sampler=val_sampler)
logger('>> N samples: Train: %d + Val: %d' % (len(idxs_train), len(idxs_val)))
# Test sets
dset_sg = ParquetDataset(
'IMG/DoublePi0Pt10To100_m0To1600_pythia8_ReAOD_PU2017_MINIAODSIM_wrapfix.tzfixed_wgts.val.parquet',
1)
sg_loader = DataLoader(dataset=dset_sg, batch_size=256, num_workers=10)
dset_bg = ParquetDataset(
'IMG/DoublePhotonPt10To100_pythia8_ReAOD_PU2017_MINIAODSIM_wrapfix.tzfixed_m0Neg%dTo0_wgts.val.parquet'
% args.neg_mass, 0)
bg_loader = DataLoader(dataset=dset_bg, batch_size=256, num_workers=10)
logger('>> N test samples: sg: %d + bg: %d' % (len(dset_sg), len(dset_bg)))
import torch_resnet_concat as networks
resnet = networks.ResNet(2, resblocks, [16, 32])
resnet.cuda()
optimizer = optim.Adam(resnet.parameters(), lr=lr_init)
#lr_scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=[10,20], gamma=0.5)
def do_eval(resnet, val_loader, mae_best, epoch, sample, tgt_label):
global expt_name
loss_ = 0.
m_pred_, m_true_, mae_, pt_, wgts_ = [], [], [], [], []
iphi_, ieta_ = [], []
label_ = []
now = time.time()
for i, data in enumerate(val_loader):
X, m0, pt, wgts = data['Xtz_aod'].cuda(), data['m'].cuda(
), data['pt'], data['w']
# !! NOTE: at fill time, need to check that ggtree.nPho corresponds to
# no. of presumed photon objects to be regressed, i.e.:
# ggtree.nPho == len(imgtree.SC_ieta) or len(imgtree.SC_iphi)
# The `ma` branch still needs to be initialized with an array of some
# arbitrarily large allocation `nPhoMax` which then gets reduced at fill time.
nPhoMax = 20
#ma = array('f', nPhoMax*[0.])
ma = np.zeros(nPhoMax, dtype='float32')
tree_out.Branch('ma', ma, 'ma[nPho]/F')
#'''
# Load mass regressor for inference
print('Loading regressor model')
import torch_resnet_concat as networks
model_file = 'Models/model_epoch80_mae0.1906.pkl'
resnet = networks.ResNet(2, 3, [16, 32], 128, 0).cuda()
resnet.load_state_dict(torch.load(model_file)['model'])
resnet.eval()
print(model_file)
#'''
eb_scale = 25.
m0_scale = 1.6
def transform_y(y):
return y / m0_scale
def inv_transform(y):
return y * m0_scale