def check_train(trainfile, classifier_file):
"""Create an instance of 'Classifier', train it with the dataset in 'trainfile'
and save the trained model to 'classifier_file" on disk."""
# Load training data
texts, labels = dataloader.load(trainfile)
# Create a Classifier instance and train it on data
classifier = Classifier()
classifier.train(texts, labels)
# Save classifier
save_classifier(classifier, classifier_file)
print("Done.\n--------------------------------------\n")
def main():
samples = []
load(samples)
space_groups = {}
for i in samples:
grp = i.space_group
if (grp in space_groups): space_groups[grp] += 1
else: space_groups[grp] = 1
print(len(space_groups))
X = []
y = []
for i in samples:
X_temp = i.image_stack
X.append(X_temp)
y_temp = i.space_group
y.append(y_temp)
X = np.array(X)
y = np.array(y)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.15, random_state=42)
model = keras.Sequential([
keras.layers.Flatten(input_shape=(3, 512, 512)),
keras.layers.Dense(1000, activation=tf.nn.relu),
keras.layers.Dense(231, activation=tf.nn.softmax)
])
model.compile(optimizer='adam',
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
model.fit(X_train, y_train, epochs=3)
test_loss, test_acc = model.evaluate(X_test, y_test)
print('Test accuracy:', test_acc)
def __init__(self, params, mode, device):
assert mode in ['train', 'test', 'valid']
np.random.seed(params['seed'])
#self._const = 0 # constrain counter
self._device = device
self._maxlen = params['maxlen']
self._benchmark = params['benchmark']
self._batchsize = params['batch_size']
# builds ad hoc dataset, the number of violated_ constraints can be tuned
(X, y) = dataloader.load(self._benchmark, maxlen=self._maxlen)
self._len, self._nfeatures = X.shape
indices = self._get_indexes(params, self._len, mode, params['seed'])
X, y = X[indices], np.reshape(y[indices], (len(indices), 1))
self._dataset = (Ten(X), Ten(y))
def eval(classifier_file, testfile):
"""Loads a classifier model from disk, and evaluate it with test data from
'testfile' file"""
try:
# load classifier
classifier = load_classifier(classifier_file)
# load test data
texts, labels = dataloader.load(testfile)
# classify input texts with the classifier (let the classifier predict classes for the input texts)
predictions = classifier.predict(texts)
# compute accuracy score by comparing the model predictions with the true labels
acc = accuracy_score(labels, predictions)
print("\nAccuracy: %.4f\n" % acc)
except FileNotFoundError:
print(" Error: cannot find model file '%s'" % classifier_file)
print("Done.\n--------------------------------------\n")
def plot_incidents_per_station(san_diego, years, with_fire_station=False):
'''
Plot number of incidents per station over years
Parameters
----------
san_diego : TYPE
DESCRIPTION.
years : TYPE
DESCRIPTION.
Returns
-------
None.
'''
assert isinstance(san_diego, gpd.GeoDataFrame)
assert hasattr(years, '__iter__')
assert all(2007<=i<=2020 and isinstance(i, int) for i in years)
assert isinstance(with_fire_station, bool)
# information of incidents over years
incidents_df = dataloader.load(years)
# get dataframe with incidents per stations
incidents_per_station_df = calculate_incidents_per_station(incidents_df)
# merge incidents_per_station with san diego map
years_san_diego = san_diego.merge(incidents_per_station_df, on = 'address_zip', how = 'inner')
# plot the heatmap
if len(years) == 1:
title = 'Number of incidents per station in %d'%(years[0]);
else:
title = 'Number of incidents per station over %d-%d'%(years[0], years[-1])
ax = years_san_diego.plot(column = 'incidents_per_station', scheme = 'quantiles', legend=True, cmap = 'OrRd', figsize=(12,20))
# combine with fire station
if with_fire_station:
fire_df = pd.read_csv('./data/fire_station_position.csv')
gdf = gpd.GeoDataFrame(fire_df, geometry=gpd.points_from_xy(fire_df.longitude, fire_df.latitude))
gdf.plot(ax = ax, color = 'blue')
plt.title(title)
#plt.savefig('incidents_per_station_%d-%d'%(years[0], years[-1])+'.png')
plt.show()
def plot_std_incidents(years):
'''
Parameters
----------
years : TYPE
DESCRIPTION.
Returns
-------
None.
'''
assert hasattr(years, '__iter__')
assert all(2007<=i<=2020 and isinstance(i, int) for i in years)
std_list = []
# information of incidents over years
incidents_df = dataloader.load(years)
# get dataframe with incidents per stations
incidents_per_station_df = calculate_incidents_per_station(incidents_df)
# remove samples with no stations
incidents_per_station_df = incidents_per_station_df[incidents_per_station_df['incidents_per_station']!=0]
std_list.append(incidents_per_station_df['incidents_per_station'].std())
for i in range(len(incidents_per_station_df)):
incidents_per_station_df.loc[i, 'num_stations'] += 1
incidents_per_station_df['incidents_per_station'] = incidents_per_station_df['num_incidents']/incidents_per_station_df['num_stations']
incidents_per_station_df.replace(np.inf, 0, inplace=True)
std_list.append(incidents_per_station_df['incidents_per_station'].std())
title = 'std of number of incidents per station %d-%d'%(years[0],years[-1])
plt.figure()
plt.plot(std_list)
plt.xlabel('number of added fire stations')
plt.ylabel('std of number of incidents per station')
plt.title(title)
#plt.savefig('std_incidents_per_station_%d-%d'%(years[0],years[-1]) + '.png')
plt.show()
def plot_fire_heatmap(san_diego, years, with_fire_station=False):
'''
Plot fire stations with number of incidents over years
Parameters
----------
san_diego : gpd.GeoDataFrame
map of san diego
years : iterable of int
recorded duration
'''
assert isinstance(san_diego, gpd.GeoDataFrame)
assert hasattr(years, '__iter__')
assert all(2007<=i<=2020 and isinstance(i, int) for i in years)
assert isinstance(with_fire_station, bool)
# incidents over years
years_df = dataloader.load(years)
years_df['address_zip'] = years_df['address_zip'].fillna(0.0).astype(int).astype(str)
years_incidents = pd.pivot_table(years_df, values='incident_number', index=['address_zip'], columns=[], aggfunc=np.ma.count, fill_value=0)
years_san_diego = san_diego.merge(years_incidents, on = 'address_zip', how = 'inner')
years_san_diego['incident_number'] = years_san_diego['incident_number'].fillna(0.0).astype(int)
# plot san diego map with incidents
if len(years) == 1:
title = 'Number of incidents in %d'%(years[0])
else:
title = 'Incidents over %d-%d'%(years[0],years[-1])
ax = years_san_diego.plot(column = 'incident_number', scheme = 'quantiles', legend=True, cmap = 'OrRd', figsize=(12,20))
# combine with fire station
if with_fire_station:
fire_df = pd.read_csv('./data/fire_station_position.csv')
gdf = gpd.GeoDataFrame(fire_df, geometry=gpd.points_from_xy(fire_df.longitude, fire_df.latitude))
gdf.plot(ax = ax, color = 'blue')
plt.title(title)
plt.axis('off')
#plt.savefig('Incidents_over_%d-%d'%(years[0],years[-1])+'.png')
plt.show()
def main():
sc = SparkContext('local[15]', 'haha')
d = load(sc)
data_train_lp, data_dev_p, label_dev_gt, test_p = d['train_tfidf_lp'], d[
'dev_tfidf'], d['dev_gt'], d['test_tfidf']
data_train_p, label_train_gt = d['train_tfidf'], d['train_gt']
# print("count =", data_train_lp.take(10))
sample_train = data_train_lp
print("sample in total: ", sample_train.count())
print("___________train______________")
sys.stdout.flush()
lg = LogisticRegressionWithSGD.train(sample_train, step=0.005)
print("___________trained____________")
sys.stdout.flush()
# lg.save(sc, 'logistic.model')
result_dev = lg.predict(data_dev_p).map(int)
result_train = lg.predict(data_train_p).map(int)
print("train info:")
valid(result_train, label_train_gt)
print("dev info:")
valid(result_dev, label_dev_gt)
dump(lg.predict(test_p).map(int).collect())
except:
pass
os.mkdir(dirname)
def visualize(discretizer, dirname):
reset_dir(dirname)
discretizer.visualize("%s/full.png" % dirname)
for i in range( len(norm_sensor_states[0]) ):
discretizer.visualize("%s/%d.png" % (dirname, i), (i, i))
if __name__ == '__main__':
output_dir = 'discretizer_visualizations'
reset_dir(output_dir)
sensor_states, behaviors = dataloader.load('test.out')
norm_sensor_states = map(normalization.normalize, sensor_states)
for data_length in [300, 1000, 10000]:
data = norm_sensor_states[:data_length]
for i in range(2, 26):
dirname = '%s/kmeans_%d_%d' % (output_dir, i, data_length)
discretizer = discretization.KMeansDiscretizer(i)
discretizer.train(data)
visualize(discretizer, dirname)
for i in range(2, 6):
dirname = '%s/som_%dx%d_%d' % (output_dir, i, i, data_length)
from model import dcnn
from time import time
batch_size = 128
num_classes = 299
epochs = 1000
size = (81, 78)
try:
x_train.shape
y_train.shape
x_test.shape
y_test.shape
except NameError:
now = time()
(x_train, y_train), (x_test, y_test) = load(size=size,
_writer_num=num_classes)
print('data loaded')
print('x_train shape:', x_train.shape)
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')
model = dcnn(input_shape=(size[0], size[1], 1), num_classes=num_classes)
save_path = 'model/model04052059.h5'
log_path = './logs/train_logs'
earlystopping = EarlyStopping(monitor='val_loss', patience=5, verbose=0)
modelcheckpoint = ModelCheckpoint(save_path,
monitor='val_loss',
save_best_only=True,
import modshogun as sg
import dataloader as loader
# load data
feature_matrix = loader.load('mnist.mat')
# create features instance
features = sg.RealFeatures(feature_matrix)
# create Linear Local Tangent Space Alignment converter instance
converter = sg.LinearLocalTangentSpaceAlignment()
# set target dimensionality
converter.set_target_dim(2)
# set number of neighbors
converter.set_k(10)
# set number of threads
converter.parallel.set_num_threads(2)
# set nullspace shift (optional)
converter.set_nullspace_shift(-1e-6)
# compute embedding with Linear Local Tangent Space Alignment method
embedding = converter.embed(features)
parser.add_argument("--lamb",
type=float,
default=0.0,
help="Regularization constant.")
parser.add_argument("--cross",
action="store_true",
help="Do cross validation or not.")
parser.add_argument("--patience",
type=int,
default=5,
help="Patience to stop.")
args = parser.parse_args()
train = {}
test = {}
train["feature"], train["label"] = load(os.path.join(args.data, "a9a"))
test["feature"], test["label"] = load(os.path.join(args.data, "a9a.t"))
# add for w_0
ones = np.ones([train["feature"].shape[0], 1])
train["feature"] = np.append(train["feature"], ones, axis=1)
ones = np.ones([test["feature"].shape[0], 1])
test["feature"] = np.append(test["feature"], ones, axis=1)
if not args.cross:
w, acc_list, ll_list = irls(train, test, args.lamb, args.patience)
print("Final step: %d" % len(acc_list))
print("Final training acc: %f" %
evaluate(train["feature"], train["label"], w))
print("Final testing acc: %f" % acc_list[-1])
print("Final L2 norm of w: %f" % np.linalg.norm(w))
else:
def main():
parser = argparse.ArgumentParser(
description='Alkane property fitting demo')
parser.add_argument('-i', '--input', type=str, help='Data')
parser.add_argument('-f', '--fp', type=str, help='Fingerprints')
parser.add_argument('-o',
'--output',
default='out',
type=str,
help='Output directory')
parser.add_argument('-t',
'--target',
default='raw_density',
type=str,
help='Fitting target')
parser.add_argument('-p',
'--part',
default='',
type=str,
help='Partition cache file')
parser.add_argument('-l',
'--layer',
default='16,16',
type=str,
help='Size of hidden layers')
parser.add_argument('--visual',
default=1,
type=int,
help='Visualzation data')
parser.add_argument('--gpu', default=1, type=int, help='Using gpu')
parser.add_argument('--epoch',
default="200",
type=str,
help='Number of epochs')
parser.add_argument('--step',
default=500,
type=int,
help='Number of steps trained for each batch')
parser.add_argument('--batch',
default=int(1e9),
type=int,
help='Batch size')
parser.add_argument('--lr',
default="0.005",
type=str,
help='Initial learning rate')
parser.add_argument('--l2', default=0.000, type=float, help='L2 Penalty')
parser.add_argument(
'--check',
default=10,
type=int,
help='Number of epoch that do convergence check. Set 0 to disable.')
parser.add_argument('--minstop',
default=0.2,
type=float,
help='Minimum fraction of step to stop')
parser.add_argument('--maxconv',
default=2,
type=int,
help='Times of true convergence that makes a stop')
parser.add_argument('--featrm',
default='',
type=str,
help='Remove features')
parser.add_argument('--optim', default='rms', type=str, help='optimizer')
parser.add_argument('--continuation',
default=False,
type=bool,
help='continue training')
parser.add_argument('--plotsize',
default=500,
type=int,
help='plotting size')
parser.add_argument('--seed',
default=233,
type=int,
help='random select samples for plotting')
opt = parser.parse_args()
if opt.layer != "":
layers = list(map(int, opt.layer.split(',')))
else:
layers = []
opt_lr = list(map(float, opt.lr.split(',')))
opt_epochs = list(map(int, opt.epoch.split(',')))
if not os.path.exists(opt.output):
os.mkdir(opt.output)
logger = logging.getLogger('train')
logger.setLevel(logging.INFO)
flog = logging.FileHandler(opt.output + '/log.txt', mode='w')
flog.setLevel(logging.INFO)
formatter = logging.Formatter(
fmt='[%(asctime)s] (%(levelname)s) %(message)s',
datefmt='%Y-%d-%m %H:%M:%S')
flog.setFormatter(formatter)
clog = logging.StreamHandler()
clog.setFormatter(formatter)
logger.addHandler(flog)
logger.addHandler(clog)
if opt.featrm == 'auto':
logger.info('Automatically remove features')
featrm = [14, 15, 17, 18, 19, 20, 21, 22]
elif opt.featrm == '':
featrm = []
else:
featrm = list(map(int, opt.featrm.split(',')))
logger.info('Remove Feature: %s' % featrm)
logger.info('Reading data...')
datax, datay, data_names = dataloader.load(filename=opt.input,
target=opt.target,
fps=opt.fp.split(','),
featrm=featrm)
logger.info('Selecting data...')
selector = preprocessing.Selector(datax, datay, data_names)
if opt.part:
logger.info('Loading partition file %s' % opt.part)
selector.load(opt.part)
else:
logger.warning(
"Partition file not found. Using auto-partition instead.")
selector.partition(0.8, 0.1)
selector.save(opt.output + '/part.txt')
trainx, trainy, trainname = selector.training_set()
validx, validy, validname = selector.validation_set()
scaler = preprocessing.Scaler()
scaler.fit(trainx)
# scaler.save(opt.output + '/scale.txt')
normed_trainx = scaler.transform(trainx)
normed_validx = scaler.transform(validx)
model = fitting.TorchMLPRegressor(len(trainx[0]),
len(trainy[0]),
layers,
batch_size=opt.batch,
batch_step=opt.step,
is_gpu=False,
args_opt={
'optimizer': torch.optim.Adam,
'lr': opt.lr,
'weight_decay': opt.l2
})
model.load(opt.output + '/model.pt')
size = 500
data_smiles = [i.split('\t')[0] for i in validname]
data_T = [i.split('\t')[1] for i in validname]
data_P = [i.split('\t')[2] for i in validname]
predy = model.predict_batch(torch.Tensor(normed_validx))
error_y = abs(predy - validy).reshape(-1)
draw(validy[:size], predy[:size], data_smiles[:size], data_T[:size],
data_P[:size])
def main_worker(gpu, args):
""" 模型训练、测试、转JIT、蒸馏文件制作
:param gpu: 运行的gpu id
:param args: 运行超参
"""
args.gpu = gpu
utils.generate_logger(
f"{datetime.datetime.now().strftime('%Y%m%d%H%M%S')}-{gpu}.log")
logging.info(f'args: {args}')
# 可复现性
if args.seed is not None:
random.seed(args.seed)
torch.manual_seed(args.seed)
cudnn.deterministic = True
logging.warning('You have chosen to seed training. '
'This will turn on the CUDNN deterministic setting, '
'which can slow down your training considerably! '
'You may see unexpected behavior when restarting '
'from checkpoints.')
if args.cuda:
logging.info(f"Use GPU: {args.gpu} ~")
if args.distributed:
args.rank = args.rank * args.gpus + gpu
dist.init_process_group(backend='nccl',
init_method=args.init_method,
world_size=args.world_size,
rank=args.rank)
else:
logging.info(f"Use CPU ~")
# 创建/加载模型，使用预训练模型时，需要自己先下载好放到 pretrained 文件夹下，以网络名词命名
logging.info(f"=> creating model '{args.arch}'")
model = my_models.get_model(args.arch,
args.pretrained,
num_classes=args.num_classes)
# 重加载之前训练好的模型
if args.resume:
if os.path.isfile(args.resume):
logging.info(f"=> loading checkpoint '{args.resume}'")
checkpoint = torch.load(args.resume,
map_location=torch.device('cpu'))
acc = model.load_state_dict(checkpoint['state_dict'])
logging.info(f'missing keys of models: {acc.missing_keys}')
del checkpoint
else:
raise Exception(
f"No checkpoint found at '{args.resume}' to be resumed")
# 模型信息
image_height, image_width = args.image_size
logging.info(
f'Model {args.arch} input size: ({image_height}, {image_width})')
utils.summary(size=(image_height, image_width), channel=3, model=model)
# 模型转换：转为 torch.jit.script
if args.jit:
if not args.resume:
raise Exception('Option --resume must specified!')
applications.convert_to_jit(model, args=args)
return
if args.criterion == 'rank':
criterion = criterions.RankingLoss(args=args) # 对比排序损失
elif args.criterion == 'emd':
criterion = criterions.EMDLoss() # 推土机距离损失
elif args.criterion == 'regress':
criterion = criterions.RegressionLoss() # MSE回归损失
else:
raise NotImplementedError(
f'Not loss function {args.criterion}，only (rank, emd, regress)!')
if args.cuda:
if args.distributed and args.sync_bn:
model = apex.parallel.convert_syncbn_model(model)
torch.cuda.set_device(args.gpu)
model.cuda(args.gpu)
criterion = criterion.cuda(args.gpu)
# 优化器：Adam > SGD > SWA(SGD > Adam)
optimizer = torch.optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
# 可尝试优化器
# optimizer = torch.optim.SGD(model.parameters(),
# args.lr, momentum=args.momentum,
# weight_decay=args.weight_decay)
# from optim.torchtools.optim import RangerLars, Ralamb, Novograd, LookaheadAdam, Ranger, RAdam, AdamW
# optimizer = RangerLars(model.parameters(), lr=args.lr, weight_decay=args.weight_decay)
# optimizer = Ralamb(model.parameters(), lr=args.lr, weight_decay=args.weight_decay)
# optimizer = Novograd(model.parameters(), lr=args.lr, weight_decay=args.weight_decay)
# optimizer = LookaheadAdam(model.parameters(), lr=args.lr, weight_decay=args.weight_decay)
# optimizer = Ranger(model_params, lr=args.lr, weight_decay=args.weight_decay)
# optimizer = RAdam(model.parameters(), lr=args.lr, weight_decay=args.weight_decay)
# optimizer = AdamW(model.parameters(), lr=args.lr, weight_decay=args.weight_decay)
# 随机均值平均优化器
# from optim.swa import SWA
# optimizer = SWA(optimizer, swa_start=10, swa_freq=5, swa_lr=0.05)
# 混合精度训练
if args.cuda:
model, optimizer = amp.initialize(model, optimizer, opt_level="O1")
model = DDP(model)
else:
model = torch.nn.DataParallel(model)
if args.train:
train_loader = dataloader.load(args, 'train')
val_loader = dataloader.load(args, 'val')
scheduler = LambdaLR(
optimizer, lambda epoch: adjust_learning_rate(epoch, args=args))
applications.train(train_loader, val_loader, model, criterion,
optimizer, scheduler, args)
args.evaluate = True
if args.evaluate:
torch.set_flush_denormal(True)
test_loader = dataloader.load(args, name='test')
acc, loss, test_results = applications.test(test_loader, model,
criterion, args)
logging.info(f'Evaluation: * Acc@1 {acc:.3f} and loss {loss:.3f}.')
logging.info(f'Evaluation results:')
for result in test_results:
logging.info(' '.join([str(r) for r in result]))
logging.info('Evaluation Over~')
import pytorch_lightning as pl
import dataloader
from config import *
from net.charrnn import CharRNN
from net.lightning import Lightning
(trl, tel, val), vocab = dataloader.load(FILE_PATH,
DEVICE,
SPLITS,
BATCH_SIZE,
SEQ_LEN,
unique=True)
for MODEL_NAME in ["rnn", "gru", "lstm"]: #["lstm"]:
for N_LAYERS in [1, 2, 3]: #[1, 2, 3, 4]:
for HIDDEN_SIZE in [32, 64, 128, 256, 512]:
net = CharRNN(len(vocab), HIDDEN_SIZE, EMBEDDING_DIM, MODEL_NAME,
DROPOUT, N_LAYERS, DEVICE)
lightning = Lightning(net, LR)
esc = pl.callbacks.EarlyStopping(monitor='val_acc',
min_delta=0.00,
patience=3,
mode="max")
trainer = pl.Trainer(gpus=int(DEVICE == "cuda"),
precision=PRECISION,
gradient_clip_val=CLIP,
max_epochs=MAX_EPOCHS,
progress_bar_refresh_rate=10,
callbacks=[esc],
benchmark=True,
timestamp = datautils.timeStamped()
datautils.createDirectory(outputDirectory)
##############################################
# This creates a logger. You can use this to create an output file
# which logs the steps of your analysis
##############################################
logger = datautils.createLogger(outputDirectory)
######################################################
# Now lets try with tfidf
######################################################
logger.info("\load data\n\n")
data_tuple = dataloader.load(dataDirectory, dataFile)
#The loaded column map
columnMap = dataloader.loadColumnMap(dataDirectory, dataFile)
#the loaded target map
targetMap = dataloader.loadTargetMap(dataDirectory, dataFile)
kmeansplots.elbow(data_tuple[0], outputDirectory, "elbow", 2, 8)
###############Run the kmeans and get back a model "kmeanstfidf"
X = conversions.convertToTfIdf(data_tuple[0])
kmeanstfidf = kmeansTools.doKmeans(X,
clusterCount=10,
maxIterations=10,
init="k-means++",
n_init=2,
precompute_distances='auto',
def main():
sc = SparkContext('local[15]', 'haha')
# sc._conf.set("spark.python.profile", "true")
print(sc.getConf().getAll())
d = load(sc)
data_train_lp, data_dev_p, label_dev_gt, test_p = d['train_tfidf_lp'], d['dev_tfidf'], d['dev_gt'], d['test_tfidf']
data_train_p, label_train_gt = d['train_tfidf'], d['train_gt']
data_train, data_dev, data_test = d['train_raw'], d['dev_raw'], d['test_raw']
data_train_lp = data_train_lp.sample(False, 0.01)
# print(sum(data_train_lp.first()[0]))
# print(data_train_lp.zipWithIndex().collect())
print(data_train_lp.take(2))
print("___________train_bayes_____________")
sys.stdout.flush()
nb = NaiveBayes.train(data_train_lp)
print("___________trained_bayes___________")
sys.stdout.flush()
# nb.save(sc, 'bayes.model')
bayes_result_dev = nb.predict(data_dev_p).map(int)
bayes_result_dev.count()
bayes_result_train = nb.predict(data_train_p).map(int)
bayes_result_train.count()
bayes_result_test = nb.predict(test_p).map(int)
bayes_result_test.count()
print("train info:")
valid(bayes_result_train, label_train_gt)
print("dev info:")
valid(bayes_result_dev, label_dev_gt)
print("___________train_logistic_____________")
sys.stdout.flush()
lg = LogisticRegressionWithSGD.train(data_train_lp, step=0.005)
print("___________trained_logisitc___________")
sys.stdout.flush()
# lg.save(sc, 'logistic.model')
logistic_result_dev = lg.predict(data_dev_p).map(int)
logistic_result_train = lg.predict(data_train_p).map(int)
logistic_result_test = lg.predict(test_p).map(int)
print("train info:")
valid(logistic_result_train, label_train_gt)
print("dev info:")
valid(logistic_result_dev, label_dev_gt)
fused_train_p = stack_label([bayes_result_train, logistic_result_train])
fused_dev_p = stack_label([bayes_result_dev, logistic_result_dev])
fused_test_p = stack_label([bayes_result_test, logistic_result_test])
fused_train_lp = label(data_train, fused_train_p)
print("___________train_GBDT___________")
sys.stdout.flush()
gbdt = GradientBoostedTrees.trainClassifier(fused_train_lp, {})
print('___________trained_GBDT_________')
sys.stdout.flush()
fused_result_train = gbdt.predict(fused_train_p)
fused_result_dev = gbdt.predict(fused_dev_p)
fused_result_test = gbdt.predict(fused_test_p)
print("train info:")
valid(fused_result_train, label_train_gt)
print("dev info:")
valid(fused_result_dev, label_dev_gt)
dump(fused_result_test.map(int).collect())
sc.show_profiles()
None.
'''
assert isinstance(fireDataFrame, pd.DataFrame)
zip_count = fireDataFrame['address_zip'].value_counts()
zip_count = zip_count[0:30]
plt.figure(figsize=(12, 14))
ax = sns.barplot(y=zip_count.index,
x=zip_count.values,
alpha=0.8,
color='orangered',
orient='h',
order=zip_count.index)
ax.xaxis.set_major_formatter(
matplotlib.ticker.FuncFormatter(lambda x, p: format(int(x), ',')))
plt.title('Number of fires by zipcode', fontsize=20)
plt.ylabel('ZIP Code(top 30)', fontsize=18)
plt.xlabel('Number of fire incidents', fontsize=18)
ax.text(123600, 1.3, 'Max at 92101(Downtown)', fontsize=20)
#plt.savefig('pic_zip-num.png')
plt.show()
if __name__ == '__main__':
fireDataFrame = dataloader.load(range(2007, 2020))
fireIncidentPerZipcode(fireDataFrame)
from cntk import Trainer, StreamConfiguration, text_format_minibatch_source, learning_rate_schedule, UnitType
from cntk.initializer import glorot_uniform
from cntk.learner import sgd
from cntk.ops import *
from cntk.utils import get_train_eval_criterion, get_train_loss
import dataloader
image_size = 28
input_dim = image_size * image_size
num_output_classes = 10
num_hidden_layers = 2
hidden_layers_dim = 400
dataloader.load()
train_file = "data/MNIST/Train-28x28_cntk_text.txt"
if os.path.isfile(train_file):
path = train_file
else:
print("Cannot find data file")
feature_stream_name = 'features'
labels_stream_name = 'labels'
mb_source = text_format_minibatch_source(path, [
StreamConfiguration(feature_stream_name, input_dim),
StreamConfiguration(labels_stream_name, num_output_classes)
])
from normalization import normalize
from graph import MarkovChainGraph
import matplotlib.pyplot as plt
stateTranslate = {
'0.0, 0.0, 0.0': 'nothing to see',
'1.0, 0.0, 0.0': 'object in front',
'1.0, 1.0, 0.0': 'facing object',
'0.0, 0.0, 1.0': 'object grabbed',
'1.0, 0.0, 1.0': 'object grabbed and object in front',
'1.0, 1.0, 1.0': 'object grabbed and facing object',
}
if __name__ == '__main__':
plt.rcParams.update({'figure.autolayout': True})
filename = sys.argv[1]
sensorStates, behaviors = dataloader.load(filename)
discretizer = SimplifyingDiscretizer()
graph = MarkovChainGraph()
result = {}
ss = []
for sensorState in sensorStates:
sensorState = normalize(sensorState)
s = discretizer.discretize(sensorState)
s = graph.state_to_key(s)
ss.append(stateTranslate[s])
if stateTranslate[s] in result:
result[stateTranslate[s]] += 1
else:
result[stateTranslate[s]] = 1
graph.construct(ss, behaviors)
{
'Emergency Medical Response': 'EMR',
'Urgent Medical Response': 'UMR',
'Non-Emergency Medical Response': 'NEMR'
},
inplace=True)
category = category.sort_values(ascending=True)
if len(years) == 1:
title = 'Call Category in %d' % (years[0])
else:
title = 'Call Category over %d-%d' % (years[0], years[-1])
fig, ax = plt.subplots(figsize=(30, 30))
ax.barh(category.index, category.values, color='orangered')
ax.xaxis.set_major_formatter(
matplotlib.ticker.FuncFormatter(lambda x, p: format(int(x), ',')))
matplotlib.rc('xtick', labelsize=30)
matplotlib.rc('ytick', labelsize=30)
plt.title(title)
#plt.savefig('call_category_%d-%d'%(years[0],years[-1]))
plt.show()
if __name__ == '__main__':
years = range(2010, 2011)
fire_df = dataloader.load(years)
plot_category_over_years(fire_df, years)
from matplotlib import pyplot as plt
from rx import operators
from rx import scheduler
from rx.scheduler.eventloop import AsyncIOThreadSafeScheduler
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument("-c", "--config", help="config file", required=True)
args = parser.parse_args()
args_config = os.path.abspath(args.config)
cfg.merge_from_file(args_config)
cfg_dir = os.path.dirname(args_config)
print("Loading training data...")
loader = cfg.train_data.loader
train_dataset = dataloader.load(loader, cfg_dir)
train_dataset.generate_original = False
if cfg.valid_data.enabled:
print("Loading validation data...")
loader = cfg.valid_data.loader
valid_dataset = dataloader.load(loader, cfg_dir)
valid_dataset.normalize_with(train_dataset.input_mean,
train_dataset.input_std,
train_dataset.target_mean,
train_dataset.target_std)
valid_dataset.generate_original = False
else:
valid_dataset = None
if cfg.test_data.enabled:
print("Loading testing data...")
'numWsClus': 10, # number of clusters of services
'saveTimeInfo': False, # whether to keep track of the running time
'saveLog': True, # whether to save log into file
'debugMode': False, # whether to record the debug info
'parallelMode': True # whether to leverage multiprocessing for speedup
}
initConfig(para)
#########################################################
startTime = time.clock() # start timing
logger.info('==============================================')
logger.info('CLUS: [Silic et al., FSE\'2013].')
# load the dataset
dataTensor = dataloader.load(para)
logger.info('Loading data done.')
# run for each density
if para['parallelMode']: # run on multiple processes
pool = multiprocessing.Pool()
for density in para['density']:
pool.apply_async(evaluator.execute, (dataTensor, density, para))
pool.close()
pool.join()
else: # run on single processes
for density in para['density']:
evaluator.execute(dataTensor, density, para)
logger.info(
time.strftime(
def main():
optparser = OptionParser()
# options for input and output
optparser.add_option('-m', '--model', action='store', dest='model_name')
# options for recomia
optparser.add_option('-e', '--embed', action='store', dest='key_embed')
optparser.add_option('-s', '--dataset', action='store', dest='key_dataset')
optparser.add_option('-p', '--prefix', action='store', dest='prefix')
# options for training
optparser.add_option('-r', '--resume', action='store_true', dest='resume', default = False)
optparser.add_option('-d', '--dim_hidden', action='store', dest='dim_hidden', type='int', default = None)
optparser.add_option('-b', '--batch_size', action='store', type='int', dest='batch_size', default = 16)
optparser.add_option('-l', '--learning_rate', action='store', type='float', dest='lrate', default = 0.05)
optparser.add_option('-c', '--decay_c', action='store', type='float', dest='decay_c', default = 1e-4)
optparser.add_option('-v', '--valid_freq', action='store', type='int', dest='valid_freq', default=5000)
opts, args = optparser.parse_args()
dir_exp = '../'
fname_dataset = dir_exp + 'dataset/%s.pkl'%(opts.key_dataset)
fname_embed = dir_exp + 'wemb/%s.txt'%(opts.key_embed)
fname_model = dir_exp + 'model/%s'%(opts.prefix)
fname_test = dir_exp + 'test/%s_test.pkl'%(opts.prefix)
fname_prec = dir_exp + 'test/%s_prec.pkl'%(opts.prefix)
dir_tokid = dir_exp + 'tokid/%s/'%(opts.key_embed)
dataset = dataloader.load(fname_dataset, dir_tokid)
wembedder = WordEmbedder.load(fname_embed)
if not opts.resume:
Wemb = wembedder.get_Wemb()
else:
Wemb = None
print >> sys.stderr, 'main: [info] start training'
clf = Classifier()
res = clf.train(
dataset = dataset,
Wemb = Wemb,
fname_model = fname_model,
resume = opts.resume,
model_name = opts.model_name,
dim_hidden = opts.dim_hidden,
batch_size = opts.batch_size,
decay_c = opts.decay_c,
lrate = opts.lrate,
validFreq = opts.valid_freq,
saveFreq = opts.valid_freq,
)
test_x, test_y = dataset[2]
proba = clf.predict_proba(test_x)
cPickle.dump((test_y, proba), open(fname_test, 'w'))
prec = precision_at_n(test_y, proba)
cPickle.dump(prec, open(fname_prec, 'w'))
def anomaly(experiment_name,
network_model,
dataset,
inside_labels,
unknown_labels,
with_unknown,
batch_size=100,
nb_epochs=200,
save_weights=True):
print('#' * 50)
print('Experiment:', experiment_name)
print('model:', network_model)
print('dataset:', dataset)
print('inside_labels:', str(inside_labels))
print('unknown_labels:', str(unknown_labels))
print('batch_size:', batch_size)
print('nb_epochs:', nb_epochs)
print('-' * 50)
inside_labels.sort()
unknown_labels.sort()
(X_train_all,
y_train_all), (X_train, y_train), (X_test, y_test) = dataloader.load(
dataset, inside_labels, unknown_labels, with_unknown)
if 'mlp' in network_model:
X_train_all = X_train_all.reshape(X_train_all.shape[0], -1)
X_train = X_train.reshape(X_train.shape[0], -1)
X_test = X_test.reshape(X_test.shape[0], -1)
input_shape = X_train.shape[1:]
nb_classes = y_train.shape[1]
nb_batchs = X_train.shape[0] // batch_size
model = create_model(network_model, batch_size, input_shape, nb_classes,
nb_batchs)
if network_model.endswith('bayesian') and 'poor' not in network_model:
mod = X_train.shape[0] % batch_size
if mod:
X_train = X_train[:-mod]
y_train = y_train[:-mod]
print('Training')
start_time = time.time()
for i in tqdm.tqdm(list(range(nb_epochs))):
model.fit(X_train,
y_train,
nb_epoch=1,
batch_size=batch_size,
verbose=0)
end_time = time.time()
if save_weights:
path = '{0}_results/weights/{1}_{2}/'
wu = 'with' if with_unknown else 'without'
path = path.format(dataset, network_model, wu)
os.makedirs(path, exist_ok=True)
model.save_weights(os.path.join(path, experiment_name + '.h5'),
overwrite=True)
print('Collecting measures of train')
measures_train, train_acc = get_measures(X_train_all, y_train_all, model,
batch_size, inside_labels)
print('Collecting measures of test')
measures_test, test_acc = get_measures(X_test, y_test, model, batch_size,
inside_labels)
print('Classification')
clf = uncertainty_classifier(measures_train, inside_labels, unknown_labels)
measures_test['classifier'] = {l: [] for l in range(10)}
for l in range(10):
n = len(measures_test['entropy_std_samples'][l])
for i in range(n):
f = [
# measures_test['variation_ratio'][l][i],
measures_test['mean_entropy'][l][i],
measures_test['pred_std_mean'][l][i],
measures_test['entropy_std_samples'][l][i],
measures_test['entropy_mean_samples'][l][i],
]
p = clf.predict_proba([f])[0, 1]
measures_test['classifier'][l].append(p)
# Anomaly detection
# by classical prediction entropy
def anomaly_detection(anomaly_score_dict, metric_name, df):
threshold = np.logspace(-10.0, 1.0, 1000)
acc = {}
for t in threshold:
tp = 0.0
tn = 0.0
for l in anomaly_score_dict:
if l in unknown_labels:
continue
if l in inside_labels:
tp += (np.array(anomaly_score_dict[l]) < t).mean()
else:
tn += (np.array(anomaly_score_dict[l]) >= t).mean()
tp /= len(inside_labels)
tn /= (10.0 - len(unknown_labels)) - len(inside_labels)
bal_acc = (tp + tn) / 2.0
f1_score = 2.0 * tp / (2.0 + tp - tn)
acc[t] = [bal_acc, f1_score, tp, tn]
trues = []
scores = []
for l in anomaly_score_dict:
if l in unknown_labels: continue
scores += anomaly_score_dict[l]
if l in inside_labels:
trues += [0] * len(anomaly_score_dict[l])
else:
trues += [1] * len(anomaly_score_dict[l])
assert len(trues) == len(scores)
auc = metrics.roc_auc_score(trues, scores)
sorted_acc = sorted(acc.items(), key=lambda x: x[1][0], reverse=True)
df.set_value(experiment_name, metric_name + '_bal_acc',
sorted_acc[0][1][0])
bal_acc = sorted_acc[0][1][0]
sorted_acc = sorted(acc.items(), key=lambda x: x[1][1], reverse=True)
df.set_value(experiment_name, metric_name + '_f1_score',
sorted_acc[0][1][1])
f1_score = sorted_acc[0][1][1]
df.set_value(experiment_name, metric_name + '_auc', auc)
msg = '{0}: (auc, {1:.2f}), (bal_acc, {2:.2f}), (f1_score, {3:.2f})'
print(msg.format(metric_name, auc, bal_acc, f1_score))
return df
print('-' * 50)
df = pd.DataFrame()
df.set_value(experiment_name, 'experiment_name', experiment_name)
df.set_value(experiment_name, 'train_time', end_time - start_time)
df.set_value(experiment_name, 'dataset', dataset)
df.set_value(experiment_name, 'test_acc', test_acc)
df.set_value(experiment_name, 'inside_labels', str(inside_labels))
df.set_value(experiment_name, 'unknown_labels', str(unknown_labels))
df.set_value(experiment_name, 'epochs', nb_epochs)
df = anomaly_detection(measures_test['pred_std_mean'], 'pred_std_', df)
df = anomaly_detection(measures_test['mean_entropy'], 'entropy_', df)
df = anomaly_detection(measures_test['entropy_mean_samples'],
'entropy_expectation_', df)
df = anomaly_detection(measures_test['variation_ratio'],
'variation_ratio_', df)
df = anomaly_detection(measures_test['classifier'], 'classifier_', df)
return df
def main_worker(gpu, args):
"""
模型训练、测试、转JIT、蒸馏文件制作
:param gpu: 运行的gpu id
:param args: 运行超参
"""
args.gpu = gpu
utils.generate_logger(f"{datetime.datetime.now().strftime('%Y%m%d%H%M%S')}-{gpu}.log")
logging.info(f'args: {args}')
# 可复现性
if args.seed is not None:
random.seed(args.seed)
torch.manual_seed(args.seed)
cudnn.deterministic = True
logging.warning('You have chosen to seed training. '
'This will turn on the CUDNN deterministic setting, '
'which can slow down your training considerably! '
'You may see unexpected behavior when restarting '
'from checkpoints.')
if args.cuda:
logging.info(f"Use GPU: {args.gpu} ~")
if args.distributed:
args.rank = args.rank * args.gpus + gpu
dist.init_process_group(backend='nccl', init_method=args.init_method,
world_size=args.world_size, rank=args.rank)
else:
logging.info(f"Use CPU ~")
# 创建/加载模型，使用预训练模型时，需要自己先下载好放到 pretrained 文件夹下，以网络名词命名
logging.info(f"=> creating model '{args.arch}'")
model = my_models.get_model(args.arch, args.pretrained, num_classes=args.num_classes)
# 重加载之前训练好的模型
if args.resume:
if os.path.isfile(args.resume):
logging.info(f"=> loading checkpoint '{args.resume}'")
checkpoint = torch.load(args.resume, map_location=torch.device('cpu'))
acc = model.load_state_dict(checkpoint['state_dict'], strict=True)
logging.info(f'missing keys of models: {acc.missing_keys}')
del checkpoint
else:
raise Exception(f"No checkpoint found at '{args.resume}' to be resumed")
# 模型信息
image_height, image_width = args.image_size
logging.info(f'Model {args.arch} input size: ({image_height}, {image_width})')
utils.summary(size=(image_height, image_width), channel=3, model=model)
# 模型转换：转为 torch.jit.script
if args.jit:
if not args.resume:
raise Exception('Option --resume must specified!')
applications.convert_to_jit(model, args=args)
return
if args.criterion == 'softmax':
criterion = criterions.HybridCELoss(args=args) # 混合策略多分类
elif args.criterion == 'bce':
criterion = criterions.HybridBCELoss(args=args) # 混合策略多标签二分类
else:
raise NotImplementedError(f'Not loss function {args.criterion}')
if args.cuda:
if args.distributed and args.sync_bn:
model = apex.parallel.convert_syncbn_model(model)
torch.cuda.set_device(args.gpu)
model.cuda(args.gpu)
criterion = criterion.cuda(args.gpu)
if args.knowledge in ('train', 'test', 'val'):
torch.set_flush_denormal(True)
distill_loader = dataloader.load(args, name=args.knowledge)
applications.distill(distill_loader, model, criterion, args, is_confuse_matrix=True)
return
if args.make_curriculum in ('train', 'test', 'val'):
torch.set_flush_denormal(True)
curriculum_loader = dataloader.load(args, name=args.make_curriculum)
applications.make_curriculum(curriculum_loader, model, criterion, args, is_confuse_matrix=True)
return
if args.visual_data in ('train', 'test', 'val'):
torch.set_flush_denormal(True)
test_loader = dataloader.load(args, name=args.visual_data)
applications.Visualize.visualize(test_loader, model, args)
return
# 优化器
opt_set = {
'sgd': partial(torch.optim.SGD, momentum=args.momentum),
'adam': torch.optim.Adam, 'adamw': AdamW,
'radam': RAdam, 'ranger': Ranger, 'lookaheadadam': LookaheadAdam,
'ralamb': Ralamb, 'rangerlars': RangerLars,
'novograd': Novograd,
}
optimizer = opt_set[args.opt](model.parameters(), lr=args.lr) # weight decay转移到train那里了
# 随机均值平均优化器
# from optim.swa import SWA
# optimizer = SWA(optimizer, swa_start=10, swa_freq=5, swa_lr=0.05)
# 混合精度训练
if args.cuda:
model, optimizer = amp.initialize(model, optimizer, opt_level="O1")
if args.distributed:
model = apex.parallel.DistributedDataParallel(model)
else:
model = torch.nn.DataParallel(model)
if args.train:
train_loader = dataloader.load(args, 'train')
val_loader = dataloader.load(args, 'val')
scheduler = LambdaLR(optimizer,
lambda epoch: adjust_learning_rate(epoch, args=args))
applications.train(train_loader, val_loader, model, criterion, optimizer, scheduler, args)
args.evaluate = True
if args.evaluate:
torch.set_flush_denormal(True)
test_loader = dataloader.load(args, name='test')
acc, loss, paths_targets_preds_probs = applications.test(test_loader, model,
criterion, args, is_confuse_matrix=True)
logging.info(f'Evaluation: * Acc@1 {acc:.3f} and loss {loss:.3f}.')
logging.info(f'Evaluation Result:\n')
for path, target, pred, prob in paths_targets_preds_probs:
logging.info(path + ' ' + str(target) + ' ' + str(pred) + ' ' + ','.join([f'{num:.2f}' for num in prob]))
logging.info('Evaluation Over~')
'saveTimeInfo': False, # whether to keep track of the running time
'saveLog': True, # whether to save log into file
'debugMode': False, # whether to record the debug info
'parallelMode': True # whether to leverage multiprocessing for speedup
}
initConfig(para)
#########################################################
startTime = time.clock() # start timing
logger.info('==============================================')
logger.info('ADF: [Wu et al., TSMC\'2013].')
# load the dataset
dataMatrix = dataloader.load(para)
logger.info('Loading data done.')
# run for each density
if para['parallelMode']: # run on multiple processes
pool = multiprocessing.Pool()
for density in para['density']:
pool.apply_async(evaluator.execute, (dataMatrix, density, para))
pool.close()
pool.join()
else: # run on single processes
for density in para['density']:
evaluator.execute(dataMatrix, density, para)
logger.info(time.strftime('All done. Total running time: %d-th day - %Hhour - %Mmin - %Ssec.',
time.gmtime(time.clock() - startTime)))
opt = Config().parse()
if not os.path.exists(opt.RESULT):
os.makedirs(opt.RESULT)
os.environ['CUDA_VISIBLE_DEVICES'] = str(opt.GPU)
print("using gpu {}".format(opt.GPU))
"""
Data preparation.
"""
opt.use_PCA = True
opt.use_SuperPCA = False
print(opt.use_PCA)
print(opt.use_SuperPCA)
assert (opt.use_PCA and opt.use_SuperPCA) == False
data, label = load(opt.DATASET)
if opt.use_PCA:
data, pca = apply_pca(data)
else:
pass
init_tr_labeled_idx, init_tr_unlabeled_idx, te_idx = get_init_indices(
data, label)
init_trl_data, init_trl_label = get_data(data, label, init_tr_labeled_idx)
init_trunl_data, init_trunl_label = get_data(data, label,
init_tr_unlabeled_idx)
te_data, te_label = get_data(data, label, te_idx)
init_trl_data = np.expand_dims(init_trl_data, axis=4)
init_trl_label = keras.utils.to_categorical(init_trl_label)
def anomaly(experiment_name, network_model, dataset,
inside_labels, unknown_labels, with_unknown,
batch_size=100, nb_epochs=200, save_weights=True):
print('#'*50)
print('Experiment:', experiment_name)
print('model:', network_model)
print('dataset:', dataset)
print('inside_labels:', str(inside_labels))
print('unknown_labels:', str(unknown_labels))
print('batch_size:', batch_size)
print('nb_epochs:', nb_epochs)
print('-'*50)
inside_labels.sort()
unknown_labels.sort()
(X_train_all, y_train_all), (X_train, y_train), (X_test, y_test) = dataloader.load(dataset,
inside_labels,
unknown_labels,
with_unknown)
if 'mlp' in network_model:
X_train_all = X_train_all.reshape(X_train_all.shape[0], -1)
X_train = X_train.reshape(X_train.shape[0], -1)
X_test = X_test.reshape(X_test.shape[0], -1)
input_shape = X_train.shape[1:]
nb_classes = y_train.shape[1]
nb_batchs = X_train.shape[0]//batch_size
model = create_model(network_model, batch_size, input_shape, nb_classes, nb_batchs)
if network_model.endswith('bayesian') and 'poor' not in network_model:
mod = X_train.shape[0]%batch_size
if mod:
X_train = X_train[:-mod]
y_train = y_train[:-mod]
print('Training')
start_time = time.time()
for i in tqdm.tqdm(list(range(nb_epochs))):
model.fit(X_train, y_train, nb_epoch=1, batch_size=batch_size, verbose=0)
end_time = time.time()
if save_weights:
path = '{0}_results/weights/{1}_{2}/'
wu = 'with' if with_unknown else 'without'
path = path.format(dataset, network_model, wu)
os.makedirs(path, exist_ok=True)
model.save_weights(os.path.join(path, experiment_name+'.h5'), overwrite=True)
print('Collecting measures of train')
measures_train, train_acc = get_measures(X_train_all, y_train_all, model, batch_size, inside_labels)
print('Collecting measures of test')
measures_test, test_acc = get_measures(X_test, y_test, model, batch_size, inside_labels)
print('Classification')
clf = uncertainty_classifier(measures_train, inside_labels, unknown_labels)
measures_test['classifier'] = {l:[] for l in range(10)}
for l in range(10):
n = len(measures_test['entropy_std_samples'][l])
for i in range(n):
f = [
# measures_test['variation_ratio'][l][i],
measures_test['mean_entropy'][l][i],
measures_test['pred_std_mean'][l][i],
measures_test['entropy_std_samples'][l][i],
measures_test['entropy_mean_samples'][l][i],
]
p = clf.predict_proba([f])[0, 1]
measures_test['classifier'][l].append(p)
# Anomaly detection
# by classical prediction entropy
def anomaly_detection(anomaly_score_dict, metric_name, df):
threshold = np.logspace(-10.0, 1.0, 1000)
acc = {}
for t in threshold:
tp = 0.0
tn = 0.0
for l in anomaly_score_dict:
if l in unknown_labels:
continue
if l in inside_labels:
tp += (np.array(anomaly_score_dict[l]) < t).mean()
else:
tn += (np.array(anomaly_score_dict[l]) >= t).mean()
tp /= len(inside_labels)
tn /= (10.0 - len(unknown_labels)) - len(inside_labels)
bal_acc = (tp + tn)/2.0
f1_score = 2.0*tp/(2.0 + tp - tn)
acc[t] = [bal_acc, f1_score, tp, tn]
trues = []
scores = []
for l in anomaly_score_dict:
if l in unknown_labels: continue
scores += anomaly_score_dict[l]
if l in inside_labels:
trues += [0]*len(anomaly_score_dict[l])
else:
trues += [1]*len(anomaly_score_dict[l])
assert len(trues) == len(scores)
auc = metrics.roc_auc_score(trues, scores)
sorted_acc = sorted(acc.items(), key=lambda x: x[1][0], reverse=True)
df.set_value(experiment_name, metric_name + '_bal_acc', sorted_acc[0][1][0])
bal_acc = sorted_acc[0][1][0]
sorted_acc = sorted(acc.items(), key=lambda x: x[1][1], reverse=True)
df.set_value(experiment_name, metric_name + '_f1_score', sorted_acc[0][1][1])
f1_score = sorted_acc[0][1][1]
df.set_value(experiment_name, metric_name + '_auc', auc)
msg = '{0}: (auc, {1:.2f}), (bal_acc, {2:.2f}), (f1_score, {3:.2f})'
print(msg.format(metric_name, auc, bal_acc, f1_score))
return df
print('-'*50)
df = pd.DataFrame()
df.set_value(experiment_name, 'experiment_name', experiment_name)
df.set_value(experiment_name, 'train_time', end_time - start_time)
df.set_value(experiment_name, 'dataset', dataset)
df.set_value(experiment_name, 'test_acc', test_acc)
df.set_value(experiment_name, 'inside_labels', str(inside_labels))
df.set_value(experiment_name, 'unknown_labels', str(unknown_labels))
df.set_value(experiment_name, 'epochs', nb_epochs)
df = anomaly_detection(measures_test['pred_std_mean'], 'pred_std_', df)
df = anomaly_detection(measures_test['mean_entropy'], 'entropy_', df)
df = anomaly_detection(measures_test['entropy_mean_samples'], 'entropy_expectation_', df)
df = anomaly_detection(measures_test['variation_ratio'], 'variation_ratio_', df)
df = anomaly_detection(measures_test['classifier'], 'classifier_', df)
return df
'saveTimeInfo': False, # whether to keep track of the running time
'saveLog': True, # whether to save log into file
'debugMode': False, # whether to record the debug info
'parallelMode': True # whether to leverage multiprocessing for speedup
}
initConfig(para)
#########################################################
startTime = time.clock() # start timing
logger.info('==============================================')
logger.info('Approach: [UMEAN, IMEAN, UPCC, IPCC, UIPCC].')
# load the dataset
dataTensor = dataloader.load(para)
logger.info('Loading data done.')
# run for each density
numTimeSlice = dataTensor.shape[2]
if para['parallelMode']: # run on multiple processes
pool = multiprocessing.Pool()
for cxtId in xrange(numTimeSlice):
dataMatrix = dataTensor[:, :, cxtId]
for density in para['density']:
pool.apply_async(evaluator.execute, (dataMatrix, density, para, cxtId))
pool.close()
pool.join()
else: # run on single processes
for cxtId in xrange(numTimeSlice):
dataMatrix = dataTensor[:, :, cxtId]
def main(unused_argv):
# Load datasets
# Get signals and labels
if FLAGS.data_type == 'all':
trn_data_list = [
dataloader.load(which_data='s{}'.format(i),
gesture=FLAGS.gesture,
for_merge=True,
train=True) for i in [1, 2, 3]
]
tst_data_list = [
dataloader.load(which_data='s{}'.format(i),
gesture=FLAGS.gesture,
for_merge=True,
train=False) for i in [1, 2, 3]
]
signals_trn = np.concatenate([data[0] for data in trn_data_list],
axis=0)
labels_trn = np.concatenate([data[1] for data in trn_data_list],
axis=0)
signals_tst = np.concatenate([data[0] for data in tst_data_list],
axis=0)
labels_tst = np.concatenate([data[1] for data in tst_data_list],
axis=0)
else:
signals_trn, labels_trn = dataloader.load(which_data=FLAGS.data_type,
gesture=FLAGS.gesture,
train=True)
signals_tst, labels_tst = dataloader.load(which_data=FLAGS.data_type,
gesture=FLAGS.gesture,
train=False)
# Set model params
model_params = dict(
learning_rate=FLAGS.learning_rate,
# regularization type and strength
wd=FLAGS.wd,
# convolutional layer
window_size=5,
n_conv=1,
n_filter=100,
# fully connected layer
n_fully_connected=1,
# pooling
pooling_size=2,
pooling_stride=1,
# n_labels
n_labels=labels_trn.shape[1],
) # model0.+n_conv1.+n_filter.+n_fc1.*/eval
model_id = 'GEST' if FLAGS.gesture is True else 'LOCO'
model_id += '_%s' % FLAGS.data_type.upper()
model_id += '_wd{}'.format(FLAGS.wd)
# model_id += '_n_conv%s' % model_params['n_conv']
# model_id += '_n_filter%s' % model_params['n_filter']
# model_id += '_n_fc%s' % model_params['n_fully_connected']
dt_now = datetime.now().strftime('%Y%m%d_%H%M%S')
fname = '{}__{}'.format(model_id, dt_now)
print('-' * 5, model_id, '-' * 5)
print('-' * 5, dt_now, '-' * 5)
# Model dir
model_dir = './tf_models/{}/{}'.format(model_id, dt_now)
# if FLAGS.restart is True:
if tf.gfile.Exists(model_dir):
tf.gfile.DeleteRecursively(model_dir)
tf.gfile.MakeDirs(model_dir)
# Instantiate Estimator
estimator = tf.estimator.Estimator(model_fn=func.model_fn,
params=model_params,
model_dir=model_dir)
# Input functions
train_input_fn = tf.estimator.inputs.numpy_input_fn(
x={'signals': signals_trn},
y=labels_trn,
batch_size=FLAGS.batch_size,
num_epochs=None,
shuffle=True)
test_input_fn = tf.estimator.inputs.numpy_input_fn(
x={'signals': signals_tst}, y=labels_tst, num_epochs=1, shuffle=False)
# Train and test ==> record test summary
# We iterate train and evaluation to save summaries
if FLAGS.train is True:
for i in range(FLAGS.steps // FLAGS.log_freq):
iter_up_to_now = i * FLAGS.log_freq
print('-' * 10, 'Begin training - iteration', iter_up_to_now,
'-' * 10)
estimator.train(input_fn=train_input_fn, steps=FLAGS.log_freq)
# Evaluate and save the result
iter_up_to_now = (i + 1) * FLAGS.log_freq
ev = estimator.evaluate(
input_fn=test_input_fn,
hooks=[_SaveWeightHook(fname, iter_up_to_now)])
save_the_evaluation(ev, fname, iter_up_to_now)
'saveTimeInfo': False, # whether to keep track of the running time
'saveLog': False, # whether to save log into file
'debugMode': False, # whether to record the debug info
'parallelMode': True # whether to leverage multiprocessing for speedup
}
initConfig(para)
#########################################################
startTime = time.clock() # start timing
logger.info('==============================================')
logger.info('Approach: HMF [He et al., ICWS\'2014].')
# load the dataset
dataMatrix = dataloader.load(para)
logger.info('Loading data done.')
# get the location groups for users as well as for services
locGroup = dataloader.getLocGroup(para)
logger.info('Clustering done.')
# run for each density
if para['parallelMode']: # run on multiple processes
pool = multiprocessing.Pool()
for density in para['density']:
pool.apply_async(evaluator.execute, (dataMatrix, locGroup, density, para))
pool.close()
pool.join()
else: # run on single processes
for density in para['density']:
def main(args=None):
parser = argparse.ArgumentParser(
description='Simple training script for training a RetinaNet network.')
parser.add_argument('--dataset',
help='Dataset type, must be one of csv or coco.')
parser.add_argument('--coco_path', help='Path to COCO directory')
parser.add_argument(
'--csv_train',
help='Path to file containing training annotations (see readme)')
parser.add_argument('--csv_classes',
help='Path to file containing class list (see readme)')
parser.add_argument(
'--csv_val',
help=
'Path to file containing validation annotations (optional, see readme)'
)
parser.add_argument(
'--depth',
help='Resnet depth, must be one of 18, 34, 50, 101, 152',
type=int,
default=50)
parser.add_argument('--epochs',
help='Number of epochs',
type=int,
default=100)
parser = parser.parse_args(args)
# Create the data loaders
if parser.dataset == 'coco':
if parser.coco_path is None:
raise ValueError('Must provide --coco_path when training on COCO,')
#dataset_train = CocoDataset(parser.coco_path, set_name='train2017', transform=transforms.Compose([Normalizer(), Augmenter(), Resizer()]))
#dataset_val = CocoDataset(parser.coco_path, set_name='val2017', transform=transforms.Compose([Normalizer(), Resizer()]))
dataloader_train = load()
else:
raise ValueError(
'Dataset type not understood (must be csv or coco), exiting.')
#sampler = AspectRatioBasedSampler(dataset_train, batch_size=2, drop_last=False)
#dataloader_train = DataLoader(dataset_train, num_workers=3, collate_fn=collater, batch_sampler=sampler)
#if dataset_val is not None:
# sampler_val = AspectRatioBasedSampler(dataset_val, batch_size=1, drop_last=False)
# dataloader_val = DataLoader(dataset_val, num_workers=3, collate_fn=collater, batch_sampler=sampler_val)
# Create the model
# returns [nms_scores, nms_class, transformed_anchors[0, anchors_nms_idx, :]]
if parser.depth == 18:
retinanet = model.resnet18(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 34:
retinanet = model.resnet34(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 50:
retinanet = model.resnet50(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 101:
retinanet = model.resnet101(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 152:
retinanet = model.resnet152(num_classes=dataset_train.num_classes(),
pretrained=True)
else:
raise ValueError(
'Unsupported model depth, must be one of 18, 34, 50, 101, 152')
use_gpu = True
if use_gpu:
retinanet = retinanet.cuda()
retinanet = torch.nn.DataParallel(retinanet).cuda()
retinanet.training = True
optimizer = optim.Adam(retinanet.parameters(), lr=1e-5)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
patience=3,
verbose=True)
loss_hist = collections.deque(maxlen=500)
retinanet.train()
# Freeze batch norm layers
retinanet.module.freeze_bn()
print('Num training images: {}'.format(len(dataset_train)))
for epoch_num in range(parser.epochs):
retinanet.train()
retinanet.module.freeze_bn()
epoch_loss = []
# ~ engine in prototypical network
for iter_num, data_temp in enumerate(
dataloader_train): #iterates through the episodes
try:
optimizer.zero_grad()
#print('size of data')
#print(data.items())
# data is a dictionary with keys: img, annot and scale
#print(data['img'])
#print(data['img'].size())
#(batch size (2), channels (3), width and height of image) (padded by 0 so every image in the batch has the same dimension)
#print(data['annot'])
#print(data['annot'].size())
#(batch size (2), maximum number of annotations per image in the batch, coordinates + class id (5))
# annotations are padded by -1 so every image in the batch has the same number of annotations
#print(data['scale'])
# vector of size 2 (size of batch) with the scale of the image
# same for when using anchors: take the mean excluding values of -1
classes_ids = dataset_train.classes
relevant_ids = [classes_ids[x] for x in data_temp['class']]
sample = []
normalizer = Normalizer()
resizer = Resizer()
for i in range(len(data_temp['x'])):
for j in range(len(data_temp['x'][0])):
idx = data_temp['x'][i][j].item()
img = load_image(idx)
annots = load_annotations(idx)
# only keep annotations for the conisidered classes
annots = annots[np.isin(annots[:, 4], relevant_ids)]
temp = {'img': img, 'annot': annots}
sample.append(resizer(normalizer(temp)))
data = collater(sample)
#print(data)
# now the data is still a dictionary with keys: img, annot and scale
#print(data['img'].size())
#print('test initial image')
#print(data['img'].sum())
#(number of images ~ batch size (=(n_support + n_query)*n_way), channels (=3), width, height)
#print(data['annot'].size())
#(number of images, max number of annotations in those images, coordinates & class (=5))
#print(len(data['scale']))
# list of length number of images, containing the scale for each
#sys.exit()
# need to change classification loss, and format of regression loss to accept the new form of the batch
classification_loss, regression_loss = retinanet(
[data['img'].cuda().float(), data['annot']])
classification_loss = classification_loss.mean()
regression_loss = regression_loss.mean()
loss = classification_loss + regression_loss
if bool(loss == 0):
continue
loss.backward()
torch.nn.utils.clip_grad_norm_(retinanet.parameters(), 0.1)
optimizer.step()
loss_hist.append(float(loss))
epoch_loss.append(float(loss))
print(
'Epoch: {} | Iteration: {} | Classification loss: {:1.5f} | Regression loss: {:1.5f} | Running loss: {:1.5f}'
.format(epoch_num, iter_num, float(classification_loss),
float(regression_loss), np.mean(loss_hist)))
del classification_loss
del regression_loss
except Exception as e:
print(e)
continue
# if parser.dataset == 'coco':
# print('Evaluating dataset')
# coco_eval.evaluate_coco(dataset_val, retinanet)
scheduler.step(np.mean(epoch_loss))
#torch.save(retinanet.module, '{}_retinanet_{}.pt'.format(parser.dataset, epoch_num))
retinanet.eval()
torch.save(retinanet, 'model_final.pt'.format(epoch_num))
