From 34ddedd68cc6322e0da78481d159d1c03f3fef81 Mon Sep 17 00:00:00 2001
From: ByUnal <mcihatunal@outlook.com>
Date: Thu, 11 Jan 2024 00:20:46 +0300
Subject: [PATCH] Notebooks and results have been added.

---
 README.md                          |   7 +-
 notebooks/hyperoptbert_training.py | 406 +++++++++++++++++++++++++++++
 results/acc-f1_scores.png          | Bin 0 -> 26358 bytes
 results/params.txt                 |  63 +++++
 4 files changed, 475 insertions(+), 1 deletion(-)
 create mode 100644 notebooks/hyperoptbert_training.py
 create mode 100644 results/acc-f1_scores.png
 create mode 100644 results/params.txt

diff --git a/README.md b/README.md
index 5fb66b6..c01401b 100755
--- a/README.md
+++ b/README.md
@@ -11,7 +11,7 @@ Overall, since data and targets are unique, the presented model in this study is
 
 
 ## Setup
-Install the requirements. I've added torch to requirements.txt but you can prefer to install by yourself according to different cuda version and resources.
+Install the requirements. I've added torch to requirements.txt, but you can prefer to install by yourself according to different cuda version and resources.
 ```commandline
 pip install -r requirements.txt
 ```
@@ -19,6 +19,11 @@ pip install -r requirements.txt
 ## Run the Code
 I've concluded hyperparameter tuning by using optuna, and therefore main.py fixed accordingly. Also, you can train standalone model by using *train_loop()*
 
+## Results
+The results that we obtained our experiments as below:
+![plot](./results/acc-f1_scores.png)
+
+You can also see the best parameters for the models after hyperparameter optimization in *results/params.txt*
 ## Acknowledgement
 Currently, I've prepared the paper of this project besides including data collection steps. However, we're doing an additional novel experiments on this topic.
 So, paper link/details will be shared as soon as the paper is published. 
diff --git a/notebooks/hyperoptbert_training.py b/notebooks/hyperoptbert_training.py
new file mode 100644
index 0000000..41a9e6e
--- /dev/null
+++ b/notebooks/hyperoptbert_training.py
@@ -0,0 +1,406 @@
+# -*- coding: utf-8 -*-
+"""hyperoptbert-newer-training.ipynb
+
+Automatically generated by Colaboratory.
+
+Original file is located at
+    https://colab.research.google.com/drive/1P4RT9W1rjk6qtuwCnV4aH9nWpieyHEeN
+"""
+
+# !pip install transformers optuna nltk scikit-learn pandas
+
+import re
+import string
+import numpy as np
+import pandas as pd
+from tqdm import tqdm
+
+import nltk
+nltk.download('stopwords')
+nltk.download('wordnet')
+nltk.download('omw-1.4')
+
+from nltk.corpus import stopwords
+from nltk.stem import WordNetLemmatizer
+
+stop_words = set(stopwords.words('english'))
+stop_words.add('subject')
+stop_words.add('http')
+
+def remove_stopwords(text):
+    return " ".join([word for word in str(text).split() if word not in stop_words])
+
+lemmatizer = WordNetLemmatizer()
+def lemmatize_words(text):
+    return " ".join([lemmatizer.lemmatize(word) for word in text.split()])
+
+
+import torch
+from torch.utils.data import TensorDataset
+from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
+from torch.optim import AdamW, SGD, RMSprop
+
+from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup
+import optuna
+
+
+from sklearn.metrics import f1_score, accuracy_score
+
+import warnings
+warnings.filterwarnings("ignore")
+
+import gc
+
+gc.collect()
+torch.cuda.empty_cache()
+
+import random, os
+
+rs = 2
+seed_val = rs
+random.seed(seed_val)
+np.random.seed(seed_val)
+os.environ['PYTHONHASHSEED'] = str(seed_val)
+torch.manual_seed(seed_val)
+torch.cuda.manual_seed_all(seed_val)
+torch.backends.cudnn.deterministic = True
+torch.backends.cudnn.benchmark = False
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+device
+
+df = pd.read_csv("cmp711_5class.csv")
+df = df.reset_index(drop=True)
+df.head()
+
+df.shape
+
+# Check how many labels are there in the dataset
+unique_labels = df.label.unique().tolist()
+
+# Map each label into its id representation and vice versa
+labels_to_ids = {k: v for v, k in enumerate(sorted(unique_labels))}
+ids_to_labels = {v: k for v, k in enumerate(sorted(unique_labels))}
+
+labels_to_ids
+
+def preprocessing(text):
+    text = re.sub('[%s]' % re.escape(string.punctuation), '' , text)
+    text = re.sub('[^A-Za-z0-9]+',' ', text)
+    text = text.replace("\t", " ")
+    text = text.replace("\n", " ")
+    text = re.sub(' +', ' ', text) # remove extra whitespaces
+    text = remove_stopwords(text)
+    text = text.lower()
+    text = lemmatize_words(text)
+
+    return text
+
+df["text"] = df["text"].apply(preprocessing)
+df["label"] = df["label"].apply(lambda x: labels_to_ids[x])
+df.head()
+
+def createBertDataset(df, tokenizer, max_length=256):
+    encoded_data_train = tokenizer.batch_encode_plus(
+        df.text.values.tolist(),
+        add_special_tokens=True,
+        return_attention_mask=True,
+        truncation=True,
+        pad_to_max_length=True,
+        max_length=max_length,
+        return_tensors='pt'
+    )
+
+    input_ids_train = encoded_data_train['input_ids']
+    attention_masks_train = encoded_data_train['attention_mask']
+    labels_train = torch.tensor(df.label.values)
+
+    return TensorDataset(input_ids_train, attention_masks_train, labels_train)
+
+df_train, df_val, df_test = np.split(df.sample(frac=1, random_state=rs), [int(.70 * len(df)), int(.80 * len(df))])
+df_train.shape, df_val.shape, df_test.shape
+
+# Function to calculate the accuracy of our predictions vs labels
+def flat_accuracy(preds, labels):
+    pred_flat = np.argmax(preds, axis=1).flatten()
+    labels_flat = labels.flatten()
+
+    # accuracy_score(pred_flat, labels_flat)
+    return np.sum(pred_flat == labels_flat) / len(labels_flat)
+
+def flat_accuracy(preds, labels):
+    pred_flat = np.argmax(preds, axis=1).flatten()
+    labels_flat = labels.flatten()
+
+    return f1_score(pred_flat, labels_flat, average='macro')
+
+def train_loop(model, optimizer, BATCH_SIZE, EPOCHS, PATIENCE):
+
+    train_dataloader = DataLoader(train_dataset, sampler=RandomSampler(train_dataset), batch_size = BATCH_SIZE)
+
+    # put device into GPU
+    model = model.to(device)
+
+    scheduler = get_linear_schedule_with_warmup(optimizer,
+                                                num_warmup_steps=0,
+                                                num_training_steps=len(train_dataloader)*EPOCHS)
+
+    last_loss = 100
+    patience = 0
+    for epoch in range(EPOCHS):
+
+        # print out active_run
+        # print("Epoch: %s\n" % (epoch + 1))
+
+        model.train()
+        loss_train_total = 0
+        acc_train_total = 0
+
+        loop = tqdm(enumerate(train_dataloader), leave=False, total=len(train_dataloader))
+        for step, batch in loop:
+
+            # clear previously calculated gradients (Zero the gradients to start fresh next time.)
+            optimizer.zero_grad()
+
+            batch = [r.to(device) for r in batch]
+            input_ids, attention_mask, labels = batch
+            outputs = model(input_ids=input_ids, attention_mask=attention_mask, labels=labels)
+
+            # Calculate loss & backpropagation
+            loss = outputs[0]
+            loss_train_total += loss.item()
+            loss.backward()
+
+            # Calculate accuracy
+            logits = outputs[1]
+            # print(logits)
+            classes = torch.argmax(logits, dim=1)
+            acc_train_total += torch.mean((classes == labels).float())
+
+            # Clip the norm of the gradients to 1.0.
+            # This is to help prevent the "exploding gradients" problem.
+            # torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
+
+            optimizer.step()
+            scheduler.step()
+
+            # Show progress while training
+            loop.set_description(f'Epoch = {epoch+1}/{EPOCHS}, training_loss: {(loss.item()/len(batch)):.3f}')
+
+        train_loss = loss_train_total / len(train_dataloader)
+
+        if train_loss >= last_loss:
+          patience += 1
+
+          if patience == PATIENCE:
+            print("Early Stopping!\n")
+            with open("es-log.txt", "a+") as es:
+              es.write(f"{MODEL_NAME} - stopped at {epoch} epoch\n")
+
+            return model
+
+        else:
+            patience = 0
+
+        last_loss = train_loss
+
+
+        ## Validation
+        model.eval()
+
+        val_dataloader = DataLoader(val_dataset, sampler = SequentialSampler(val_dataset), batch_size = 32)
+
+        loss_val_total = 0
+        total_eval_accuracy = 0
+        predictions, true_vals = [], []
+
+        # print("\nValidating...")
+        for batch in tqdm(val_dataloader,  total = len(val_dataloader), leave=False):
+
+            batch = tuple(b.to(device) for b in batch)
+
+            input_ids, attention_mask, labels = batch
+
+            with torch.no_grad():
+                outputs = model(input_ids, attention_mask, labels)
+
+            loss = outputs[0]
+            logits = outputs[1]
+            loss_val_total += loss.item()
+
+            logits = logits.detach().cpu().numpy()
+            label_ids = labels.cpu().numpy()
+
+            total_eval_accuracy += flat_accuracy(logits, label_ids)
+
+        if epoch == EPOCHS-1:
+            print(f'Training Loss: {train_loss: .3f}')
+            print(f'Train Acc.: {acc_train_total / len(train_dataloader)}')
+
+            # Report the final accuracy for this validation run.
+            val_acc_avg = total_eval_accuracy / len(val_dataloader)
+            loss_val_avg = loss_val_total / len(val_dataloader)
+
+            print('Val. Average loss: {:.3f}'.format(loss_val_avg))
+            print('Val. Average Acc.: {:.3f}\n'.format(val_acc_avg))
+
+    return model
+
+def evaluate(model, test_dataset):
+
+    model.eval()
+
+    test_dataloader = DataLoader(test_dataset,
+                                sampler = SequentialSampler(test_dataset),
+                                batch_size = 32)
+
+    loss_val_total = 0
+    predictions, true_vals = [], []
+
+    for batch in tqdm(test_dataloader):
+
+        batch = tuple(b.to(device) for b in batch)
+
+        input_ids, attention_mask, labels = batch
+
+        with torch.no_grad():
+            outputs = model(input_ids, attention_mask, labels)
+
+        loss = outputs[0]
+        logits = outputs[1]
+        loss_val_total += loss.item()
+
+        logits = logits.detach().cpu().numpy()
+        label_ids = labels.cpu().numpy()
+        predictions.append(logits)
+        true_vals.append(label_ids)
+
+    loss_val_avg = loss_val_total/len(test_dataloader)
+
+    predictions = np.concatenate(predictions, axis=0)
+    preds_flat = np.argmax(predictions, axis=1).flatten()
+
+    true_vals = np.concatenate(true_vals, axis=0)
+    labels_flat = true_vals.flatten()
+
+    f1 = f1_score(preds_flat, labels_flat, average='macro')
+    acc = accuracy_score(preds_flat, labels_flat)
+
+    return loss_val_avg, f1, acc
+
+class BertModel(torch.nn.Module):
+    def __init__(self, checkpoint, unique_labels):
+        super(BertModel, self).__init__()
+        self.bert = AutoModelForSequenceClassification.from_pretrained(checkpoint,
+                                                                  num_labels = len(unique_labels),
+                                                                  output_attentions=False,
+                                                                  output_hidden_states=False)
+
+    def forward(self, input_ids, attention_mask, labels=None):
+        output = self.bert(input_ids=input_ids, attention_mask=attention_mask, labels=labels, return_dict=False)
+        return output
+
+    def save(self, path="model"):
+        self.bert.save_pretrained(f"./{path}")
+
+import torch.optim as optim
+def objective(trial):
+    gc.collect()
+    torch.cuda.empty_cache()
+
+    raw_model = BertModel(MODEL_NAME, unique_labels)
+    print(f"{MODEL_NAME} - {trial.number} is training...")
+
+    EPOCHS = trial.suggest_int("EPOCHS", low=4, high=7)
+    BATCH_SIZE = trial.suggest_categorical('BATCH_SIZE', [16, 32, 48, 64])
+    LR = trial.suggest_categorical("LR", [1e-3, 3e-3, 3e-4, 1e-5, 3e-5, 5e-5])
+    WD = trial.suggest_categorical("WD", [1e-4, 1e-5, 2e-5])
+    OPT = trial.suggest_categorical("OPT", ["AdamW", "SGD", "RMSprop"])
+
+    optimizer = getattr(optim, OPT)(raw_model.parameters(), lr = LR, weight_decay = WD)
+
+    parameters = {
+      "model": raw_model,
+      "optimizer" : optimizer,
+      "BATCH_SIZE" : BATCH_SIZE,
+      "EPOCHS" : EPOCHS,
+      "PATIENCE": 2
+      }
+
+    trained_model = train_loop(**parameters)
+    loss, f1, acc = evaluate(trained_model, test_dataset)
+
+    if f1 > 0.63:
+        trained_model.save(f"./{MODEL_NAME}-{trial.number}")
+
+    return f1
+
+model_names = ["distilbert-base-uncased", "albert-base-v2", "huawei-noah/TinyBERT_General_6L_768D", "roberta-base",
+               "bert-base-uncased", "google/electra-base-discriminator", "YituTech/conv-bert-base"]
+
+for model in model_names[5:]:
+    MODEL_NAME = model
+
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
+
+    # Create datasets For BERT
+    train_dataset = createBertDataset(df_train, tokenizer)
+    val_dataset = createBertDataset(df_val, tokenizer)
+    test_dataset = createBertDataset(df_test, tokenizer)
+
+    # We want to maximize the f1
+    study = optuna.create_study(study_name='airbert-hyperopt',
+                              direction='maximize',
+                              sampler=optuna.samplers.TPESampler(seed=rs))
+
+    # Optimize the objective using 50 different trials
+    study.optimize(objective, n_trials=50)
+
+    with open("param_new.txt", "a+") as file:
+        trial = study.best_trial
+        file.write(f"Model Name: {MODEL_NAME}\n")
+        file.write(f"Best Score: {trial.value}\n")
+        file.write("Best Params: \n")
+        for key, value in trial.params.items():
+          file.write("  {}: {}\n".format(key, value))
+        file.write("*"*60)
+        file.write("\n")
+
+    del train_dataset, val_dataset, test_dataset, tokenizer, MODEL_NAME
+
+"""- I realized in DistilBert training, Model overtfits the training data up to %93 accuracy score however it generalizes badly.
+- Also, torch.clip_norm function demolishes the model success rate, it shows that additional alghoritmhs are unnecessar for bert base models.
+
+## Save model
+"""
+
+# model.bert.save_pretrained("./f35-bert-uncased")
+
+tokenizer = AutoTokenizer.from_pretrained("YituTech/conv-bert-base")
+test_dataset = createBertDataset(df_test, tokenizer)
+
+model_list = [dir for dir in os.listdir("/home/YituTech") if "bert" in dir]
+model_list
+
+for mdl in model_list:
+    print(f"*** {mdl} ***")
+    load_model = BertModel("/home/YituTech/"+mdl, unique_labels)
+    load_model = load_model.to(device)
+
+    test_loss, test_f1, test_acc = evaluate(load_model, test_dataset)
+
+    print(f'Test loss: {test_loss}')
+    print(f'F1 Score (Weighted): {test_f1}')
+    print(f'Acc Score: {test_acc}')
+    print("-------------")
+
+# load_model = BertModel("bert-base-uncased-41", unique_labels)
+# load_model = load_model.to(device)
+# evaluate(load_model, test_dataset)
+
+# test_loss, test_f1, test_acc = evaluate(load_model, test_dataset)
+
+# print(f'Test loss: {test_loss}')
+# print(f'F1 Score (Weighted): {test_f1}')
+# print(f'Acc Score: {test_acc}')
+
diff --git a/results/acc-f1_scores.png b/results/acc-f1_scores.png
new file mode 100644
index 0000000000000000000000000000000000000000..ca324d2aaf81fc188581fe480170170c8bc2aa75
GIT binary patch
literal 26358
zcmd442~>{l+Bbfuq$G`$2F*z{l1PIFm6YZ|(tr#_bDAd&3Q2S$L#d>B5Ji+mB4acm
z4U$F;l%nbXJM--Q?!DjrzTdaj|6SjDpXFH|s{6j~>pIWl_)W+8#2D?;VP49!ltQ5}
z@6gpUp->h%P$)Eo4D|S$u8<-d{I<nwyOo!z>p?Fc2M<SzfrFRZQCF{{NA?LGcl7W)
z;_9+NT2WeVz2IRlFE>vm8JS~${|0GSk3%vamol8hi!i$B?)IcmSn%pB4m6o+8Am9T
zgZp=Ashjy;`}Qfw=H=wP+Gt&tX~rPa+KgPU?I}eJX)o_ysh@ZpAy(SP%~P^0IZ<41
z(<RMYhU<7e#)LNTN*&`7;bhkI;Mcn#Z6<pC$8=4>WXtXplk%GG;D_a|$*=t?>_%Gr
z%3pna@AuQcynQ4?E6Yq$T?_xj>KP2tv$C<Vm7n5{z;A3Zq3YcDEuP`O|KrVlik;>+
zj3nHCd^PAB9Mlw!kN(6bEG(Q&lWr38&}XE&x_Xh5(}bGKW5HauwlY7Cix)5MQw)!e
zjvf$Pw~ncx{HJw!%X54UKTPZE#m2mBnU)JP-ndE2*qCj-lvHHL*zYI!y4fyu9&T<~
z1`%Ol|Ei6RCc@G$*5G~EX~Jdkn=sqLSMbLca`jE&r%o+RxOD0Cr)Ak<%*@Q0IXS1l
zw=b*4+a`sZWNGc%wS4K)rL`_go8sf+*}i!ArKvr?Kt6ASXXdiX7cXAiNK30Z=~1yQ
zK~GF?=g#=6SD7fXii)m{lZRNfwY1E1R;*Z|K{>c@pYU<d%;k5Bi)kpOrKPePHaItC
z?O3?5?6i6XdnF|$;d>EDIz1E4LkdBE-ku{v(C+BIy=wFh<hBdjlh@M{U9|8+PL$?5
z!7g_v9~5O8rj>Iw37jP#ro4K&fa)@DZ}0K`nr->F<rZ#1<aQmM{R53zQ{8cD<lZ-Y
z80)N}ZD`-K{@U52TDyL*-w^LU_CjlXpv!CbJ2M-bGg5m>1QirkY}v9UIyUz4wJn|z
z-D<Pr9A952$C+mAQa3bYaV?)%`MJBBhT`bxIQ99oOyQAt;$Oagt$zEK{m!A+Pwu;S
zv-0wuKJxBX)&aKGQqMK?%sh1ICX5sYM#idF*SE?!)^o16xzB1@U>m8CQrTV}aNE9W
z(KXc|o`HdZzUDk$i(D(3ty{OI<PWwKh-}`>&B(|ov2LB_$C6{W{U@C69&T7vQc|+n
ze_W!itu5mF^Gh;yMUKzietmb$%u={MmT~JHJ088`r~B@-B06UyVqzF^-@Ar}&NMbQ
zIz8WYZ8IBt3>|(d#Ob<_kkC0@QOT=mX=!?iENU~X$B!R>)7%`Dm9>_B8F#2>@2e)u
zdsoWx_S~y(Xkc5ud^z_P579@{!&`-PfB))DmOK1LaD4a^@0wlLwwai)Kk(`ozP8n=
zQKI(oV~3u4>gIr->mSYj+|$?J-!w2QV{<?B+sAwR>(27)8X7L4BqSu*KRL5X*T7&q
zeiqjzi&vB2TuO;sRCli<m{-xItulY_!_gMI5QncX5{)D{3rk9(E?wf<w{M?W`x<d^
z`ZS$57ZV!p%6Ken_T9n3s%@V>b#--xsRaD2Y;5FAOiYyBv?<?Yld9_KwQJWF78Yh2
zf5H{iJqntRXfeqecx{`p>w4$N2+O*4>)J{^m*8&nHRW(g+dW!uZPNRMlb=5}UPNaR
zB``41$jU1E%a;S=L+?45xTI@GyK2UIF4-;Z+3M9tp==JBQ!a7&bgH&i%h%73hH~`S
zG3VJimetCYlN})}rlzJ29p{8J&FTXuKl#WiC_GNwaH#Xkmk5Kao4S%i=BjS*uX~fb
zdjlo3yu92fcxGzs@@9A8i_&|SChA9Z*Tr==W|>Zn)d@I%dKA@sS3?8Owg)?*601#4
zPOk55I@8gkN8jXF%24~)bT%A*k8i*2-mQK3@ZrWxmn^fTr8d$>UtU$T+%i|SMBs)_
z+-lrLZ9D|B=ngnKcGXkYO*EB@uM-iW#TOq(vx!iYf9FOq^G5ac9j0va_g7L@QOQwJ
zR8ir_`t+Utb}u=HInTP-gt3w$xM2f_i1wM#ygaF8%a++6JH|}8fBf5((f0iOd=1J^
zSK1}ZBOW|>AbX65mzRl!g(W5FRs^XYF-|S{d)+)eYu==2IvhMmZ)s_%wS9Y5g0TcA
z0|Uc0H#gZ2_gw9|A8gpL!NI|SX3w5IQ`6H-YG0c09M$#pm*ErJA33tScdkq#H7#vy
zv`f><%4(Y=RYQ_0EHBUg@t&)WgG0(W)7wghTj?oIPEOT#52g>8Eb1Q^uz256V5deN
zx3RJDW{)onXZf})>YQIWtTtG#pr9Z}mTzrsXI1!;*1JbIo}A^MZuG3J7CaLjt>xYS
zR!mD<JJoxzSyxxL^5%{Utel(?6N4>Gjwq~xR6RCMPP1<RhsVF2YHhU|`t*<sMe}~(
zFZp)=L8-R#0L%2YD+<mxGBTE^sHhCRzq2GJCgx3JWA42<h3zZ?s<l`xabi0!3aY5=
zZJql)D1*-=C@sBQT3T9HPcQpt;Gsi@us=lCtg#Ge!}|6aY2UtMM@(UX$IF&?UwV3k
z@C31*w%~C!lpNdhrMEXH!QI_m9Zzt$t@KB&|C2a@xS;8u4ll1L_)d)q(X*_cc`><?
zmNbI+OwOf?QBGJ_uB4~9d3!(Zf0uFl_%~}hdU|0gDdu$+xs_wxb<Wi8*DWoUQ}ei_
zz7NIgtc#r@Q5*KP-FL5+3keA+n;w4n^wp~;*s{!OYHBkm5pUkU&92zBNe&%O8$0w#
zEbkIcO-<^4&8TJSUS1m?{T^87>+fIv>ecd^lPq=kckH=LMdzK@s^P`O3X5oI<&F<o
z<=T|8^DB8(Hsx5727_1j`u1T3ddB4qhnAl>al-A(Q?}YCPgL*hee&dqtejlc1Mfk5
z7nifnE%^@3xw|DDwa;*e6&A`a<J-)N`hfma7sIi}-oYWv<<kQi%7>C;rxn|MX_JF~
zt|HI-MZ<>M3r*5E@#@vIy$_E|PLppz<)Ec3=i40G)@Eb1dpBd<I!VdVXKRw}e|_&!
zu&r=Ba3EsSu@9ZqkxY`)=%fyZ4@WoW?dgtHY!AJCTMqpqve?ACDM8>l8$Um@j7)(f
z4fm)7!=5r<IV_5(V;@U|Hg4oRy;AB#SQw4}c%M+h=`&|)(SAl7bOeNCWLPXLEULSu
z$oDeyZW2UW!1^5>AJ@{=Wl){!49_UpX<)#7Rnb-3$cXjYwQHT7ouSlqmZ^tcCuyuQ
zFff?;Jy<aH<$~dxhd!>#fm5vCJ`~psH078N@F=?QE$3HiJdhzwvn_RI@}^YN!_-t^
zZXTX$j}I2AYtBU5?l5H1Z4a3bBITkojJEm_R!8^Kc#+LseXQTVe?NEqdY;GGd-oI<
zEm|ad@CEng&6^uDnIAuXTv$}3p{x6pBSYbNeZBqBqh}PHo9)N@8eZrAn46s@I|Ccp
zb|C%^m+|4l5+%nzMqOBAsCsAULjL}q5g!K!ArzH?@>o7bU>ReBWL5r84}Em3thi@p
zX57a6jC1$gTRzyFM@zYU`En(SZx0{t`_mo|D&@tycO2pL%+(2NO(~S*;NNS1PYj;4
z$g}SH`ts^!Mc37hO-+$(`8zL3J3PNAMQXmjfr0z%Z}0WicNso>_|S*<5t&FD*yHXl
zmv2*g!aT>qq3XoqeTuzV_hpk&37Seh57pMyiC}l`+O>;;nfdYep66X7BX&9d(za#M
zs6R{v1$Gr@*+jG}tEy;%gM+aw)%EmFjq1`#NlEc-cB7OBOh%QLtKebM2L%Q3DS0s*
zJa};8gN~E3=`DSAG^RJe65=W=7~PbLjNO#ZN85TE_yslo==A#XTs^m6cLDJLd^%V2
zG#trump#`s^V3N<mCIH)rBzDMsc17h@KUN%^p{U4LT>KvVOUnUop18?D5CV}Y!2>y
z#jk%|X$6`(>G)LY3hX2!tEV^Od02FS|8zGpQ_H4{TWo7mx1WCVz}s1Zs+Vh5v6_~C
zStWMQn@2$b1f?pvw%202#~%Oop$p52v@VzSvY7eVkr3tv^Hj-!jqDdMTnGm^uWoK$
zDWA9uu(q$Ipo8Nb@5TsJNnKHPhj;JZY2pJH?yp-;^+5ABTWeWraBZ`(#E+?9T3TAK
z;kMmROXhC=Sa+{ZKu||ZtN5MKvxn1oqJU0mD8`51-C|xTZNq@7dh_N@3I(gbFlhSt
z_+SeQ@QiQob)Q1FE{&P-cZMi?4&md}z78i9_TxSEEABshSo8ELgI&<Xifh-e8{68R
zjfkMrPf}jh`|7%L^l^*lm6eoLt5#7$QNDzMIPf5+rbjBwllV6{RMWjkR<rA|;pcC9
z$hm*tzRvFMaO|7h-FN6m$|qN!6;M0Nzx6o74$V77MQ_q~at(dFXQ(SUUDV4ZBJ8A#
zCUF8gCTcnVDd38`N8hsx1dX$a>t7a3$a4NZtOwlOT5^n6Pfsu6?xv=!GTL2|J)&$1
zXRYUVPWPq73C8P^9Y7uGMOorXs&oJLL4iEY9XoaynVN>j#j!m4^_7=EMG7lB`-$sY
zeLi0>RINhWi%(8wv9q&7yH&Td6TtS-L;=aSt5~j3JkOiUD{D{D*ViBY<kLapyYYnB
zrROvhk5WFSmFp|XQ^xkprgK2$)hJyfCog}(M*5_Uv<~(zAZu7h$o$@(toq53jwPdg
zRbhrX1-ZG)0FkQ)Q&UqjLNc|q3Nq`zeA;mEMJ39b+fb_<-QvaSrluTNnv{WyWxSMK
z*OekslDkkG-e64tcTvZ6sV*IlV&A-ZbLdu67WG{A+Atn{Dc4YL75}?>C&u*v@(S-^
zU4+bp6+HUA^vIDTXh7P5fy(sD`K%0M)Afq4oGtV0UH)_U)9Ggy*5uIHQ_q|^10E)8
zUr8Y-;Muchy88OH={qmU9&MGS7IDeg)&iv84w!TqYA@%-m#yv|W~12my-9ic{CUkt
z#k_4#dGh$^BM_%90W$z8=XX*o(7v_M&O#r=c|@K($yEAu89D$#C_h=*%3q+jDjgq+
zMhm+C;KAo`7S+lKMvg+XR|@NSA4_TVEaoLkmfXx(sHvAX9$`n{a|h&=>^XJn)SC=L
zfxFHv%K=U&8uGYK;L#S{xnsVGmzS5aZO0BqC9ggsF@gQ2DQ1#(r;c=6FLvOPwh1-K
zibzZA=6EhDCe{T&NT7_8_do>j^k=MJ9S;wAEKczRQwdIxoX86oI0!_cqoZSRu0InS
z+ch|N8t9a<XMVC`KE?0H*U`~Y)}Jd@uG9eg@%Q%!XsHGIDa5mqbY?`06-mfx+vL)g
zX>zbOnhlRGTwlCqy#Jj80HK;%#fL!U6YuXFN-HchYAq@%8ZcvFYj{4q<JrT4y$|WP
zDfrzplZ-Tze2n%!hJIT4;luANZ)ae(_V(4+mAo~4eK+GnhOSs=k(p2$Fe!~?aigFh
zZfa_ZCr>e8_tOLN4)mEGrk0jVC{w?G??a*H-gww3c5r&IprW&@iw5PDo16Ov$eW~3
z=^QXg-}^gjyu7^7ke0c+x&kNB`fd#8<m99S+&p~bNY_in4zrk5KSw%PVzf@69+4kC
z_w2xC-%&w4$?BRKhyD9$j~zSa)%T_=^1<y?i+Aw~e)s3_5%I{g<bS`*-@BE2)v7Mk
z*}k^Y4HSGLp5MQJYpnA0^aM{_S};31D~I|)N(Cw+m<}Nn=C{^NOiZ}XOu19G9XYc0
z=eLh1QD>`bYw2cxPnc#6s?Gn3)7!c(Lwg+<>M~nf+xuSq?6(Y)&yp^``_4gyUsl<;
zO8oLxx)j}K{&{#<I`><i*W8TXU8km{6joMNij0g5wL5OjNIr9<hOuK-G%C;m`BkXh
zkJq%TtpS3=`Xm4vKpz~9>%@ogaWBBLeM7Cct#0pIeA8rY@cgWgnVA^`m{-?g2%W>l
z9d62Au~NpC+)SJMZt6Yve6b{^fHxeRotgIU-yd=E<o=o{W~+zCrL6C{#H!6rEf7<X
z%=X~vW5%i{KV+J-HlxMpik|o19XLNyu}PGjk#lL}=mWcoU?WND6dqawS}y6EXxOG(
zbO3dJ{`^_(%*Dk;fDPmdBgSg%IrDmR9qCIoSULD|bWDu;NP9W9(yFg4D7xFw(E@%B
zy&s*Jh(x~|ZZ9XR++luh7I2<2DnS_p5`c<v+o6V@bHxhAK_B3zNVb$q?}tA<Ov%Vt
zQ}%OW;xS;#fK5lSa|<gU-})C9i(NihZ!Ym2+jqaTlpyQsNA*|NqinOTS`~x-h84Mp
zg8Fyv^5tmoDY_*~n0nDS1V$cBONZ*D9Rj^+z2}<jRuc&hwS2X5v}?ue&mN{~;5>TV
z&s^Jb9(8qf^XAAEoSZiR5}(%BYgkxZl;t^&PlYwB?5=*m`|jPlvc9ACN89eRk<A|-
zt{L0xpOur-jRslVZ?R(U122Lo&;_d{HCrIdgh*cSd7Y%f`=j^u$=Fzfa?z1jFJC^+
zG)@DQV+KCIfB(LqkdXQmjau-`W<roC=5x9gfxo_Q$WXGQU%h&D^qDi9wwCvrA8tDM
z;&K?MLH9|4;82irG&ITeEP959b)a@RDkGmCpO%o3Ir;G6!$Lz%O<G(@750gF{e#j{
z#`$@GWeF*%75PWfm$0m^Y-*Y){r%R;DahvwphVa3@LB1-5A5e=egVEzff~f)9zK1l
zi{?T9>eZ{!$;mpPE<*kS#~XAAF*R2$yMn*g>39@01~Rq(%hP!Cv+JmOU%q^yQ1A>-
z0wjuxiguosd1SQK(%M=wAxHGWv+N*l`*f%^4sUL#U*6&=Ze4P8cg}GRHshBQ-#&7I
zFy*=bT4$NB>}l7-Y+?8ou#89QwqGK+q^jq|%uN~^8l!X5Bi+5pAy1N20%Tp=m5;WU
ztpfZ!QBrceysWM}Y#(qWr4!$f|9$Ih>UzfPi5hFv53<*NeSJI8nZrm44w~GHdM$+g
z2Z08=L<FU{Ax(z`qCyqG)1H!}S>E04e!ct~otu{dmiW%kPUi5*NJ^e4E#30s<x9f)
z*6%K&!&5CNngc!2ldvp}<(B)LHS(>yn-(1dvQuI6H*rV^D~gZ2qkorSSXWRncM4=z
z1n8zRtw+Pq|Hi}X*ZDUpD2(BHmrmC|f9{UmpYiGjR{U~QD-d^%iyY=IGWDj6qV<-u
zhGE)9W@dC)G+pYUT3X{C_e17YA)zee<-MBCG<|i-vl<FwA*3XyIPBJLJaP&Je1HAk
z2b@?)hnsU(<>uyQCsg>1tUeJQKGq&EQuX8sJxIEay}e-6vQ;OdqZzKLg{;OS=LS^~
z-*t_F&>7rvZY3{?D~hgX$+Lqv7Jc?Cfw8DWXF%Gzu<{`PJOP#<n$XNlAWAah$TUjH
zesCYFq63WWfj<m5SFZg0T5Uc87-aNX=@s9%-bed7Tvb<IMoBTddFxiTGP@I0DoSW&
z<u-6ki?s#$`De?@l+W-eEWPLQN#n=YmvY#8fxmh#QC7#En?*4qtQ1n%BGMtgeK5aF
zwZ3~e{@kiAEI>#eni?9_rk9v$wgEMPb+D~oeHJ{V3Mdq4k?_5N{tJh#_!yn(2<`FT
zqT(%j#+G|uY(&%AFhz)*Y$^i(qKN+EAIGox#(g|(kULmM6pPHWl0O75UA%azjop7W
zwPmEkFOQcenPWkiMTwV{m)CG{k)~|Z(OJSCqxHqY*kCPHFX7@vVMRqQZEfw?r2E(Z
zXuap2_fQ1s_KklA=wr{{zkeSQW5@YWrLpWF;6#1@?s5C7X7SCGl!iVLd+LG|-b4#2
zwjuaM|Kr-oU;m`x{^J!x()}O!<GE9xLM>z7mC<EB^Qv{BXN&L4Uq2vnoFiJRj;Mo9
z=Uv+Fhe6CE?Gyyh?f7wRC#P(;K#<V|0W>boFm})mKo1}3mci9qMQUr-ETN&H!3TfS
z)U=(*Whex{2Uf<1ia?KU`~uWa^XU^mR&eW`Lo9fZ@rj9y=yg;NRaI3q9yZ(O2a0%_
zLV$7uQJGM|vOJD9W*SRKNDx)?O_`tP(GPbSK`0=AM&{%U{_sB2Cr6a;%x5vt(V^+-
z>3X4#`}dosU(**4w~;O|{{%u!ikq1k=gRd~Mm<w0nkPw<L5-wPROe?r*8>&+|1PV}
zv~>qxjJ|RW+o0;@OJ>cJ%OevLSJ~{{%eG>Ld3qbR1QFeUIrXn>Wc2g%dz~Bb{mZky
zcNr{YWo6ka=o&sp-xn?R8Ey-MKn>39K0g-(?emA<!|gXu{J8-(54;$`EHBILrw3FG
zhpa0kA;ECv$`yuX%jiZAOL~lURu$g9UG?}hm)xOOeCUIp0dkBiEelH8AxCykjdgE6
zKGe$55daRtz3C`7Xa^*;l(e+)sVU#qj~^9v?fRP@lVyB`FH(a>QBg4>GLnXZPvQQ1
z{7CNJhZYRgQ7i(|2|mY<N1Z;+G&D3+7&0FWxMGpD0geeMfgC#rM@oA7MD1kKRv&3n
zhFm*TH@dd-ZS@(BNKO`jnnO8ssMmjZZF&7>{vBFc24_)8KNLG>6cwP?LdyKW@%nag
z@pkNT_pa(lr?`ytbRldGY}JSb(7GK++9KbEG0O~4eW^{EFI$d9UbRn0&=NyKL*PRi
z{ABpmtGpHkwtEYLL_h07{8ZR7JUsmL=~JWkK7X*;KuoVqMKC`e1f;2gpZ9?@z=YC^
zTwh$4bMx}5J(L_okFA7_xrox(XjUfy`69(gV&LtH)HQ#-A}YC#p59WJNa*p4$;!fA
zt@-lo0*&&68aaqX5KixV^ejDp{yZT*xV<$WKkmt_gHD1AUPOVZRP*fF4##&8Mqn6#
zK~pGa&z>dJ3XMG)O8aNjQ&Kc>OXb`;E4AnO`g)!$Je_NxOCr$Np*6aBcxa%mU6ip~
z4lqfH)4sJUNjV1VEe$}6RBu8;APQ$HuN%Va-3LpL?XUzcsF?ZAgiX`{X%^U{6C9-u
zl1&3>xBtkR@(-eV=_YHBi`X9DzJ0sw#*Gm@ck}f3Xlc>il1k;${0FHh_;0ER(+}Q1
zs>qIoYA~fiiQqi-X{oTWLaXj&En#BA3<d^+`PzoedUs9qHKdhFoR98X@y=*R8L!fo
z&Bg|eN(=!r0=J@~=uy7I@I<*ff+nxme|z};Jpqt;#@Cw`A0^TfZmCn_@))f}92^|r
zB%QeF<^%aIAItnUgzBu7S>564<a7#Kis_!5TK`PFl&I)U?I2O%)DKzupYu{uh+EP5
zB5?!m?T+FdWq0n}iIkF--j(1V<CL@OT|9O7AR4*bV6zm_lgXN%11EDp7e1$#Uc4*w
zw=3ae@80DYZ>0#LW)T7S_s?&8dK`k`4w%oKK*cukyB;4;10{tBHdR$sWTQ*NjTss`
zn)2eJ)Sl7DTy~kt55Mj{hRb?zBvZ7VIg#&L=B9y-{JVd#+;Mz<v9I;E2-)R$e>Na!
z^`l4EQ<@s`0NU*96V_tYp1gQ*B~gP%N8Ko)P;M#q08^-})q^31I;wg6&=3GLln%bU
z5}BL3e&)w(HOU_E@CXm;#|l%6%ZZ8V(wEXuLoRPR#siAzHr%#_>IpnA6t8RclNfnU
z+|1P4he8*^IhwS~^B4ml>HuIa0W?XgzYPYo4|k;Xf%iIjdHF)PA409!`l&n|9O0-6
z8RF(aS3ZB<ceJ&T2CM)|T$ZQWQYNOzEIk8*mr6~>78X$uZm6-*(M}oiF+SKi@#oHk
z?zkXMDjGaEu6puZ32%l`1*QMowQI{Dp@0T$^e@*cHZSnefh1zF`@&)hu0j((Klb&i
ztA=oV=FRKydlrv|q?s2JLBT;Z`{qr$uV25mqQX+W7OWH4BlV9q9C}q*U$5so`@y+@
zae}f9l%%V#&noAq*-{pkx=VJ!yE8mvqWj{bbU%MOH^JUyQQpoELs}@_EQqcKi#I+g
zX*tR{^pq~hG<laMe~gY2=jQFZckYmhd;<b%T3WbZIGV4$cl5n{icQZ!_#-A?A?E7t
z+^LDj`Qh%7a3H+q(!Id9+W2B2l$)kfY2j2$b4}4Pm#1Vs+UiulN>bVL?K5QZt)BtD
zN$&$uT>?oSAhj!Ge!9c6JUKsKdg$A?$QB|35*a)*Q)0bU;kH(pOV7d2>VWP5F5KMQ
zwgG$)p9xSmZP&N+&kjf)*?|7{lSeqlN3(gS*`A3_>_0!cRuBb)TmFD<aUD!=qS8RT
zBEoI!qoA#DHfTGI^YedSK6FV}6xvq}gw>>+BxpU|P#~h?;zEWuv9TF{+Ni}o@;moZ
zpI_4vQ2@?m@OLHSG8tG{2wQ;blPTGA37#7$IykTf_$3e}`YzXM4yYKhH*bnUE7TOf
zVD4w}{LTj&W&h#l+2%P$@QZ_Ie`h7ElaOGoH-FpGaxy>PrFdeXug|_Zf{{qgM2Gsu
zxejGNw_dazH}}cRuOAQ)OpmdTz&c91af7mQ<w~MKYiMdFU#WfgXh6*0)_Z9+6e}iN
zjc&RLH!p7*dbm}ohZvNhob{)ctX`~qnoCCe;6ZV;0>X@EzCV}Icn|Z4SgO&dPHEuw
zxWj_AE>8Qd!pF~l``E`#XyB&lZD`op*LZ<CAsMLL9$ssD%fA0@I?)=Tsyj8VMk2uu
zP-i>tN{-5{yu8?mh=@bY#-CjEf*N;@m;JpGGhdIqPQnRxn)AVmR>Qj37sCGX{{v9&
zZ$|(BvzN+hW%i|Rez`OBioVX!K~}by&-JOQ`xL`6cXW|<{a3QX_=0=)6Xr8~TiC_L
z#ff5m=gtNbQ_~!&t?r+-pklALdo+{fZDD3+k0lA<6IN2P3B`w%hi55zR{}izu(UK0
zFK=%L$WG8<__lc5RDGyzwm3s1COUJ+)7P&-lD8@=D;FMmz54Q%E5fi=VP~1wKY`C6
zCbk%h<aMs1u<#;)E~|=Q6&R#(00U=MZJ>{eiegy0bP?q)C>^@_2`CAK?4oX`V9Vck
z>so}?I#^zyMIjLoC-S7=p2l3h%<Yg2_D}U7Ry+0>sd}nGN#(DGP+GZV?L&>4p<JP@
zBWsXp^mTT!!*a_ZEEW*0C)i(I{r$UJq}fzf{P_f}#m-!C5M`l9qMlbZHLZX_CFuyo
z47L!ok~1h2Et&1OGLf}w8GBzRFQLp#wAc|<Eifo3M`|9i4C3hE(?Dyj`Qd%4Z|#DK
zVNeZmmLT0#8xGpPy0!(dksDscjjXH~kTQUb=QmBVWMi#3?i_rnfh-J`dkRPz@EA#1
z%+1Y7_8cE+kL{{CnRfRsC#o;>UM=9V!Qwcrxc!957@W{Z2?a_6e4<cbU2hKhy|Foe
z?`gdJ&GdAK$EQ|oh7Kaiz9V+<6M77k&S>D76l@cyWOcah@DhdSd?-*G!=RY8JqqF_
z=G>1THJ}{!2M=0|1c@#B127~01z_?}I#Gn5zI;g(Dd=nXm2X>q%24$eU2}7@VXh*u
z8$MSxurrBpgoaY|;e&zEO7@Vo<mwtntjmUm2EYM4m;A|KcJn`&knvwo=07Na?d|Qb
z*(*^BlB`^zH4;$C_XkuaeOMi-5r!80!7D7&g77%w3$j7Ky}U-+`Jq*j&jd;0w#z3a
z@~kNK`}PssL-pLaaiid+=Utthf<r&X$7^q>hsz@7LfKXkJWI?M7^dWQZFc`mhh)Pk
zL>?AVAd-<UjgNE);GqBuZURb<g&In<_SnV%{}rV1r;4XQJ3t+;gPbZCG8fp|+Da@}
z^ObMZo`YPtPyamDHKEuM7+&hx`#CL6l~CIC>(`HVofJSe=MiP}u-OoVezIraBrxM@
zVJNLyvYan-jQ_Hhw~(>w+*5le*8f13FTex})_*;Flgr?rp`}9VJL_(H!JJ?GDq31H
z)N}RoQk-{IXc^FtoE62M`jfD7GD)J=2K)!Fd~G^7gXqy4AX}1A5RSKpw`5)~8>6UD
z;?Gi@J?)`9xA=BRN$|cW(@>er0`j$liBo7a3{}Po?l81(3O_$Tp)Gi#MAP*1`^7Ps
z*#_Y$yUY@cV-!uY8>^_P<Xa7k8Uhze&?qfB1Xi)J&4y5Slqqc7NQw&4&30Z~pLM?W
zV_n@6z}V`QTwIfl{yX*b2$X`<_&df?@`7)&lJ`;|y9mHI|A~RfkhxzOw!gK<U73r-
zU)0s<h|#R0JO#H+u5j&mB<XY0to`B)wk|M6Hy|E?|A^r@*xQFDBya}KPP8CRu>JV)
z61VohBh3J9kfZyax4xGXYbTP4D;2LqLNq|yLJ2ZbQjbA7<vhRYlfX=wV`6k>SmVZ+
zdw+Pad-<bBj|^X|*d4Q@EbF#{E~~{ynYxO^%a@hMW_q4q!mXa$=Hw&^nix5Ym=w{7
zRxA5b_U_&5P?3@%giTESEg4=|KY9Ln4^e71xwfA{OzK9+{G7&nkiHa<=F1AsE3t&m
zQK^e}?tHE^85bR$eb2c?8lQ1L)~4>RUE1icV=(3&(B&n2h*bdZ0DZdqx=%X{2&Qah
z%^^SXjH9?*3|f(T=<4Z7XAd=h`{s=iW2F-jW~ZQF(NI8`CF_u3k@NVXODuQ5wA;6D
z8|LKzqF_l%@<Q<0<H$aOD1nHO5X}W~0}c55%>{Nc31eepRU|S3Yyx#faLblexHFlV
znZCo59pwRv<KyEu($gc+1>|rYne?j|b7pjA?r-t@N{Xn0QboYzu(^poCl?njMSjDE
z20#=4BfZ@t=GNBeSTSIrS<AK_I(SeR&1dWJAqs``K4RWM7o4l5D?$yL0_qsdi-oaw
z>X6tNeJDySQ_uD5*SkJ{UW5SKP)EglX$Es*?NuJr@G}wSb0WckS2w%wi+$Zu<hYai
zV_nna9Y^5N6C7)H*-!uc2DYgKD2h@N(G9Z2UshfoSrOqsIUHMEtucc|ZkX!{`CelE
zdd?C+IR9fC;3Ey}-cgpb-`V*z)(jCBU(e3ze>RBCQrE7QoM&sHrl<FCymo~{49aKO
z`yeNv0JoC3#tRV`W}XO8XW+V&)YL_Q`Z@?a-h4YhohgZM2TV|vF&XJZeTy7kwrId_
ze8~J)0B!M?E-mr&{NeA)a>3VB)4!`ZqI`aC8aloJ(sjl?R(^h4P}~M=CId#wmp~do
z?14fF{?)^U5D*u_az?ADr%#`*0f;ujKfRN}!l~KB$-$0+#sQ&$i5Pac&XZ`fw7z#Y
ziv6D$Ji9Hl)tnc(Hje^V_4oHzGwgviNEjPl(x%*h1#}_qt;g5na>Jd@3H}ZLe(Zi`
z=Gu1O*ztj!nws6_q)V49duFC5M`$U~!03H^e2DimG0-#xC(RAvr=y=9tZ*$GqsMX?
zY;%D*cH@R1iA+GOLhsQ|Rt+MBWb>9Sj8~zqfZu_NJ$VqK3&&wQ*6{uNoAEWoT+GB`
z?x9Nk7uvv!Z8KYAe0umNsEXb%FU!%QA)BA~WmzRpckIJmbu^T7moC+bwm04wPzjp^
z`7s~fm-Pi{7qp2g;1xigJdYWu7L`|BAMO73!=8c=)h??uyFv4?864i-GI;vpMa0>&
z%Soz-kwd()qr=W|09x@%d?`oX$s54{G?~uY7!F_5ugF<^Jl}7{Eu-%bC3k_j*5T%4
z&#r?Duc`sf1xEkei`QDZ6s4WXrNDBRx1P6UWxihdusvKhJ3B>27!b0lB5H0n80&-x
zxaG_K`=s#y&aW=ymp^Zmp>dMuzkO9*OZ3#b)4!uX(WQz<E(n4L5CZ!@=1q}mxC|u@
zsh=BpdCVvg5m8YZ$S69(Nsa`rs6hxCRC;u5OdV;MpyRQbONpXmeBm5Y3uHa|`r5;{
z0;%2(?gM)yCm}X2jwoL!*GaowAqfeKh};gI^}p-du{tX&i+jVtWsngN%tHb<CX99o
zNER)zRuF|L^y)VutQF!5G~Q#;!|V%q<CYLWHc~1?#Qpicwbg8`4}d=jR{+EJz3Y8%
zdhO(L{_0Q~+K3`Rv}kN(TsaK|Nd(;;JE{<kYCtxIvhBbD(JvlbK7al!i%jS5zSYJn
zt^p>biWBYTzl%N=vdXILoE-iXpO6p%Cx@0oqJVtLzO1P3l+dCgc}gg*&h6}#D?IRS
zVHsKp>=nsgxd83iAqGdlBD?GIi3e9pvIIcNg}5In`nWd$m-wTT%?O-l*}y5c%?NVf
zeqe(U-DPHWmS#cpD=Ml6)<RGu4lqCvcY~NQNYa!~UdUws1NM=|C0+)ySQOZHD>4j|
zpFlQc<K~XOdv{~5)$K)faRMi`gtOes_uO-NjQ)2U8ENuNP@Kqn;c>BYb0av#3<Za_
z`$?=OaLkST{N-RqkczX{x39@|B?d^UcvV{)9}FAQ_bsruY#ttGra*qAA$jZi`lTjW
zocM(;|9DMu;38+|!8Ogiuw~CgZP~h2L{N}IbWva5`$o55;RCDdsH@W;J#pJ-#1?@`
zm}({JzxC+cM2X1Efc16Yz=6!6tpte_J1wRFBg4|XRPPQW7TE}U$gzgG_rM`w9qeBY
z0|j@%yq<@LXOZt+D>Y=sQ&LjA2bhcvNO}+t`uBK$Uwiphu!_#!-jmQ!_klfF-E-lm
zAgx9R8Y$U}l<qe8)m2sMt-byIG3w#;4iMBRp%4`u&<99>4vo?Mb1ftB%OH@gc=aAs
z9emEg(J>MqstbVU39w&7j%9&%-<I<1Kal;+pMctaw9CtCA*$GLg@M039B6_J_Clk+
z=i0H-Hsr@kMY0)HfA@)yyNRtQ5IiOHXzDXlivIEA3&t)U9=ZK#+OgGu(1UrdT21j1
zr2~ak*vM6H-mqSLS0aB56|LxUNwn=oYDGt6WTdRDZ06+eQj(L4)}<pM`;6izRog6n
ze9VrCwv~?Drx^MdvsPHe9A7yNLP%Bwx#Pq!1J8il!f_TizMI5AL9+?bL0Ho`O`GKR
zCf+NgO8)tjMv0YsuzsOPRv}Yec;MOj-d&w4!85+1;^Lo42SDtMB%{eL#~P%pR;m6B
z-+_GUzv5Yf$jr`8PZmt7va95)hZkZeV=@CMeH&hP)22<t4<7EQuuJ<<U{fl5O(tF!
zw39I)Ab<#3J!_vJl_S|Bsi>o)lcJBp6gg{cZB6262#Fw&lk7X!4FIQ*TgTz+)Bkoc
zUXTbqkOj=@&BaIC@4Jft>?B$DLbD=?we_}T$Tf{42(AKfZx3>Zy-bNqV!n9!G6OU=
z1S>m{C#{6?bnfCsab<~%nL-<<uE+`q2q3e3qpXY%eIEWx1SkX`VeBA*Bt$4c)gToF
zY4VfcpSTrk5{`5{QUN2+@JKSt{!orKNQ`!j8Pp(v5XGW8MS(^)g~*&7<c$;?aF#K|
z$L#@ky@wRg&_ItLc7FR73lmBh$P<y2S{;F)o8LP=e-IxHM!FaP%Iv%+ZD>i!PtyyY
zT*wXy$5%dmn)zy0$p0ceunwXs&wC2*-mS&z1v-v|H&2O^Z&|W<an*|#%Wxlku|Ung
z_OAoB#~w*T2ov@f5lQgu@=dgn_Ljbsvk+}xi}?kSMqpEJ5?qjzV}m9x*^2;#CS+RQ
za+Ds2JUD7%WT^x1s@;AKaZ^H4(&Rl*T|bX-*u6)OR1q4E0y-xeISJntI?!U8WTy5@
z8v<>0uwbGP=G^#nKV!6{!oZb66qas)9VDUXC@-#U;Xve*I2Odx|6{K3!aeib$B!OF
z6js#y!M5dSK`=i$v6VFTBxyrq1mUYiY%Ug<vKuoGHE2c2R;tR%d=NpZ50d;?2|HW4
z<Ui%dmTq82uz4{>D9i`jI`3l4%xDb@LNbTn7m1-Ez>SK*m748vq*J!OInvv^1oD%<
zCA&fbJxeuo5N4z#5pt?}|9%xFaTt~_*9KBV(;REEw5tO9K&jP)#fXs+<dpMFbkK<q
zeOUs_ozQ0ZmD~zW33`brJs5>CbPNbQ#YREghZZl3Bo=rt?b66p?GOV9B+PI}F>LT;
z>quGnVy2b$5XKGHs|EZ#3e$b*9^{#&ROO&An`(-UmfX3c2`lZ?>C^F4Y9vZ!Dx63?
zkKuaC5TYOuY`#cvE8M5I<M;}R1)G8lkGwB(gy;cbn1KLEIM7gNnH)iON!s$7N`OWB
zSu9Hunn41k4t-h9d%z4FZGS`Rc8yh#%ZEM`@9ZfH2?;~z8%v&_S@P@GFCy7K@EJLa
zs-$*@1+`Zb@iS0|lW+(~Y#NI}2a1W8SGFpRo%jWH@e56TKZJx+b$BGhe8j*q`vwMt
z)zk#geL>!DyXDGLp1*ieOJaj!I%39}herkVEiJj=Lr_AI0YdQFYa4=tN<PCj-AZJN
zONpRTh`akBBVu5PZRH+|_AS3QVnJvmk~thin*hPJsWKDYB$|v_mVIw;?y$P=CPbp)
z;Y(OZ8sB%cQ@xm2_f59Axw~|ANgyyhoR+vgot^gR`%A<2fTbX8p~9yY%tzAs+qW<F
z@bI8ekYUPcC|Z&F;BcdY14btZNhbDk>-+a4tV!emH#{6ZROuXfzFAi?6oUN@l5Nym
zf>ED^j+Bucd@3XTwf@eX;c$@8i|rhJ!s~hiYa%dIfX$dTQ@@j+i)+c6HERF>n|Ch4
z^qU5W-NJIj^5P`H4CHhe;_hLViXJX(B&b<8#6ncB(n8#_OxtPQ@%!3_dp1ICM5amv
zu>?YmF*rhkQ<&!=r4+bbGA@Y{ins<6K%HQxo4p4($#fM)QK<b-PO=Cr4JkD{uPb`E
zA@woG$M2Aggb~6gdl}5l^4KaAdJM;PgGdnGF-j6Kr?nv02LDSkC>eAI@_CZg$>QP(
z(|PVSP7<-JBX*>l=qJiA#>0F9Wg_y>BSQaNTwIVg%lz*3rM_PSuI^>b7+{hE4=_b9
zv0{FX7h0I)^a1(zrUJ`8Vl0RRP0(eym3`fced>v&jJU{lfVqROlVVq^_zNm4m(|D8
zpxkQ}3mMDLxVrsOva}#cj+6k)`m1joA5rwwVF&g2znoS4-^D<-Ngk)Q6<e6pd&l;0
zm6jIsz#oCWFLGh@%gKxA*BD1|fEWn~u^G@$LBV|UYIgQTXw2A<)vc}E2uxzy=c#AQ
z^nJuC(Ts?QIMF1PYQ3QHD`X=fgaT4iQ$z4X7{ZBQxzOs@Cm>$LjM;(3m_llM=)(f#
zN3+1TocQdJ>6=U2^&rFm`G-YDZV#Wh(6@U5o1~pu2yuvESS=z5eP7a%3{X&8$iu^a
zN5x_^G<>bp!lcY2bydk*lAWFXO$JNGt{ut?EI}JyoHR)kLT<veWOz=_I)n(5tmcR)
z2m}dE(}337n{An5BnX@fSNAX+oAQa6OzA)Gj|37s%6bE0mDo!5s1fi8s?nm%>%Sq_
zOkCr)Z>uplNRo+wRce;ySRg2<_%^4)IkqOw1*C7hmUAhtLumc_Wt7f=fm7&vBt8X@
z!CCL;>x)?>exONhbPO_-fk7N-|4mu#|HJ~0Or2#{p(R5Fng2XegFcouL?Q<OJng>S
z%vw4+nwa2HQhM-eVl}d!WVaU;o%q*y@Pl9_9W$l{cA(vZl;D{&%cwyx6m(#IB>5+4
zq-@*0d-wH0=Kl=ei~AQDmoFGhu}=RrI==yv7E^v&aA6%VMuc*(A0h}TMI_zW*!>?s
zT_L;WFC>4g;mw;+#Nm;>#N<fdJek;t<}n9uViORE`@8bNiMh*@6%}ffZF~3f<BK&I
z419CTfS1fFApf!1tuwTqKbB<Bk*#&Mkuf$EF_xt2-R2!DLJZ&p1kCjr#~eRkSNF>u
zV=G@T8B)Q-#00g~Xk@~Z&=^2&8;<7aUCy!LVQ21b=V*AIZ-jZ}a#q{u_(C9EpxYZa
zZfqO6#mZ)^ayXi<9TXNmy?F_gCMHKf-8lS54L%eW2x+Y#=1umqqW)LJS48lu26Cg?
zzkbnUzAC}o1vx6LxpKE9`sskpmeY~rV`FUf%5&*W4Gm#{?Q_!|^Qsdw^|7z1Lc-bQ
z^;#c6m55^u`7s<;rP<ZWX}C_r0LsnF1A&ntxd6l5$=2KZ5#nE2LJ*-*X_^O~)Kq^)
zav?lS48L__v65<7<9)Gj?UhEa1K3Wtc3nS(bS<#kkHq)iHpt7{Lw|<8*@Zq-jikE$
zp+gzx+nR`aok-Mcd1?!&bsbXpM%LD+fDnj|O8^v^W7vLrC1VeAZEPGIB6RNL1AlY`
zOxcA_x(Ri~ul#U>Fq&c4w|gD*6hIlQSDvezj#UHZ&{Bv#59c`k{P`0wX1YtKwas`9
z4vOa9zt4^QW)-f51p3Le+7CWR#+X)u)Fg<U0cr$Hru%xU67}gzZSsp$=2}*gSdJ#>
zA;O<zIE^%T+!M&e+1Hj3e2#QQNEdPe?*Tq0D_*|`W`#4q?xqAn%$k0FrS-IX;2aR9
zi1e(QT}Jf4jc0TYAa#+z%Y1x%vJjn##-7>#ZUY9m$&3WHWD5TEb|HNm{D6#$kd#D=
z>3^goRv*cvYmfVg$!0{VWc|kvJJ>-sMP$_Bz=4}4CAfc2pFQ(ycVTO?zQ9=hm$*%d
zyTyv78h`EDOrxi5z&8az1%fAneven-vAsdgl`2oQ`&@Ug{!2bpb$Vz6`3CU|UFDPQ
z;TWL9@S<JsBSgt-0Mc85lm>0MApdHo7~KFZW?*6>H4=kxwECB2Y48cm*IFi4P6MJt
zObSDL_VDy{9$yWk4B?%iuSF|`<rHEB*Mt83ZRnK(Apt4%0H;)oPGk9I_;_R@qhfwm
z2Gkh#W0uLS?>lFvHh~XO<xy{m^+&i8h|oSjQHa+Co*|?u1IsvI-Ywg6FI=LlQF@(w
z`+;FW#q6kRw|YFb5Tw+zYZ{M%ap$gR9@7g}mjy6t3^bBG^Eef$jtkEUotm=OBpuqn
zUu4-jl-JcdXZfn23d<dNE1II;b7|@bVz0{qPCzyGW3?g97<KjP9_qYqV}ibTPSF&&
z_rgP>qWuv65pM@QUmctl143k5hw6kcWK|Ghb$ZW~+PmhrZYvqSgvY!8*?BR^H1m+z
z2`-}2aB*37KY%yrjwF4rQyf?niEJRlt!4=q%&Z)^n%L7I=)`Eo1Y`uXb;j2}B-gI3
zg5z`!8v)V;86w4F%G-o>I&Wl&u@meQ<a=+LG`Z6wJA8&$aXIl8fRoRncMw+_T1ofy
zc+sT|I)t{wqCJyu0uGj}BdK6)_rQrJ%ci8OSFfIiW}ODZggi=2v4>8!2N<U7qX|$K
zUoy!W&pN#olHY!!Jwv1c0P`M#HD;b6-vD{!Z=NEmQ2=Hchz*G(BJ`8qL*k?Kp)f7G
zplm@rKUzFej_gckR+hulb3$Os;h3HP^LgiXUCe)JLwF&s2ttE7V`Zq0dMnax?#NU+
zz4mr<!)QTf5r{5H(w#fUN|i@tiYzVlqy7M~n5^AIlGWw@6V+UHfnj#RztS(Nzz{gp
z|M8v-NpgLY!^9v|vhGB0n^KST`bW0q{wfo%{PDMBIsi8}-+fOE&nd0l?|QcxU1!|g
zdZ>Sck0@LDj0TaSmseww^3Nb;Fm9#o+B*fb%x-MRnD#vu`CT0d<1*Ef?O)D@_+R};
z-N!DKIkiV@7d0MGA}3^I?kFwFjB7OzWM51$2F5ch7jlVZ#7f2p00LhmjB)My^>~<V
zd;%)|8%5b;>c4h(|7cXgswO8jq(EIjs6!iU7jbtvs=RmBrOaD@EPLz!;$#7fwY?bm
zLyDrSSYJHAsT}GGVVPs_U@IY}@oqXw1|mqagY(&3YMXjxAvSGb4e*ZoWxhxV5qZfT
zaCV%l5j#+Z0DOsS$61oGg9t>SAb8V>{-psSnFM|_kiZIm7mx7qO#m<^EMwwr;s_1`
z%V7Kc-G+opMBpB#)dKciCxdBVc9;+``|0fL+>OC@@BB-doBjxB$N%ezc1&4~q8gB_
z25=!P&!~a|j2cL0=b2(|5SvpIPA?htM9zXF>rlS5p%Gsn6w79gF@fpzDf-f0q%nY0
z9kv`>Ct$3uH7@hoUn}y~Epfy|-1^=KIoQF<gb5af%3~Su3Mmit>66%TRqx*2`5;eF
z5TF2#J_sRb37K0hv{?vWQLrj;KmGm8=7J$CRaaKtXdmodh}Iqfdpg{bf4%+@Eb!0R
zxQ5Dg=$;51*TH8_vU-9+bBy#!NK0#5Sgbs|TICdUgVqlpR4vQ36#j&bH~xzbJW=b3
z@ko;S2USzM6AX@z{}B1Z@sX0*R*qcM`i127j;pFcN+u>ItSeVW0YKWrPa<Gl+%Rbw
zg}f!`DZc6ke+hFQ4wa!aXc{N)-MeQLhsZA|hFBS>Iyn%;9fx^9!-R@RODOS%gZg;`
zV%h6RB&zYIH*^?K4j2#Nm>PkAe;d*+&PSl7kTU)K%MOyLfK^rn#QG-7bPcAaw)G_1
z{sk%{TcFWiR`!!eOokkGQ1<KVRS?q9s3`4)_mF<!N@sJ~P;2}Eau5dQnldkXYpnkz
zOuO{43r&3A!tVGhzTYUSp!IgRP6}oqW>@gJP2VSD>=$trARI8TR1|;2^#X#tPwS3q
zEWymxqAM8MMh;<(VG?Kk%rujA=`l>rRHk`x#*HY#-ywrrIJPAt$*wdqA|krMTR8QS
zwc}eCVj1A*ghQ{cRY6wBHuwO)S?VZ4N($NYSS?eqiD5sR^t6&`rug_UG73WSgUY_L
zgrR({jq&g2HBRSJVxTpv2t_*t<>~u)uRsqnBR5s%2$XL`_A*Vx0S%Cmz^wtBSOo|Q
zX=QchwL7@im?q2|nua#Yd;1=|@2I*9$N#gs+mhcl$A8m21HcavBt-JEOJ=~>WT8QM
zK?+7RVG3{;h$#Yl(|=HaZIqqPAufc*4uRwsR}`4gy{4cyM92f@N!po|+SrO6$w7)A
zo)VcQMa>dI$Q=~IbW0tSc+77&CMPA??)OQI?n`fB-SKdp0R$DGt1!5_#dT14h#J<3
zjsw@C72#xp^)auc>UuBQtjx?u2Vs!?z&~V3#>DKnK0s{tTaq{uW6(?occ8w$KB5KE
zf_L^w3>TUw-VH#~kZ40>1H%VC?BpaYE6ZA$mUs>vwFV(ivfzLKIx$R&4X(L&Z}|s*
zjEkQkXPV$i83vA!vm+q1k^&8m2S*mCBp}k%AsCcy*}^0c0iwExI~GVH45~HJF3F7Y
z7hbu2T2z@YQ2gy39mUFmVa-M;QAmITVfkFstJpQ<aF*tN8(3G6PXEoZ7D5cXs)6^7
z+987CD2u{7cWRJP&oRiq4)GBI5^fyKVSZ3^wIWnG%-Gq3Ss!jM+j4xkjS(t5_7ere
zNE2<Q{?zi(ckxg;iM52dE$YcqctvE^0f<i{5q2}9qog8WFV&x%906nl<ev%|46~5O
zu7BM44_fH;K{Ay^b`ib{ehk?$Sep2S+WQ2Dv<Ol45Go5J(qL1z`PA^kkxmd%2#EIW
zWq!m^#9NX`o6L;vWjK2EtYe7d+JP(xOG+-q05=6&@%y3WA|G2@p~FM?U5?b*<3fj;
zQ54w5ECcz(Gi#-A?31RZ=(nHa<DA2R1slm>4oXU&i4Oz69;||9o8zX%z6sU3DMpRy
z>$D8VtK%YOuWAX0!HtAuYT2uVVs{220R{pdt=@xdokA1VA_iN#zI?e^gqb-qfrB#>
zC=^IDfUn)?mbIXd?qi>~li4+L=ZCY`_}j_^vO*z_x^|6^Xm{B1m!XnCA0{ILDf;T_
zVMez+a%`r}L2&p9<VT>7*3_AuP2o|L-iO)spqf)y1op_@!%M4r{CF{Rb`|7i;!s{&
z0@V7OepVamkW1$FZ$@kvga{qumdCoWg@7#A-vYAj{~VAVDF1}PColA<yzgywdZ554
z5WWZbE%R;s2FAdzE#J35aZwdY=84pLmXbu@@@97{uvX@r!uZG}lNuQ^(Ha+C00RFr
zlTZK8;<Eo~#47Wz5vwb7e>{56f0vJq?1q>hr6p|q9)Kdr`Vhd3jg1ZHIfb0;Fxbot
zvCVj`C+3TY{!5~@@7@tn1QQQ=^!*|Wz|_g)`;Q+O2VMTpwIB1_xhAjvDavLCUP%He
zu)#1^^_UnBG_>@cU==ItNp6xH<b}r^g(DY;&x4ZgaxR$t&$Ltm<_xW@t#QnQb}N|I
z-v_VU%gOw;9|S)Q9R5SnP>^PDypRxxEpfHWe!SXB2}Sd`d+g&%jM@0&$d`duk<<bA
z1zS1c|8@vVAvlgeDKtSbF)?xs7)21Isi&uB^!s=A`YlRI#P0^m3Z6U5Hp{eN_nZ9J
zcK;c21jPJY02wGn83g75??w53Y&xh45gg$#!3)o1!DVaf>c|;4lu+z6)bMJUD?Qg%
zN?Bu=B%^(LM{Pfn-A*R|zKr5@phc8_KZB)u+XW5_7&NB-Suqwc54)e7pe4jA;-vq_
z*niyCR;Fjv_{Y`%7SjtDmWs0mKqIS72b&ukBcKYBc_GY{5ZN1N5Rnq*q3k<KS-g0$
zWN+I8?;H!~*U?U=h|{64_FwX@|2$RI6aXPJfl8&J;3es>kZ@xd>!1|5;p7a5Pu#{q
zpS?uMtB}M1SO6OiLL#+Z#h()^u?FTKg@OS*l6aFl@RSwTVFM!_a~tF;Nl?un$1aic
z(1_wskwsxg7a&gz+Wun%hRA3dbXW=ncv#M*Z8Ne2v~XI&aViTDuOPFHz0!4J{R1QW
zacTfsp&VX?%!)%AvU(J_Wg(vdIcl@tcTrNvPMlbTMClELACRMn)EDO%DD@o^Mx;`z
z?B7ju)4*=ZH#!KZf=o$Ziy2v1ER2}%zH@6#$*_R)-=Ow)E7=;>#T_}_x83;H^j<3~
z23p#1|AJ)azi3wfQ*AG6XTD_z*cTE(N#`_f@7dOP;7^_+;gHx!a2rBYnBhVA;Cn{1
z-{kOKeb|CHJPf1Jn~2SZCYO~^>OHtpJ8m^I`ea{U9|48P(hzB0l1Gez7jzAPtd-jp
zriV=774Z@JPz(s-pcjy=&xsQ-gIA+zOu@Tx8)#fRKG4KO&PIS(OWBrXwh!rQa!x|<
z%$Ux#&F+iHTZ5F5bY6L~);1M72lkjRS$i-DMwCBDasG#g=U3ruk@{(JsL&sLnrl3b
zl<2Zf<Kf$NlUJ`?NyRgTlVlb<jl*ary2rkJ!dQD)Y%DY4vWpRF`g@@MuXEUxi(i59
zJHSjNd0w3DL<8vt?5+}C9@6}j&=)WC@p%McV&R~cDEM>`?xDA+;2b>#s}QuRHgzjR
zXbYw69`VuB(@RdP&W@X+vs*lSY%YMh8Zo&;PY*Kr!q7&->6jSfBc|YD)c8_J;|C5M
zBFRxW-em5Jh)-#Ah*ZsH?)dJ4s5c-w474<KG@Qx=iwsPX<csQq;iQnv03~8ciK;Cz
ztK}2*B{a=ClD0Bjc-~We@bF=V7_HPIn7{?~UI!3<K*%Zrpp2YF0cT%QM2`qQ?E2J8
zmug_>LwT!KP1F}JdyOz589pKKb$;eYTzx4)xW^F3HGgylsW1F288^e8iNdJZwXHt+
z+OyO5kYK=(3|U9fJ&!8vXy<2RqXEZ+l+DwWg~LFwf5`D!fR}{rZaUf;3dkw4bt^BW
zv#*Z<e&7q7b^v=9#wZOG4>z3kl(Nh?h~t>U94wI(oO+22u>IKQ$0RL-5l0&3D-Sy!
z`V`cw&EPC3GB*HdoKy1XU&hED{=*m<Cb^eG$#S@J_inCdKy<XN*wV;FgtQ_gg*5Wt
z;uGW)3{u$*tUe3JH_o{HK<I`>FP$CWJOZbt0O-l>t7XW{Qu)2A@y(k`pli3e89y@d
zk6_#s7&#7SS(hM`?V39K+ymz#5bYF*gV=<S1T-Gy1JfW+M58t{8g8VDseVijnXmHu
z@K+)jRzA){f%DBojuZGExk_$dgi*f7w2yFWi){ciRDVo`SL5ctsR#)H_G`{^{l|f@
ztkqD9UfwzHBk*6_GrwadKD>V~gj5I$CeEv<2E4=3O3UG&rRd`rigQUxXHmPbF871v
zL8dTWt7(?36|n0vLb3JEe>)@v39>jyiRA3Z`zR2Q>cTsMrww}eVBR_yJ3^Epu>%A=
zP@~Pn8U|rybs}W#F<dcnYl~Ya1t+`^@ezmabQR=3z_BQjS!BBGU=p=#tDxYG`jFQ+
zZ;_3aHB@f@6XgY3|4-JvJ5?T9BW1Nzvms8XL{4k(L(Ors)Wr_-kFIV<J}AYAp+F~=
zClV{U3&R=YC^W<|Mw|DzW?f6m$zj44@0$2$1kU>NISwcDXhOg{45z6=fIkQA(xE1*
z3!K9@Nt&I_SW0D`-#ySTs@NJJSZoE1$+7;N-RkF~p|J=OJ${B5kJzy_4GqR(ZV-R}
zn=sg^D??g%HAd9WXaE#&kLErZK8&h@)xpXfDQzcn-(=ba-6lDQ%zj<R2~B?=<jvIk
zK78f3&-dYfYtLscV<5E!?jtk`xP}_rwowEO)tR$*WJvL-1uH|mJPkwxkYKUJ_79YY
z2_RybXB~-c1joxT-G<=!{PRfPAmf!2WcMFfz6Plv=i;dgAS8xl6fLOj%>~GxJ&Ds`
z@rcu5Zh=G@1M9?$gANyoI8gBH7h~xlqHj5^^#y^dY6RDy29a|L2}LE*ew>9T*#p{2
zjw{0XSXVP1P2z+&x1U41kA5u4Q+-HGbUMm8oOG7n8tZ!=S4j@$!W=EsgTmIAS$7Zu
zLD@|Oq(K?pr+5-#qtlI)kcz8cyl%oT0WiFM;2Asjs#Pgqqx%#A{G&5%6`l!Vzd4g(
zc;|xhD01jox<OGQc{A^3(KKGjavQsLn-C$kxA`X5R}PX9&i$YF;BV|eGV8sWxCmrC
z_EIsty_qSs;v==9dm>idN(38`mI45b!y<4g*?ADW@L8J!%9mC1s3V<8A%=nTxzoDE
zra13o=gyr69387LAvN^1?5(fVmF2dlZ-uv{_Cq#^pi-%$TWveIBNP@X6zl990p|G?
z>7;?ot*^=zhQGeCv-q%*WJuj$c{_a;mB_Y93pj3J+)iN;&OQYeg#Q!siE~kbj%l@X
zgV(6_M*gyu20EH{Off|%!w<>O&0iDn|IP8sRw7pA{t9pyaS}izIRFch@qLQ@>4f4z
z4M9Xq09}WBvm5%ayjgu)MTh{7Ig7%!0Z5hw)&tJmd2?lj5<oE>GVf4c7^i)ucwS>v
zi5O_W3ETGU;e{dq%CUkRHxoGan2XFr1LGhlb1pG)?Y`4?yZ&AFIs1#mm*T^lLuSU7
zpbaen0Kz08rj<;4d<|5maY_NQ=>k}4q$UH>VBFLu=eR6`vpw+=VUm!;W=MVqk(f}p
z5%9PkyL{Lwh0=?Upp95GS{XUj4wga6&6`>h8!*X5&cP$&p64%IaL0Mb%Mc*0wWydO
zdjxjYO6Vby(|%GsI)7AeL|Xr0LY)zj>qvPLXhEMM=YwL3kq#5gqd$K>0r}m2^w)Eo
zHN7yRSnRX{v%;rAsz?;4Bk)J8QB+H67lK<i$SiBQzjtw+^Y{1D)6?CcXWHb1Xt+G^
zX=nukN7cXaq~4@&LXN8k<Mfd&xOaNEIgW_(M(6oYKpv5#k+^W-0ye72ZX=@`=#)k-
zA4?1*Fdwp8jM4)f^VsR@&Tr_2^%$dAM@H1};Yz+EEo&i3En6Ae4^6|A)Br=}d?Sej
zk5Zdo2*5E;!_iCt&{@UB)xvEM=!g+woh>+mlo|xCXz_b`8jsiZ^b@*ms{{mO%JQEb
z3c)c<ctZ@dMaUB>t0&6dgN5NV(~cutwgGN!v$p2JDu@U|lnP9)8ke^H==+=u)R4+~
zF;m08oZ4b{FE8wy6$J*gi4bR@H>Q10q{W3E`FC+~#yT0hM`ys;sLo`-8ms!db>kg;
z+J84%Ni%94oh6<;i+6_HNQZ+oQqBMa<HTj;HR5?HxHx<jZy@B5h7lMDxJK_gQnxb!
zckQ}qQSJ}o2QWZpL13TZEI<zr5`YyE5uq00d$;Y{b(SG7a6!67zple=$mZ)VKCFyU
zYT{dv!=W&6)Csu{4dMi(|3$R4k*^;<m_ImX-B^I$nUa~wfcAlNL39Az$SIEG;E2v*
zR%2`cd_crWviL6PrQ@8S-Ff*e;j;eyMAsyNbI6R3UFz4Z1&$==B4DXwhl?Hev^@gg
z1uO(ln3h5g+`~g8s0(ks-Q`n`Bluh^lDEXFguyL>1Pk=)@W{x;1&`N7yo+bfMByRl
zlz3cffk26=#_di`IgW^Bw*o_v^UunDesHc1F10xeMVN#s(v=GgW>F9|F?k4WmH`6|
zz@{L(<ltJk%D7P|cB21kO;7{_+h<%YYM7QUfAY6L{y(PHH&bkx<eI+RxddeFQW$`Y
z9U_6IIfFHVxim+doKi`oT~CkwCx|V`Yzq8j;gX@A3^Pe`z&L0<NtA}i#>Uo<{qcS^
zdp;oF-icOJwrPT#kP0lO;qU*T5HXu7%)m^|W4Yl_oAUZJ>sQHYaWEusmi5u2xC7?-
z+hOJ4Ah2p@<JXV@Q*Pb*&5?_^J`QFHZPCzI_%-wp@mZic%*!%522X!e#7V_C{iyh>
z>(91(E8dtw+~2+S!m|WmpL(!n2{kzN=^#JyYu4xt-70SsH@I>WS!gp7woW99KEit%
ziAZWm=H=!l;7qiu^2lDi$LXbx@T|3tPULitJrr5kj2<}QAl*#T9N|VI5XUU!AJa^r
zE|xmf(5Jk>>7@xE81+~p#(pqIdN8<=07hL8N*2@l{gR=*{S83Ji}k?y>#(bG+ivea
zS%f7}4=%IR5rzI0zCtBaTS=SUH5j**Vr--b7nT6xZ7xYwm>voW3``)3D=35nc6=tr
zq>4Ro5Mr^@lJG+9Q5+{IK|TaV{pumuNSK<K)VjRg$+E3@kG}pHqpuJo>*G{^>mGrg
z)d*C#oM)phxdpaqc0%*Zt4XUdRGYo&<>6sce?UGRS!WN}Z<2K@mPUGDMbx7&NjS<U
z;;aYL#Q1n26K6YH2tD<{lvYO^9n%}%zBPePDLMZmd0qm!gb4YvmCU^u94z>$2*utg
z%dC0m&TF9Xv)pnTSs57_8}XNCS8dP+UH@cVteI-2`qS4Ityb8H(}tWK_5WYD9M{7s
zTp_YZGhi+lJsyXD;5Zsy5-n{t5J|N>V@CcL3ktV{G^5Bt++=hYXR5p75ELS3;p}Af
z_7VI&!q-R7J&5Uh`~+t_`8z_XC$rQ*5}_wgE+MCp;BMQ4H}rK>R2Y4Q(5!{yl`#{!
z7`aoN?YSp&KPWjOOe(Nofe#3DqfV03(cx@r$MP(OLK!$(wS*Fi+W$H4UON#jVB8X(
zd;UCoR~g3oZh)a6N2h@o9z~c|p?Zym);^-{7A!*`8VME$I5fx#<apSm7*!%(jnBA-
z&{OOId*qN&XO>W~XR2PmW&vX*{}T&xUMtcE!kac(q$gpbmIuw9oVY6sGlFPNIA7{-
zndkqnv~zoDIt=6Z&{XiyPle1O24xuRLd%;B#1kIU3t>Z05JWLW5Q2!XU~0_66ueOs
zPhGT0LRjgIGC~<%=#>x*B!oahq)25h9Oz2t=R55$=yEq>zjyENd7tNddPV54K$Cf`
z@h06_KvJOQS7o5!a6yYbyH;e}NuAK;73)v0rC3|GiOe8mR@x{R<GDif9$*Ag<_^Ij
z{ZPC*v?-0XkG|4+FrORUXTiGW9hB(-)LA;KV~^VJZesd@gDjmQ`VVF&lgaGkqMzTQ
zZ?xKM4s;ALEgQpYDBE2N3p=k@@44stFp3>lo24p)Hx@1mzJhe45-D*CHzpvtQEA<Z
zZ8E7RM~yju@Vv9LvaKx};<>Mjnr>;88;U_DKI<8?Dt*TQeK+db0DOoW5`U4jCgu=X
z^nE^G$X@`Y2dMUQe3tG{DL%La*TusjTiTp?Xz*qPZC<^-`?ULZU?jMlo6FwwPy)3C
zW~i6pL}n2re9*qTILeE$f62F*dwAPDHw5;$`kR{SWJ_on>Ep06+n4sxaX0K))AWZE
zu9Ed3W4?yxR2U_+N8Pr?|7NWEojQU`sYn&mO$>Fe;q1jIiC2Yg$6%uwNuWr6L|qB^
zJ5hrnY`vlOiy|i1H`w|uZ3f*Iz`)6&eI5Fh9TG|U#sh@BS`1sh{p3F7yq_Ig5bP>K
zT(uI|$k4t)R%O6w%Nq`gR!`TaHVl=ji?3FGUat3V6jeSh&RG$9GWa+#vY0pmoL9Oh
z511n%fh3;}$wY$OftmSRSk4h}4B9=RT)ZApN9|=D9KE_QL!lVVYQi>og6>~`K4aUU
zX$_x#>CpsT^1e2b2tSY%ZAv$rn^`yAok52wIV-4`vILSsuaZYRFHyQEqQjGYDSi}y
i`cDCZ7V>{ShK(G{){p7$(Qmnf##B{aQ}(H}$#DS+x)%)q

literal 0
HcmV?d00001

diff --git a/results/params.txt b/results/params.txt
new file mode 100644
index 0000000..572b16d
--- /dev/null
+++ b/results/params.txt
@@ -0,0 +1,63 @@
+Model Name: distilbert-base-uncased
+Best Score: 0.6430863652840643
+Best Params: 
+  EPOCHS: 5
+  BATCH_SIZE: 32
+  LR: 5e-05
+  WD: 0.0001
+  OPT: RMSprop
+************************************************************
+Model Name: albert-base-v2
+Best Score: 0.6370035931881839
+Best Params: 
+  EPOCHS: 5
+  BATCH_SIZE: 48
+  LR: 5e-05
+  WD: 0.0001
+  OPT: AdamW
+************************************************************
+Model Name: huawei-noah/TinyBERT_General_6L_768D
+Best Score: 0.6499193438968445
+Best Params: 
+  EPOCHS: 7
+  BATCH_SIZE: 16
+  LR: 3e-05
+  WD: 0.0001
+  OPT: AdamW
+************************************************************
+Model Name: roberta-base
+Best Score: 0.6491600098598747
+Best Params: 
+  EPOCHS: 5
+  BATCH_SIZE: 48
+  LR: 5e-05
+  WD: 0.0001
+  OPT: AdamW
+************************************************************
+bert-base-uncased
+Best Score: 0.6548374336934382
+Best Params: 
+  EPOCHS: 5
+  BATCH_SIZE: 16
+  LR: 3e-05
+  WD: 0.0001
+  OPT: AdamW
+************************************************************
+Model Name: google/electra-base-discriminator
+Best Score: 0.6513194910894985
+Best Params: 
+  EPOCHS: 7
+  BATCH_SIZE: 64
+  LR: 5e-05
+  WD: 0.0001
+  OPT: AdamW
+************************************************************
+Model Name: YituTech/conv-bert-base
+Best Score: 0.6431987265042431
+Best Params: 
+  EPOCHS: 5
+  BATCH_SIZE: 16
+  LR: 5e-05
+  WD: 0.0001
+  OPT: AdamW
+************************************************************