def __init__(self,
path='',
title='',
params=None,
resume=False,
data_format='csv',
epochs=0,
index='epoch'):
"""
Parameters
----------
path: string
path to directory to save data files
plot_path: string
path to directory to save plot files
title: string
title of HTML file
params: Namespace
optionally save parameters for results
resume: bool
resume previous logging
data_format: str('csv'|'json')
which file format to use to save the data
"""
if data_format not in ResultsLog.supported_data_formats:
raise ValueError('data_format must of the following: ' + '|'.join(
['{}'.format(k) for k in ResultsLog.supported_data_formats]))
if data_format == 'json':
self.data_path = '{}.json'.format(path)
else:
self.data_path = '{}.csv'.format(path)
self.data_format = data_format
if params is not None:
export_args_namespace(params, '{}.json'.format(path))
self.plot_path = '{}.html'.format(path)
self.results = None
self.clear()
self.first_save = True
if os.path.isfile(self.data_path):
if resume:
self.load(self.data_path)
self.first_save = False
else:
os.remove(self.data_path)
self.results = pd.DataFrame()
else:
self.results = pd.DataFrame()
self.title = title
self.epochs = epochs
self.index = index
if HYPERDASH_AVAILABLE:
name = self.title if title != '' else path
self.hd_experiment = hyperdash.Experiment(name)
if params is not None:
for k, v in params._get_kwargs():
self.hd_experiment.param(k, v, log=False)
def hyper_dash_manager(exp_name, debug=False):
"""
debug=True の時に、hyperdashにログを送信しないダミーインスタンスを返す。
"""
if debug:
exp = debugHyperDash()
else:
exp = hyperdash.Experiment(exp_name)
try:
yield exp
finally:
exp.end()
def test(arguments):
log.debug('Creating network')
net = ln.models.Yolo(CLASSES, arguments.weight, CONF_THRESH, NMS_THRESH)
net.postprocess.append(
ln.data.transform.TensorToBrambox(NETWORK_SIZE, LABELS))
net.eval()
if arguments.cuda:
net.cuda(non_blocking=PIN_MEM)
log.debug('Creating dataset')
loader = torch.utils.data.DataLoader(
CustomDataset(TESTFILE, net),
batch_size=MINI_BATCH,
shuffle=False,
drop_last=False,
num_workers=WORKERS if arguments.cuda else 0,
pin_memory=PIN_MEM if arguments.cuda else False,
collate_fn=ln.data.list_collate,
)
if arguments.visdom:
log.debug('Creating visdom visualisation wrappers')
vis = visdom.Visdom(port=VISDOM_PORT)
plot_pr = ln.engine.VisdomLinePlotter(vis,
'pr',
opts=dict(
xlabel='Recall',
ylabel='Precision',
title='Precision Recall',
xtickmin=0,
xtickmax=1,
ytickmin=0,
ytickmax=1,
showlegend=True))
if arguments.hyperdash:
log.debug('Creating hyperdash visualisation wrappers')
hd = hyperdash.Experiment('YOLOv2 Test')
hyperdash_plot_pr = ln.engine.HyperdashLinePlotter(hd)
log.debug('Running network')
tot_loss = []
coord_loss = []
conf_loss = []
cls_loss = []
anno, det = {}, {}
for idx, (data, box) in enumerate(tqdm(loader, total=len(loader))):
if arguments.cuda:
data = data.cuda(non_blocking=PIN_MEM)
data = torch.autograd.Variable(data).no_grad()
output, loss = net(data, box)
if torch.__version__.startswith('0.3'):
data = torch.autograd.Variable(data).no_grad()
output, loss = net(data, box)
else:
with torch.no_grad():
output, loss = net(data, box)
if torch.__version__.startswith('0.3'):
tot_loss.append(net.loss.loss_tot.data[0] * len(box))
coord_loss.append(net.loss.loss_coord.data[0] * len(box))
conf_loss.append(net.loss.loss_conf.data[0] * len(box))
if net.loss.loss_cls is not None:
cls_loss.append(net.loss.loss_cls.data[0] * len(box))
else:
tot_loss.append(net.loss.loss_tot.item() * len(box))
coord_loss.append(net.loss.loss_coord.item() * len(box))
conf_loss.append(net.loss.loss_conf.item() * len(box))
if net.loss.loss_cls is not None:
cls_loss.append(net.loss.loss_cls.item() * len(box))
key_val = len(anno)
anno.update(
{loader.dataset.keys[key_val + k]: v
for k, v in enumerate(box)})
det.update({
loader.dataset.keys[key_val + k]: v
for k, v in enumerate(output)
})
pr_dict, ap_dict, m_ap, all_key = bbb.pr_ap_dicts(det, anno, LABELS,
IGNORE)
pr = pr_dict[all_key]
log.info('Computed statistics')
for label in sorted(ap_dict.keys()):
log.info('\tLabel %r: m_ap = %0.04f' % (
label,
ap_dict[label],
))
tot = round(sum(tot_loss) / len(anno), 5)
coord = round(sum(coord_loss) / len(anno), 2)
conf = round(sum(conf_loss) / len(anno), 2)
if len(cls_loss) > 0:
cls = round(sum(cls_loss) / len(anno), 2)
log.test(
'\n{seen} mAP:{m_ap}% Loss:{tot} (Coord:{coord} Conf:{conf} Cls:{cls})'
.format(seen=net.seen // BATCH,
m_ap=m_ap,
tot=tot,
coord=coord,
conf=conf,
cls=cls))
else:
log.test('\n{seen} mAP:{m_ap}% Loss:{tot} (Coord:{coord} Conf:{conf})'.
format(seen=net.seen // BATCH,
m_ap=m_ap,
tot=tot,
coord=coord,
conf=conf))
name = 'mAP: {m_ap}%'.format(m_ap=m_ap)
if arguments.visdom:
plot_pr(np.array(pr[0]), np.array(pr[1]), name=name)
if arguments.hyperdash:
now = time.time()
re_seen = None
for index, (re_, pr_) in enumerate(sorted(zip(pr[1], pr[0]))):
re_ = round(re_, 2)
if re_ != re_seen:
re_seen = re_
re_ = int(re_ * 100.0)
hyperdash_plot_pr(name, pr_, now + re_, log=False)
if arguments.save_det is not None:
# Note: These detection boxes are the coordinates for the letterboxed images,
# you need ln.data.ReverseLetterbox to have the right ones.
# Alternatively, you can save the letterboxed annotations, and use those for statistics later on!
bbb.generate('det_pickle', det,
Path(arguments.save_det).with_suffix('.pkl'))
#bbb.generate('anno_pickle', det, Path('anno-letterboxed_'+arguments.save_det).with_suffix('.pkl'))
if arguments.hyperdash:
hyperdash_plot_pr.close()
# Parse arguments
if args.cuda:
if not torch.cuda.is_available():
log.debug('CUDA not available')
args.cuda = False
else:
log.debug('CUDA enabled')
if args.visdom:
args.visdom = visdom.Visdom(port=VISDOM_PORT)
else:
args.visdom = None
if args.hyperdash:
args.hyperdash = hyperdash.Experiment('YOLOv2 Pascal VOC Train')
else:
args.hyperdash = None
if not os.path.isdir(args.backup):
if not os.path.exists(args.backup):
log.warn('Backup folder does not exist, creating...')
os.makedirs(args.backup)
else:
raise ValueError('Backup path is not a folder')
# Train
eng = VOCTrainingEngine(args)
b1 = eng.batch
t1 = time.time()
eng()
# Parse arguments
if args.cuda:
if not torch.cuda.is_available():
log.error('CUDA not available')
args.cuda = False
else:
log.debug('CUDA enabled')
if args.visdom:
args.visdom = visdom.Visdom(port=VISDOM_PORT)
else:
args.visdom = None
if args.hyperdash:
args.hyperdash = hyperdash.Experiment('YOLOv2 Train')
else:
args.hyperdash = None
if not os.path.isdir(args.backup):
if not os.path.exists(args.backup):
log.warn('Backup folder does not exist, creating...')
os.makedirs(args.backup)
else:
raise ValueError('Backup path is not a folder')
# Train
eng = TrainingEngine(args)
b1 = eng.batch
t1 = time.time()
eng()