def compress(name, k=8, force_update=False):
result = []
block_list = []
elements_file_name = name + '.pickle'
compressed_elements_file_name = name + '_compressed.pickle'
if not force_update and os.path.exists(compressed_elements_file_name):
return load_data(compressed_elements_file_name)
else:
elements = load_data(elements_file_name)
for i in xrange(0, len(elements), k):
block_list.append(
(0,
elements[i].term,
elements[i].count,
elements[i].posting_lists
)
)
last_index = len(elements) - i
for bi in xrange(1, min(k, last_index)):
c = compare(elements[i + bi - 1].term, elements[i + bi].term)
block_list.append(
(c,
elements[i + bi].term[c:],
elements[i + bi].count,
# elements[i + bi].term,
elements[i + bi].posting_lists
)
)
result.append(block_list)
block_list = []
save_data(result, compressed_elements_file_name)
return result
def hand_generator(self):
for root, directories, files in walk(self.color_folder):
for c_filename in files:
if self.c_file_ext in c_filename:
try:
c_filepath = path.join(root, c_filename)
d_filepath = c_filepath.replace('color', 'depth').replace(self.c_file_ext, self.d_file_ext)
_, c_data = utils.load_data(c_filepath, self.data_type)
d_data, d_data_norm = utils.load_data(d_filepath, 'depth')
c_fgmask = self.c_fgbg.apply(c_data)
c_fgmask = cv2.morphologyEx(c_fgmask, cv2.MORPH_OPEN, self.kernel)
img_bs = d_data_norm.copy()
img_bs[c_fgmask==0] = 0
img_bs = ip.smooth_image(img_bs)
boxes = ip.parse_hands(img_bs, display=False)
hands = Hands(c_fgmask, d_data, boxes, d_filepath, 'depth')
yield hands
except Exception as e:
print e
def initialize_detector(self, background_folder, video_folder, file_ext):
self.c_fgbg = cv2.BackgroundSubtractorMOG2(300, 20, True)
self.d_fgbg = cv2.BackgroundSubtractorMOG2(300, 10, False)
color_bg_folder = background_folder + "/color"
self.color_folder = video_folder + "/color"
self.depth_folder = video_folder + "/depth"
self.c_file_ext = file_ext
self.d_file_ext = 'bin'
count = 0
thresh_total = 0
for root, directories, files in walk(color_bg_folder):
for c_filename in files:
if self.c_file_ext in c_filename:
try:
c_filepath = path.join(root, c_filename)
# d_filepath = c_filepath.replace('color', 'depth').replace(self.c_file_ext, self.d_file_ext)
_, c_data = utils.load_data(c_filepath, self.data_type)
# d_data, d_data_norm = utils.load_data(d_filepath, 'depth')
c_fgmask = self.c_fgbg.apply(c_data)
# d_fgmask = self.d_fgbg.apply(d_data)
except Exception as e:
print e
self.initialized = True
def hand_generator(self): for root, directories, files in walk(self.folder): for filename in files: if self.file_ext in filename: filepath = path.join(root, filename) data, data_norm = utils.load_data(filepath, self.data_type) fgmask = self.fgbg.apply(data_norm) fgmask = cv2.morphologyEx(fgmask, cv2.MORPH_OPEN, self.kernel) img_bs = data.copy() img_bs[fgmask==0] = 0 img_bs = ip.smooth_image(img_bs) orig_mask = fgmask.copy() fgmask, thresh_val = self.remove_legs(img_bs) img_fg = data_norm.copy() img_fg[fgmask==0] = 0 boxes = ip.parse_hands(fgmask, display=False) hands = Hands(fgmask, data, boxes, filepath, 'depth') yield hands
def get_index(folder_name, force_update=False):
index_file_name = folder_name + '.pickle'
if not force_update and os.path.exists(index_file_name):
return load_data(index_file_name)
else:
elements = []
documents = get_file_list(folder_name)
for doc_id in xrange(len(documents)):
elements += map(lambda x: Element(x, doc_id), get_tokens(documents[doc_id]))
elements.sort()
result = []
for el in elements:
if result and result[-1] == el:
result[-1].update(el)
else:
result.append(el)
save_data(result, name=index_file_name)
def img_generator(self): for root, directories, files in walk(self.folder): for filename in files: if self.file_ext in filename: filepath = path.join(root, filename) try: data, data_norm = utils.load_data(filepath, self.data_type) fgmask = self.fgbg.apply(data_norm) fgmask = cv2.morphologyEx(fgmask, cv2.MORPH_OPEN, self.kernel) img_bs = data_norm.copy() img_bs[fgmask==0] = 0 img_bs = ip.smooth_image(img_bs) yield img_bs, filepath except Exception as e: print e
def initialize_detector(self, background_folder, video_folder, file_ext): self.fgbg = cv2.BackgroundSubtractorMOG2(300,.2,False) self.folder = video_folder self.file_ext = file_ext count = 0 thresh_total = 0 for root, directories, files in walk(background_folder): for filename in files: if self.file_ext in filename: try: filepath = path.join(root, filename) data, data_norm = utils.load_data(filepath, self.data_type) fgmask = self.fgbg.apply(data_norm) except Exception as e: print e self.initialized = True
if __name__ == "__main__":
args = parse_args()
# fix random seed
set_random_seed(args.seed)
# use cuda or not
if args.use_cuda:
device = torch.device('cuda')
else:
device = torch.device('cpu')
# load data
data, n_fields, n_features = load_data(args.dataset,
device=device,
use_content=args.use_content,
use_rating=False,
print_info=True)
# create model
model = AttentionalFM(n_features=n_features,
n_fields=n_fields,
embedding_dim=args.embedding_dim)
model.to(device=device)
# output dir
output_dir = "./results/{:%Y%m%d_%H%M%S}/".format(datetime.now())
output_path = output_dir + "model.weights"
if not os.path.exists(output_dir):
os.makedirs(output_dir)
############### PREPROCESSING ###############
classes = results.classes.replace(" ", "").split(',')
preprocess_dir(TRAIN_DIR,
PREPROCESSED_DIR,
REORIENT_SCRIPT_PATH,
ROBUSTFOV_SCRIPT_PATH,
classes,
results.numcores,
verbose=0)
############### DATA IMPORT ###############
X, y, filenames, num_classes, img_shape = load_data(
PREPROCESSED_DIR, classes)
print("Finished data processing")
############### MODEL SELECTION ###############
LR = 1e-3
LOAD_WEIGHTS = False
MODEL_NAME = "phinet_model_" + "-".join(results.classes.split(","))
MODEL_PATH = os.path.join(WEIGHT_DIR, MODEL_NAME + ".json")
if not os.path.exists(WEIGHT_DIR):
os.makedirs(WEIGHT_DIR)
if LOAD_WEIGHTS:
weight_files = os.listdir(WEIGHT_DIR)
cont_channels = [
i for (i, x) in enumerate(discretizer_header) if x.find("->") == -1
]
normalizer = Normalizer(
fields=cont_channels) # choose here which columns to standardize
normalizer_state = 'ihm_normalizer'
normalizer_state = os.path.join(os.path.dirname(data_path), normalizer_state)
normalizer.load_params(normalizer_state)
# %%
n_trained_chunks = 0
train_raw = utils.load_data(train_reader,
discretizer,
normalizer,
small_part,
return_names=True)
val_raw = utils.load_data(val_reader,
discretizer,
normalizer,
small_part,
return_names=True)
# %%
demographic_data = []
diagnosis_data = []
idx_list = []
demo_path = data_path + 'demographic/'
import sys from utils.utils import load_data, datset_gen, savepickle from utils.trainer import train_with_checkpoin_tensorboard from Models import svhn_model_simple import tensorflow as tf import numpy as np import pandas as pd root_path = "I:\\Files_ML\\Coursera\\Dl_ON_ud\\dataset" test_path = os.path.join(root_path, "test") train_path = os.path.join(root_path, "train") extra_path = os.path.join(root_path, "extra") X_test, y_test = load_data(test_path, num_only=True) X_train, y_train = load_data(train_path, num_only=True) X_extra, y_extra = load_data(extra_path, num_only=True) #X_train = np.concatenate([X_train,X_extra]) #y_train = np.concatenate([y_train,y_extra]) y_train = y_train.reshape((-1, 5, 1)) y_test = y_test.reshape((-1, 5, 1)) # %% batch_size = 32 ds_train = datset_gen(X_train, y_train, batch_size=batch_size, buffer_size=100) ds_test = datset_gen(X_test, y_test, batch_size=batch_size, buffer_size=100) # %%
def initial_load():
kwargs = request.args.to_dict()
print kwargs
return load_data(**kwargs)
def main():
logger = logMaster.get_logger('main')
logger.info('loading data...')
att_feats, train_data, val_data, test_data, test_s_data, classes = load_data(
att_path=att_path, res_path=res_path)
logger.info('building model...')
gen = Generator(x_dim=args.x_dim,
s_dim=args.s_dim,
z_dim=args.z_dim,
layers=args.dec)
# gen.train()
# states = torch.load(args.vae_ckpt)
# gen.load_state_dict(states['model'])
dis = Discriminator(x_dim=args.x_dim, s_dim=args.s_dim, layers=args.dis)
reg = Regressor(x_dim=args.x_dim, s_dim=args.s_dim, layers=args.reg)
gen.cuda()
dis.cuda()
reg.cuda()
mse_loss = nn.MSELoss()
l1_loss = nn.L1Loss()
adam_betas = (0.8, 0.999)
gen_opt = optim.Adam(gen.parameters(),
lr=args.learning_rate,
weight_decay=0.01,
betas=adam_betas)
dis_opt = optim.Adam(dis.parameters(),
lr=args.learning_rate,
weight_decay=0.01,
betas=adam_betas)
reg_opt = optim.Adam(reg.parameters(),
lr=args.learning_rate,
weight_decay=0.01,
betas=adam_betas)
train_manager = DataManager(train_data,
args.epoch,
args.batch,
infinite=True)
ones = Variable(torch.ones([args.batch, 1]),
requires_grad=False).float().cuda()
zeros = Variable(torch.zeros([args.batch, 1]),
requires_grad=False).float().cuda()
loss_history = []
logger.info('start training...')
for epoch in range(args.epoch):
running_loss = 0
t1 = time.time()
d_total_loss = 0.0
g_total_loss = 0.0
# cyc_total_loss = 0.0
r_total_loss = 0.0
# rd_total_loss = 0.0
# vae_total_loss = 0.0
g_scores = 0.0
if args.steps == -1:
steps = train_manager.num_batch
else:
steps = args.steps
for batch in tqdm(range(steps), leave=False, ncols=70, unit='b'):
for i in range(args.d_iter):
dis.zero_grad()
# get true data
data = train_manager.get_batch()
X = Variable(
torch.from_numpy(np.asarray([item[0] for item in data
]))).float().cuda()
Y = [item[1] for item in data]
S = Variable(torch.from_numpy(att_feats[Y])).float().cuda()
Yc = get_negative_samples(Y, classes['train'])
Sc = Variable(torch.from_numpy(att_feats[Yc])).float().cuda()
# get fake data
# Xp = gen.forward(X, S)
# Xp = Xp.detach() # fix the generator
Xpp = gen.sample(S).detach()
Sp = reg.forward(X).detach() # fix the regressor
# get scores
true_scores = dis.forward(X, S)
# fake_scores = dis.forward(Xp, S)
fake_scores2 = dis.forward(Xpp, S)
# reg_scores = dis.forward(X, Sp)
# ctrl_scores = dis.forward(X, Sc)
# calculate loss
d_loss = mse_loss(true_scores, ones) + mse_loss(
fake_scores2,
zeros) # + args.theta3 * mse_loss(reg_scores, zeros) \
# + mse_loss(ctrl_scores, zeros)
d_loss.backward()
dis_opt.step()
d_total_loss += d_loss.cpu().data.numpy()
for i in range(args.g_iter):
gen.zero_grad()
reg.zero_grad()
# get true data
data = train_manager.get_batch()
X = Variable(
torch.from_numpy(np.asarray([item[0] for item in data
]))).float().cuda()
Y = [item[1] for item in data]
S = Variable(torch.from_numpy(att_feats[Y])).float().cuda()
# get fake data
# Xp, mu, log_sigma = gen.forward(X, S)
Xp2 = gen.sample(S)
Sp = reg.forward(X)
# Spp = reg.forward(Xp)
# Xpp, _, _ = gen.forward(X, Sp)
# get scores
# fake_scores = dis.forward(Xp, S)
fake_scores2 = dis.forward(Xp2, S)
# reg_scores = dis.forward(X, Sp)
# calculate loss
# vae_loss = gen.vae_loss(X=X, Xp=Xp, mu=mu, log_sigma=log_sigma)
# cyc_loss = mse_loss(Spp, S) + mse_loss(Xpp, X)
g_loss = mse_loss(fake_scores2, ones)
r_loss = mse_loss(Sp, S)
# rd_loss = mse_loss(reg_scores, ones)
# total_loss = vae_loss + g_loss + args.theta1 * cyc_loss + args.theta2 * r_loss + args.theta3 * rd_loss
total_loss = g_loss + args.theta2 * r_loss
total_loss.backward()
gen_opt.step()
reg_opt.step()
# vae_total_loss += vae_loss.cpu().data.numpy()
g_total_loss += g_loss.cpu().data.numpy()
# cyc_total_loss += cyc_loss.cpu().data.numpy()
r_total_loss += r_loss.cpu().data.numpy()
# rd_total_loss += rd_loss.cpu().data.numpy()
g_scores += np.mean(fake_scores2.cpu().data.numpy())
g_total_steps = steps * args.g_iter
d_total_steps = steps * args.d_iter
# vae_avg_loss = vae_total_loss / g_total_steps
g_avg_loss = g_total_loss / g_total_steps
# cyc_avg_loss = cyc_total_loss / g_total_steps
r_avg_loss = r_total_loss / g_total_steps
# rd_avg_loss = rd_total_loss / g_total_steps
d_avg_loss = d_total_loss / d_total_steps
g_avg_score = g_scores / g_total_steps
loss_history.append(
f'{g_avg_loss:.4}\t{d_avg_loss:.4}\t{r_avg_loss:.4}\t'
f'{g_avg_score:.4}\n')
elapsed = (time.time() - t1) / 60.0
if (epoch + 1) % 10 == 0 or epoch == 0:
filename = 'gdan_' + str(epoch + 1) + '.pkl'
save_path = save_dir / Path(filename)
states = dict()
states['epoch'] = epoch + 1
states['gen'] = gen.state_dict()
states['dis'] = dis.state_dict()
states['reg'] = reg.state_dict()
# states['enc_layers'] = args.enc
states['gen_layers'] = args.dec
states['reg_layers'] = args.reg
states['dis_layers'] = args.dis
states['z_dim'] = args.z_dim
states['x_dim'] = args.x_dim
states['s_dim'] = args.s_dim
states['gen_opt'] = gen_opt.state_dict()
states['dis_opt'] = dis_opt.state_dict()
states['reg_opt'] = reg_opt.state_dict()
states['theta1'] = args.theta1
states['theta2'] = args.theta2
states['theta3'] = args.theta3
torch.save(states, str(save_path))
logger.info(
f'epoch: {epoch+1:4}, g_loss: {g_avg_loss: .4}, d_loss: {d_avg_loss: .4}, \n'
f'r_loss: {r_avg_loss: .4}, '
f'g_score: {g_avg_score:.4}')
with result_path.open('w') as fout:
for s in loss_history:
fout.write(s)
logger.info('program finished')
def partOne(data):
valid = 0
for n in data:
l, u, k, p = processItem(n)
count = utils.count_characters_in_string(p, k)
if utils.int_in_range(count, l, u):
valid += 1
return valid
def partTwo(data):
valid = 0
for n in data:
p1, p2, k, p = processItem(n)
if (p[int(p1) - 1] is k) ^ (p[int(p2) - 1] is k):
valid += 1
return valid
if __name__ == "__main__":
# Load Data
data = utils.load_data("day2.txt")
# Do puzzle
print("---- Day 2 ----")
print("Part 1: " + str(partOne(data)))
print("Part 2: " + str(partTwo(data)))
def plot_power_sector(s=None, c=None, order=None):
data = load_data()
data = data[data.cost != 'none']
if s is None:
all_cost = sorted(list(set(data.cost)))
s = [
min(all_cost), all_cost[int((len(all_cost) - 1) / 2)],
max(all_cost)
]
if c is None:
all_cost = sorted(list(set(data.tax)))
c = [
min(all_cost), all_cost[int((len(all_cost) - 1) / 2)],
max(all_cost)
]
data = data[data.cost.isin(s)]
data = data[data.tax.isin(c)]
synonyms = {
'Solar': 'Renewable',
'Hydro': 'Renewable',
'Wind': 'Renewable',
'Biomass': 'Renewable',
'Import': 'Others',
'Oil': 'Others'
}
col_dic = {
'Coal|w/o CCS': '#000000',
'Coal|w/ CCS': '#918F88',
'Gas|w/o CCS': '#A3CFD6',
'Gas|w/ CCS': '#D3EAED',
'Renewable': '#4EB378',
'Nuclear': '#724ac1',
'Others': '#b2b2b2'
}
years = [2020, 2030, 2040, 2050]
activity = get_plot_data(data,
keyword='Secondary Energy|Electricity',
synonyms=synonyms,
col_dic=col_dic)
activity[years] = activity[years] * 8.760
plot_facet_grids(activity,
y_title='PPL Activity [TWh]',
figure_title='Power_Activity_TWh',
col_dic=col_dic,
y_max=880,
order=order)
capacity = get_plot_data(data,
keyword='Capacity|Electricity',
synonyms=synonyms,
col_dic=col_dic)
plot_facet_grids(capacity,
y_title='PPL Capacity [GW]',
figure_title='Power_Capacity_GW',
col_dic=col_dic,
y_max=280,
order=order)
parser = argparse.ArgumentParser()
parser.add_argument('--dataset', type=str, default='MUTAG')
parser.add_argument('--hidden', type=int, default=32)
parser.add_argument('--idx', type=int, default=1)
parser.add_argument('--batch_size', type=int, default=128)
parser.add_argument('--epochs', type=int, default=300)
parser.add_argument('--lr', type=float, default=0.01)
args = parser.parse_args()
np.random.seed(0)
torch.manual_seed(0)
torch.cuda.manual_seed_all(0)
# writer = SummaryWriter('runs/la_PROTEINS')
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
dataset = load_data(args.dataset)
index_train = []
index_test = []
with open(
osp.join(osp.dirname(osp.realpath(__file__)), 'datasets',
'%s' % args.dataset, '10fold_idx',
'train_idx-%d.txt' % args.idx), 'r') as f_train:
for line in f_train:
index_train.append(int(line.split('\n')[0]))
with open(
osp.join(osp.dirname(osp.realpath(__file__)), 'datasets',
'%s' % args.dataset, '10fold_idx',
'test_idx-%d.txt' % args.idx), 'r') as f_test:
for line in f_test:
index_test.append(int(line.split('\n')[0]))
import os
from S_and_R import main
from utils.utils import load_data
if __name__ == '__main__':
input_parameters = dict(number_of_candles=60,
minimum_window_size=5,
maximum_window_size=20,
tolerance=0.001)
path = os.path.join("data", "EURUSD1440.csv")
dataframe = load_data(path)
unique_resistances_supports_list, scaled_power_resistances_supports_list = main(dataframe, input_parameters)
def load_data_from_json(self):
title, description = load_data()
self.set_title(title)
self.set_description(description)
import pandas as pd
from sklearn.preprocessing import MinMaxScaler
from sklearn.metrics import roc_auc_score, log_loss
from sklearn.model_selection import StratifiedKFold
from keras.models import load_model
from keras.models import Sequential, Model
from keras.layers import Input, Dense, Dropout, Activation
from keras.layers.normalization import BatchNormalization
from keras.callbacks import EarlyStopping, ModelCheckpoint, Callback, CSVLogger
from keras.wrappers.scikit_learn import KerasClassifier
MODEL_NAME = 'keras_1'
combined = utils.load_data()
combined = utils.cat_transform(combined, 'onehot')
train, test = utils.recover_train_test_na(combined, fillna=True)
# Fillna for minmax scaler
train = train.replace(np.NaN, -1)
test = test.replace(np.NaN, -1)
X_train = train.drop('target', axis=1)
y_train = train.target
X_test = test
scaler = MinMaxScaler()
X_train = scaler.fit_transform(X_train)
X_test = scaler.transform(X_test)
args = get_args()
if (args.model_type != const.GMVAE and args.model_type != const.GMVAECNN):
print('Choose a valid model_type!')
sys.exit()
# odel checkpoint experiments/checkpoint/GMV_FREY_0001_8_2_8_64_2/-1029 ..
config, flags = get_config_and_flags(args)
# create the experiments dirs
utils.create_dirs(
[config.summary_dir, config.checkpoint_dir, config.results_dir])
utils.save_args(args, config.summary_dir)
''' ------------------------------------------------------------------------------
GET DATA
------------------------------------------------------------------------------ '''
print('\n Loading data...')
data_train, data_valid, data_test = utils.load_data(config.dataset_name)
''' ------------------------------------------------------------------------------
GET NETWORK PARAMS
------------------------------------------------------------------------------ '''
network_params = Bunch()
network_params.input_height = data_train.height
network_params.input_width = data_train.width
network_params.input_nchannels = data_train.num_channels
network_params.hidden_dim = config.hidden_dim
network_params.z_dim = config.z_dim
network_params.w_dim = config.w_dim
network_params.K = config.K_clusters
network_params.num_layers = config.num_layers
''' -----------------------------------------------------------------------------
COMPUTATION GRAPH (Build the model)
def run(args):
print("\nInput args:")
pprint(vars(args))
t0 = time()
te_size = verify_size(args.te_size)
datapath = Path(args.datapath).resolve()
# Hard split
# split_on = None if args.split_on is None else args.split_on.upper()
cv_method = args.cv_method
te_method = cv_method
# Specify ML task (regression or classification)
if cv_method == "strat":
mltask = "cls" # cast mltask to cls in case of stratification
else:
mltask = args.ml_task
# Target column name
trg_name = str(args.trg_name)
# assert args.trg_name in data.columns, f'The prediction target ({args.name}) \
# was not found in the dataset.'
# import ipdb; ipdb.set_trace()
# -----------------------------------------------
# Create outdir
# -----------------------------------------------
if args.gout is not None:
gout = Path(args.gout).resolve()
sufx = "none" if args.split_on is None else args.split_on
gout = gout / datapath.with_suffix(".splits")
if args.split_on is not None:
gout = gout / f"split_on_{sufx}"
else:
gout = gout / f"split_on_none"
else:
# Note! useful for drug response
sufx = "none" if args.split_on is None else args.split_on
gout = datapath.with_suffix(".splits")
outfigs = gout / "outfigs"
os.makedirs(gout, exist_ok=True)
os.makedirs(outfigs, exist_ok=True)
# -----------------------------------------------
# Create logger
# -----------------------------------------------
lg = Logger(gout / "data.splitter.log")
print_fn = get_print_func(lg.logger)
print_fn(f"File path: {fdir}")
print_fn(f"\n{pformat(vars(args))}")
dump_dict(vars(args), outpath=gout / "data.splitter.args.txt")
# -----------------------------------------------
# Load data
# -----------------------------------------------
print_fn("\nLoad master dataset.")
data = load_data(datapath)
print_fn("data.shape {}".format(data.shape))
# ydata = data[trg_name] if trg_name in data.columns else None
# if (cv_method == "strat") and (ydata is None):
# raise ValueError("Prediction target column must be available if splits need to be stratified.")
if (cv_method == "strat") and (trg_name not in data.columns):
raise ValueError(
"Prediction target column must be available if splits need to be stratified."
)
# if ydata is not None:
# plot_hist(ydata, title=f"{trg_name}", fit=None, bins=100,
# path=outfigs/f"{trg_name}_hist_all.png")
if trg_name in data.columns:
plot_hist(data[trg_name],
title=f"{trg_name}",
fit=None,
bins=100,
path=outfigs / f"{trg_name}_hist_all.png")
# -----------------------------------------------
# Generate splits (train/val/test)
# -----------------------------------------------
print_fn("\n{}".format("-" * 50))
print_fn("Split data into hold-out train/val/test")
print_fn("{}".format("-" * 50))
kwargs = {
"cv_method": cv_method,
"te_method": te_method,
"te_size": te_size,
"mltask": mltask,
"split_on": args.split_on
}
data_splitter(
data=data,
n_splits=args.n_splits,
gout=gout,
outfigs=outfigs,
# ydata = ydata,
target_name=trg_name,
print_fn=print_fn,
seed=seed,
**kwargs)
print_fn("Runtime: {:.1f} min".format((time() - t0) / 60))
print_fn("Done.")
lg.close_logger()
def gen_ml_data(
fpath,
common_samples,
# fea_type,
# drg_set,
dd_fea=None,
fps_fea=None,
img_fea=None,
ID='TITLE',
fea_sep='_',
score_name='reg',
n_samples=None,
n_top=None,
sampling=None,
q_cls=0.025,
binner=False,
bin_th=2.0,
baseline=False,
print_fn=print,
outdir=Path('out'),
outfigs=Path('outfigs')):
""" Generate a single set of ML data for the loaded target from fpath.
This func was specifically created to process the new LARGE DOE-MD datasets
with ZINC drugs that contains >6M molecules.
Args:
fpath: path to load docking scores file
common_samples : list of drug names that are commong to all features
types including dd_fea, fps_fea, and img_fea
dd_fea : df of Mordred descriptors
fps_fea : df pf ecfp2 fingerprints
img_fea : image data (TODO: this is not supported yet!)
fea_sep : separator between feature prefix string and feature name
score_name : rename the docking score col with score_name
n_samples : total number of samples in the final ml_df
n_top : keep this number of top-most dockers
sampling : specify the method to use when sampling samples from df
q_cls : quantile value to compute along the docking scores to generate
the 'cls' col
bin_th : threshold value of docking score to generate the 'binner' col
binner : add binner column
baseline : whether to compute ML baseline scores
Returns:
res : results summary
"""
print_fn(f'\nProcess {fpath.name} ...')
res = {}
trg_name = fpath.with_suffix('').name # note! depends on dock file names
res['target'] = trg_name
# Load docking
dock = load_data(fpath)
if dock.empty:
print_fn('Empty file')
return None
if (n_samples is not None) and (dock.shape[0] <= n_samples):
print_fn("n_samples is larger than len(dock), skip this receptor")
return res
# Pre-proc the dock file
## ID = 'TITLE'
scoring_func = 'Chemgauss4'
dock = proc_dock_score(dock,
ID=ID,
score_name=score_name,
scoring_func=scoring_func)
# Plot histogram of all (raw) scores
plot_hist_dock_scores(dock,
outfigs=outfigs,
subdir_name='all.raw',
trg_name=trg_name,
scoring_func=scoring_func)
# Convert and bound scores to >=0
dock[score_name] = abs(np.clip(dock[score_name], a_min=None, a_max=0))
print_fn('dock: {}'.format(dock.shape))
# Plot histogram of all (transformed) scores
plot_hist_dock_scores(dock,
outfigs=outfigs,
subdir_name='all.transformed',
trg_name=trg_name,
scoring_func=scoring_func)
# -----------------------------------------
# Sample a subset of scores
# -------------------------
# Extract samples that are common to all feature types
aa = dock[dock[ID].isin(common_samples)].reset_index(drop=True)
# Extract subset of samples
if (n_samples is not None) and (n_top is not None):
n_bot = n_samples - n_top
aa = aa.sort_values('reg', ascending=False).reset_index(drop=True)
df_top = aa[:n_top].reset_index(drop=True) # e.g. 100K
df_rest = aa[n_top:].reset_index(drop=True)
# if flatten:
# df_bot = flatten_dist(df=df_rest, n=n_bot, score_name=score_name)
# else:
# df_bot = df_rest.sample(n=n_bot, replace=False)
if sampling == 'flatten':
df_bot = flatten_dist(df=df_rest, n=n_bot, score_name=score_name)
elif sampling == 'random':
df_bot = df_rest.sample(n=n_bot, replace=False)
else:
raise ValueError("'sampling' arg must be specified.")
assert df_top.shape[1] == df_bot.shape[
1], 'Num cols must be the same when concat.'
aa = pd.concat([df_top, df_bot], axis=0).reset_index(drop=True)
# Plot histogram of sampled scores
outfigs_dir = outfigs / 'sampled.transformed'
os.makedirs(outfigs_dir, exist_ok=True)
fig, ax = plt.subplots()
ax.hist(df_top[score_name],
bins=100,
facecolor='r',
alpha=0.7,
label='Top 10K Docking Ligands')
ax.hist(df_bot[score_name],
bins=100,
facecolor='b',
alpha=0.7,
label='Other Ligands (balanced)')
ax.set_xlabel(f'Docking Score ({scoring_func})')
ax.set_ylabel('Count')
plt.grid(True)
plt.legend(loc='best', framealpha=0.5)
plt.title(f'sampled.transformed; Samples {n_samples}; n_top {n_top}')
plt.savefig(outfigs_dir / f'dock.dist.{trg_name}.png', dpi=150)
del df_top, df_bot, df_rest
elif (n_samples is not None):
# if flatten:
# aa = flatten_dist(df=aa, n=n_samples, score_name=score_name)
# else:
# aa = aa.sample(n=n_samples, replace=False)
if sampling == 'flatten':
aa = flatten_dist(df=aa, n=n_samples, score_name=score_name)
elif sampling == 'random':
aa = aa.sample(n=n_samples, replace=False)
else:
raise ValueError("'sampling' arg must be specified.")
plot_hist_dock_scores(dock,
outfigs=outfigs,
subdir_name='sampled.transformed',
trg_name=trg_name,
scoring_func=scoring_func)
dock = aa
del aa
# -----------------------------------------
# Create cls col
# --------------
# Find quantile value
if dock[score_name].min() >= 0: # if scores were transformed to >=0
q_cls = 1.0 - q_cls
cls_th = dock[score_name].quantile(q=q_cls)
res['cls_th'] = cls_th
print_fn('Quantile score (q_cls={:.3f}): {:.3f}'.format(q_cls, cls_th))
# Generate a classification target col
if dock[score_name].min() >= 0: # if scores were transformed to >=0
value = (dock[score_name] >= cls_th).astype(int)
else:
value = (dock[score_name] <= cls_th).astype(int)
dock.insert(loc=1, column='cls', value=value)
# print_fn('Ratio {:.2f}'.format( dd['dock_bin'].sum() / dd.shape[0] ))
# Plot
hist, bin_edges = np.histogram(dock[score_name], bins=100)
x = np.ones((10, )) * cls_th
y = np.linspace(0, hist.max(), len(x))
fig, ax = plt.subplots()
plt.hist(dock[score_name],
bins=200,
density=False,
facecolor='b',
alpha=0.7)
plt.title(f'Scores clipped to 0: {trg_name}')
plt.xlabel(f'Docking Score ({scoring_func})')
plt.ylabel('Count')
plt.plot(x, y, 'm--', alpha=0.7, label=f'{q_cls}-th quantile')
plt.grid(True)
plt.tight_layout()
plt.savefig(outfigs / f'dock.dist.cls.{trg_name}.png')
# -----------------------------------------
# Save dock scores
cols = ['Inchi-key', 'SMILES', 'TITLE', 'reg', 'cls']
dock = dock[[c for c in cols if c in dock.columns]]
trg_outdir = outdir / f'DIR.ml.{trg_name}'
outpath = trg_outdir / f'docks.df.{trg_name}.csv'
os.makedirs(trg_outdir, exist_ok=True)
dock.to_csv(outpath, index=False)
# Add binner (note! may not be necessary since we get good dock scores)
# if binner:
# dock = add_binner(dock, score_name=score_name, bin_th=bin_th)
# Merge only on TITLE (when including also SMILES, there is a mismatch on
# certain samples; maybe smiles that come with features are canonicalied)
merger = ID
def merge_dock_and_fea(dock,
fea_df,
fea_prfx,
fea_sep,
merger='TITLE',
fea_name=None,
baseline=False):
""" ... """
# drug_names = set(common_samples).intersection(set(dock[ID].values))
ml_df = pd.merge(dock, fea_df, how='inner',
on=merger).reset_index(drop=True)
del fea_df
# bb = fea_df[ fea_df[merger].isin(dock[merger].tolist()) ].reset_index(drop=True)
# xdata = extract_subset_fea(bb, fea_list=[fea_prfx], fea_sep=fea_sep)
# bb = pd.concat([bb[merger], xdata], axis=1) # keep only the merger meta col from fea_df
# xdata = extract_subset_fea(fea_df, fea_list=[fea_prfx], fea_sep=fea_sep)
# fea_df = pd.concat([fea_df[merger], xdata], axis=1) # keep only the merger meta col from fea_df
# ml_df = pd.merge(dock, fea_df, how='inner', on=merger).reset_index(drop=True)
# del fea_df, xdata
# Re-org cols
fea_cols = extract_subset_fea_col_names(ml_df,
fea_list=[fea_prfx],
fea_sep=fea_sep)
meta_cols = ['Inchi-key', 'SMILES', 'TITLE', 'CAT', 'reg', 'cls']
cols = meta_cols + fea_cols
# ml_df = ml_df[cols]
ml_df = ml_df[[c for c in cols if c in ml_df.columns]]
print_fn('{}: {}'.format(fea_name, ml_df.shape))
# Save
outpath = trg_outdir / f'ml.{trg_name}.{fea_name}'
ml_df.to_parquet(str(outpath) + '.parquet')
# Compute baseline if specified
if baseline:
te_scr = trn_baseline(ml_df, fea_list=[fea_prfx], fea_sep=fea_sep)
res[f'{fea_prfx}_r2'] = te_scr['r2']
res[f'{fea_prfx}_mae'] = te_scr['median_absolute_error']
del te_scr
del ml_df
if dd_fea is not None:
merge_dock_and_fea(dock,
fea_df=dd_fea,
fea_prfx='dd',
fea_sep=fea_sep,
merger=ID,
fea_name='descriptors',
baseline=baseline)
if fps_fea is not None:
merge_dock_and_fea(dock,
fea_df=fps_fea,
fea_prfx='ecfp2',
fea_sep=fea_sep,
merger=ID,
fea_name='ecfp2',
baseline=baseline)
if img_fea is not None:
pass
# if n_samples is not None:
# assert n_samples == ml_df.shape[0], 'Final ml_df size must match n_samples {}'.format(fpath)
return res
from time import time
from optparse import OptionParser
from multiprocessing import Process
from mne import Epochs, find_events
from time import time, strftime, gmtime
import os
from stimulus_presentation import auditory_p300
from utils import utils
from collections import OrderedDict
import numpy as np
from pandas import DataFrame
from psychopy import visual, core, event, sound, monitors
from pylsl import StreamInfo, StreamOutlet, resolve_byprop, StreamInlet
raw = utils.load_data('auditory/P300',
sfreq=256.,
subject_nb=subject,
session_nb=session)
raw.plot_psd()
raw.filter(1, 30, method='iir')
events = find_events(raw)
event_id = {'Non-Target': 1, 'Target': 2}
epochs = Epochs(raw,
events=events,
event_id=event_id,
tmin=-0.1,
tmax=0.8,
baseline=None,
def train(network='rnn'):
word2id, id2word = load_data(TOKEN_DATA)
tag2id, id2tag = load_data(TAG_DATA)
x_train, y_train, seq_lens, _, _ = generate_data(TRAIN_DATA,
word2id,
tag2id,
max_len=hp.max_len)
x_dev, y_dev, dev_seq_lens, _, source_tag = generate_data(
DEV_DATA, word2id, tag2id, max_len=hp.max_len)
vocab_size = len(word2id)
num_tags = len(tag2id)
if network == "transformer":
model = TransformerCRFModel(vocab_size, num_tags, is_training=True)
elif network == 'rnn':
model = BiRnnCRF(vocab_size, num_tags)
elif network == 'cnn':
model = CnnCRF(vocab_size, num_tags)
elif network == 'match-pyramid':
model = CnnCRF(vocab_size, num_tags)
else:
return
sv = tf.train.Supervisor(graph=model.graph,
logdir=logdir,
save_model_secs=0)
with sv.managed_session() as sess:
for epoch in range(1, hp.num_epochs + 1):
if sv.should_stop():
break
train_loss = []
for x_batch, y_batch, len_batch in batch_data(
x_train, y_train, seq_lens, hp.batch_size):
feed_dict = {
model.x: x_batch,
model.y: y_batch,
model.seq_lens: len_batch
}
loss, _ = sess.run([model.loss, model.train_op],
feed_dict=feed_dict)
train_loss.append(loss)
dev_loss = []
predict_lists = []
for x_batch, y_batch, len_batch in batch_data(
x_dev, y_dev, dev_seq_lens, hp.batch_size):
feed_dict = {
model.x: x_batch,
model.y: y_batch,
model.seq_lens: len_batch
}
loss, logits = sess.run([model.loss, model.logits], feed_dict)
dev_loss.append(loss)
transition = model.transition.eval(session=sess)
pre_seq = model.predict(logits, transition, len_batch)
pre_label = recover_label(pre_seq, len_batch, id2tag)
predict_lists.extend(pre_label)
train_loss_v = np.round(float(np.mean(train_loss)), 4)
dev_loss_v = np.round(float(np.mean(dev_loss)), 4)
print('****************************************************')
acc, p, r, f = get_ner_fmeasure(source_tag, predict_lists)
print('epoch:\t{}\ttrain loss:\t{}\tdev loss:\t{}'.format(
epoch, train_loss_v, dev_loss_v))
print('acc:\t{}\tp:\t{}\tr:\t{}\tf:\t{}'.format(acc, p, r, f))
print('****************************************************\n\n')
def main(
data_dir: str = '/project/cq-training-1/project2/teams/team12/data/',
model_name: str = 'seq2seqgru',
epochs: int = 20,
optimizer: str = 'adam',
lr: float = 1e-3,
batch_size: int = 32,
vocab_size: int = None, # If None all tokens of will be in vocab
seq_len:
int = None, # If None the seq len is dynamic (might not work with all models)
seed: bool = True,
model_config: dict = None,
embedding: str = None,
embedding_dim: int = 128,
back_translation_model: str = 'saved_model/<model_folder_name>',
back_translation: bool = False,
back_translation_ratio: float = 1.0,
fr_to_en: bool = False):
# Call to remove tensorflow warning about casting float64 to float32
tf.keras.backend.set_floatx('float32')
# Set random seed
if seed:
tf.random.set_seed(SEED)
np.random.seed(SEED)
# Data paths
path_en = os.path.join(data_dir, 'train.lang1')
path_fr = os.path.join(data_dir, 'train.lang2')
path_unaligned_en = os.path.join(data_dir, 'unaligned-tok.en')
path_unaligned_fr = os.path.join(data_dir, 'unaligned-tok.fr')
if fr_to_en: # Switch paths
tmp = path_en
path_en = path_fr
path_fr = tmp
# Create vocabs
logger.info('Creating vocab...')
word2idx_en, idx2word_en = utils.create_vocab(path_en, vocab_size)
word2idx_fr, idx2word_fr = utils.create_vocab(path_fr, vocab_size)
logger.info(
f'Size of english vocab : {len(word2idx_en)}, size of french vocab : {len(word2idx_fr)}'
)
# Back translation
prediction_file = None
if back_translation:
prediction_file = os.path.join(utils.SHARED_PATH,
'translated_unaligned.en')
if os.path.exists(prediction_file):
logger.info(
f'Using translation from {prediction_file} for back-translation.'
)
else:
logger.info(
f'Translating {path_unaligned_fr} for back-translation...')
# Load data
data = utils.load_data(path_unaligned_fr, word2idx_fr)
dataset = tf.data.Dataset.from_generator(
lambda: [ex for ex in data],
tf.int64,
output_shapes=tf.TensorShape([None])).padded_batch(
128, padded_shapes=[None])
# Load model
model_config = {
'num_layers': 2,
'd_model': 128,
'dff': 512,
'num_heads': 8
}
model = Transformer(model_config, len(word2idx_fr), word2idx_en)
model.load_weights(os.path.join(back_translation_model, "model"))
# Write prediction to file
with open(prediction_file, 'w') as f:
print('opening file and writing predictions...')
for batch in tqdm(dataset,
desc='Translating...',
total=len(data) // 128 + 1):
preds = model({
'inputs': batch,
'labels': tf.zeros_like(batch)
})
for pred in preds:
sentence = utils.generate_sentence(
np.argmax(pred.numpy(), axis=1).astype('int'),
idx2word_en)
f.writelines([sentence, '\n'])
# Load datasets
logger.info('Loading datasets...')
train_dataset, valid_dataset, nb_train_ex, nb_valid_ex = utils.load_training_data(
path_en,
path_fr,
word2idx_en,
word2idx_fr,
seq_len,
batch_size,
en_back_translated_path=prediction_file,
fr_unaligned_path=path_unaligned_fr,
back_translation_ratio=back_translation_ratio)
logger.info(
f'Number of training examples : {nb_train_ex}, number of valid examples : {nb_valid_ex}'
)
# Load embeddings
embedding_matrix = None
if embedding:
logger.info(f'Loading embedding {embedding} ...')
if embedding == 'fasttext':
embedding_matrix = utils.create_fasttext_embedding_matrix(
path_unaligned_en, word2idx_en, embedding_dim)
elif embedding == 'word2vec':
raise Exception(f'Embedding "{embedding}" not implemented yet')
elif embedding == 'glove':
raise Exception(f'Embedding "{embedding}" not implemented yet')
else:
raise Exception(f'Embedding "{embedding}" not recognized.')
# Create model
if model_name == 'gru':
model = baselines.GRU(len(word2idx_fr), batch_size)
elif model_name == 'seq2seqgru':
if model_config is None:
model_config = {
'embedding_dim': 256,
'encoder_units': 512,
'decoder_units': 512,
'n_layers': 1
}
model = Seq2SeqGRU(len(word2idx_en),
word2idx_fr,
batch_size,
model_config,
embedding_matrix=embedding_matrix)
elif model_name == 'transformer':
if model_config is None:
model_config = {
'num_layers': 2,
'd_model': 128,
'dff': 512,
'num_heads': 8
}
model = Transformer(model_config,
len(word2idx_en),
word2idx_fr,
embedding_matrix=embedding_matrix)
else:
raise Exception(f'Model "{model}" not recognized.')
# Optimizer
if optimizer == 'adam':
if model_name == 'transformer': # Use adam according to transformer paper
optimizer = tf.keras.optimizers.Adam(utils.CustomSchedule(
model_config['d_model']),
beta_1=0.9,
beta_2=0.98,
epsilon=1e-9)
logger.info(
'Using custom scheduler for learning rate, --lr argument ignored.'
)
else:
optimizer = tf.keras.optimizers.Adam(lr)
elif optimizer == 'sgd':
optimizer = tf.keras.optimizers.SGD(lr)
else:
raise Exception(f'Optimizer "{optimizer}" not recognized.')
# Training loop
logger.info(f'Training with model {model.get_name()} ...')
metrics = {
'train_accuracy': [],
'valid_accuracy': [],
'train_loss': [],
'valid_loss': [],
'train_bleu': [],
'valid_bleu': []
}
model_path = model.get_name() + f'_fr_to_en_{fr_to_en}_embedding_{embedding}_embedding_dim_{embedding_dim}'\
f'_back_translation_{back_translation}_ratio_{back_translation_ratio}'
best_valid_bleu = 0
for epoch in range(epochs):
train_epoch(model, train_dataset, optimizer,
np.ceil(nb_train_ex / batch_size), idx2word_fr)
test_epoch(model, valid_dataset, np.ceil(nb_valid_ex / batch_size),
idx2word_fr, idx2word_en)
train_accuracy = train_accuracy_metric.result().numpy()
valid_accuracy = valid_accuracy_metric.result().numpy()
train_loss = train_loss_metric.result().numpy()
valid_loss = valid_loss_metric.result().numpy()
train_bleu = train_bleu_metric.result()
valid_bleu = valid_bleu_metric.result()
if valid_bleu > best_valid_bleu:
best_valid_bleu = valid_bleu
utils.save_model(model, model_path)
# Logs
logger.info(f'Epoch {epoch}\n'\
f' Train BLEU : {train_bleu:.4f} - Valid BLEU : {valid_bleu:.4f}\n'\
f' Train Accuracy : {train_accuracy:.4f} - Valid Accuracy : {valid_accuracy:.4f}\n'\
f' Train Loss : {train_loss:.4f} - Valid Loss : {valid_loss:.4f}')
metrics['train_accuracy'].append(train_accuracy)
metrics['valid_accuracy'].append(valid_accuracy)
metrics['train_loss'].append(train_loss)
metrics['valid_loss'].append(valid_loss)
metrics['train_bleu'].append(train_bleu)
metrics['valid_bleu'].append(valid_bleu)
# If using back translation, sample new generated examples for next epoch
if back_translation:
train_dataset, _, _, _ = utils.load_training_data(
path_en,
path_fr,
word2idx_en,
word2idx_fr,
seq_len,
batch_size,
en_back_translated_path=prediction_file,
fr_unaligned_path=path_unaligned_fr,
back_translation_ratio=back_translation_ratio)
# If training with embeddings, unfreeze embedding layer at 50th epoch
if epoch == 48 and embedding and model_name == 'transformer':
model.unfreeze_embedding_layer()
# save metrics
utils.save_metrics(metrics, model_path)
# Plot accuracy
plots.plot_accuracy(metrics['train_accuracy'], metrics['valid_accuracy'])
def ensemble_methods_regressor_forest_dataset():
data = utils.load_data('forestfires.csv')
new_data = utils.convert_data_to_numeric(data, [2, 3])
feature_vector = new_data[:, 0:-1]
targets = new_data[:, -1]
# Data normalization
data_features_normalized = normalization.z_score_normalization(feature_vector)
data_features_train, data_features_test, data_targets_train, data_targets_test = \
train_test_split(data_features_normalized,
targets,
test_size=0.25)
# Model declaration
"""
Parameters to select:
n_estimators: The number of base estimators in the ensemble.
Values: Random Forest and Bagging. Default 10
AdaBoost. Default: 50
###Only for Bagging and Boosting:###
base_estimator: Base algorithm of the ensemble. Default: DecisionTree
###Only for Random Forest:###
criterion: "entropy" or "gini": default: gini
max_depth: maximum depth of tree, default: None
"""
names = ["Bagging Regressor", "AdaBoost Regressor", "Random Forest Regressor"]
models = [
BaggingRegressor(
base_estimator=tree.DecisionTreeRegressor(
criterion='mse',
max_depth=10)
),
AdaBoostRegressor(
base_estimator=tree.DecisionTreeRegressor(
criterion='mse',
max_depth=10)
),
RandomForestRegressor(
criterion='mse',
max_depth=10
)
]
for name, em_reg in zip(names, models):
logger.info("###################---" + name + "---###################")
em_reg.fit(data_features_train, data_targets_train)
# Model evaluation
test_data_predicted = em_reg.predict(data_features_test)
error = metrics.mean_absolute_error(data_targets_test, test_data_predicted)
logger.debug('Total Error: %s', error)
def incremental_load():
kwargs = request.args.to_dict()
return load_data("incremental", **kwargs)
from utils import utils
from sklearn.tree import DecisionTreeClassifier
from sklearn.ensemble import RandomForestClassifier
from sklearn.naive_bayes import MultinomialNB, GaussianNB
import numpy as np
from sklearn import svm
from sklearn.metrics import roc_curve, auc, accuracy_score, recall_score, precision_score
from sklearn.decomposition import TruncatedSVD
features, labels, valid_adj, valid_graph_labels, train_adj, train_graph_labels = utils.load_data(
cuda=False)
svd = TruncatedSVD(300)
tr = [[v for row in graph for v in row] for graph in train_adj]
#tr = svd.fit_transform(tr)
vr = [[v for row in graph for v in row] for graph in valid_adj]
#vr = svd.fit_transform(vr)
tl = np.argmax(train_graph_labels, axis=1)
vl = np.argmax(valid_graph_labels, axis=1)
print('Decision Tree:')
cl = DecisionTreeClassifier()
cl = cl.fit(tr, tl)
pl = cl.predict(vr)
print(accuracy_score(vl, pl), precision_score(vl, pl), recall_score(vl, pl))
print('Random Forest:')
cl = RandomForestClassifier(oob_score=True, random_state=10)
cl = cl.fit(tr, tl)
pl = cl.predict(vr)
print(accuracy_score(vl, pl), precision_score(vl, pl), recall_score(vl, pl))
])
device = torch.device(
'cuda', args.cuda) if torch.cuda.is_available() else torch.device('cpu')
# Reset random state for reproducibility
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
if args.multinet:
args.sym = 0
args.embedding = 0
# Load dataset
graph, adj, features, labels, idx_train, idx_val, idx_test = load_data(
path=args.data_dir, percent=args.train_percent, sym=args.sym)
embedding = node2vec(graph, args.data_dir, args.sym)
if args.multinet:
embedding1 = node2vec(graph, args.data_dir, 1)
del graph
gc.collect()
if args.pca > 0:
aff_features = PCA(args.pca, whiten=True).fit_transform(features.numpy())
features = torch.FloatTensor(aff_features)
if args.embedding == 0:
del features
features = embedding.to(device)
msg = 'Uses only the embedding features (extracting from Node2Vec model).'
if args.multinet:
features1 = embedding1.to(device)
msg = '`multinet` specified, load embeddings for both adj and adj.T.'