def recall(model, loader, args, it):
split = loader.dataset.split
keys = []
if args.dtp_only:
features = edtp.extract_features(model, loader, args, args.numpy)
res = postprocessing_dtp(features, loader, args)
for ot in args.overlap_thresholds:
keys += ['%d_total_recall_%d' % (i, ot * 100) for i in range(1, 4)]
keys += ['%d_total_proposals' % i for i in range(1, 4)]
else:
features = et.extract_features(model, loader, args, args.numpy)
res = postprocessing(features, loader, args)
for i in range(1, 4):
for ot in args.overlap_thresholds:
keys += [
'%d_dtp_recall_%d' % (i, ot * 100),
'%d_rpn_recall_%d' % (i, ot * 100),
'%d_total_recall_%d' % (i, ot * 100)
]
for i in range(1, 4):
keys += [
'%d_total_proposals' % i,
'%d_dtp_proposals' % i,
'%d_rpn_proposals' % i
]
avg = average_dictionary(res, keys, scale=True)
log = '\n\n Fold %d\n' % loader.dataset.fold
log += '--------------------------------\n'
if args.dtp_only:
for ot in args.overlap_thresholds:
for i, rtype in enumerate(['a', 't', 'n']):
total_recall = avg['%d_total_recall_%d' % (i + 1, ot * 100)]
total_np = avg['%d_total_proposals' % (i + 1)] / 100
pargs = (rtype, total_recall, total_np, ot * 100)
rs2 = '%s: total: %.2f, %.2f using %d%% overlap' % pargs
log += '[%s set] %s\n' % ('test', rs2)
else:
for ot in args.overlap_thresholds:
for i, rtype in enumerate(['a', 't', 'n']):
rpn_recall = avg['%d_rpn_recall_%d' % (i + 1, ot * 100)]
dtp_recall = avg['%d_dtp_recall_%d' % (i + 1, ot * 100)]
total_recall = avg['%d_total_recall_%d' % (i + 1, ot * 100)]
rpn_np = avg['%d_rpn_proposals' % (i + 1)] / 100
total_np = avg['%d_total_proposals' % (i + 1)] / 100
dtp_np = avg['%d_dtp_proposals' % (i + 1)] / 100
# pargs = (rtype, rpn_recall, dtp_recall, total_recall, ot*100)
pargs = (rtype, rpn_recall, rpn_np, dtp_recall, dtp_np,
total_recall, total_np, ot * 100)
# rs2 = '%s: rpn_recall: %.2f, dtp_recall: %.2f, total recall: %.2f using %d%% overlap' % pargs
rs2 = '%s: rpn: %.2f, %d, dtp: %.2f, %d, total: %.2f, %d using %d%% overlap' % pargs
log += '[%s set] %s\n' % (split, rs2)
log += '--------------------------------\n'
return log, avg
def final_log(dicts, keys, title):
res = average_dictionary(dicts, keys, False, True)
pargs = (res.mAP_qbe_25, res.mAP_qbs_25, res.mR_qbe_25, res.mR_qbs_25)
s1 = 'QbE mAP: %.1f, QbS mAP: %.1f, QbE mR: %.1f, QbS mR: %.1f, 25%% overlap' % pargs
pargs = (res.mAP_qbe_50, res.mAP_qbs_50, res.mR_qbe_50, res.mR_qbs_50)
s2 = 'QbE mAP: %.1f, QbS mAP: %.1f, QbE mR: %.1f, QbS mR: %.1f, 50%% overlap' % pargs
log = '%s\n--------------------------------\n' % title
log += '[test set] %s\n' % s1
log += '[test set] %s\n' % s2
log += '--------------------------------\n'
return log
opt.weights = opt.weights[:s] + ('fold%d' % fold) + opt.weights[e:]
opt.fold = fold
if opt.h5:
testset = h5_dataset.H5Dataset(opt, split=2)
opt.num_workers = 0
else:
testset = datasets.Dataset(opt, 'test')
loader = DataLoader(testset,
batch_size=1,
shuffle=False,
num_workers=0)
model.load_weights(opt.weights)
model.cuda()
log, rf, rt = mAP(model, loader, args, 0)
rt['log'] = average_dictionary(rt['log'], r_keys)
copy_log(rt)
print(log)
rts.append(rt)
else:
log, _, avg = mAP(model, loader, args, 0)
avg['log'] = average_dictionary(avg['log'], r_keys)
copy_log(avg)
rts = [avg]
print(log)
def final_log(dicts, keys, title):
res = average_dictionary(dicts, keys, False, True)
pargs = (res.mAP_qbe_25, res.mAP_qbs_25, res.mR_qbe_25, res.mR_qbs_25)
if opt.h5:
testset = h5_dataset.H5Dataset(opt, split=2)
opt.num_workers = 0
else:
testset = datasets.Dataset(opt, 'test')
loader = DataLoader(testset,
batch_size=1,
shuffle=False,
num_workers=0)
model = ctrlf.CtrlFNet(opt)
model.load_state_dict(torch.load(opt.weights))
model.cuda()
log, rf, rt = mAP(model, loader, args, 0)
print(log)
rt['log'] = average_dictionary(rt['log'], r_keys)
rf['log'] = average_dictionary(
rf['log'], ['3_total_recall_50', '3_total_recall_25'])
copy_log(rt, rf)
rts.append(rt)
rfs.append(rf)
avg_rts = average_dictionary(rts, keys + r_keys)
avg_rfs = average_dictionary(
rfs, keys + ['3_total_recall_50', '3_total_recall_25'])
else:
log, rf, rt = mAP(model, loader, args, 0)
print(log)
#average pagewise recalls
rt['log'] = average_dictionary(rt['log'], r_keys)
testset = h5_dataset.H5Dataset(opt, split=2)
opt.num_workers = 0
else:
testset = datasets.Dataset(opt, 'test')
loader = DataLoader(testset,
batch_size=1,
shuffle=False,
num_workers=0)
model.load_state_dict(torch.load(opt.weights))
model.cuda()
log, avg = recall(model, loader, args, 0)
avg.append(avg)
avg = average_dictionary(avg, r_keys)
else:
if opt.h5:
testset = h5_dataset.H5Dataset(opt, split=2)
opt.num_workers = 0
else:
testset = datasets.Dataset(opt, 'test')
loader = DataLoader(testset, batch_size=1, shuffle=False, num_workers=0)
log, avg = recall(model, loader, args, 0)
log = '\n\n'
log += 'Averaged over folds\n'
log += '--------------------------------\n'
if dtp_only or not args.use_external_proposals:
for ot in args.overlap_thresholds: