def proc_dcrf():
names = os.listdir(path_save_train_voc_prob)
pool = Pool(4)
pool.map(one_dcrf, names)
#for name in tqdm(names):
# one_dcrf(name)
miou = evaluate_iou(path_save_train_voc_crf, voc_save_gt_dir, c_output)
return miou
def test(model):
print("============================= TEST ============================")
model.switch_to_eval()
for i, (img, name, WW, HH) in tqdm(enumerate(voc_val_loader), desc='testing'):
model.test(img, name, WW, HH)
model.switch_to_train()
miou = evaluate_iou(opt.results_dir, voc_val_gt_dir, c_output)
model.performance = {'miou': miou}
return miou
def test(model):
print("============================= TEST ============================")
model.switch_to_eval()
for i, (img, name, WW, HH) in tqdm(enumerate(voc_test_loader), desc='testing'):
model.test(img, name, WW, HH)
model.switch_to_train()
miou = evaluate_iou(opt.results_dir, voc_val_gt_dir, c_output)
model.performance = {'miou': miou}
print(miou)
with open('val_voc.json', 'w') as f:
json.dump(model.performance, f)
return model.performance
def val_voc():
net.eval()
with torch.no_grad():
for t in range(10):
for it, (img, gt, batch_name, WW,
HH) in tqdm(enumerate(voc_loader), desc='train'):
gt_cls = gt[:, None, ...] == torch.arange(c_output)[None, ...,
None, None]
gt_cls = (gt_cls.sum(3, keepdim=True).sum(2, keepdim=True) >
0).float().cuda()
raw_img = img
img = (img.cuda() - mean) / std
batch_seg, _, seg32x = net(img)
batch_seg[:, 1:] *= gt_cls[:, 1:]
batch_prob = F.softmax(batch_seg, 1)
_, batch_seg = batch_seg.detach().max(1)
for n, name in enumerate(batch_name):
msk = batch_seg[n]
prob = batch_prob[n]
prob = prob.detach().cpu().numpy()
np.save(
'{}/{}_{}.npy'.format(path_save_train_voc_prob, name,
t), prob)
_img = raw_img[n] * 255
_img = _img.numpy().astype(np.uint8)
_img = _img.transpose((1, 2, 0))
_img = Image.fromarray(_img)
_img.save('{}/{}_{}.jpg'.format(voc_save_img_dir, name, t))
_gt = gt[n]
_gt = _gt.numpy()
_gt = _gt.astype(np.uint8)
_gt = Image.fromarray(_gt.astype(np.uint8))
_gt = _gt.convert('P')
_gt.putpalette(palette_voc)
_gt.save('{}/{}_{}.png'.format(voc_save_gt_dir, name, t),
'PNG')
msk = msk.detach().cpu().numpy()
msk = Image.fromarray(msk.astype(np.uint8))
msk = msk.convert('P')
msk.putpalette(palette_voc)
msk.save(
'{}/{}_{}.png'.format(path_save_train_voc, name, t),
'PNG')
miou = evaluate_iou(path_save_train_voc, voc_save_gt_dir, c_output)
net.train()
return miou
def val_voc():
net.eval()
with torch.no_grad():
for it, (img, gt, batch_name, WW,
HH) in tqdm(enumerate(voc_val_loader), desc='train'):
img = (img.cuda() - mean) / std
outputs = net(img)
batch_seg = outputs[0]
_, batch_seg = batch_seg.detach().max(1)
for n, name in enumerate(batch_name):
msk = batch_seg[n]
msk = msk.detach().cpu().numpy()
w = WW[n]
h = HH[n]
msk = Image.fromarray(msk.astype(np.uint8))
msk = msk.convert('P')
msk.putpalette(palette_voc)
msk = msk.resize((w, h))
msk.save('{}/{}.png'.format(path_save_valid_voc, name), 'PNG')
miou = evaluate_iou(path_save_valid_voc, voc_val_gt_dir, c_output)
net.train()
return miou
def syn(model):
print("============================= TEST ============================")
model.switch_to_eval()
_rdir = model.opt.results_dir
model.opt.results_dir = '/'.join(model.opt.results_dir.split('/')[:-1])+'/syn_train'
if not os.path.exists(model.opt.results_dir):
os.mkdir(model.opt.results_dir)
for i, (img, gt, name, WW, HH) in tqdm(enumerate(voc_train_loader), desc='testing'):
bsize = img.size(0)
gt[gt == -1] = c_output
gt = torch.zeros(bsize, c_output + 1, opt.imageSize, opt.imageSize).scatter(1, gt.unsqueeze(1), 1)
gt = gt.sum(3).sum(2)
gt = gt[:, :c_output]
gt = gt[:, 1:]
gt[gt>0] = 1
model.test(img, name, WW, HH, gt, save_prob=True)
model.switch_to_train()
miou = evaluate_iou(model.opt.results_dir, voc_train_gt_dir, c_output)
model.opt.results_dir = _rdir
model.performance = {'miou': miou}
print(miou)
with open(model.save_dir+'/'+'syn-log.json', 'w') as outfile:
json.dump(model.performance, outfile)
return miou