def predict(predict_data):
machine = predict_data.get('machineId')
first_date = predict_data.get('first_date')
last_date = predict_data.get('last_date')
analysis.predict(machine, first_date, last_date)
return make_response("Successfully predicted on given period", 200)
def analysismenu_print(v, f, d):
print(
"=============MENU=============\n1 - Regression Analysis Summary\n2 - Predict a Players Dispossession "
"Value\n3 - Change Data")
choice = int(input("============CHOICE============ \nInput - "))
if choice == 1:
from analysis import summary
summary(v, f, d)
if choice == 2:
from analysis import predict
predict(v, f, d)
def predict(target_data=[],
features=[],
labels=[],
dataset_path='models/data.csv',
model_name='fraud_model'):
return analysis.predict(target_data, features, labels, dataset_path,
model_name)
parser.add_argument('--gpus', default='0',
type=str, help='visible devices for CUDA')
args = parser.parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpus
test_gts = torch.load({
'VOC': 'truths\\gts_voc07test.pth',
'COCO': 'truths\\gts_coco_17val.pth',
'SHWD': 'truths\\gts_shwd_test.pth'
}[args.dataset]).float().cuda()
print('name\tloss\tpower1/3\tgeomean\tmean\trecall\tpower3\tbestgts')
template = '\t%.2f\t%.4f\t%.4f\t%.4f\t%.4f\t%.4f\t%.4f'
bl = torch.load('anchors/ssd_coco_anchors.pth')
bl = bl.cuda()
_, bl_a = predict(bl, test_gts, True)
print('ssdcoco' + template % tuple(bl_a))
bl = torch.load('anchors/voc_baseline.pth')
bl = bl.cuda()
_, bl_a = predict(bl, test_gts, True)
print('ssdvoc' + template % tuple(bl_a))
bl = torch.load('anchors/yolov3_anchors.pth')
bl = bl.detach().cuda()
_, bl_a = predict(bl, test_gts, True)
print('yolo' + template % tuple(bl_a))
test_lst = [tuple(o.split(',')) for o in args.test_pths.split(';')]
for i, (algo, pth) in enumerate(test_lst):
if algo == 'SSD300':
cfg = config_dict[args.dataset]
def train():
if args.algo == 'SSD300':
apt = IOAdapterSSD(config_dict[args.dataset], args.random_init)
# args.algo == 'YOLOv3'
else:
apt = IOAdapterYOLOv3(random_range=args.random_init)
# init params
anchs, anch2fmap, fmap2locs, msks = apt.fit_input((1,))
# create data loader
data_loader = BoundingBoxesLoader(dataset, None, args.batch_size, shuffle=True,
drop_last=True, cache_pth=args.truths_pth)
b_iter = iter(data_loader)
# create optimizer
# optimizer = optim.SGD(params + [alphas], lr=args.lr, momentum=args.momentum,
# weight_decay=args.weight_decay)
optimizer = optim.SGD([anchs], lr=args.lr, momentum=args.momentum,
weight_decay=args.weight_decay)
# create loss function
loss_fn = MixedIOULoss(ignore_size=args.ignore_size, lambda_=args.l)
gen_fn = AnchorsGenerator(anchs, anch2fmap, fmap2locs)
step = 0
# train
for iteration in range(6000):
try:
truths = next(b_iter)
except StopIteration:
b_iter = iter(data_loader)
truths = next(b_iter)
if iteration in (4000, 5000):
step += 1
adjust_learning_rate(optimizer, 0.1, step)
truths = truths.float().cuda() if args.cuda else truths.float()
optimizer.zero_grad()
loss = torch.zeros(8)
for i, msk in enumerate(msks):
tmp_anchs = gen_fn(msk)
loss[i] = loss_fn(tmp_anchs, truths)
loss = loss.sum()
loss.backward()
optimizer.step()
with torch.no_grad():
anchs.clamp_(0., 1.)
if args.log:
writer.add_scalar(args.save_pth, loss.item(), iteration + 1)
if (iteration + 1) % 10 == 0:
print('iter %d: loss=%.4f' % (iteration + 1, loss.item()))
if (iteration + 1) % args.cache_interval == 0:
if not os.path.exists('./cache/'):
os.mkdir('./cache/')
pth = './cache/%s_iter%d.pth' % (file_name, iteration + 1)
torch.save((anchs, anch2fmap, fmap2locs, msks), pth)
print('save cache to %s ' % pth)
if args.test_per_cache:
print(bl_results)
with torch.no_grad():
maps = []
for i, msk in enumerate(msks):
tmp_anchs = gen_fn(msk)
maps.append(predict(tmp_anchs, test_gts, True))
print('\n'.join(['%dxAnchs = %.4f [loss:%.2f|power1/3:%.4f|geo mean:%.4f|'
'mean:%.4f|recall:%.4f|power3:%.4f|best gt:%.4f]'
% (i + 1, o, *l) for i, (o, l) in enumerate(maps)]))
for fmap, locs in fmap2locs.items():
fmap2locs[fmap] = locs.cpu()
msks = [msk.cpu() for msk in msks]
torch.save((anchs.detach().cpu(), anch2fmap, fmap2locs, msks), args.save_pth)
if args.algo == 'YOLOv3':
print(apt.fit_output())
file_name = get_file_name_from_path(args.save_pth)
if args.log and 'train' in modes:
from datetime import datetime
writer = SummaryWriter('runs/adaptive_priors_loss/%s/' % datetime.now().strftime("%Y%m%d-%H%M%S"))
if args.cuda:
torch.set_default_tensor_type('torch.cuda.FloatTensor')
dataset = None
clamp = True
test_gts = torch.load(args.eval_pth).float().cuda()
bl = torch.load('anchors/yolov3_anchors.pth')
bl = bl.detach().cuda()
bl_y, bl_a = predict(bl, test_gts, True)
bl_results = 'baseline = %.4f [loss:%.2f|power1/3:%.4f|geo mean:%.4f|mean:%.4f|recall:%.4f|power3:%.4f|best gt:%.4f]'\
% (bl_y, *bl_a)
def train():
if args.algo == 'SSD300':
apt = IOAdapterSSD(config_dict[args.dataset], args.random_init)
# args.algo == 'YOLOv3'
else:
apt = IOAdapterYOLOv3(random_range=args.random_init)
# init params
anchs, anch2fmap, fmap2locs, msks = apt.fit_input((1,))
# create data loader
import numpy as np
import analysis
dataset_path = 'models/data.csv'
model_name = 'fraud_model'
features = [
'userId', 'profileCompleted', 'visitedCountries', 'depositAmount',
'totalSpendingsInt', 'spendings', 'maxDailySpending'
]
labels = []
test_data = [191670, 1, 2, 2492270, 14, 82939, 20000]
target_data = np.array(test_data).reshape(-1, len(test_data))
x = analysis.predict(target_data, features, labels, dataset_path, model_name,
True)
y = x["prediction"]
print(y)
z = x["pred"]