def main():
data = pd.read_csv('data/svm-data.csv', header=None)
X = data.iloc[:, 1:].values
y = data.iloc[:, 0].values
clf = SVC(C=100000, random_state=241)
clf.fit(X, y)
result = str([x+1 for x in clf.support_])
write_submission(result, '61')
def main():
model = utils.load_model()
valid_df = utils.get_valid_df()
predictions = model.predict(valid_df)
predictions = predictions.reshape(len(predictions), 1)
utils.write_submission(predictions)
def main():
boston = load_boston()
X = scale(boston.data)
y = boston.target
classifier = KNeighborsRegressor(n_neighbors=5, weights='distance')
kf = KFold(len(X), n_folds=5, shuffle=True, random_state=42)
neighbors_range = np.linspace(1.0, 10.0, num=200)
write_submission(
int(classifier_choice_cv(
X, y, classifier, 'p', neighbors_range, kf,
scoring='mean_squared_error')[0]),
'41')
def main():
wine_file = 'data/wine.data'
url = 'https://archive.ics.uci.edu/ml/machine-learning-databases/wine/wine.data'
try:
data = pd.read_csv(wine_file, header=None)
except IOError:
print ('No {0} file found, downloading it from {1}'
.format(wine_file, url))
wine_reques = requests.get(url).content
data = pd.read_csv(io.StringIO(wine_reques.decode('utf-8')),
header=None)
X = data.iloc[:, 1:].values
X_scaled = scale(X)
y = data.iloc[:, 0].values
kf = KFold(len(data.index), n_folds=5, shuffle=True, random_state=42)
neighbors_range = [x for x in range(1, 51)]
write_submission(
classifier_choice_cv(
X, y, KNeighborsClassifier, 'n_neighbors', neighbors_range, kf)[0],
'31')
write_submission(
classifier_choice_cv(
X, y, KNeighborsClassifier, 'n_neighbors', neighbors_range, kf)[1],
'32')
write_submission(
classifier_choice_cv(
X_scaled, y, KNeighborsClassifier, 'n_neighbors', neighbors_range, kf)[0],
'33')
write_submission(
classifier_choice_cv(
X_scaled, y, KNeighborsClassifier, 'n_neighbors', neighbors_range, kf)[1],
'34')
def generate_predict(self):
# Train model and create a submission
if not self.hasTrained:
train_x, y = self.load_train()
self.fit(train_x, y)
self.hasTrained = True
print("Creating submission")
test_features = self.load_test()
preds = self.compute_predict(test_features)
preds = self.transform(preds)
u.write_submission(preds.astype(int), self.subm + self.name + "_submission.csv")
print("Submission successfully created")
def trees_assignment():
data = pd.read_csv("data/titanic.csv", index_col="PassengerId")
data21 = data.dropna(subset=["Pclass", "Fare", "Age", "Survived", "Sex"])
pd.options.mode.chained_assignment = None # suppress false positive warn
data21["Sex"] = data21["Sex"].map({"female": 0, "male": 1})
pd.options.mode.chained_assignment = "warn" # turning warn back
feature_names = ["Pclass", "Fare", "Age", "Sex"]
X = data21[feature_names]
y = data21[["Survived"]]
clf = DecisionTreeClassifier(random_state=241)
clf.fit(X, y)
importances = clf.feature_importances_
indices = np.argsort(importances)[::-1]
subm21 = [feature_names[f] for f in indices][:2]
write_submission([x for x in subm21], "21")
def main():
data = {}
scaler = StandardScaler()
for data_type in ['train', 'test']:
df = pd.read_csv('data/perceptron-{0}.csv'.format(data_type),
header=None)
data['X_' + data_type] = df.iloc[:, 1:].values
data['y_' + data_type] = df.iloc[:, 0].values
data['X_train_scaled'] = scaler.fit_transform(data['X_train'])
data['X_test_scaled'] = scaler.transform(data['X_test'])
acc = get_accuracy(data['X_train'], data['y_train'],
data['X_test'], data['y_test'],)
acc_scaled = get_accuracy(data['X_train_scaled'], data['y_train'],
data['X_test_scaled'], data['y_test'])
write_submission(round(abs(acc - acc_scaled), 3), '51')
def main():
if not os.path.exists(ckpt_path + 'checkpoint'):
print('there is not saved model, please check the ckpt path')
exit()
print('Loading model...')
W_embedding = np.load(embedding_path)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
model = network.TextCNN(W_embedding, settings)
model.saver.restore(sess, tf.train.latest_checkpoint(ckpt_path))
#print('Local predicting...')
#print ('valid batches:%d'%n_va_batches)
#local_predict(sess, model)
print('Test predicting...')
print('test batches:%d' % n_tr_batches)
results = predict(sess, model)
sub_path_name = sub_path + model_name + str(
strftime("%m%d%H%M")) + '.csv'
id_list = get_id_list(id_list_path)
# id_list=id_list[:len(results)]
write_submission(sub_path_name, id_list, results, sr_id2title)
def main():
logging.basicConfig(level=logging.INFO,
format='%(asctime)s %(levelname)s %(message)s')
newsgroups = fetch_20newsgroups(subset='all',
categories=['alt.atheism', 'sci.space'])
vectorizer = TfidfVectorizer()
X = newsgroups.target
y = newsgroups.data
X_train = vectorizer.fit_transform(y)
grid = {'C': np.power(10.0, np.arange(-5, 6))}
cv = KFold(X_train.shape[0], n_folds=5, shuffle=True, random_state=241)
clf = SVC(kernel='linear', random_state=241)
gs = GridSearchCV(clf, grid, scoring='accuracy', cv=cv)
gs.fit(X_train, X)
clf.set_params(**gs.best_params_)
clf.fit(X_train, X)
result = (show_top10(clf, vectorizer))
result.sort()
write_submission(str(
[x for x in result]).lower().encode('ascii', 'ignore'),
'71') # still need some work to get rid of unicode problem
def main():
data = np.genfromtxt("data/data-logistic.csv", delimiter=",")
y = data[:, 0]
X = data[:, 1:]
e_convergence = 0.00001
max_iteration = 10000
c_reg = 10
k_step = 0.1
weights_start = [[0.0, 0.0]]
euclidean_distance = maxint
iteration_count = 0
while math.sqrt(euclidean_distance) > e_convergence and iteration_count < max_iteration:
weights_gradient = update_weights(X, y, weights_start, k_step)
euclidean_distance = 0
for w in xrange(len(weights_gradient[0])):
euclidean_distance += (weights_gradient[-1][w] - weights_gradient[-2][w]) ** 2
iteration_count += 1
final_w = weights_gradient[-1]
euclidean_distance = maxint
iteration_count = 0
while math.sqrt(euclidean_distance) > e_convergence and iteration_count < max_iteration:
weights_gradient_reg = update_weights_reg(X, y, weights_start, c_reg, k_step)
euclidean_distance = 0
for w in xrange(len(weights_gradient_reg[0])):
euclidean_distance += (weights_gradient_reg[-1][w] - weights_gradient_reg[-2][w]) ** 2
iteration_count += 1
final_w_reg = weights_gradient_reg[-1]
y_scores = [sigmoid(X[i].tolist(), final_w) for i in xrange(len(X))]
y_scores_reg = [sigmoid(X[i].tolist(), final_w_reg) for i in xrange(len(X))]
write_submission([round(roc_auc_score(y, y_scores), 3), round(roc_auc_score(y, y_scores_reg), 3)], "81")
X, y, X_test, y_label_encoder = utils.training_data()
# X, X_test = utils.extract_feature(X, X_test, "bypass", False)
X_train, X_cv, y_train, y_cv, cv_indices = utils.split_data(X, y)
y_train = to_categorical(y_train)
y_cv = to_categorical(y_cv)
y_full = to_categorical(y_cv)
lb = LabelBinarizer()
lb.fit(y_train)
y_train = lb.transform(y_train)
y_test = lb.transform(y_cv)
num_classes = y_train.shape[1]
X_train = np.transpose(X_train, (0, 2, 1))
X_cv = np.transpose(X_cv, (0, 2, 1))
X_test = np.transpose(X_test, (0, 2, 1))
X_full = np.transpose(X, (0, 2, 1))
X_train = X_train[..., np.newaxis]
X_cv = X_cv[..., np.newaxis]
X_test = X_test[..., np.newaxis]
X_full = X_full[..., np.newaxis]
model = load_model("trained_model.h5")
y_pred_test = model.predict_classes(X_test)
y_pred_labels = list(y_label_encoder.inverse_transform(y_pred_test))
utils.write_submission("neural_prediction", y_pred_labels)
def main():
input_dir = "/amit/kaggle/tgs"
output_dir = "/artifacts"
image_size_target = 128
batch_size = 32
epochs_to_train = 300
bce_loss_weight_gamma = 0.98
sgdr_min_lr = 0.0001 # 0.0001, 0.001
sgdr_max_lr = 0.001 # 0.001, 0.03
sgdr_cycle_epochs = 20
sgdr_cycle_epoch_prolongation = 3
sgdr_cycle_end_patience = 3
train_abort_epochs_without_improval = 30
ensemble_model_count = 3
swa_epoch_to_start = 30
model_dir = sys.argv[1] if len(sys.argv) > 1 else None
train_data = TrainData(input_dir)
train_set = TrainDataset(train_data.train_set_df, image_size_target, augment=True)
train_set_data_loader = DataLoader(train_set, batch_size=batch_size, shuffle=True, num_workers=8)
val_set = TrainDataset(train_data.val_set_df, image_size_target, augment=False)
val_set_data_loader = DataLoader(val_set, batch_size=batch_size, shuffle=False, num_workers=2)
if model_dir:
model = create_model(pretrained=False).to(device)
model.load_state_dict(torch.load("{}/model.pth".format(model_dir), map_location=device))
else:
model = create_model(pretrained=True).to(device)
torch.save(model.state_dict(), "{}/model.pth".format(output_dir))
swa_model = create_model(pretrained=False).to(device)
print("train_set_samples: %d, val_set_samples: %d" % (len(train_set), len(val_set)))
global_val_precision_best_avg = float("-inf")
global_swa_val_precision_best_avg = float("-inf")
sgdr_cycle_val_precision_best_avg = float("-inf")
epoch_iterations = len(train_set) // batch_size
# optimizer = optim.SGD(model.parameters(), lr=sgdr_max_lr, weight_decay=0, momentum=0.9, nesterov=True)
optimizer = optim.Adam(model.parameters(), lr=sgdr_max_lr)
lr_scheduler = CosineAnnealingLR(optimizer, T_max=sgdr_cycle_epochs, eta_min=sgdr_min_lr)
optim_summary_writer = SummaryWriter(log_dir="{}/logs/optim".format(output_dir))
train_summary_writer = SummaryWriter(log_dir="{}/logs/train".format(output_dir))
val_summary_writer = SummaryWriter(log_dir="{}/logs/val".format(output_dir))
swa_val_summary_writer = SummaryWriter(log_dir="{}/logs/swa_val".format(output_dir))
sgdr_iterations = 0
sgdr_reset_count = 0
batch_count = 0
epoch_of_last_improval = 0
sgdr_next_cycle_end_epoch = sgdr_cycle_epochs + sgdr_cycle_epoch_prolongation
swa_update_count = 0
ensemble_model_index = 0
for model_file_path in glob.glob("{}/model-*.pth".format(output_dir)):
model_file_name = os.path.basename(model_file_path)
model_index = int(model_file_name.replace("model-", "").replace(".pth", ""))
ensemble_model_index = max(ensemble_model_index, model_index + 1)
print('{"chart": "best_val_precision", "axis": "epoch"}')
print('{"chart": "val_precision", "axis": "epoch"}')
print('{"chart": "val_loss", "axis": "epoch"}')
print('{"chart": "sgdr_reset", "axis": "epoch"}')
print('{"chart": "precision", "axis": "epoch"}')
print('{"chart": "loss", "axis": "epoch"}')
print('{"chart": "swa_val_precision", "axis": "epoch"}')
print('{"chart": "swa_val_loss", "axis": "epoch"}')
train_start_time = time.time()
criterion = nn.BCEWithLogitsLoss()
for epoch in range(epochs_to_train):
epoch_start_time = time.time()
model.train()
train_loss_sum = 0.0
train_precision_sum = 0.0
train_step_count = 0
for batch in train_set_data_loader:
images, masks, mask_weights = \
batch[0].to(device, non_blocking=True), \
batch[1].to(device, non_blocking=True), \
batch[2].to(device, non_blocking=True)
lr_scheduler.step(epoch=min(sgdr_cycle_epochs, sgdr_iterations / epoch_iterations))
optimizer.zero_grad()
prediction_logits = model(images)
predictions = torch.sigmoid(prediction_logits)
criterion.weight = mask_weights
loss = criterion(prediction_logits, masks)
loss.backward()
optimizer.step()
train_loss_sum += loss.item()
train_precision_sum += np.mean(precision_batch(predictions, masks))
sgdr_iterations += 1
train_step_count += 1
batch_count += 1
optim_summary_writer.add_scalar("lr", get_learning_rate(optimizer), batch_count + 1)
train_loss_avg = train_loss_sum / train_step_count
train_precision_avg = train_precision_sum / train_step_count
val_loss_avg, val_precision_avg = evaluate(model, val_set_data_loader, criterion)
model_improved_within_sgdr_cycle = val_precision_avg > sgdr_cycle_val_precision_best_avg
if model_improved_within_sgdr_cycle:
torch.save(model.state_dict(), "{}/model-{}.pth".format(output_dir, ensemble_model_index))
sgdr_cycle_val_precision_best_avg = val_precision_avg
model_improved = val_precision_avg > global_val_precision_best_avg
ckpt_saved = False
if model_improved:
torch.save(model.state_dict(), "{}/model.pth".format(output_dir))
global_val_precision_best_avg = val_precision_avg
ckpt_saved = True
swa_model_improved = False
if epoch + 1 >= swa_epoch_to_start:
if model_improved_within_sgdr_cycle:
swa_update_count += 1
moving_average(swa_model, model, 1.0 / swa_update_count)
bn_update(train_set_data_loader, swa_model)
swa_model_improved = val_precision_avg > global_swa_val_precision_best_avg
if swa_model_improved:
torch.save(swa_model.state_dict(), "{}/swa_model.pth".format(output_dir))
global_swa_val_precision_best_avg = val_precision_avg
if model_improved or swa_model_improved:
epoch_of_last_improval = epoch
sgdr_reset = False
if (epoch + 1 >= sgdr_next_cycle_end_epoch) and (epoch - epoch_of_last_improval >= sgdr_cycle_end_patience):
sgdr_iterations = 0
sgdr_next_cycle_end_epoch = epoch + 1 + sgdr_cycle_epochs + sgdr_cycle_epoch_prolongation
ensemble_model_index += 1
sgdr_cycle_val_precision_best_avg = float("-inf")
sgdr_reset_count += 1
sgdr_reset = True
swa_val_loss_avg, swa_val_precision_avg = evaluate(swa_model, val_set_data_loader, criterion)
optim_summary_writer.add_scalar("sgdr_reset", sgdr_reset_count, epoch + 1)
train_summary_writer.add_scalar("loss", train_loss_avg, epoch + 1)
train_summary_writer.add_scalar("precision", train_precision_avg, epoch + 1)
val_summary_writer.add_scalar("loss", val_loss_avg, epoch + 1)
val_summary_writer.add_scalar("precision", val_precision_avg, epoch + 1)
swa_val_summary_writer.add_scalar("loss", swa_val_loss_avg, epoch + 1)
swa_val_summary_writer.add_scalar("precision", swa_val_precision_avg, epoch + 1)
epoch_end_time = time.time()
epoch_duration_time = epoch_end_time - epoch_start_time
print(
"[%03d/%03d] %ds, lr: %.6f, loss: %.3f, val_loss: %.3f|%.3f, prec: %.3f, val_prec: %.3f|%.3f, ckpt: %d, rst: %d" % (
epoch + 1,
epochs_to_train,
epoch_duration_time,
get_learning_rate(optimizer),
train_loss_avg,
val_loss_avg,
swa_val_loss_avg,
train_precision_avg,
val_precision_avg,
swa_val_precision_avg,
int(ckpt_saved),
int(sgdr_reset)),
flush=True)
print('{"chart": "best_val_precision", "x": %d, "y": %.3f}' % (epoch + 1, global_val_precision_best_avg))
print('{"chart": "val_precision", "x": %d, "y": %.3f}' % (epoch + 1, val_precision_avg))
print('{"chart": "val_loss", "x": %d, "y": %.3f}' % (epoch + 1, val_loss_avg))
print('{"chart": "sgdr_reset", "x": %d, "y": %.3f}' % (epoch + 1, sgdr_reset_count))
print('{"chart": "precision", "x": %d, "y": %.3f}' % (epoch + 1, train_precision_avg))
print('{"chart": "loss", "x": %d, "y": %.3f}' % (epoch + 1, train_loss_avg))
print('{"chart": "swa_val_precision", "x": %d, "y": %.3f}' % (epoch + 1, swa_val_precision_avg))
print('{"chart": "swa_val_loss", "x": %d, "y": %.3f}' % (epoch + 1, swa_val_loss_avg))
if sgdr_reset and sgdr_reset_count >= ensemble_model_count and epoch - epoch_of_last_improval >= train_abort_epochs_without_improval:
print("early abort")
break
optim_summary_writer.close()
train_summary_writer.close()
val_summary_writer.close()
train_end_time = time.time()
print()
print("Train time: %s" % str(datetime.timedelta(seconds=train_end_time - train_start_time)))
eval_start_time = time.time()
print()
print("evaluation of the training model")
model.load_state_dict(torch.load("{}/model.pth".format(output_dir), map_location=device))
analyze(Ensemble([model]), train_data.val_set_df, use_tta=False)
analyze(Ensemble([model]), train_data.val_set_df, use_tta=True)
score_to_model = {}
ensemble_model_candidates = glob.glob("{}/model-*.pth".format(output_dir))
ensemble_model_candidates.append("{}/swa_model.pth".format(output_dir))
for model_file_path in ensemble_model_candidates:
model_file_name = os.path.basename(model_file_path)
m = create_model(pretrained=False).to(device)
m.load_state_dict(torch.load(model_file_path, map_location=device))
val_loss_avg, val_precision_avg = evaluate(m, val_set_data_loader, criterion)
print("ensemble '%s': val_loss=%.3f, val_precision=%.3f" % (model_file_name, val_loss_avg, val_precision_avg))
if len(score_to_model) < ensemble_model_count or min(score_to_model.keys()) < val_precision_avg:
del score_to_model[min(score_to_model.keys())]
score_to_model[val_precision_avg] = m
ensemble_models = list(score_to_model.values())
for ensemble_model in ensemble_models:
val_loss_avg, val_precision_avg = evaluate(ensemble_model, val_set_data_loader, criterion)
print("ensemble: val_loss=%.3f, val_precision=%.3f" % (val_loss_avg, val_precision_avg))
model = Ensemble(ensemble_models)
mask_threshold_global, mask_threshold_per_cc = analyze(model, train_data.val_set_df, use_tta=True)
eval_end_time = time.time()
print()
print("Eval time: %s" % str(datetime.timedelta(seconds=eval_end_time - eval_start_time)))
print()
print("submission preparation")
submission_start_time = time.time()
test_data = TestData(input_dir)
calculate_predictions(test_data.df, model, use_tta=True)
calculate_prediction_masks(test_data.df, mask_threshold_global)
print()
print(test_data.df.groupby("predictions_cc").agg({"predictions_cc": "count"}))
write_submission(test_data.df, "prediction_masks", "{}/{}".format(output_dir, "submission.csv"))
write_submission(test_data.df, "prediction_masks_best", "{}/{}".format(output_dir, "submission_best.csv"))
submission_end_time = time.time()
print()
print("Submission time: %s" % str(datetime.timedelta(seconds=submission_end_time - submission_start_time)))
with open('result_statistics_cross_val_marleen.txt', mode='a+') as f:
f.write(
'%s\t%s\t%s\t%s\t%.2f\t%.2f\t%.2f\t%.2f\t%.2f\t%.5f\t%s\t%.0f\n' %
(k, rand_seed,
len(feats), n_estimators, all_n, len(df_to_train) / 100000,
len(df_to_test) / 100000, learning_rate, downsampling_rate,
metric.calc_mean(teid_arr, terelevance_arr,
tepred), balance_flag, min_samples_leaf))
results_df = df_to_test[['srch_id', 'prop_id', 'relevance']].copy()
results_df['score'] = -1 * tepred
# predictions = list(-1.0*predictions)
recommendations = zip(results_df["srch_id"], results_df["prop_id"],
results_df['relevance'], results_df['score'])
utils.write_submission(recommendations, "lambdamart_test.csv")
path_results = "lambdamart_test.csv"
nDCG_result = nDCG.compute_ndcg(path_results)
print(nDCG_result)
with open('result_statistics_our_nDCG_cross_val_marleen.txt',
mode='a+') as f:
f.write(
'%s\t%s\t%s\t%s\t%.2f\t%.2f\t%.2f\t%.2f\t%.2f\t%.5f\t%s\t%.0f\n' %
(k, rand_seed, len(feats), n_estimators, all_n, len(df_to_train) /
100000, len(df_to_test) / 100000, learning_rate,
downsampling_rate, nDCG_result, balance_flag, min_samples_leaf))
import pickle
model_save = pickle.dumps(model)
new_model = pickle.loads(model_save)
print "cross validation results:"
print roc_aucs
############################################################################
# fit and predict
############################################################################
result = fb_funcs.fit_and_predict(info_humans, info_bots, info_test,
params=params_ens, scale='log')
y_test_proba = result['y_test_proba']
ytps = result['ytps']
# feature_importances = pd.DataFrame(np.array([result['features'],
# result['importances']]).T)
############################################################################
# submission file generation
############################################################################
submissionfile = 'data/submi/sub_ens.csv'
testfile = 'data/test.csv'
print "writing a submission file..."
write_submission(y_test_proba, info_test.index, testfile, submissionfile)
print "writing results from different models..."
for i in range(len(ytps)):
submf = 'data/submi/sub_%s.csv' %(params_ens[i]['model'])
write_submission(ytps[i], info_test.index, testfile, submf)
def generate_predict(self):
# Train model and create a submission
# If checkpoint available do not train anew
if self.checkpoint_dir:
text_path = os.path.join(os.path.curdir, self.checkpoint_dir, "..",
"text_vocab")
self.text_vocab_processor = tf.contrib.learn.preprocessing.VocabularyProcessor.restore(
text_path)
checkpoint_file = tf.train.latest_checkpoint(self.checkpoint_dir)
graph = tf.Graph()
with graph.as_default():
session_conf = tf.ConfigProto(
allow_soft_placement=self.allow_soft_placement,
log_device_placement=self.log_device_placement)
sess = tf.Session(config=session_conf)
with sess.as_default():
# Load the saved meta graph and restore variables
saver = tf.train.import_meta_graph(
"{}.meta".format(checkpoint_file))
saver.restore(sess, checkpoint_file)
self.session = sess
# Get params
predictions = graph.get_operation_by_name(
"output/predictions").outputs[0]
input_text = graph.get_operation_by_name(
"input_text").outputs[0]
dropout_keep_prob = graph.get_operation_by_name(
"dropout_keep_prob").outputs[0]
else:
train_x, y = self.load_train()
self.fit(train_x, y)
# Get params
predictions = self.model.predictions
input_text = self.model.input_text
dropout_keep_prob = self.model.dropout_keep_prob
test_x = self.load_test()
batches = u.batch_iter(list(test_x), self.batch_size, 1, shuffle=False)
print("Creating submission")
all_predictions = []
# Collect the predictions here
for x_batch in batches:
batch_predictions = self.session.run(predictions, {
input_text: x_batch,
dropout_keep_prob: 1.0
})
all_predictions = np.concatenate(
[all_predictions, batch_predictions])
# Map to correct labels
preds = np.asarray(
[-1 if pred == 0 else pred for pred in all_predictions])
u.write_submission(preds.astype(int),
self.subm + self.name + "_submission.csv")
print("Submission successfully created")
def pandas_assignment():
data = pd.read_csv("data/titanic.csv", index_col="PassengerId")
subm11 = data["Sex"].value_counts()
write_submission(" ".join([str(x) for x in subm11]), "11")
write_submission(int(data["Survived"].value_counts(normalize=True).to_dict()[1] * 100), "12")
write_submission(int(data["Pclass"].value_counts(normalize=True).to_dict()[1] * 100), "13")
subm14 = []
subm14.append(round(float(data["Age"].mean()), 1))
subm14.append(int(data["Age"].median()))
write_submission(" ".join([str(x) for x in subm14]), "14")
write_submission(round(data.corr("pearson")["SibSp"]["Parch"], 2), "15")
write_submission(
data["Name"]
.str.extract("(Miss\. |Mrs\.[A-Za-z ]*\()([A-Za-z]*)")[1]
.value_counts()
.head(n=1)
.to_string()
.split(" ", 1)[0],
"16",
)
