def valid(model, device, dataloader):
sigmod = nn.Sigmoid()
pbar = tqdm(enumerate(dataloader), total=len(dataloader))
ajiMeanList = []
ajiList = []
for idx, (images, labels, bounds) in pbar:
images = images.float().to(device)
sout, sout_0, sout_1, sout_2, sout_3, cout, cout_0, cout_1, cout_2, cout_3 = model(
images)
# 计算损失函数
# 计算aji
# print('输出图片的大小: ',sout.shape)
for index in range(sout.shape[0]):
pred_sout = sigmod(sout[index][0]).cpu().data.numpy()
pred_cout = sigmod(cout[index][0]).cpu().data.numpy()
gt_label0 = labels[index][0]
# 格式化label
gt_label0[gt_label0 != 0] = 1
# 后处理图像并且计算值
# 这个是TAFE需要进行的操作
postProcImg = post_proc(pred_sout - pred_cout,
post_dilation_iter=2)
# 解决一些训练的时候产生的错误
if postProcImg is not None:
aji = get_fast_aji(remap_label(gt_label0),
remap_label(postProcImg))
print(aji)
ajiList.append(aji)
# 全部计算完成之后计算一次数据
ajiArray = np.array(ajiList)
ajiMean = np.mean(ajiArray)
pbar.set_description(f'valid: ----ajiMean---- {ajiMean}')
ajiMeanList.append(ajiMean)
return ajiMeanList
ajis = []
for i, gt in enumerate(gts):
spred = np.zeros(gt.shape, dtype=np.uint16)
for isz, cur_sz in enumerate(size_list):
pred_dir = os.path.join(
pred_root_dir, 'noaug_nodrop_' + losstype + '_iouthsp' +
str(iou_ths) + '_dilation_' + str(idilation) + '_sz' + cur_sz +
'_ds0_dc2' + '_wtta', 'rnd_0', 'pred_res')
dmat = scio.loadmat(os.path.join(pred_dir, str(i) + '.mat'))
s = dmat['instance'].astype(np.uint16)
s_mask = np.logical_and(spred == 0, s > 0)
max_intance_id = spred.max()
s[s_mask] += max_intance_id
spred[s_mask] = s[s_mask]
#spred = s
spred = remap_label(spred)
gt = remap_label(gt)
aji = get_fast_aji(gt, spred)
print(i + 1, aji)
ajis.append(aji)
if if_save:
save_dir = os.path.join(pred_root_dir, 'final_res')
confirm_loc(save_dir)
scio.savemat(os.path.join(save_dir,
str(i) + '.mat'), {
'instance': spred,
'gt': gt
})
avg_aji = np.mean(ajis)
print(avg_aji)
net_val_dir = '/data0/cong/workplace/kumar_5fold/val/baseline_4fold_epoch600'
gts = np.load('/home/cong/workplace/kumar/gt.npy').astype(np.uint16)
for itrain in range(ntrain):
ajis_allcp = []
pred_dir = os.path.join(net_val_dir, 'rnd_' + str(itrain))
for i_cp in range(save_step, max_epoch + 1, save_step):
ajis = []
for i in kfold_val_idx[itrain]:
dmat = scio.loadmat(
os.path.join(pred_dir, 'snapshot_' + str(i_cp),
str(i) + '.mat'))
s = dmat['s']
c = dmat['c']
if len(s.shape) > 2:
s = np.reshape(s, [s.shape[1], s.shape[2]])
if len(c.shape) > 2:
c = np.reshape(c, [c.shape[1], c.shape[2]])
lbl = post_proc(s - c, post_dilation_iter=2).astype(np.uint16)
aji = get_fast_aji(remap_label(gts[i]), remap_label(lbl))
#print(i, aji)
ajis.append(aji)
avg_aji = np.array(ajis).mean()
ajis_allcp.append(avg_aji)
print('Itrain {0:d}: Validation AJI of snapshot {1:d}: {2:.3f}'.format(
itrain, i_cp, avg_aji))
maxaji_allcp = np.array(ajis_allcp).max()
icp = np.array(ajis_allcp).argmax()
np.save(os.path.join(pred_dir, 'val_aji.npy'), np.array(ajis_allcp))
print('Itrain {0:d}: the max validation AJI {1:.3f} in {2:d} snapshot'.
format(itrain, maxaji_allcp, icp))
dir_checkpoint = './weight/noaug_nodrop_focalloss_iouthsp5_dilation_2_sz' + str(
cur_sz) + '_ds0_dc2_ifold' + str(ifold) + '/rnd_0/'
snapshot_list = glob.glob(dir_checkpoint + '*.pth')
for isnapshot_name in snapshot_list:
isnapshot = int(
isnapshot_name.split('/')[-1].split('.')[0].split('_')[-1])
# if isnapshot <= 10:
# continue
val_pred_filefold = './val/noaug_nodrop_focalloss_iouthsp5_dilation_2_sz' + str(
cur_sz) + '_ds0_dc2_ifold' + str(
ifold) + '_wtta/rnd_0/snapshot_' + str(isnapshot)
ajis = []
for i_data in valset[0]:
pred_dmat = scio.loadmat(
os.path.join(val_pred_filefold,
str(i_data) + '.mat'))
pred_lbl_map = pred_dmat['instance'].copy()
if len(np.unique(pred_lbl_map[pred_lbl_map > 0])) >= 1:
pred_lbl_map = remap_label(pred_lbl_map)
aji = get_fast_aji(remap_label(gts[i_data]), pred_lbl_map)
ajis.append(aji)
avg_aji = np.mean(ajis)
if avg_aji >= cur_bset_aji:
cur_bset_aji = avg_aji
cur_best_aji_isnapshot = isnapshot
info2print = 'Size: {:s}, snapshot: {:d}, cur aji: {:.5f}, best aji: {:.5f} in isnapshot: {:d}'.format(
cur_sz, isnapshot, avg_aji, cur_bset_aji,
cur_best_aji_isnapshot)
print(info2print)
f.write(info2print + '\n')
for i in valset:
lbl = None
for cursz in [48, 176]:
dir_savefile = './val/noaug_nodrop_bceloss_iouthsp' + 'ths' + str(
ths) + '_dilation_' + str(dilation_rate) + '_sz' + str(
cursz) + '_ds' + str(dilation_s) + '_dc' + str(
dilation_c) + '_wtta/rnd_' + str(irnd) + '/'
dmat = scio.loadmat(
os.path.join(dir_savefile, 'snapshot_' + str(i_cp),
str(i) + '.mat'))
if lbl is None:
lbl = dmat['instance']
else:
lbl_tmp = lbl.copy()
max_num = instance_tmp.max()
lbl[instance_tmp ==
0] = dmat['instance'][lbl_tmp == 0] + max_num
lbl = remap_label(lbl.astype(np.uint16))
aji = get_fast_aji(remap_label(gts[i]), remap_label(lbl))
#print(i, aji)
ajis.append(aji)
avg_aji = np.array(ajis).mean()
ajis_allcp.append(avg_aji)
print('Itrain {0:d}: Validation AJI of snapshot {1:d}: {2:.3f}'.format(
itrain, i_cp, avg_aji))
maxaji_allcp = np.array(ajis_allcp).max()
icp = np.array(ajis_allcp).argmax()
np.save(os.path.join(pred_dir, 'val_aji.npy'), np.array(ajis_allcp))
print('Itrain {0:d}: the max validation AJI {1:.3f} in {2:d} snapshot'.format(
itrain, maxaji_allcp, icp))