/fWtTXdRKzJQCAABoJKqqqpSTk6PTp08rKipK+/btU0lJiWJjY80Yq9Wq6OhobdiwQZJUWFioyspKp5jg4GCFh4ebMRs3bpTNZjMbUpLUq1cv2Ww2p5jw8HCzISVJcXFxcjgcKiwsPG/OU6ZMMX8SaLPZFBISUjsXAwAANHo0pQAAABq4L7/8Utdcc42sVqsef/xxrVy5Up07d1ZJSYkkKTAw0Ck+MDDQHCspKZGnp6d8fX0vGBMQEFDtuAEBAU4x5x7H19dXnp6eZkxNxo8fL7vdbi4HDx68zLMHAABNlXt9JwAAAIALCwsLU1FRkU6cOKEVK1bo4YcfVn5+vjlusVic4g3DqLbuXOfG1BR/JTHnslqtslqtF8wFAABcnZgpBQAA0MB5enrqpptuUo8ePTRlyhTdeuuteuONNxQUFCRJ1WYqHT161JzVFBQUpIqKCpWWll4w5siRI9WOe+zYMaeYc49TWlqqysrKajOoAAAALgVNKQAAgEbGMAw5HA516NBBQUFBWrNmjTlWUVGh/Px89e7dW5IUEREhDw8Pp5ji4mLt2LHDjImKipLdbteWLVvMmM2bN8tutzvF7NixQ8XFxWZMXl6erFarIiIi6vR8AQBA08TP9wAAABqwF154QYMGDVJISIhOnjypnJwcrV+/Xrm5ubJYLEpLS1NmZqZCQ0MVGhqqzMxMtWjRQomJiZIkm82mkSNHKj09XX5+fmrdurUyMjLUpUsXDRgwQJLUqVMnDRw4UMnJyZo/f74kadSoUYqPj1dYWJgkKTY2Vp07d1ZSUpKmT5+u48ePKyMjQ8nJybxFDwAAXBGaUgAAAA3YkSNHlJSUpOLiYtlsNnXt2lW5ubmKiYmRJI0bN07l5eVKSUlRaWmpIiMjlZeXp5YtW5r7mD17ttzd3TV8+HCVl5erf//+Wrx4sdzc3MyYZcuWKTU11XxLX0JCgrKyssxxNzc3rV69WikpKerTp4+8vLyUmJioGTNmuOhKAACApsZiGIZR30k0JWVlZbLZbLLb7bX+reHsNXtrbV9Px3SstX0BANAU1OXfcPx/XGcAwIXU5r97Jf7tW5vq4m84z5QCAAAAAACAy9GUAgAAAAAAgMvRlAIAAAAAAIDL0ZQCAAAAAACAy9GUAgAAAAAAgMvVe1Pqr3/9q4YMGaLg4GBZLBZ99NFH5lhlZaWee+45denSRd7e3goODtZDDz2kw4cPO+3D4XBo7Nix8vf3l7e3txISEnTo0CGnmNLSUiUlJclms8lmsykpKUknTpxwijlw4ICGDBkib29v+fv7KzU1VRUVFXV16gAAAAAAAFetem9KnT59WrfeequysrKqjX3//ff6+9//rpdeekl///vf9ec//1l79+5VQkKCU1xaWppWrlypnJwcFRQU6NSpU4qPj1dVVZUZk5iYqKKiIuXm5io3N1dFRUVKSkoyx6uqqjR48GCdPn1aBQUFysnJ0YoVK5Senl53Jw8AAAAAAHCVcq/vBAYNGqRBgwbVOGaz2bRmzRqndXPnzlXPnj114MABtWvXTna7XYsWLdLSpUs1YMAASVJ2drZCQkK0du1axcXFaffu3crNzdWmTZsUGRkpSVq4cKGioqK0Z88ehYWFKS8vT7t27dLBgwcVHBwsSZo5c6ZGjBihyZMny8fHpw6vAgAAAAAAwNWl3mdKXS673S6LxaJWrVpJkgoLC1VZWanY2FgzJjg4WOHh4dqwYYMkaePGjbLZbGZDSpJ69eolm83mFBMeHm42pCQpLi5ODodDhYWF583H4XCorKzMaQEAAAAAAMCFNaqm1A8//KDnn39eiYmJ5sylkpISeXp6ytfX1yk2MDBQJSUlZkxAQEC1/QX8v/buPayqOu///2vHYYMM7ATilKg0t5AKnTARbdJSUUckc7qtYW5Gy7D5UhopNZn3TNTtYbJSG0xTx7REpeuesvGQDDip5XiM0bs8DHYwlQIxxW06zAZx/f7ocv3a4gncbDbwfFzXui73Wu/9We/PXiCf670+e33CwpxiwsPDnY63b99evr6+ZszFTJ8+3XxOlc1mU3R09DX1EQAAAAAAoC1oMUWp2tpaPfTQQzp37pzmzp17xXjDMGSxWMzXP/73tcRcaNKkSbLb7eZ25MiRK+YGAAAAAADQ1rWIolRtba1GjhypgwcPqri42On5ThEREaqpqVFVVZXTeyorK82ZTxERETp69Gi9do8dO+YUc+GMqKqqKtXW1tabQfVjVqtVQUFBThsAAAAAAAAuz+OLUucLUp9//rnWr1+vkJAQp+OJiYny8fFxeiB6eXm59uzZo969e0uSkpOTZbfbtWPHDjNm+/btstvtTjF79uxReXm5GVNUVCSr1arExMSm7CIAAAAAAECb0+yr750+fVpffPGF+frgwYPavXu3goODFRUVpQceeED/+Mc/tGbNGtXV1ZmzmYKDg+Xr6yubzaYxY8Zo4sSJCgkJUXBwsHJycpSQkGCuxte1a1cNHjxYmZmZmj9/viRp7NixSk1NVVxcnCQpJSVF3bp1U0ZGhl5++WWdOHFCOTk5yszMZPYTAAAAAACAizV7UeqTTz7RPffcY76eMGGCJGnUqFHKzc3VqlWrJEm33Xab0/s2bNigfv36SZJmzZolb29vjRw5UtXV1erfv7+WLFkiLy8vM37ZsmUaP368uUpfWlqa5syZYx738vLS2rVrlZWVpT59+sjf31/p6el65ZVXmqLbAAAAAAAAbVqzF6X69esnwzAuefxyx87z8/NTXl6e8vLyLhkTHBys/Pz8y7bTsWNHrVmz5ornAwAAAAAAwLXx+GdKAQAAAAAAoPWhKAUAAAAAAAC3oygFAAAAAAAAt6MoBQAAAAAAALejKAUAAAAAAAC3oygFAAAAAAAAt6MoBQAAAAAAALejKAUAAAAAAAC3oygFAAAAAAAAt6MoBQAAAAAAALejKAUAAAAAAAC3827uBAAAAAAAgDSr+IBL23tqYKxL2wNcjZlSAAAAAAAAcDuKUgAAAAAAAHA7ilIAAAAAAABwO4pSAAAAAAAAcDuKUgAAAAAAAHA7ilIAAAAAAABwO4pSAAAAAAAAcDuKUgAAAB5s+vTpuvPOOxUYGKiwsDANHz5cpaWlTjGjR4+WxWJx2nr16uUU43A4NG7cOIWGhiogIEBpaWkqKytziqmqqlJGRoZsNptsNpsyMjJ08uRJp5jDhw9r2LBhCggIUGhoqMaPH6+ampom6TsAAGjdKEoBAAB4sE2bNunxxx/Xtm3bVFxcrLNnzyolJUVnzpxxihs8eLDKy8vN7YMPPnA6np2drZUrV6qgoECbN2/W6dOnlZqaqrq6OjMmPT1du3fvVmFhoQoLC7V7925lZGSYx+vq6jR06FCdOXNGmzdvVkFBgd59911NnDixaT8EAADQKnk3dwIAAAC4tMLCQqfXixcvVlhYmEpKSnT33Xeb+61WqyIiIi7aht1u16JFi7R06VINGDBAkpSfn6/o6GitX79egwYN0v79+1VYWKht27YpKSlJkrRw4UIlJyertLRUcXFxKioq0r59+3TkyBFFRUVJkl599VWNHj1aU6dOVVBQUFN8BAAAoJViphQAAEALYrfbJUnBwcFO+zdu3KiwsDDFxsYqMzNTlZWV5rGSkhLV1tYqJSXF3BcVFaX4+Hht2bJFkrR161bZbDazICVJvXr1ks1mc4qJj483C1KSNGjQIDkcDpWUlFw0X4fDoVOnTjltAAAAEkUpAACAFsMwDE2YMEF33XWX4uPjzf1DhgzRsmXL9OGHH+rVV1/Vzp07de+998rhcEiSKioq5Ovrq/bt2zu1Fx4eroqKCjMmLCys3jnDwsKcYsLDw52Ot2/fXr6+vmbMhaZPn24+o8pmsyk6OrrxHwAAAGhV+PoeAABAC/HEE0/o008/1ebNm532P/jgg+a/4+Pj1aNHD3Xq1Elr167ViBEjLtmeYRiyWCzm6x//+1pifmzSpEmaMGGC+frUqVMUpgAAgCRmSgEAALQI48aN06pVq7RhwwZ16NDhsrGRkZHq1KmTPv/8c0lSRESEampqVFVV5RRXWVlpznyKiIjQ0aNH67V17Ngxp5gLZ0RVVVWptra23gyq86xWq4KCgpw2AAAAiaIUAACARzMMQ0888YTee+89ffjhh4qJibnie44fP64jR44oMjJSkpSYmCgfHx8VFxebMeXl5dqzZ4969+4tSUpOTpbdbteOHTvMmO3bt8tutzvF7NmzR+Xl5WZMUVGRrFarEhMTXdJfAADQdvD1PQAAAA/2+OOPa/ny5frLX/6iwMBAc6aSzWaTv7+/Tp8+rdzcXP3iF79QZGSkvv76az333HMKDQ3V/fffb8aOGTNGEydOVEhIiIKDg5WTk6OEhARzNb6uXbtq8ODByszM1Pz58yVJY8eOVWpqquLi4iRJKSkp6tatmzIyMvTyyy/rxIkTysnJUWZmJjOgAABAgzFTCgAAwIPNmzdPdrtd/fr1U2RkpLm98847kiQvLy999tlnuu+++xQbG6tRo0YpNjZWW7duVWBgoNnOrFmzNHz4cI0cOVJ9+vRRu3bttHr1anl5eZkxy5YtU0JCglJSUpSSkqJbbrlFS5cuNY97eXlp7dq18vPzU58+fTRy5EgNHz5cr7zyivs+EAAA0GowUwoAAMCDGYZx2eP+/v7661//esV2/Pz8lJeXp7y8vEvGBAcHKz8//7LtdOzYUWvWrLni+QAAaAtmFR9waXtPDYx1aXuejplSAAAAAAAAcDuKUgAAAAAAAHA7ilIAAAAAAABwu2YvSn300UcaNmyYoqKiZLFY9P777zsdNwxDubm5ioqKkr+/v/r166e9e/c6xTgcDo0bN06hoaEKCAhQWlqaysrKnGKqqqqUkZEhm80mm82mjIwMnTx50inm8OHDGjZsmAICAhQaGqrx48erpqamKboNAAAAAADQpjV7UerMmTO69dZbNWfOnIsenzFjhmbOnKk5c+Zo586dioiI0MCBA/X999+bMdnZ2Vq5cqUKCgq0efNmnT59WqmpqaqrqzNj0tPTtXv3bhUWFqqwsFC7d+9WRkaGebyurk5Dhw7VmTNntHnzZhUUFOjdd9/VxIkTm67zAAAAAAAAbVSzr743ZMgQDRky5KLHDMPQ7NmzNXnyZI0YMUKS9NZbbyk8PFzLly/XY489JrvdrkWLFmnp0qUaMGCAJCk/P1/R0dFav369Bg0apP3796uwsFDbtm1TUlKSJGnhwoVKTk5WaWmp4uLiVFRUpH379unIkSOKioqSJL366qsaPXq0pk6dqqCgoIvm6HA45HA4zNenTp1y2WcDAAAAAIAnYtU5uEKzz5S6nIMHD6qiokIpKSnmPqvVqr59+2rLli2SpJKSEtXW1jrFREVFKT4+3ozZunWrbDabWZCSpF69eslmsznFxMfHmwUpSRo0aJAcDodKSkoumeP06dPNrwTabDZFR0e7pvMAAAAAAACtmEcXpSoqKiRJ4eHhTvvDw8PNYxUVFfL19VX79u0vGxMWFlav/bCwMKeYC8/Tvn17+fr6mjEXM2nSJNntdnM7cuRIA3sJAAAAAADQ9jT71/euhsVicXptGEa9fRe6MOZi8Y2JuZDVapXVar1sLgAAAAAAAHDm0TOlIiIiJKneTKXKykpzVlNERIRqampUVVV12ZijR4/Wa//YsWNOMReep6qqSrW1tfVmUAEAAAAAAODaeHRRKiYmRhERESouLjb31dTUaNOmTerdu7ckKTExUT4+Pk4x5eXl2rNnjxmTnJwsu92uHTt2mDHbt2+X3W53itmzZ4/Ky8vNmKKiIlmtViUmJjZpPwEAAAAAANqaZv/63unTp/XFF1+Yrw8ePKjdu3crODhYHTt2VHZ2tqZNm6YuXbqoS5cumjZtmtq1a6f09HRJks1m05gxYzRx4kSFhIQoODhYOTk5SkhIMFfj69q1qwYPHqzMzEzNnz9fkjR27FilpqYqLi5OkpSSkqJu3bopIyNDL7/8sk6cOKGcnBxlZmZecuU9AABwaa5clYcVeQAAAFqfZi9KffLJJ7rnnnvM1xMmTJAkjRo1SkuWLNEzzzyj6upqZWVlqaqqSklJSSoqKlJgYKD5nlmzZsnb21sjR45UdXW1+vfvryVLlsjLy8uMWbZsmcaPH2+u0peWlqY5c+aYx728vLR27VplZWWpT58+8vf3V3p6ul555ZWm/ggAAAAAAADanGYvSvXr10+GYVzyuMViUW5urnJzcy8Z4+fnp7y8POXl5V0yJjg4WPn5+ZfNpWPHjlqzZs0VcwYAAAAAAMC18ehnSgEAAAAAAKB1oigFAAAAAAAAt6MoBQAAAAAAALejKAUAAAAAAAC3oygFAAAAAAAAt6MoBQAAAAAAALejKAUAAAAAAAC3oygFAAAAAAAAt6MoBQAAAAAAALejKAUAAAAAAAC3oygFAAAAAAAAt6MoBQAAAAAAALejKAUAAAAAAAC3oygFAAAAAAAAt6MoBQAAAAAAALejKAUAAAAAAAC3oygFAAAAAAAAt6MoBQAAAAAAALfzbu4EAAAAAABoSWYVH3Bpe08NjHVpe0BLQVEKAAAAANAqUTwCPBtf3wMAAAAAAIDbUZQCAAAAAACA21GUAgAA8GDTp0/XnXfeqcDAQIWFhWn48OEqLS11ijEMQ7m5uYqKipK/v7/69eunvXv3OsU4HA6NGzdOoaGhCggIUFpamsrKypxiqqqqlJGRIZvNJpvNpoyMDJ08edIp5vDhwxo2bJgCAgIUGhqq8ePHq6ampkn6DgAAWjeeKQUAAODBNm3apMcff1x33nmnzp49q8mTJyslJUX79u1TQECAJGnGjBmaOXOmlixZotjYWE2ZMkUDBw5UaWmpAgMDJUnZ2dlavXq1CgoKFBISookTJyo1NVUlJSXy8vKSJKWnp6usrEyFhYWSpLFjxyojI0OrV6+WJNXV1Wno0KG64YYbtHnzZh0/flyjRo2SYRjKy8trhk+nPp4fAwBAy0FRCgAAwIOdLxCdt3jxYoWFhamkpER33323DMPQ7NmzNXnyZI0YMUKS9NZbbyk8PFzLly/XY489JrvdrkWLFmnp0qUaMGCAJCk/P1/R0dFav369Bg0apP3796uwsFDbtm1TUlKSJGnhwoVKTk5WaWmp4uLiVFRUpH379unIkSOKioqSJL366qsaPXq0pk6dqqCgIDd+MgAAoKXj63sAAAAtiN1ulyQFBwdLkg4ePKiKigqlpKSYMVarVX379tWWLVskSSUlJaqtrXWKiYqKUnx8vBmzdetW2Ww2syAlSb169ZLNZnOKiY+PNwtSkjRo0CA5HA6VlJRcNF+Hw6FTp045bQAAABJFKQAAgBbDMAxNmDBBd911l+Lj4yVJFRUVkqTw8HCn2PDwcPNYRUWFfH191b59+8vGhIWF1TtnWFiYU8yF52nfvr18fX3NmAtNnz7dfEaVzWZTdHR0Q7sNAABaKYpSAAAALcQTTzyhTz/9VCtWrKh3zGKxOL02DKPevgtdGHOx+MbE/NikSZNkt9vN7ciRI5fNCQAAtB0UpQAAAFqAcePGadWqVdqwYYM6dOhg7o+IiJCkejOVKisrzVlNERERqqmpUVVV1WVjjh49Wu+8x44dc4q58DxVVVWqra2tN4PqPKvVqqCgIKcNAABAoigFAADg0QzD0BNPPKH33ntPH374oWJiYpyOx8TEKCIiQsXFxea+mpoabdq0Sb1795YkJSYmysfHxymmvLxce/bsMWOSk5Nlt9u1Y8cOM2b79u2y2+1OMXv27FF5ebkZU1RUJKvVqsTERNd3HgAAtGqsvgcAAODBHn/8cS1fvlx/+ctfFBgYaM5Ustls8vf3l8ViUXZ2tqZNm6YuXbqoS5cumjZtmtq1a6f09HQzdsyYMZo4caJCQkIUHBysnJwcJSQkmKvxde3aVYMHD1ZmZqbmz58vSRo7dqxSU1MVFxcnSUpJSVG3bt2UkZGhl19+WSdOnFBOTo4yMzOZAQUAABqMohQAAIAHmzdvniSpX79+TvsXL16s0aNHS5KeeeYZVVdXKysrS1VVVUpKSlJRUZECAwPN+FmzZsnb21sjR45UdXW1+vfvryVLlsjLy8uMWbZsmcaPH2+u0peWlqY5c+aYx728vLR27VplZWWpT58+8vf3V3p6ul555ZUm6j0AAGjNKEoBAAB4MMMwrhhjsViUm5ur3NzcS8b4+fkpLy9PeXl5l4wJDg5Wfn7+Zc/VsWNHrVmz5oo5AQAAXInHP1Pq7Nmz+u///m/FxMTI399fN910k1588UWdO3fOjDEMQ7m5uYqKipK/v7/69eunvXv3OrXjcDg0btw4hYaGKiAgQGlpaSorK3OKqaqqUkZGhrlkcUZGhk6ePOmObgIAAAAAALQpjSpKeXl5OT0E88dKSkqcpoFfq5deeklvvPGG5syZo/3792vGjBl6+eWXne7yzZgxQzNnztScOXO0c+dORUREaODAgfr+++/NmOzsbK1cuVIFBQXavHmzTp8+rdTUVNXV1Zkx6enp2r17twoLC1VYWKjdu3crIyPDZX0BAABtgzvHSgAAAC1Vo76+d7lp5OfOnZPFYml0QhfaunWr7rvvPg0dOlSS1LlzZ61YsUKffPKJmcvs2bM1efJkjRgxQpL01ltvKTw8XMuXL9djjz0mu92uRYsWaenSpebDPPPz8xUdHa3169dr0KBB2r9/vwoLC7Vt2zYlJSVJkhYuXKjk5GSVlpaaD/i8kMPhkMPhMF+fOnXKZX0HAAAtkzvHSgAAAC1Vo7++d6nBVElJiWw2W6MTutBdd92lv/3tbzpw4IAk6f/+7/+0efNm/fznP5ckHTx4UBUVFeYDOSXJarWqb9++2rJli5lTbW2tU0xUVJTi4+PNmK1bt8pms5kFKUnq1auXbDabGXMx06dPN7/uZ7PZFB0d7bK+AwCAlstdYyUAAICW6qpnSr322mt67bXXJP0wyBo+fLisVqtTTHV1tSorK/XAAw+4LMHf/va3stvtuvnmm+Xl5aW6ujpNnTpVv/zlLyXJXBY5PDzc6X3h4eE6dOiQGePr66v27dvXizn//oqKCoWFhdU7f1hYmBlzMZMmTdKECRPM16dOnaIwBQBAG9RcYyUAAICW6qqLUmFhYerevbsk6euvv9ZNN92k66+/3inGarUqISFBTz75pMsSfOedd5Sfn6/ly5ere/fu2r17t7KzsxUVFaVRo0aZcRfejTQM44pT4y+MuVj8ldqxWq31BpwAAKDtaa6xEgAAQEt11UWpX/7yl+bspHvuuUfz5s3TzTff3GSJnff000/r2Wef1UMPPSRJSkhI0KFDhzR9+nSNGjVKERERkn6Y6RQZGWm+r7Ky0pw9FRERoZqaGlVVVTnNlqqsrFTv3r3NmKNHj9Y7/7Fjx+rNwgIAALhQc42VAAAAWqpGPVNqw4YNbhtk/etf/9J11zmn6eXlpXPnzkmSYmJiFBERoeLiYvN4TU2NNm3aZBacEhMT5ePj4xRTXl6uPXv2mDHJycmy2+1OK+Vs375ddrvdjAEAALga7hwrAQAAtFSNWn1P+uFrbTt37tShQ4dUXV1d7/ivf/3ra0rsvGHDhmnq1Knq2LGjunfvrl27dmnmzJl65JFHJP3wlbvs7GxNmzZNXbp0UZcuXTRt2jS1a9dO6enpkiSbzaYxY8Zo4sSJCgkJUXBwsHJycpSQkGCuxte1a1cNHjxYmZmZmj9/viRp7NixSk1NveTKewAAAJfirrESAABAS9WootSBAweUlpamzz///KJLHlssFpcNtPLy8vS73/1OWVlZqqysVFRUlB577DH9/ve/N2OeeeYZVVdXKysrS1VVVUpKSlJRUZECAwPNmFmzZsnb21sjR45UdXW1+vfvryVLlsjLy8uMWbZsmcaPH2+u0peWlqY5c+a4pB8AAKDtcOdYCQAAoKVqVFHq8ccf17///W+98847uuWWW5r0Qd+BgYGaPXu2Zs+efckYi8Wi3Nxc5ebmXjLGz89PeXl5ysvLu2RMcHCw8vPzryFbAAAA946VAAAAWqpGFaV27NihhQsXspwxAADARTBWAgAAuLJGFaV+8pOfKCgoyNW5AACayKziAy5r66mBsS5rC2itGCsBAABcWaNW33v44Ye1fPlyV+cCAADQKjBWAgAAuLJGzZSKj4/XihUrlJaWpmHDhikkJKRezIgRI645OQAAgJaIsRIAAMCVNaoolZ6eLkk6ePCg1qxZU++4xWJRXV3dtWUGAADQQjFWAoCG43EDQNvTqKLUhg0bXJ0HAABAq8FYCQAA4MoaVZTq27evq/MAAABoNRgrAQAAXFmjHnQOAAAAAAAAXItGzZS69957L3vcYrHob3/7W6MSAgAAaOkYKwEAAFxZo4pS586dk8Vicdr33XffqbS0VGFhYYqN5aFyAACg7WKsBAAAcGWNKkpt3LjxovsPHDig++67T88///y15AQAANCiMVYCAAC4Mpc+Uyo2NlZPP/20nnnmGVc2CwAA0CowVgIAAPj/ufxB5507d9aePXtc3SwAAECrwFgJAADgBy4vSr377ruKiopydbMAAACtAmMlAACAHzTqmVKPPPJIvX0Oh0Offvqp9u3bpxkzZlxzYgAAAC0VYyUAAIAra1RR6sMPP6y3ooyfn586d+6sSZMmKT093SXJAQAAtESMlQAAAK6sUUWpr7/+2sVpAAAAtB6MlQAAAK7M5c+UAgAAAAAAAK6kUTOlJOnEiROaNWuW/va3v+n48eMKDQ3VgAEDlJ2drfbt27syRwAAgBaHsRIAAMDlNWqm1DfffKM77rhDU6dOld1uV8eOHXXy5En9z//8j+644w59++23rs4TAACgxWCsBAAAcGWNKko999xzqq6u1vbt27V3714VFxdr79692r59u6qrq/Xcc8+5Ok8AAIAWg7ESAADAlTWqKFVYWKgpU6bozjvvdNp/55136sUXX9S6detckhwAAEBLxFgJAADgyhpVlLLb7ercufNFj8XExMhut19LTgAAAC0aYyUAAIAra1RRKiYmRmvXrr3osXXr1ikmJuaakgIAAGjJGCsBAABcWaNW33v44Yf17LPP6ty5cxo1apQiIyNVXl6u/Px85eXl6Q9/+IOr8wQAAGgxGCsBAABcWaNmSj399NPKzMzUnDlz1LNnT0VHR6tnz5567bXX9OijjyonJ8fVeQIAALQYrhwrffTRRxo2bJiioqJksVj0/vvvOx0fPXq0LBaL09arVy+nGIfDoXHjxik0NFQBAQFKS0tTWVmZU0xVVZUyMjJks9lks9mUkZGhkydPOsUcPnxYw4YNU0BAgEJDQzV+/HjV1NQ06LMBAAA4r1EzpSwWi+bPn68JEyZow4YNOn78uEJCQnTvvfcqNjbW1TkCAAC0KK4cK505c0a33nqrHn74Yf3iF7+4aMzgwYO1ePFi87Wvr6/T8ezsbK1evVoFBQUKCQnRxIkTlZqaqpKSEnl5eUmS0tPTVVZWpsLCQknS2LFjlZGRodWrV0uS6urqNHToUN1www3avHmzjh8/rlGjRskwDOXl5TWoTwAAAFIDilJVVVV69NFH9fDDDys1NVWSFBcXp7i4ODNmzZo1mjRpkhYsWKCQkBDXZwsAAOChmmqsNGTIEA0ZMuSyMVarVRERERc9ZrfbtWjRIi1dulQDBgyQJOXn5ys6Olrr16/XoEGDtH//fhUWFmrbtm1KSkqSJC1cuFDJyckqLS1VXFycioqKtG/fPh05ckRRUVGSpFdffVWjR4/W1KlTFRQUdNHzOxwOORwO8/WpU6euqt9txaziAy5r66mB3ByGe7jy51biZxdoy67663t/+tOf9H//938aPHjwJWMGDx6szz77TK+//rpLkgMAAGgpmnOstHHjRoWFhSk2NlaZmZmqrKw0j5WUlKi2tlYpKSnmvqioKMXHx2vLli2SpK1bt8pms5kFKUnq1auXbDabU0x8fLxZkJKkQYMGyeFwqKSk5JK5TZ8+3fxKoM1mU3R0tMv6DQAAWrarLkoVFBQoMzNT3t6Xnlzl7e2tzMxMrVq1yiXJAQAAtBTNNVYaMmSIli1bpg8//FCvvvqqdu7cqXvvvdecnVRRUSFfX1+1b9/e6X3h4eGqqKgwY8LCwuq1HRYW5hQTHh7udLx9+/by9fU1Yy5m0qRJstvt5nbkyJFr6i8AAGg9rroodeDAAfXo0eOKcXfccYcOHHDtdE4AAABP11xjpQcffFBDhw5VfHy8hg0bpnXr1unAgQNau3btZd9nGIYsFov5+sf/vpaYC1mtVgUFBTltAAAAUgOKUmfPnpWPj88V43x8fFRbW3tNSQEAALQ0njJWioyMVKdOnfT5559LkiIiIlRTU6OqqiqnuMrKSnPmU0REhI4ePVqvrWPHjjnFXDgjqqqqSrW1tfVmUAEAAFyNqy5KRUZGat++fVeM27t37yUftAkAANBaecpY6fjx4zpy5IgiIyMlSYmJifLx8VFxcbEZU15erj179qh3796SpOTkZNntdu3YscOM2b59u+x2u1PMnj17VF5ebsYUFRXJarUqMTGxyfoDAABar6tefa9v376aO3euxowZc8m7gLW1tZo3b57uuecelyUoSd98841++9vfat26daqurlZsbKwWLVpkDoAMw9ALL7ygBQsWqKqqSklJSXr99dfVvXt3sw2Hw6GcnBytWLFC1dXV6t+/v+bOnasOHTqYMVVVVRo/frz5nIe0tDTl5eXp+uuvd2l/AADwFKz85TpNNVY6ffq0vvjiC/P1wYMHtXv3bgUHBys4OFi5ubn6xS9+ocjISH399dd67rnnFBoaqvvvv1+SZLPZNGbMGE2cOFEhISEKDg5WTk6OEhISzNX4unbtqsGDByszM1Pz58+XJI0dO1apqanm6oEpKSnq1q2bMjIy9PLLL+vEiRPKyclRZmYmX8kDAACNctVFqaeeeko9evTQ/fffrwULFjitvCJJ3377rTIzM1VaWqply5a5LMGqqir16dNH99xzj9atW6ewsDB9+eWXToWiGTNmaObMmVqyZIliY2M1ZcoUDRw4UKWlpQoMDJQkZWdna/Xq1SooKFBISIgmTpyo1NRUlZSUyMvLS5KUnp6usrIyFRYWSvphMJaRkaHVq1e7rD8AAKB1aqqx0ieffOJUxJowYYIkadSoUZo3b54+++wzvf322zp58qQiIyN1zz336J133jHHQJI0a9YseXt7a+TIkebNuSVLlphjIElatmyZxo8fb67Sl5aWpjlz5pjHvby8tHbtWmVlZalPnz7y9/dXenq6XnnllYZ9UADcxpU3HiRuPgBwvasuSt1yyy16/fXXlZWVpZiYGCUmJiomJkbSD3fsSkpKdO7cOc2bN08JCQkuS/Cll15SdHS0Fi9ebO7r3Lmz+W/DMDR79mxNnjxZI0aMkCS99dZbCg8P1/Lly/XYY4/Jbrdr0aJFWrp0qXlHMD8/X9HR0Vq/fr0GDRqk/fv3q7CwUNu2bTOXQ164cKGSk5NVWlpq3iW8kMPhMFe3kaRTp065rO8AAKDlaKqxUr9+/WQYxiWP//Wvf71iG35+fsrLy1NeXt4lY4KDg5Wfn3/Zdjp27Kg1a9Zc8XwAAABX46qfKSVJmZmZ+uijj5SSkqJPP/1UK1as0IoVK/Tpp59q8ODB+vjjj/Xoo4+6NMFVq1apR48e+s///E+FhYXp9ttv18KFC83jBw8eVEVFhXlXT/phlZe+fftqy5YtkqSSkhLV1tY6xURFRSk+Pt6M2bp1q2w2m1mQkqRevXrJZrOZMRczffp02Ww2c4uOjnZZ3wEAQMvSHGMlAACAluqqZ0qdl5ycrNWrV+vcuXP67rvvJEmhoaG67roG1beu2ldffaV58+ZpwoQJeu6557Rjxw6NHz9eVqtVv/71r81VYC5c9SU8PFyHDh2SJFVUVMjX11ft27evF3P+/RUVFQoLC6t3/rCwsHorzfzYpEmTzGn00g8zpShMAWgreB6R+/BZtxzuHisBAAC0VA0uSp133XXXXbSI42rnzp1Tjx49NG3aNEnS7bffrr1792revHn69a9/bcZZLBan9xmGUW/fhS6MuVj8ldqxWq2yWq1X1RcAANB2uGusBAAA0FJ5/C27yMhIdevWzWlf165ddfjwYUkyl1S+cDZTZWWlOXsqIiJCNTU1qqqqumzM0aNH653/2LFj9WZhAQAAAAAA4Np4fFGqT58+Ki0tddp34MABderUSZIUExOjiIgIFRcXm8dramq0adMm9e7dW5KUmJgoHx8fp5jy8nLt2bPHjElOTpbdbteOHTvMmO3bt8tut5sxAAAAAAAAcI1Gf33PXZ566in17t1b06ZN08iRI7Vjxw4tWLBACxYskPTDV+6ys7M1bdo0denSRV26dNG0adPUrl07paenS5JsNpvGjBmjiRMnKiQkRMHBwcrJyVFCQoK5Gl/Xrl01ePBgZWZmav78+ZKksWPHKjU19ZIr7wEAAAAAAKBxPL4odeedd2rlypWaNGmSXnzxRcXExGj27Nn61a9+ZcY888wzqq6uVlZWlqqqqpSUlKSioiIFBgaaMbNmzZK3t7dGjhyp6upq9e/fX0uWLJGXl5cZs2zZMo0fP95cpS8tLU1z5sxxX2cBAAAAAADaCI8vSklSamqqUlNTL3ncYrEoNzdXubm5l4zx8/NTXl6e8vLyLhkTHBys/Pz8a0kVAAAAAAAAV6FFFKUAAAAAoLWaVXzApe09NTDWpe0BQFPx+AedAwAAAAAAoPWhKAUAAAAAAAC34+t7AACP5MqvMvA1BgAAAMDzMFMKAAAAAAAAbsdMKQAAAAAAWjEepg9PxUwpAAAAAAAAuB0zpQAAbQrPqgIAAAA8AzOlAAAAAAAA4HYUpQAAAAAAAOB2fH0Pkvg6CwAAAAAAcC9mSgEAAAAAAMDtKEoBAAAAAADA7ShKAQAAAAAAwO0oSgEAAAAAAMDtKEoBAAAAAADA7ShKAQAAAAAAwO0oSgEAAAAAAMDtKEoBAAAAAADA7ShKAQAAAAAAwO0oSgEAAAAAAMDtKEoBAAAAAADA7ShKAQAAAAAAwO28mzsBAABai1nFB1zW1lMDY13WFgAAAOCJKEoBAAAAaHVceaNA+uFmQVO0CQBtGV/fAwAAAAAAgNtRlAIAAAAAAIDbUZQCAADwYB999JGGDRumqKgoWSwWvf/++07HDcNQbm6uoqKi5O/vr379+mnv3r1OMQ6HQ+PGjVNoaKgCAgKUlpamsrIyp5iqqiplZGTIZrPJZrMpIyNDJ0+edIo5fPiwhg0bpoCAAIWGhmr8+PGqqalpim4DAIA2gKIUAACABztz5oxuvfVWzZkz56LHZ8yYoZkzZ2rOnDnauXOnIiIiNHDgQH3//fdmTHZ2tlauXKmCggJt3rxZp0+fVmpqqurq6syY9PR07d69W4WFhSosLNTu3buVkZFhHq+rq9PQoUN15swZbd68WQUFBXr33Xc1ceLEpus8AABo1XjQOQAAgAcbMmSIhgwZctFjhmFo9uzZmjx5skaMGCFJeuuttxQeHq7ly5frsccek91u16JFi7R06VINGDBAkpSfn6/o6GitX79egwYN0v79+1VYWKht27YpKSlJkrRw4UIlJyertLRUcXFxKioq0r59+3TkyBFFRUVJkl599VWNHj1aU6dOVVBQkBs+DVwNHsYNAGgpmCkFAADQQh08eFAVFRVKSUkx91mtVvXt21dbtmyRJJWUlKi2ttYpJioqSvHx8WbM1q1bZbPZzIKUJPXq1Us2m80pJj4+3ixISdKgQYPkcDhUUlJyyRwdDodOnTrltAEAAEgUpQAAAFqsiooKSVJ4eLjT/vDwcPNYRUWFfH191b59+8vGhIWF1Ws/LCzMKebC87Rv316+vr5mzMVMnz7dfE6VzWZTdHR0A3sJAABaK76+hybF9HEAAJqexWJxem0YRr19F7ow5mLxjYm50KRJkzRhwgTz9alTpyhMAQAASS1wptT06dNlsViUnZ1t7nPnqjMAAACeIiIiQpLqzVSqrKw0ZzVFRESopqZGVVVVl405evRovfaPHTvmFHPheaqqqlRbW1tvBtWPWa1WBQUFOW0AAABSC5sptXPnTi1YsEC33HKL0/7zq84sWbJEsbGxmjJligYOHKjS0lIFBgZK+mHVmdWrV6ugoEAhISGaOHGiUlNTVVJSIi8vL0k/rDpTVlamwsJCSdLYsWOVkZGh1atXu7ejAAAAVyEmJkYREREqLi7W7bffLkmqqanRpk2b9NJLL0mSEhMT5ePjo+LiYo0cOVKSVF5erj179mjGjBmSpOTkZNntdu3YsUM9e/aUJG3fvl12u129e/c2Y6ZOnary8nJFRkZKkoqKimS1WpWYmOjWfqN5MAMeAOBqLaYodfr0af3qV7/SwoULNWXKFHO/O1eduRiHwyGHw2G+5uGdAADAlU6fPq0vvvjCfH3w4EHt3r1bwcHB6tixo7KzszVt2jR16dJFXbp00bRp09SuXTulp6dLkmw2m8aMGaOJEycqJCREwcHBysnJUUJCgjku6tq1qwYPHqzMzEzNnz9f0g8351JTU80xUEpKirp166aMjAy9/PLLOnHihHJycpSZmcnsJwAA0Cgt5ut7jz/+uIYOHWoOns5z56ozF8PDOwEAQFP65JNPdPvtt5szoSZMmKDbb79dv//97yVJzzzzjLKzs5WVlaUePXrom2++UVFRkTlbXJJmzZql4cOHa+TIkerTp4/atWun1atXm7PFJWnZsmVKSEhQSkqKUlJSdMstt2jp0qXmcS8vL61du1Z+fn7q06ePRo4cqeHDh+uVV15x0ycBAABamxYxU6qgoED/+Mc/tHPnznrHLrfqzKFDh8wYV6w6czE8vBMAADSlfv36yTCMSx63WCzKzc1Vbm7uJWP8/PyUl5envLy8S8YEBwcrPz//srl07NhRa9asuWLOAAAAV8Pji1JHjhzRk08+qaKiIvn5+V0yzl2rzlzIarXKarVe9jwAAAAAAABw5vFf3yspKVFlZaUSExPl7e0tb29vbdq0SX/84x/l7e1tzpByx6ozAAAAAAAAcA2PnynVv39/ffbZZ077Hn74Yd1888367W9/q5tuusltq84AQFNy5apGrGgEAGgpWNUPANoujy9KBQYGKj4+3mlfQECAQkJCzP3uWnUGAAAAAAAAruHxRamr8cwzz6i6ulpZWVmqqqpSUlLSRVed8fb21siRI1VdXa3+/ftryZIl9VadGT9+vLlKX1pamubMmeP2/gAAAAAAALR2LbIotXHjRqfX7lx1BgAAAAAAANfO4x90DgAAAAAAgNaHohQAAAAAAADcjqIUAAAAAAAA3I6iFAAAAAAAANyOohQAAAAAAADcjqIUAAAAAAAA3I6iFAAAAAAAANzOu7kTABprVvEBl7X11MBYl7UFAAAAAACujJlSAAAAAAAAcDuKUgAAAAAAAHA7ilIAAAAAAABwO4pSAAAAAAAAcDuKUgAAAAAAAHA7ilIAAAAAAABwO4pSAAAAAAAAcDuKUgAAAAAAAHA7ilIAAAAAAABwO4pSAAAAAAAAcDuKUgAAAAAAAHA7ilIAAAAAAABwO4pSAAAAAAAAcDvv5k4AAAAA8HSzig+4tL2nBsa6tD0AAFoiZkoBAAAAAADA7ShKAQAAAAAAwO0oSgEAAAAAAMDteKYUcBGufG4Ez4wAAAAAAKA+ZkoBAAAAAADA7ShKAQAAAAAAwO0oSgEAAAAAAMDteKYU4EY8qwoAAAAAgB8wUwoAAKAFy83NlcVicdoiIiLM44ZhKDc3V1FRUfL391e/fv20d+9epzYcDofGjRun0NBQBQQEKC0tTWVlZU4xVVVVysjIkM1mk81mU0ZGhk6ePOmOLgIAgFaKmVIA0ADMdgPgibp3767169ebr728vMx/z5gxQzNnztSSJUsUGxurKVOmaODAgSotLVVgYKAkKTs7W6tXr1ZBQYFCQkI0ceJEpaamqqSkxGwrPT1dZWVlKiwslCSNHTtWGRkZWr16tRt7CgAAWhOPL0pNnz5d7733nv75z3/K399fvXv31ksvvaS4uDgzxjAMvfDCC1qwYIGqqqqUlJSk119/Xd27dzdjHA6HcnJytGLFClVXV6t///6aO3euOnToYMZUVVVp/PjxWrVqlSQpLS1NeXl5uv76693WX6CxKJYAQNvl7e3tNDvqPMMwNHv2bE2ePFkjRoyQJL311lsKDw/X8uXL9dhjj8lut2vRokVaunSpBgwYIEnKz89XdHS01q9fr0GDBmn//v0qLCzUtm3blJSUJElauHChkpOTVVpa6jQuu5DD4ZDD4TBfnzp1ypVdBwAALZjHf31v06ZNevzxx7Vt2zYVFxfr7NmzSklJ0ZkzZ8yY83cA58yZo507dyoiIkIDBw7U999/b8ZkZ2dr5cqVKigo0ObNm3X69Gmlpqaqrq7OjElPT9fu3btVWFiowsJC7d69WxkZGW7tLwAAQEN9/vnnioqKUkxMjB566CF99dVXkqSDBw+qoqJCKSkpZqzValXfvn21ZcsWSVJJSYlqa2udYqKiohQfH2/GbN26VTabzSxISVKvXr1ks9nMmEuZPn26+ZU/m82m6Ohol/UbAAC0bB4/U+r8FPHzFi9erLCwMJWUlOjuu+/2iDuAAAAAzSUpKUlvv/22YmNjdfToUU2ZMkW9e/fW3r17VVFRIUkKDw93ek94eLgOHTokSaqoqJCvr6/at29fL+b8+ysqKhQWFlbv3GFhYWbMpUyaNEkTJkwwX586dYrCFAAAkNQCZkpdyG63S5KCg4MlNf8dQIfDoVOnTjltAAAA7jJkyBD94he/UEJCggYMGKC1a9dK+uEm3XkWi8XpPYZh1Nt3oQtjLhZ/Ne1YrVYFBQU5bQAAAFILK0oZhqEJEyborrvuUnx8vCRd9g7gj+/uNdUdQKakAwAATxIQEKCEhAR9/vnn5nOmLhzLVFZWmmOniIgI1dTUqKqq6rIxR48erXeuY8eO1RuDAQAAXK0WVZR64okn9Omnn2rFihX1jjXXHcBJkybJbreb25EjR67UDQAAgCbjcDi0f/9+RUZGKiYmRhERESouLjaP19TUaNOmTerdu7ckKTExUT4+Pk4x5eXl2rNnjxmTnJwsu92uHTt2mDHbt2+X3W43YwAAABrK458pdd64ceO0atUqffTRR04r5v34DmBkZKS5/1J3AH88W6qystIcSDX2DqDVapXVar22zgEAADRSTk6Ohg0bpo4dO6qyslJTpkzRqVOnNGrUKFksFmVnZ2vatGnq0qWLunTpomnTpqldu3ZKT0+XJNlsNo0ZM0YTJ05USEiIgoODlZOTY34dUJK6du2qwYMHKzMzU/Pnz5ckjR07VqmpqTx3EwAANJrHz5QyDENPPPGE3nvvPX344YeKiYlxOs4dQAAA0JaVlZXpl7/8peLi4jRixAj5+vpq27Zt6tSpkyTpmWeeUXZ2trKystSjRw998803KioqUmBgoNnGrFmzNHz4cI0cOVJ9+vRRu3bttHr1anl5eZkxy5YtU0JCglJSUpSSkqJbbrlFS5cudXt/AQBA6+HxM6Uef/xxLV++XH/5y18UGBhoPhPBZrPJ39+fO4AAAKBNKygouOxxi8Wi3Nxc5ebmXjLGz89PeXl5ysvLu2RMcHCw8vPzG5smAABAPR5flJo3b54kqV+/fk77Fy9erNGjR0v64Q5gdXW1srKyVFVVpaSkpIveAfT29tbIkSNVXV2t/v37a8mSJfXuAI4fP95cpS8tLU1z5sxp2g4CAAAAAAC0QR5flDIM44ox3AEEWqZZxQdc1tZTA2Nd1hYAAAAAoOl5/DOlAAAAAAAA0PpQlAIAAAAAAIDbUZQCAAAAAACA23n8M6UAAAAAAACuhSufZyvxTFtXoSgF4LJa6n/ePEQdAAAAADwbX98DAAAAAACA21GUAgAAAAAAgNtRlAIAAAAAAIDbUZQCAAAAAACA21GUAgAAAAAAgNtRlAIAAAAAAIDbUZQCAAAAAACA21GUAgAAAAAAgNtRlAIAAAAAAIDbUZQCAAAAAACA21GUAgAAAAAAgNtRlAIAAAAAAIDbUZQCAAAAAACA21GUAgAAAAAAgNtRlAIAAAAAAIDbUZQCAAAAAACA21GUAgAAAAAAgNtRlAIAAAAAAIDbUZQCAAAAAACA21GUAgAAAAAAgNtRlAIAAAAAAIDbUZQCAAAAAACA21GUAgAAAAAAgNtRlAIAAAAAAIDbUZQCAAAAAACA21GUAgAAAAAAgNtRlAIAAAAAAIDbUZQCAAAAAACA21GUuoi5c+cqJiZGfn5+SkxM1Mcff9zcKQEAAHgMxkoAAMAVKEpd4J133lF2drYmT56sXbt26Wc/+5mGDBmiw4cPN3dqAAAAzY6xEgAAcBWKUheYOXOmxowZo0cffVRdu3bV7NmzFR0drXnz5jV3agAAAM2OsRIAAHAV7+ZOwJPU1NSopKREzz77rNP+lJQUbdmy5aLvcTgccjgc5mu73S5JOnXqlMvz+/eZ0y5r68L8mqptV7bblG23xM+jJebclG23xM/DXTk3ZdvkzDW8VLtN3XZTOH8OwzCa/FwtlaePlaSm+7vVEtptSbm21HZbUq5tvd2WlGtLbbcl5doS2/VETTJWMmD65ptvDEnG3//+d6f9U6dONWJjYy/6nueff96QxMbGxsbGxtZKtiNHjrhj2NEiMVZiY2NjY2Nj+/LLL102tmCm1EVYLBan14Zh1Nt33qRJkzRhwgTz9blz53TixAmFhIRc8j2udurUKUVHR+vIkSMKCgpyyzk9RVvuu0T/23L/23LfJfrflvvflH03DEPff/+9oqKiXNpua3QtY6WTJ0+qU6dOOnz4sGw2W5Pmictry/+XeCKuh+fgWngOroVnsdvt6tixo4KDg13WJkWpHwkNDZWXl5cqKiqc9ldWVio8PPyi77FarbJarU77rr/++qZK8bKCgoLa7C9qW+67RP/bcv/bct8l+t+W+99UfadIcnmuGitJP3zWbfXn19O05f9LPBHXw3NwLTwH18KzXHed6x5PzoPOf8TX11eJiYkqLi522l9cXKzevXs3U1YAAACegbESAABwJWZKXWDChAnKyMhQjx49lJycrAULFujw4cP6zW9+09ypAQAANDvGSgAAwFUoSl3gwQcf1PHjx/Xiiy+qvLxc8fHx+uCDD9SpU6fmTu2SrFarnn/++YtOjW/t2nLfJfrflvvflvsu0f+23P+23HdPca1jJa6h5+BaeBauh+fgWngOroVnaYrrYTEM1j0GAAAAAACAe/FMKQAAAAAAALgdRSkAAAAAAAC4HUUpAAAAAAAAuB1FKQAAAAAAALgdRakWYO7cuYqJiZGfn58SExP18ccfXzJ28+bN6tOnj0JCQuTv76+bb75Zs2bNcmO2rteQ/v/Y3//+d3l7e+u2225r2gSbWEP6v3HjRlkslnrbP//5Tzdm7DoNvfYOh0OTJ09Wp06dZLVa9dOf/lRvvvmmm7J1vYb0f/To0Re99t27d3djxq7V0Ou/bNky3XrrrWrXrp0iIyP18MMP6/jx427K1vUa2v/XX39dXbt2lb+/v+Li4vT222+7KVPX+uijjzRs2DBFRUXJYrHo/fffv+J7Nm3apMTERPn5+emmm27SG2+80fSJ4pIa+rPL9WtaDbke7733ngYOHKgbbrhBQUFBSk5O1l//+lc3Ztu6tfUxradp6+NMT9LWx3yeotnGYAY8WkFBgeHj42MsXLjQ2Ldvn/Hkk08aAQEBxqFDhy4a/49//MNYvny5sWfPHuPgwYPG0qVLjXbt2hnz5893c+au0dD+n3fy5EnjpptuMlJSUoxbb73VPck2gYb2f8OGDYYko7S01CgvLze3s2fPujnza9eYa5+WlmYkJSUZxcXFxsGDB43t27cbf//7392Ytes0tP8nT550uuZHjhwxgoODjeeff969ibtIQ/v/8ccfG9ddd53x2muvGV999ZXx8ccfG927dzeGDx/u5sxdo6H9nzt3rhEYGGgUFBQYX375pbFixQrjJz/5ibFq1So3Z37tPvjgA2Py5MnGu+++a0gyVq5cedn4r776ymjXrp3x5JNPGvv27TMWLlxo+Pj4GH/+85/dkzCcNPRnl+vXtBp6PZ588knjpZdeMnbs2GEcOHDAmDRpkuHj42P84x//cHPmrU9bH9N6mrY+zvQkbX3M50maawxGUcrD9ezZ0/jNb37jtO/mm282nn322atu4/777zf+67/+y9WpuUVj+//ggw8a//3f/208//zzLfoPeEP7f74oVVVV5YbsmlZD+75u3TrDZrMZx48fd0d6Te5af/dXrlxpWCwW4+uvv26K9JpcQ/v/8ssvGzfddJPTvj/+8Y9Ghw4dmizHptTQ/icnJxs5OTlO+5588kmjT58+TZajO1zNgOiZZ54xbr75Zqd9jz32mNGrV68mzAyX0tCfXa5f03LFOLJbt27GCy+84OrU2py2Pqb1NG19nOlJ2vqYz1O5cwzG1/c8WE1NjUpKSpSSkuK0PyUlRVu2bLmqNnbt2qUtW7aob9++TZFik2ps/xcvXqwvv/xSzz//fFOn2KSu5frffvvtioyMVP/+/bVhw4amTLNJNKbvq1atUo8ePTRjxgzdeOONio2NVU5Ojqqrq92Rsku54nd/0aJFGjBggDp16tQUKTapxvS/d+/eKisr0wcffCDDMHT06FH9+c9/1tChQ92Rsks1pv8Oh0N+fn5O+/z9/bVjxw7V1tY2Wa6eYOvWrfU+q0GDBumTTz5p9X33NI352eX6NR1X/C05d+6cvv/+ewUHBzdFim1GWx/Tepq2Ps70JG19zNfSuepvuLerE4PrfPfdd6qrq1N4eLjT/vDwcFVUVFz2vR06dNCxY8d09uxZ5ebm6tFHH23KVJtEY/r/+eef69lnn9XHH38sb++W/ePdmP5HRkZqwYIFSkxMlMPh0NKlS9W/f39t3LhRd999tzvSdonG9P2rr77S5s2b5efnp5UrV+q7775TVlaWTpw40eK+738tv/uSVF5ernXr1mn58uVNlWKTakz/e/furWXLlunBBx/Uv//9b509e1ZpaWnKy8tzR8ou1Zj+Dxo0SH/60580fPhw3XHHHSopKdGbb76p2tpafffdd4qMjHRH6s2ioqLiop/V2bNnW33fPU1jfna5fk3nWv+WSNKrr76qM2fOaOTIkU2RYpvR1se0nqatjzM9SVsf87V0rvobzkypFsBisTi9Ngyj3r4Lffzxx/rkk0/0xhtvaPbs2VqxYkVTptikrrb/dXV1Sk9P1wsvvKDY2Fh3pdfkGnL94+LilJmZqTvuuEPJycmaO3euhg4dqldeecUdqbpcQ/p+7tw5WSwWLVu2TD179tTPf/5zzZw5U0uWLGmxd7Ea87svSUuWLNH111+v4cOHN1Fm7tGQ/u/bt0/jx4/X73//e5WUlKiwsFAHDx7Ub37zG3ek2iQa0v/f/e53GjJkiHr16iUfHx/dd999Gj16tCTJy8urqVNtdhf7rC62H+7R0P+7uH5Nq7F/S1asWKHc3Fy98847CgsLa6r02pS2Pqb1NG19nOlJ2vqYryVzxd9wyu4eLDQ0VF5eXvWqxJWVlfUqkheKiYmRJCUkJOjo0aPKzc3VL3/5yybLtSk0tP/ff/+9PvnkE+3atUtPPPGEpB/+gBiGIW9vbxUVFenee+91S+6ucC3X/8d69eql/Px8V6fXpBrT98jISN14442y2Wzmvq5du8owDJWVlalLly5NmrMrXcu1NwxDb775pjIyMuTr69uUaTaZxvR/+vTp6tOnj55++mlJ0i233KKAgAD97Gc/05QpU1rUbIvG9N/f319vvvmm5s+fr6NHj5qzJgMDAxUaGuqOtJtNRETERT8rb29vhYSENFNWbVNjfna5fk3nWv6WvPPOOxozZoz+93//VwMGDGjKNNuEtj6m9TRtfZzpSdr6mK+lc9XfcGZKeTBfX18lJiaquLjYaX9xcbF69+591e0YhiGHw+Hq9JpcQ/sfFBSkzz77TLt37za33/zmN4qLi9Pu3buVlJTkrtRdwlXXf9euXS3uP+fG9L1Pnz769ttvdfr0aXPfgQMHdN1116lDhw5Nmq+rXcu137Rpk7744guNGTOmKVNsUo3p/7/+9S9dd53zn7TzM4TO37FpKa7l+vv4+KhDhw7y8vJSQUGBUlNT630urU1ycnK9z6qoqEg9evSQj49PM2XVNjXmZ5fr13Qa+3/JihUrNHr0aC1fvpxntLhIWx/Tepq2Ps70JG19zNfSuexveIMeiw63O79E5qJFi4x9+/YZ2dnZRkBAgLmi1rPPPmtkZGSY8XPmzDFWrVplHDhwwDhw4IDx5ptvGkFBQcbkyZObqwvXpKH9v1BLX6mkof2fNWuWsXLlSuPAgQPGnj17jGeffdaQZLz77rvN1YVGa2jfv//+e6NDhw7GAw88YOzdu9fYtGmT0aVLF+PRRx9tri5ck8b+7P/Xf/2XkZSU5O50Xa6h/V+8eLHh7e1tzJ071/jyyy+NzZs3Gz169DB69uzZXF24Jg3tf2lpqbF06VLjwIEDxvbt240HH3zQCA4ONg4ePNhMPWi877//3ti1a5exa9cuQ5Ixc+ZMY9euXebS0Bf2/fxyxE899ZSxb98+Y9GiRY1ajhiu0dCfXa5f02ro9Vi+fLnh7e1tvP7660Z5ebm5nTx5srm60Gq09TGtp2nr40xP0tbHfJ6kucZgFKVagNdff93o1KmT4evra9xxxx3Gpk2bzGOjRo0y+vbta77+4x//aHTv3t1o166dERQUZNx+++3G3Llzjbq6umbI3DUa0v8LtYY/4A3p/0svvWT89Kc/Nfz8/Iz27dsbd911l7F27dpmyNo1Gnrt9+/fbwwYMMDw9/c3OnToYEyYMMH417/+5easXaeh/T958qTh7+9vLFiwwM2ZNo2G9v+Pf/yj0a1bN8Pf39+IjIw0fvWrXxllZWVuztp1GtL/ffv2Gbfddpvh7+9vBAUFGffdd5/xz3/+sxmyvnYbNmwwJNXbRo0aZRjGxa/9xo0bjdtvv93w9fU1OnfubMybN8/9icPU0N9drl/Tasj16Nu372V//3Bt2vqY1tO09XGmJ2nrYz5P0VxjMIthMMcNAAAAAAAA7tW6HzQBAAAAAAAAj0RRCgAAAAAAAG5HUQoAAAAAAABuR1EKAAAAAAAAbkdRCgAAAAAAAG5HUQoAAAAAAABuR1EKAAAAAAAAbkdRCgAAAAAAAG5HUQpAs/v000/18MMPKyYmRn5+fvrJT36iO+64QzNmzNCJEyea7Ly7du1S3759ZbPZZLFYNHv2bG3cuFEWi0UbN2684vtHjx6tzp07N1l+AAAAEmMlAK2Xd3MnAKBtW7hwobKyshQXF6enn35a3bp1U21trT755BO98cYb2rp1q1auXNkk537kkUd05swZFRQUqH379urcubPatWunrVu3qlu3bk1yTgAAgIZgrASgNbMYhmE0dxIA2qatW7fqZz/7mQYOHKj3339fVqvV6XhNTY0KCwuVlpbWJOf38fFRZmam5s6d26j3jx49Whs3btTXX3/t2sQAAADEWAlA68fX9wA0m2nTpslisWjBggX1BlmS5Ovraw6yzp07pxkzZujmm2+W1WpVWFiYfv3rX6usrMzpPf369VN8fLx27typn/3sZ2rXrp1uuukm/eEPf9C5c+ckSUuWLJHFYtHZs2c1b948WSwWWSwWSbrklPQlS5YoLi5OVqtVXbt21dtvv33RPtXU1GjKlClmnjfccIMefvhhHTt2zCmuc+fOSk1NVWFhoe644w75+/vr5ptv1ptvvlmvzW+++UZjx45VdHS0fH19FRUVpQceeEBHjx41Y06dOqWcnBzFxMTI19dXN954o7Kzs3XmzJkrXAUAAOCpGCsxVgJaPQMAmsHZs2eNdu3aGUlJSVcVP3bsWEOS8cQTTxiFhYXGG2+8Ydxwww1GdHS0cezYMTOub9++RkhIiNGlSxfjjTfeMIqLi42srCxDkvHWW28ZhmEYlZWVxtatWw1JxgMPPGBs3brV2Lp1q2EYhrFhwwZDkrFhwwazzcWLFxuSjPvuu89YvXq1kZ+fb/zHf/yHER0dbXTq1MmMq6urMwYPHmwEBAQYL7zwglFcXGz86U9/Mm688UajW7duxr/+9S8ztlOnTkaHDh2Mbt26GW+//bbx17/+1fjP//xPQ5KxadMmM66srMyIjIw0QkNDjZkzZxrr16833nnnHeORRx4x9u/fbxiGYZw5c8a47bbbnGJee+01w2azGffee69x7ty5Bl8fAADQvBgrMVYC2gKKUgCaRUVFhSHJeOihh64Yu3//fkOSkZWV5bR/+/bthiTjueeeM/f17dvXkGRs377dKbZbt27GoEGDnPZJMh5//HGnfRcOtOrq6oyoqCjjjjvucBqwfP3114aPj4/TQGvFihWGJOPdd991anPnzp2GJGPu3Lnmvk6dOhl+fn7GoUOHzH3V1dVGcHCw8dhjj5n7HnnkEcPHx8fYt2/fJT+f6dOnG9ddd52xc+dOp/1//vOfDUnGBx98cMn3AgAAz8RYibES0Bbw9T0AHm/Dhg2SfnguwY/17NlTXbt21d/+9jen/REREerZs6fTvltuuUWHDh1q8LlLS0v17bffKj093Zy2LkmdOnVS7969nWLXrFmj66+/XsOGDdPZs2fN7bbbblNERES9ae633XabOnbsaL728/NTbGysU57r1q3TPffco65du14yxzVr1ig+Pl633Xab03kHDRp01avjAACAlouxEmMloKVi9T0AzSI0NFTt2rXTwYMHrxh7/PhxSVJkZGS9Y1FRUfUGUCEhIfXirFarqqurG5zn+XNHRETUOxYREeH04M6jR4/q5MmT8vX1vWhb3333XYPzPHbsmDp06HDZHI8ePaovvvhCPj4+V3VeAADg+RgrMVYC2gKKUgCahZeXl/r3769169aprKzssoOJ8wOS8vLyenHffvutQkNDmyzP8+euqKiod+zCfaGhoQoJCVFhYeFF2woMDGzw+W+44YZ6Dyi9UGhoqPz9/S/64M/zxwEAQMvCWOnqMFYCWja+vgeg2UyaNEmGYSgzM1M1NTX1jtfW1mr16tW69957JUn5+flOx3fu3Kn9+/erf//+TZZjXFycIiMjtWLFChmGYe4/dOiQtmzZ4hSbmpqq48ePq66uTj169Ki3xcXFNfj8Q4YM0YYNG1RaWnrJmNTUVH355ZcKCQm56Hk7d+7c4PMCAIDmx1jpyhgrAS0bM6UANJvk5GTNmzdPWVlZSkxM1P/7f/9P3bt3V21trXbt2qUFCxYoPj5eK1eu1NixY5WXl6frrrtOQ4YM0ddff63f/e53io6O1lNPPdVkOV533XX6n//5Hz366KO6//77lZmZqZMnTyo3N7feNPWHHnpIy5Yt089//nM9+eST6tmzp3x8fFRWVqYNGzbovvvu0/3339+g87/44otat26d7r77bj333HNKSEjQyZMnVVhYqAkTJujmm29Wdna23n33Xd1999166qmndMstt+jcuXM6fPiwioqKNHHiRCUlJbnyYwEAAG7AWOnKGCsBLRtFKQDNKjMzUz179tSsWbP00ksvqaKiQj4+PoqNjVV6erqeeOIJSdK8efP005/+VIsWLdLrr78um82mwYMHa/r06Rd93oArjRkzRpL00ksvacSIEercubOee+45bdq0yenBmF5eXlq1apVee+01LV26VNOnT5e3t7c6dOigvn37KiEhocHnvvHGG7Vjxw49//zz+sMf/qDjx4/rhhtu0F133aXg4GBJUkBAgD7++GP94Q9/0IIFC3Tw4EH5+/urY8eOGjBgAHf/AABowRgrXR5jJaBlsxg/nmMJAAAAAAAAuAHPlAIAAAAAAIDbUZQCAAAAAACA21GUAgAAAAAAgNtRlAIAAAAAAIDbUZQCAAAAAACA21GUAgAAAAAAgNtRlAIAAAAAAIDbUZQCAAAAAACA21GUAgAAAAAAgNtRlAIAAAAAAIDbUZQCAAAAAACA2/1/+LejZ63Ou8kAAAAASUVORK5CYII=", 73 | "text/plain": [ 74 | "
" 75 | ] 76 | }, 77 | "metadata": {}, 78 | "output_type": "display_data" 79 | } 80 | ], 81 | "source": [ 82 | "import matplotlib.pyplot as plt\n", 83 | "\n", 84 | "# 假设 data1 到 data4 是你的四个数据列表\n", 85 | "# 设置子图布局为 2x2\n", 86 | "fig, axs = plt.subplots(2, 2, figsize=(12, 8)) # 2 行 2 列\n", 87 | "# 设置 x 轴坐标范围从 0 到 1\n", 88 | "plt.xlim(0, 1)\n", 89 | "\n", 90 | "# 绘制第一个子图（左上角）\n", 91 | "axs[0, 0].hist(data1, bins=20, alpha=0.5, width=0.04)\n", 92 | "axs[0, 0].set_title('CN15K')\n", 93 | "axs[0, 0].set_xlabel('Confidence', fontsize=12)\n", 94 | "axs[0, 0].set_ylabel('Count', fontsize=12)\n", 95 | "\n", 96 | "# 绘制第二个子图（右上角）\n", 97 | "axs[0, 1].hist(data2, bins=20, alpha=0.5, width=0.04)\n", 98 | "axs[0, 1].set_title('NL27K')\n", 99 | "axs[0, 1].set_xlabel('Confidence', fontsize=12)\n", 100 | "axs[0, 1].set_ylabel('Count', fontsize=12)\n", 101 | "\n", 102 | "# 绘制第三个子图（左下角）\n", 103 | "axs[1, 0].hist(data3, bins=20, alpha=0.5, width=0.03)\n", 104 | "axs[1, 0].set_title('UWN18RR')\n", 105 | "axs[1, 0].set_xlabel('Confidence', fontsize=12)\n", 106 | "axs[1, 0].set_ylabel('Count', fontsize=12)\n", 107 | "\n", 108 | "# 绘制第四个子图（右下角）\n", 109 | "axs[1, 1].hist(data4, bins=20, alpha=0.5, width=0.04)\n", 110 | "axs[1, 1].set_title('UFB15K237')\n", 111 | "axs[1, 1].set_xlabel('Confidence', fontsize=12)\n", 112 | "axs[1, 1].set_ylabel('Count', fontsize=12)\n", 113 | "\n", 114 | "# 调整子图之间的间距\n", 115 | "plt.tight_layout()\n", 116 | "\n", 117 | "plt.savefig('histogram.eps', format='eps')\n", 118 | "# 显示图形\n", 119 | "plt.show()\n" 120 | ] 121 | }, 122 | { 123 | "cell_type": "code", 124 | "execution_count": null, 125 | "metadata": {}, 126 | "outputs": [], 127 | "source": [] 128 | } 129 | ], 130 | "metadata": { 131 | "kernelspec": { 132 | "display_name": "pytorch", 133 | "language": "python", 134 | "name": "python3" 135 | }, 136 | "language_info": { 137 | "codemirror_mode": { 138 | "name": "ipython", 139 | "version": 3 140 | }, 141 | "file_extension": ".py", 142 | "mimetype": "text/x-python", 143 | "name": "python", 144 | "nbconvert_exporter": "python", 145 | "pygments_lexer": "ipython3", 146 | "version": "3.8.18" 147 | } 148 | }, 149 | "nbformat": 4, 150 | "nbformat_minor": 2 151 | } 152 | --------------------------------------------------------------------------------