def test():
model.eval()
test_loss = 0
total_length = 0
for data, target in test_loader:
if CUDA_ON:
data, target = data.cuda(), target.cuda()
data, target = Var(data, volatile=True), Var(target)
output = nn.functional.sigmoid(model(data))
for i in range(output.size(0)):
try:
if target.data[i].sum() == 0: continue
if target.data[i].sum() >= 12: continue
test_loss += accuracy(output.data[i], target.data[i])
total_length += 1
except Exception as e:
print(e)
print(target.data[i])
sys.exit()
return test_loss / total_length
def test_3rd_snapshot():
te_set = PosterSet(POSTER_PATH, split, 'test', gen_d=gen_d, augment=False, resize=None, ten_crop=None)#, debug=True)
te_load = DataLoader(te_set, batch_size=64, shuffle=False, num_workers=3, drop_last=True)
model = SmallerNetwork(INP_SIZE, 23)
state = torch.load(SNAP_PATH + "snap3rd.nn")
model.load_state_dict(state['state_dict'])
if CUDA_ON: model.cuda()
model.eval()
loss = 0
skipped = 0
for X, y in tqdm(te_load, desc='3rd'):
X, y = Var(X, volatile=True), Var(y)
if CUDA_ON:
X, y = X.cuda(), y.cuda()
out = model(X)
for i in range(out.size(0)):
try:
loss += accuracy(out.data[i], y.data[i])
except:
skipped += 1
return loss / (len(te_set) - skipped)
def test_5th_snapshot():
te_set = PosterSet(POSTER_PATH, split, 'test', gen_d=gen_d, augment=False, resize=None, ten_crop=CROP_SIZE)#, debug=True)
te_load = DataLoader(te_set, batch_size=64, shuffle=False, num_workers=3, drop_last=True)
model = MidrangeNetwork(CROP_SIZE, 23)
state = torch.load(SNAP_PATH + "snap5th.nn")
model.load_state_dict(state['state_dict'])
if CUDA_ON: model.cuda()
model.eval()
loss = 0
skipped = 0
for X, y in tqdm(te_load, desc='5th'):
X, y = Var(X, volatile=True), Var(y)
bs, ncrops, c, h, w = X.size()
if CUDA_ON:
X, y = X.cuda(), y.cuda()
out = model(X.view(-1, c, h, w))
out = out.view(bs, ncrops, -1).mean(1)
for i in range(out.size(0)):
try:
loss += accuracy(out.data[i], y.data[i])
except:
skipped += 1
return loss / (len(te_set) - skipped)
def test_distribution():
te_set = PosterSet(POSTER_PATH, split, 'test', gen_d=gen_d, augment=False, resize=None, ten_crop=None)#, debug=True)
te_load = DataLoader(te_set, batch_size=64, shuffle=False, num_workers=3, drop_last=True)
pred = torch.Tensor([0.28188283, 0.23829031, 0.27430824, 0.39426496, 0.38900912, 0.22754676, 0.24563302, 0.11153192, 0.29865512, 0.14260318, 0.17011903, 0.0307621 , 0.20026279, 0.2485701 , 0.24037718, 0.17645695, 1. , 0.42100788, 0.11593755, 0.31264492, 0.62699026, 0.1946205 , 0.27446282])
loss = 0
skipped = 0
for X, y in tqdm(te_load, desc='dis'):
for i in range(X.size(0)):
try:
loss += accuracy(pred, y[i])
except:
skipped += 1
return loss / (len(te_set) - skipped)
def loss(self,
cls_score,
bbox_pred,
labels,
label_weights,
bbox_targets,
bbox_weights,
reduce=True):
losses = dict()
if cls_score is not None:
losses['loss_cls'] = weighted_cross_entropy(cls_score,
labels,
label_weights,
reduce=reduce)
losses['acc'] = accuracy(cls_score, labels)
if bbox_pred is not None:
losses['loss_reg'] = weighted_smoothl1(
bbox_pred,
bbox_targets,
bbox_weights,
avg_factor=bbox_targets.size(0))
return losses