def __init__(self, log_path, resume=True):
self.log_path = log_path
check_mkdir(dir_path=os.path.dirname(self.log_path))
if not resume:
open(self.log_path, encoding="utf-8", mode="w").close()
self.record_str(f"=== {datetime.now()} ===")
def main():
check_mkdir(path_config["save"])
check_mkdir(path_config["pth"])
# shutil.copy(f"{user_config['proj_root']}/config.py", path_config["cfg_log"])
# shutil.copy(f"{user_config['proj_root']}/train.py", path_config["trainer_log"])
if user_config["is_distributed"]:
construct_print("We will use the distributed training.")
args.world_size = args.ngpus_per_node * args.nodes
torch.multiprocessing.spawn(
main_worker,
nprocs=args.ngpus_per_node,
args=(args.ngpus_per_node, args.world_size),
)
else:
construct_print("We will not use the distributed training.")
main_worker(
local_rank=0,
ngpus_per_node=1,
world_size=1,
)
tb_recorder.close_tb()
def __init__(self, tb_path):
check_mkdir(tb_path)
self.tb = SummaryWriter(tb_path)
def validate(epoch):
net.eval()
val_loss = AverageMeter()
inputs_all, labels_all, predictions_all = [], [], []
for i, (inputs, labels) in enumerate(valloader):
if args.cuda:
inputs, labels = inputs.cuda(), labels.cuda()
N = inputs.size(0)
outputs = net(inputs)
# predictions 为输入图片尺寸大小的对应每一像素点的分类值
predictions = outputs.data.max(1)[1].squeeze_(1).squeeze_(
0).cpu().numpy()
loss = criterion(outputs, labels) / N
val_loss.update(loss.item(), N)
if random.random() > args.valImgSampleRate:
inputs_all.append(None)
else:
inputs_all.append(inputs.data.squeeze_(0).cpu())
labels_all.append(labels.data.squeeze_(0).cpu().numpy())
predictions_all.append(predictions)
# 计算本次epoch之后对验证集的正确率等评价指标
acc, acc_cls, mean_iu, fwavacc = evaluate(predictions_all, labels_all,
num_classes)
if mean_iu > best_record['mean_iu']:
best_record['val_loss'] = val_loss.avg
best_record['epoch'] = epoch
best_record['acc'] = acc
best_record['acc_cls'] = acc_cls
best_record['mean_iu'] = mean_iu
best_record['fwavacc'] = fwavacc
snapshot_name = 'epoch_%d_loss_%.5f_acc_%.5f_acc-cls_%.5f_mean-iu_%.5f_fwavacc_%.5f_lr_%.10f' % (
epoch, val_loss.avg, acc, acc_cls, mean_iu, fwavacc, args.lr)
torch.save(net.state_dict(),
os.path.join(ckpt_path, exp_name, snapshot_name + '.pth'))
#torch.save(optimizer.state_dict(), os.path.join(ckpt_path, exp_name, 'opt_' + snapshot_name + '.pth'))
if args.val_save_to_img_file:
to_save_dir = os.path.join(ckpt_path, exp_name, str(epoch))
check_mkdir(to_save_dir)
#val_visual = []
for idx, data in enumerate(zip(inputs_all, labels_all,
predictions_all)):
if data[0] is None:
continue
input_pil = restore_transform(data[0])
labels_pil = colorize_mask(data[1])
predictions_pil = colorize_mask(data[2])
if args.val_save_to_img_file:
input_pil.save(os.path.join(to_save_dir, '%d_input.png' % idx))
predictions_pil.save(
os.path.join(to_save_dir, '%d_prediction.png' % idx))
labels_pil.save(os.path.join(to_save_dir,
'%d_label.png' % idx))
# val_visual.extend([visualize(input_pil.convert('RGB')), visualize(labels_pil.convert('RGB')),
# visualize(predictions_pil.convert('RGB'))])
# val_visual = torch.stack(val_visual, 0)
# val_visual = vutils.make_grid(val_visual, nrow=3, padding=5)
# writer.add_image(snapshot_name, val_visual)
print(
'--------------------------------------------------------------------')
print(
'[epoch %d], [val loss %.5f], [acc %.5f], [acc_cls %.5f], [mean_iu %.5f], [fwavacc %.5f]'
% (epoch, val_loss.avg, acc, acc_cls, mean_iu, fwavacc))
print(
'best record: [val loss %.5f], [acc %.5f], [acc_cls %.5f], [mean_iu %.5f], [fwavacc %.5f], [epoch %d]'
%
(best_record['val_loss'], best_record['acc'], best_record['acc_cls'],
best_record['mean_iu'], best_record['fwavacc'], best_record['epoch']))
print(
'--------------------------------------------------------------------')
# writer.add_scalar('val_loss', val_loss.avg, epoch)
# writer.add_scalar('acc', acc, epoch)
# writer.add_scalar('acc_cls', acc_cls, epoch)
# writer.add_scalar('mean_iu', mean_iu, epoch)
# writer.add_scalar('fwavacc', fwavacc, epoch)
# writer.add_scalar('lr', optimizer.param_groups[1]['lr'], epoch)
return val_loss.avg
# betas=(args.momentum, 0.999))
scheduler = ReduceLROnPlateau(optimizer,
'min',
patience=args.lr_patience,
min_lr=1e-10,
verbose=True)
# resume 优化器参数加载
if len(args.resume) > 0:
optimizer.load_state_dict(
torch.load(os.path.join(ckpt_path, exp_name, 'opt_' + args.resume)))
optimizer.param_groups[0]['lr'] = 2 * args.lr
optimizer.param_groups[1]['lr'] = args.lr
check_mkdir(ckpt_path)
check_mkdir(os.path.join(ckpt_path, exp_name))
#open(os.path.join(ckpt_path, exp_name, str(datetime.datetime.now()) + '.txt'), 'w').write(str(args) + '\n\n')
def train(epoch):
net.train()
# 计算平均损失，每个epoch更新为0
train_loss = AverageMeter()
#每次迭代调用 _getitem_ 方法，进行transform变换。
curr_iter = (epoch - 1) * len(trainloader)
for i, (inputs, labels) in enumerate(trainloader):
if args.cuda:
inputs, labels = inputs.cuda(), labels.cuda()
N = inputs.size(0)
# 清空梯度
def find_non_stationary_clusters(args):
if args['use_gpu']:
print("Using CUDA" if torch.cuda.is_available() else "Using CPU")
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
else:
device = "cpu"
network_folder_name = args['folder'].split('/')[-1]
tmp = re.search(r"cn(\d+)-", network_folder_name)
n_clusters = int(tmp.group(1))
save_folder = os.path.join(args['dest_root'], network_folder_name)
if os.path.exists(
os.path.join(save_folder, 'cluster_histogram_for_corr.npy')):
print('{} already exists. skipping'.format(
os.path.join(save_folder, 'cluster_histogram_for_corr.npy')))
return
check_mkdir(save_folder)
with open(os.path.join(args['folder'], 'bestval.txt')) as f:
best_val_dict_str = f.read()
bestval = eval(best_val_dict_str.rstrip())
# Network and weight loading
model_config = model_configs.PspnetCityscapesConfig()
net = model_config.init_network(
n_classes=n_clusters,
for_clustering=False,
output_features=False,
use_original_base=args['use_original_base']).to(device)
net.load_state_dict(
torch.load(os.path.join(args['folder'],
bestval['snapshot'] + '.pth'))) # load weights
net.eval()
# copy network file to save location
copyfile(os.path.join(args['folder'], bestval['snapshot'] + '.pth'),
os.path.join(save_folder, 'weights.pth'))
if args['only_copy_weights']:
print('Only copying weights')
return
# Data loading setup
if args['corr_set'] == 'rc':
corr_set_config = data_configs.RobotcarConfig()
elif args['corr_set'] == 'cmu':
corr_set_config = data_configs.CmuConfig()
elif args['corr_set'] == 'both':
corr_set_config1 = data_configs.CmuConfig()
corr_set_config2 = data_configs.RobotcarConfig()
sliding_crop_im = joint_transforms.SlidingCropImageOnly(
713, args['stride_rate'])
input_transform = model_config.input_transform
pre_validation_transform = model_config.pre_validation_transform
if args['corr_set'] == 'both':
corr_set_val1 = correspondences.Correspondences(
corr_set_config1.correspondence_path,
corr_set_config1.correspondence_im_path,
input_size=(713, 713),
input_transform=None,
joint_transform=None,
listfile=corr_set_config1.correspondence_val_list_file)
corr_set_val2 = correspondences.Correspondences(
corr_set_config2.correspondence_path,
corr_set_config2.correspondence_im_path,
input_size=(713, 713),
input_transform=None,
joint_transform=None,
listfile=corr_set_config2.correspondence_val_list_file)
corr_set_val = merged.Merged([corr_set_val1, corr_set_val2])
else:
corr_set_val = correspondences.Correspondences(
corr_set_config.correspondence_path,
corr_set_config.correspondence_im_path,
input_size=(713, 713),
input_transform=None,
joint_transform=None,
listfile=corr_set_config.correspondence_val_list_file)
# Segmentor
segmentor = Segmentor(net, n_clusters, n_slices_per_pass=4)
# save args
open(os.path.join(save_folder,
str(datetime.datetime.now()) + '.txt'),
'w').write(str(args) + '\n\n')
cluster_histogram_for_correspondences = np.zeros((n_clusters, ),
dtype=np.int64)
cluster_histogram_non_correspondences = np.zeros((n_clusters, ),
dtype=np.int64)
for i in range(0, len(corr_set_val), args['step']):
img1, img2, pts1, pts2, _ = corr_set_val[i]
seg1 = segmentor.run_and_save(
img1,
None,
pre_sliding_crop_transform=pre_validation_transform,
input_transform=input_transform,
sliding_crop=sliding_crop_im,
use_gpu=args['use_gpu'])
seg1 = np.array(seg1)
corr_loc_mask = np.zeros(seg1.shape, dtype=np.bool)
valid_inds = (pts1[0, :] >= 0) & (pts1[0, :] < seg1.shape[1]) & (
pts1[1, :] >= 0) & (pts1[1, :] < seg1.shape[0])
pts1 = pts1[:, valid_inds]
for j in range(pts1.shape[1]):
pt = pts1[:, j]
corr_loc_mask[pt[1], pt[0]] = True
cluster_ids_corr = seg1[corr_loc_mask]
hist_tmp, _ = np.histogram(cluster_ids_corr, np.arange(n_clusters + 1))
cluster_histogram_for_correspondences += hist_tmp
cluster_ids_no_corr = seg1[~corr_loc_mask]
hist_tmp, _ = np.histogram(cluster_ids_no_corr,
np.arange(n_clusters + 1))
cluster_histogram_non_correspondences += hist_tmp
if ((i + 1) % 100) < args['step']:
print('{}/{}'.format(i + 1, len(corr_set_val)))
np.save(os.path.join(save_folder, 'cluster_histogram_for_corr.npy'),
cluster_histogram_for_correspondences)
np.save(os.path.join(save_folder, 'cluster_histogram_non_corr.npy'),
cluster_histogram_non_correspondences)
frac = cluster_histogram_for_correspondences / \
(cluster_histogram_for_correspondences +
cluster_histogram_non_correspondences)
stationary_inds = np.argwhere(frac > 0.01)
np.save(os.path.join(save_folder, 'stationary_inds.npy'), stationary_inds)
print('{} stationary clusters out of {}'.format(
len(stationary_inds), len(cluster_histogram_for_correspondences)))
def train_with_correspondences(save_folder, startnet, args):
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
check_mkdir(save_folder)
writer = SummaryWriter(save_folder)
# Network and weight loading
model_config = model_configs.PspnetCityscapesConfig()
net = model_config.init_network().to(device)
if args['snapshot'] == 'latest':
args['snapshot'] = get_latest_network_name(save_folder)
if len(args['snapshot']) == 0: # If start from beginning
state_dict = torch.load(startnet)
# needed since we slightly changed the structure of the network in
# pspnet
state_dict = rename_keys_to_match(state_dict)
net.load_state_dict(state_dict) # load original weights
start_iter = 0
args['best_record'] = {
'iter': 0,
'val_loss': 1e10,
'acc': 0,
'acc_cls': 0,
'mean_iu': 0,
'fwavacc': 0
}
else: # If continue training
print('training resumes from ' + args['snapshot'])
net.load_state_dict(
torch.load(os.path.join(save_folder,
args['snapshot']))) # load weights
split_snapshot = args['snapshot'].split('_')
start_iter = int(split_snapshot[1])
with open(os.path.join(save_folder, 'bestval.txt')) as f:
best_val_dict_str = f.read()
args['best_record'] = eval(best_val_dict_str.rstrip())
net.train()
freeze_bn(net)
# Data loading setup
if args['corr_set'] == 'rc':
corr_set_config = data_configs.RobotcarConfig()
elif args['corr_set'] == 'cmu':
corr_set_config = data_configs.CmuConfig()
sliding_crop_im = joint_transforms.SlidingCropImageOnly(
713, args['stride_rate'])
input_transform = model_config.input_transform
pre_validation_transform = model_config.pre_validation_transform
target_transform = extended_transforms.MaskToTensor()
train_joint_transform_seg = joint_transforms.Compose([
joint_transforms.Resize(1024),
joint_transforms.RandomRotate(10),
joint_transforms.RandomHorizontallyFlip(),
joint_transforms.RandomCrop(713)
])
train_joint_transform_corr = corr_transforms.Compose([
corr_transforms.CorrResize(1024),
corr_transforms.CorrRandomCrop(713)
])
# keep list of segmentation loaders and validators
seg_loaders = list()
validators = list()
# Correspondences
corr_set = correspondences.Correspondences(
corr_set_config.correspondence_path,
corr_set_config.correspondence_im_path,
input_size=(713, 713),
mean_std=model_config.mean_std,
input_transform=input_transform,
joint_transform=train_joint_transform_corr)
corr_loader = DataLoader(corr_set,
batch_size=args['train_batch_size'],
num_workers=args['n_workers'],
shuffle=True)
# Cityscapes Training
c_config = data_configs.CityscapesConfig()
seg_set_cs = cityscapes.CityScapes(
c_config.train_im_folder,
c_config.train_seg_folder,
c_config.im_file_ending,
c_config.seg_file_ending,
id_to_trainid=c_config.id_to_trainid,
joint_transform=train_joint_transform_seg,
sliding_crop=None,
transform=input_transform,
target_transform=target_transform)
seg_loader_cs = DataLoader(seg_set_cs,
batch_size=args['train_batch_size'],
num_workers=args['n_workers'],
shuffle=True)
seg_loaders.append(seg_loader_cs)
# Cityscapes Validation
val_set_cs = cityscapes.CityScapes(
c_config.val_im_folder,
c_config.val_seg_folder,
c_config.im_file_ending,
c_config.seg_file_ending,
id_to_trainid=c_config.id_to_trainid,
sliding_crop=sliding_crop_im,
transform=input_transform,
target_transform=target_transform,
transform_before_sliding=pre_validation_transform)
val_loader_cs = DataLoader(val_set_cs,
batch_size=1,
num_workers=args['n_workers'],
shuffle=False)
validator_cs = Validator(val_loader_cs,
n_classes=c_config.n_classes,
save_snapshot=False,
extra_name_str='Cityscapes')
validators.append(validator_cs)
# Vistas Training and Validation
if args['include_vistas']:
v_config = data_configs.VistasConfig(
use_subsampled_validation_set=True, use_cityscapes_classes=True)
seg_set_vis = cityscapes.CityScapes(
v_config.train_im_folder,
v_config.train_seg_folder,
v_config.im_file_ending,
v_config.seg_file_ending,
id_to_trainid=v_config.id_to_trainid,
joint_transform=train_joint_transform_seg,
sliding_crop=None,
transform=input_transform,
target_transform=target_transform)
seg_loader_vis = DataLoader(seg_set_vis,
batch_size=args['train_batch_size'],
num_workers=args['n_workers'],
shuffle=True)
seg_loaders.append(seg_loader_vis)
val_set_vis = cityscapes.CityScapes(
v_config.val_im_folder,
v_config.val_seg_folder,
v_config.im_file_ending,
v_config.seg_file_ending,
id_to_trainid=v_config.id_to_trainid,
sliding_crop=sliding_crop_im,
transform=input_transform,
target_transform=target_transform,
transform_before_sliding=pre_validation_transform)
val_loader_vis = DataLoader(val_set_vis,
batch_size=1,
num_workers=args['n_workers'],
shuffle=False)
validator_vis = Validator(val_loader_vis,
n_classes=v_config.n_classes,
save_snapshot=False,
extra_name_str='Vistas')
validators.append(validator_vis)
else:
seg_loader_vis = None
map_validator = None
# Extra Training
extra_seg_set = cityscapes.CityScapes(
corr_set_config.train_im_folder,
corr_set_config.train_seg_folder,
corr_set_config.im_file_ending,
corr_set_config.seg_file_ending,
id_to_trainid=corr_set_config.id_to_trainid,
joint_transform=train_joint_transform_seg,
sliding_crop=None,
transform=input_transform,
target_transform=target_transform)
extra_seg_loader = DataLoader(extra_seg_set,
batch_size=args['train_batch_size'],
num_workers=args['n_workers'],
shuffle=True)
seg_loaders.append(extra_seg_loader)
# Extra Validation
extra_val_set = cityscapes.CityScapes(
corr_set_config.val_im_folder,
corr_set_config.val_seg_folder,
corr_set_config.im_file_ending,
corr_set_config.seg_file_ending,
id_to_trainid=corr_set_config.id_to_trainid,
sliding_crop=sliding_crop_im,
transform=input_transform,
target_transform=target_transform,
transform_before_sliding=pre_validation_transform)
extra_val_loader = DataLoader(extra_val_set,
batch_size=1,
num_workers=args['n_workers'],
shuffle=False)
extra_validator = Validator(extra_val_loader,
n_classes=corr_set_config.n_classes,
save_snapshot=True,
extra_name_str='Extra')
validators.append(extra_validator)
# Loss setup
if args['corr_loss_type'] == 'class':
corr_loss_fct = CorrClassLoss(input_size=[713, 713])
else:
corr_loss_fct = FeatureLoss(
input_size=[713, 713],
loss_type=args['corr_loss_type'],
feat_dist_threshold_match=args['feat_dist_threshold_match'],
feat_dist_threshold_nomatch=args['feat_dist_threshold_nomatch'],
n_not_matching=0)
seg_loss_fct = torch.nn.CrossEntropyLoss(
reduction='elementwise_mean',
ignore_index=cityscapes.ignore_label).to(device)
# Optimizer setup
optimizer = optim.SGD([{
'params': [
param for name, param in net.named_parameters()
if name[-4:] == 'bias' and param.requires_grad
],
'lr':
2 * args['lr']
}, {
'params': [
param for name, param in net.named_parameters()
if name[-4:] != 'bias' and param.requires_grad
],
'lr':
args['lr'],
'weight_decay':
args['weight_decay']
}],
momentum=args['momentum'],
nesterov=True)
if len(args['snapshot']) > 0:
optimizer.load_state_dict(
torch.load(os.path.join(save_folder, 'opt_' + args['snapshot'])))
optimizer.param_groups[0]['lr'] = 2 * args['lr']
optimizer.param_groups[1]['lr'] = args['lr']
open(os.path.join(save_folder,
str(datetime.datetime.now()) + '.txt'),
'w').write(str(args) + '\n\n')
if len(args['snapshot']) == 0:
f_handle = open(os.path.join(save_folder, 'log.log'), 'w', buffering=1)
else:
clean_log_before_continuing(os.path.join(save_folder, 'log.log'),
start_iter)
f_handle = open(os.path.join(save_folder, 'log.log'), 'a', buffering=1)
##########################################################################
#
# MAIN TRAINING CONSISTS OF ALL SEGMENTATION LOSSES AND A CORRESPONDENCE LOSS
#
##########################################################################
softm = torch.nn.Softmax2d()
val_iter = 0
train_corr_loss = AverageMeter()
train_seg_cs_loss = AverageMeter()
train_seg_extra_loss = AverageMeter()
train_seg_vis_loss = AverageMeter()
seg_loss_meters = list()
seg_loss_meters.append(train_seg_cs_loss)
if args['include_vistas']:
seg_loss_meters.append(train_seg_vis_loss)
seg_loss_meters.append(train_seg_extra_loss)
curr_iter = start_iter
for i in range(args['max_iter']):
optimizer.param_groups[0]['lr'] = 2 * args['lr'] * (
1 - float(curr_iter) / args['max_iter'])**args['lr_decay']
optimizer.param_groups[1]['lr'] = args['lr'] * (
1 - float(curr_iter) / args['max_iter'])**args['lr_decay']
#######################################################################
# SEGMENTATION UPDATE STEP
#######################################################################
#
for si, seg_loader in enumerate(seg_loaders):
# get segmentation training sample
inputs, gts = next(iter(seg_loader))
slice_batch_pixel_size = inputs.size(0) * inputs.size(
2) * inputs.size(3)
inputs = inputs.to(device)
gts = gts.to(device)
optimizer.zero_grad()
outputs, aux = net(inputs)
main_loss = args['seg_loss_weight'] * seg_loss_fct(outputs, gts)
aux_loss = args['seg_loss_weight'] * seg_loss_fct(aux, gts)
loss = main_loss + 0.4 * aux_loss
loss.backward()
optimizer.step()
seg_loss_meters[si].update(main_loss.item(),
slice_batch_pixel_size)
#######################################################################
# CORRESPONDENCE UPDATE STEP
#######################################################################
if args['corr_loss_weight'] > 0 and args[
'n_iterations_before_corr_loss'] < curr_iter:
img_ref, img_other, pts_ref, pts_other, weights = next(
iter(corr_loader))
# Transfer data to device
# img_ref is from the "good" sequence with generally better
# segmentation results
img_ref = img_ref.to(device)
img_other = img_other.to(device)
pts_ref = [p.to(device) for p in pts_ref]
pts_other = [p.to(device) for p in pts_other]
weights = [w.to(device) for w in weights]
# Forward pass
if args['corr_loss_type'] == 'hingeF': # Works on features
net.output_all = True
with torch.no_grad():
output_feat_ref, aux_feat_ref, output_ref, aux_ref = net(
img_ref)
output_feat_other, aux_feat_other, output_other, aux_other = net(
img_other
) # output1 must be last to backpropagate derivative correctly
net.output_all = False
else: # Works on class probs
with torch.no_grad():
output_ref, aux_ref = net(img_ref)
if args['corr_loss_type'] != 'hingeF' and args[
'corr_loss_type'] != 'hingeC':
output_ref = softm(output_ref)
aux_ref = softm(aux_ref)
# output1 must be last to backpropagate derivative correctly
output_other, aux_other = net(img_other)
if args['corr_loss_type'] != 'hingeF' and args[
'corr_loss_type'] != 'hingeC':
output_other = softm(output_other)
aux_other = softm(aux_other)
# Correspondence filtering
pts_ref_orig, pts_other_orig, weights_orig, batch_inds_to_keep_orig = correspondences.refine_correspondence_sample(
output_ref,
output_other,
pts_ref,
pts_other,
weights,
remove_same_class=args['remove_same_class'],
remove_classes=args['classes_to_ignore'])
pts_ref_orig = [
p for b, p in zip(batch_inds_to_keep_orig, pts_ref_orig)
if b.item() > 0
]
pts_other_orig = [
p for b, p in zip(batch_inds_to_keep_orig, pts_other_orig)
if b.item() > 0
]
weights_orig = [
p for b, p in zip(batch_inds_to_keep_orig, weights_orig)
if b.item() > 0
]
if args['corr_loss_type'] == 'hingeF':
# remove entire samples if needed
output_vals_ref = output_feat_ref[batch_inds_to_keep_orig]
output_vals_other = output_feat_other[batch_inds_to_keep_orig]
else:
# remove entire samples if needed
output_vals_ref = output_ref[batch_inds_to_keep_orig]
output_vals_other = output_other[batch_inds_to_keep_orig]
pts_ref_aux, pts_other_aux, weights_aux, batch_inds_to_keep_aux = correspondences.refine_correspondence_sample(
aux_ref,
aux_other,
pts_ref,
pts_other,
weights,
remove_same_class=args['remove_same_class'],
remove_classes=args['classes_to_ignore'])
pts_ref_aux = [
p for b, p in zip(batch_inds_to_keep_aux, pts_ref_aux)
if b.item() > 0
]
pts_other_aux = [
p for b, p in zip(batch_inds_to_keep_aux, pts_other_aux)
if b.item() > 0
]
weights_aux = [
p for b, p in zip(batch_inds_to_keep_aux, weights_aux)
if b.item() > 0
]
if args['corr_loss_type'] == 'hingeF':
# remove entire samples if needed
aux_vals_ref = aux_feat_ref[batch_inds_to_keep_orig]
aux_vals_other = aux_feat_other[batch_inds_to_keep_orig]
else:
# remove entire samples if needed
aux_vals_ref = aux_ref[batch_inds_to_keep_aux]
aux_vals_other = aux_other[batch_inds_to_keep_aux]
optimizer.zero_grad()
# correspondence loss
if output_vals_ref.size(0) > 0:
loss_corr_hr = corr_loss_fct(output_vals_ref,
output_vals_other, pts_ref_orig,
pts_other_orig, weights_orig)
else:
loss_corr_hr = 0 * output_vals_other.sum()
if aux_vals_ref.size(0) > 0:
loss_corr_aux = corr_loss_fct(
aux_vals_ref, aux_vals_other, pts_ref_aux, pts_other_aux,
weights_aux) # use output from img1 as "reference"
else:
loss_corr_aux = 0 * aux_vals_other.sum()
loss_corr = args['corr_loss_weight'] * \
(loss_corr_hr + 0.4 * loss_corr_aux)
loss_corr.backward()
optimizer.step()
train_corr_loss.update(loss_corr.item())
#######################################################################
# LOGGING ETC
#######################################################################
curr_iter += 1
val_iter += 1
writer.add_scalar('train_seg_loss_cs', train_seg_cs_loss.avg,
curr_iter)
writer.add_scalar('train_seg_loss_extra', train_seg_extra_loss.avg,
curr_iter)
writer.add_scalar('train_seg_loss_vis', train_seg_vis_loss.avg,
curr_iter)
writer.add_scalar('train_corr_loss', train_corr_loss.avg, curr_iter)
writer.add_scalar('lr', optimizer.param_groups[1]['lr'], curr_iter)
if (i + 1) % args['print_freq'] == 0:
str2write = '[iter %d / %d], [train corr loss %.5f] , [seg cs loss %.5f], [seg vis loss %.5f], [seg extra loss %.5f]. [lr %.10f]' % (
curr_iter, len(corr_loader), train_corr_loss.avg,
train_seg_cs_loss.avg, train_seg_vis_loss.avg,
train_seg_extra_loss.avg, optimizer.param_groups[1]['lr'])
print(str2write)
f_handle.write(str2write + "\n")
if val_iter >= args['val_interval']:
val_iter = 0
for validator in validators:
validator.run(net,
optimizer,
args,
curr_iter,
save_folder,
f_handle,
writer=writer)
# Post training
f_handle.close()
writer.close()
def segment_images_in_folder(network_file, img_folder, save_folder, args):
# get current available device
if args['use_gpu']:
print("Using CUDA" if torch.cuda.is_available() else "Using CPU")
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
else:
device = "cpu"
# Network and weight loading
model_config = model_configs.PspnetCityscapesConfig()
if 'n_classes' in args:
print('Initializing model with %d classes' % args['n_classes'])
net = model_config.init_network(
n_classes=args['n_classes'],
for_clustering=False,
output_features=False,
use_original_base=args['use_original_base']).to(device)
else:
net = model_config.init_network().to(device)
print('load model ' + network_file)
state_dict = torch.load(network_file,
map_location=lambda storage, loc: storage)
# needed since we slightly changed the structure of the network in pspnet
state_dict = rename_keys_to_match(state_dict)
net.load_state_dict(state_dict)
net.eval()
# data loading
input_transform = model_config.input_transform
pre_validation_transform = model_config.pre_validation_transform
# make sure crop size and stride same as during training
sliding_crop = joint_transforms.SlidingCropImageOnly(
713, args['sliding_transform_step'])
check_mkdir(save_folder)
t0 = time.time()
# get all file names
filenames_ims = list()
filenames_segs = list()
print('Scanning %s for images to segment.' % img_folder)
for root, subdirs, files in os.walk(img_folder):
filenames = [f for f in files if f.endswith(args['img_ext'])]
if len(filenames) > 0:
print('Found %d images in %s' % (len(filenames), root))
seg_path = root.replace(img_folder, save_folder)
check_mkdir(seg_path)
filenames_ims += [os.path.join(root, f) for f in filenames]
filenames_segs += [
os.path.join(seg_path, f.replace(args['img_ext'], '.png'))
for f in filenames
]
# Create segmentor
if net.n_classes == 19: # This could be the 19 cityscapes classes
segmentor = Segmentor(net,
net.n_classes,
colorize_fcn=cityscapes.colorize_mask,
n_slices_per_pass=args['n_slices_per_pass'])
else:
segmentor = Segmentor(net,
net.n_classes,
colorize_fcn=None,
n_slices_per_pass=args['n_slices_per_pass'])
count = 1
for im_file, save_path in zip(filenames_ims, filenames_segs):
tnow = time.time()
print("[%d/%d (%.1fs/%.1fs)] %s" %
(count, len(filenames_ims), tnow - t0,
(tnow - t0) / count * len(filenames_ims), im_file))
segmentor.run_and_save(
im_file,
save_path,
pre_sliding_crop_transform=pre_validation_transform,
sliding_crop=sliding_crop,
input_transform=input_transform,
skip_if_seg_exists=True,
use_gpu=args['use_gpu'])
count += 1
tend = time.time()
print('Time: %f' % (tend - t0))
def train_with_clustering(save_folder, tmp_seg_folder, startnet, args):
print(save_folder.split('/')[-1])
skip_clustering = False
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
check_mkdir(save_folder)
writer = SummaryWriter(save_folder)
check_mkdir(tmp_seg_folder)
# Network and weight loading
model_config = model_configs.PspnetCityscapesConfig()
net = model_config.init_network(
n_classes=args['n_clusters'],
for_clustering=True,
output_features=True,
use_original_base=args['use_original_base']).to(device)
state_dict = torch.load(startnet)
if 'resnet101' in startnet:
load_resnet101_weights(net, state_dict)
else:
# needed since we slightly changed the structure of the network in pspnet
state_dict = rename_keys_to_match(state_dict)
# different amount of classes
init_last_layers(state_dict, args['n_clusters'])
net.load_state_dict(state_dict) # load original weights
start_iter = 0
args['best_record'] = {
'iter': 0,
'val_loss_feat': 1e10,
'val_loss_out': 1e10,
'val_loss_cluster': 1e10
}
# Data loading setup
if args['corr_set'] == 'rc':
corr_set_config = data_configs.RobotcarConfig()
elif args['corr_set'] == 'pola':
corr_set_config = data_configs.PolaConfig()
elif args['corr_set'] == 'cmu':
corr_set_config = data_configs.CmuConfig()
elif args['corr_set'] == 'both':
corr_set_config1 = data_configs.CmuConfig()
corr_set_config2 = data_configs.RobotcarConfig()
ref_image_lists = corr_set_config.reference_image_list
# ref_image_lists = glob.glob("/media/HDD1/datasets/Creusot_Jan15/Creusot_3/*.jpg", recursive=True)
# print(f'ici on print ref image list ---------------------------------------------------- {ref_image_lists}')
# print(corr_set_config)
# corr_im_paths = [corr_set_config.correspondence_im_path]
# ref_featurs_pos = [corr_set_config.reference_feature_poitions]
input_transform = model_config.input_transform
#corr_set_train = correspondences.Correspondences(corr_set_config.correspondence_path,
# corr_set_config.correspondence_im_path,
# input_size=(713, 713),
# input_transform=input_transform,
# joint_transform=train_joint_transform_corr,
# listfile=corr_set_config.correspondence_train_list_file)
scales = [0, 1, 2, 3]
# corr_set_train = Poladata.MonoDataset(corr_set_config,
# seg_folder = "media/HDD1/NsemSEG/Result_fold/" ,
# im_file_ending = ".jpg" )
train_joint_transform = joint_transforms.Compose([
# train_joint_transform_corr = corr_transforms.Compose([
# corr_transforms.CorrResize(1024),
# corr_transforms.CorrRandomCrop(713)
joint_transforms.Resize(1024),
joint_transforms.RandomCrop(713)
])
sliding_crop = joint_transforms.SlidingCrop(713, 2 / 3., 255)
# corr_set_train = correspondences.Correspondences(corr_set_config.train_im_folder,
# corr_set_config.train_im_folder,
# input_size=(713, 713),
# input_transform=input_transform,
# joint_transform=train_joint_transform,
# listfile=None)
corr_set_train = Poladata.MonoDataset(
corr_set_config.train_im_folder,
corr_set_config.train_seg_folder,
im_file_ending=".jpg",
id_to_trainid=None,
joint_transform=train_joint_transform,
sliding_crop=sliding_crop,
transform=input_transform,
target_transform=None, #train_joint_transform,
transform_before_sliding=None #sliding_crop
)
#print (corr_set_train)
# print(corr_set_train.mask)
corr_loader_train = DataLoader(corr_set_train,
batch_size=1,
num_workers=args['n_workers'],
shuffle=True)
# corr_loader_train = input_transform(corr_loader_train)
# print(corr_loader_train)
seg_loss_fct = torch.nn.CrossEntropyLoss(reduction='elementwise_mean')
# Optimizer setup
optimizer = optim.SGD([{
'params': [
param for name, param in net.named_parameters()
if name[-4:] == 'bias' and param.requires_grad
],
'lr':
2 * args['lr']
}, {
'params': [
param for name, param in net.named_parameters()
if name[-4:] != 'bias' and param.requires_grad
],
'lr':
args['lr'],
'weight_decay':
args['weight_decay']
}],
momentum=args['momentum'],
nesterov=True)
# Clustering
deepcluster = clustering.Kmeans(args['n_clusters'])
if skip_clustering:
deepcluster.set_index(cluster_centroids)
open(os.path.join(save_folder,
str(datetime.datetime.now()) + '.txt'),
'w').write(str(args) + '\n\n')
f_handle = open(os.path.join(save_folder, 'log.log'), 'w', buffering=1)
# clean_log_before_continuing(os.path.join(save_folder, 'log.log'), start_iter)
# f_handle = open(os.path.join(save_folder, 'log.log'), 'a', buffering=1)
val_iter = 0
curr_iter = start_iter
while curr_iter <= args['max_iter']:
net.eval()
net.output_features = True
# max_num_features_per_image = args['max_features_per_image']
# print('-----------------------------------------------------------------')
# print (f'ref_image_lists est: {ref_image_lists},model_config es : {model_config} , net es: {net} , max feature par image es : {max_num_features_per_image} ')
# print('-----------------------------------------------------------------')
# print('le next du loader es : ---------------')
# print(next(iter(corr_loader_train)))
# features, _ = extract_features_for_reference(net, model_config, ref_image_lists,
# corr_im_paths, ref_featurs_pos,
# max_num_features_per_image=args['max_features_per_image'],
# fraction_correspondeces=0.5)
print(
'ici on a la len de la ref im list --------------------------------------------------------'
)
print(len(ref_image_lists))
features = extract_features_for_reference_nocorr(
net,
model_config,
corr_set_train,
10,
max_num_features_per_image=args['max_features_per_image'])
cluster_features = np.vstack(features)
del features
# cluster the features
cluster_indices, clustering_loss, cluster_centroids, pca_info = deepcluster.cluster_imfeatures(
cluster_features, verbose=True, use_gpu=False)
# save cluster centroids
h5f = h5py.File(
os.path.join(save_folder, 'centroids_%d.h5' % curr_iter), 'w')
h5f.create_dataset('cluster_centroids', data=cluster_centroids)
h5f.create_dataset('pca_transform_Amat', data=pca_info[0])
h5f.create_dataset('pca_transform_bvec', data=pca_info[1])
h5f.close()
# Print distribution of clusters
cluster_distribution, _ = np.histogram(
cluster_indices,
bins=np.arange(args['n_clusters'] + 1),
density=True)
str2write = 'cluster distribution ' + \
np.array2string(cluster_distribution, formatter={
'float_kind': '{0:.8f}'.format}).replace('\n', ' ')
print(str2write)
f_handle.write(str2write + "\n")
# set last layer weight to a normal distribution
reinit_last_layers(net)
# make a copy of current network state to do cluster assignment
net_for_clustering = copy.deepcopy(net)
optimizer.param_groups[0]['lr'] = 2 * args['lr'] * (
1 - float(curr_iter) / args['max_iter'])**args['lr_decay']
optimizer.param_groups[1]['lr'] = args['lr'] * (
1 - float(curr_iter) / args['max_iter'])**args['lr_decay']
net.train()
freeze_bn(net)
net.output_features = False
cluster_training_count = 0
# Train using the training correspondence set
corr_train_loss = AverageMeter()
seg_train_loss = AverageMeter()
feature_train_loss = AverageMeter()
while cluster_training_count < args[
'cluster_interval'] and curr_iter <= args['max_iter']:
# First extract cluster labels using saved network checkpoint
print(
'on rentre dans la boucle extract cluster_______________________________________________'
)
net.to("cpu")
net_for_clustering.to(device)
net_for_clustering.eval()
net_for_clustering.output_features = True
data_samples = []
extract_label_count = 0
while (extract_label_count < args['chunk_size']) and (
cluster_training_count + extract_label_count <
args['cluster_interval']
) and (val_iter + extract_label_count < args['val_interval']) and (
extract_label_count + curr_iter <= args['max_iter']):
# img_ref, img_other, pts_ref, pts_other, _ = next(iter(corr_set_train))
corr_loader_train = input_transform(corr_loader_train)
print(
f'la valeur de corr loader train es de {corr_loader_train} lors de l iteration : {curr_iter}'
)
img_ref, img_other, pts_ref, pts_other, _ = next(
iter(corr_loader_train))
# print('le next du loader es : ---------------')
# print(next(iter(corr_loader_train)))
# print(img_ref)
# Transfer data to device
img_ref = img_ref.to(device)
with torch.no_grad():
features = net_for_clustering(img_ref)
# assign feature to clusters for entire patch
output = features.cpu().numpy()
output_flat = output.reshape(
(output.shape[0], output.shape[1], -1))
cluster_image = np.zeros(
(output.shape[0], output.shape[2], output.shape[3]),
dtype=np.int64)
for b in range(output_flat.shape[0]):
out_f = output_flat[b]
out_f2, _ = preprocess_features(np.swapaxes(out_f, 0, 1),
pca_info=pca_info)
cluster_labels = deepcluster.assign(out_f2)
cluster_image[b] = cluster_labels.reshape(
(output.shape[2], output.shape[3]))
cluster_image = torch.from_numpy(cluster_image).to(device)
# assign cluster to correspondence positions
cluster_labels = assign_cluster_ids_to_correspondence_points(
features,
pts_ref, (deepcluster, pca_info),
inds_other=pts_other,
orig_im_size=(713, 713))
# Transfer data to cpu
img_ref = img_ref.cpu()
cluster_labels = [p.cpu() for p in cluster_labels]
cluster_image = cluster_image.cpu()
data_samples.append((img_ref, cluster_labels, cluster_image))
extract_label_count += 1
net_for_clustering.to("cpu")
net.to(device)
for data_sample in data_samples:
img_ref, cluster_labels, cluster_image = data_sample
# Transfer data to device
img_ref = img_ref.to(device)
cluster_labels = [p.to(device) for p in cluster_labels]
cluster_image = cluster_image.to(device)
optimizer.zero_grad()
outputs_ref, aux_ref = net(img_ref)
seg_main_loss = seg_loss_fct(outputs_ref, cluster_image)
seg_aux_loss = seg_loss_fct(aux_ref, cluster_image)
loss = args['seg_loss_weight'] * \
(seg_main_loss + 0.4 * seg_aux_loss)
loss.backward()
optimizer.step()
cluster_training_count += 1
if type(seg_main_loss) == torch.Tensor:
seg_train_loss.update(seg_main_loss.item(), 1)
####################################################################################################
# LOGGING ETC
####################################################################################################
curr_iter += 1
val_iter += 1
writer.add_scalar('train_seg_loss', seg_train_loss.avg,
curr_iter)
writer.add_scalar('lr', optimizer.param_groups[1]['lr'],
curr_iter)
if (curr_iter + 1) % args['print_freq'] == 0:
str2write = '[iter %d / %d], [train seg loss %.5f], [train corr loss %.5f], [train feature loss %.5f]. [lr %.10f]' % (
curr_iter + 1, args['max_iter'], seg_train_loss.avg,
optimizer.param_groups[1]['lr'])
print(str2write)
f_handle.write(str2write + "\n")
if curr_iter > args['max_iter']:
break
# Post training
f_handle.close()
writer.close()
def train_with_clustering(save_folder, tmp_seg_folder, startnet, args):
print(save_folder.split('/')[-1])
skip_clustering = False
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
check_mkdir(save_folder)
writer = SummaryWriter(save_folder)
check_mkdir(tmp_seg_folder)
# Network and weight loading
model_config = model_configs.PspnetCityscapesConfig()
net = model_config.init_network(
n_classes=args['n_clusters'],
for_clustering=True,
output_features=True,
use_original_base=args['use_original_base']).to(device)
if args['snapshot'] == 'latest':
args['snapshot'] = get_latest_network_name(save_folder)
if len(args['snapshot']) == 0: # If start from beginning
state_dict = torch.load(startnet)
if 'resnet101' in startnet:
load_resnet101_weights(net, state_dict)
else:
# needed since we slightly changed the structure of the network in pspnet
state_dict = rename_keys_to_match(state_dict)
init_last_layers(state_dict,
args['n_clusters']) # different amount of classes
net.load_state_dict(state_dict) # load original weights
start_iter = 0
args['best_record'] = {
'iter': 0,
'val_loss_feat': 1e10,
'val_loss_out': 1e10,
'val_loss_cluster': 1e10
}
else: # If continue training
print('training resumes from ' + args['snapshot'])
net.load_state_dict(
torch.load(os.path.join(save_folder,
args['snapshot']))) # load weights
split_snapshot = args['snapshot'].split('_')
start_iter = int(split_snapshot[1])
with open(os.path.join(save_folder, 'bestval.txt')) as f:
best_val_dict_str = f.read()
args['best_record'] = eval(best_val_dict_str.rstrip())
if start_iter >= args['max_iter']:
return
if (start_iter % args['cluster_interval']) == 0:
skip_clustering = False
else:
skip_clustering = True
last_cluster_network_snapshot_iter = (
start_iter //
args['cluster_interval']) * args['cluster_interval']
# load cluster info
cluster_info = {}
f = h5py.File(
os.path.join(
save_folder, 'centroids_{}.h5'.format(
last_cluster_network_snapshot_iter)), 'r')
for k, v in f.items():
cluster_info[k] = np.array(v)
cluster_centroids = cluster_info['cluster_centroids']
pca_info = [
cluster_info['pca_transform_Amat'],
cluster_info['pca_transform_bvec']
]
# load network that was used for last clustering
net_for_clustering = model_config.init_network(
n_classes=args['n_clusters'],
for_clustering=True,
output_features=True,
use_original_base=args['use_original_base'])
if last_cluster_network_snapshot_iter == 0:
state_dict = torch.load(
startnet, map_location=lambda storage, loc: storage)
if 'resnet101' in startnet:
load_resnet101_weights(net_for_clustering, state_dict)
else:
# needed since we slightly changed the structure of the network in pspnet
state_dict = rename_keys_to_match(state_dict)
init_last_layers(
state_dict,
args['n_clusters']) # different amount of classes
net_for_clustering.load_state_dict(
state_dict) # load original weights
else:
cluster_network_weights = get_network_name_from_iteration(
save_folder, last_cluster_network_snapshot_iter)
net_for_clustering.load_state_dict(
torch.load(os.path.join(save_folder,
cluster_network_weights),
map_location=lambda storage, loc: storage)
) # load weights
# Data loading setup
if args['corr_set'] == 'rc':
corr_set_config = data_configs.RobotcarConfig()
elif args['corr_set'] == 'cmu':
corr_set_config = data_configs.CmuConfig()
elif args['corr_set'] == 'both':
corr_set_config1 = data_configs.CmuConfig()
corr_set_config2 = data_configs.RobotcarConfig()
if args['corr_set'] == 'both':
ref_image_lists = [
corr_set_config1.reference_image_list,
corr_set_config2.reference_image_list
]
corr_im_paths = [
corr_set_config1.correspondence_im_path,
corr_set_config2.correspondence_im_path
]
ref_featurs_pos = [
corr_set_config1.reference_feature_poitions,
corr_set_config2.reference_feature_poitions
]
else:
ref_image_lists = [corr_set_config.reference_image_list]
corr_im_paths = [corr_set_config.correspondence_im_path]
ref_featurs_pos = [corr_set_config.reference_feature_poitions]
input_transform = model_config.input_transform
train_joint_transform_corr = corr_transforms.Compose([
corr_transforms.CorrResize(1024),
corr_transforms.CorrRandomCrop(713)
])
# Correspondences for training
if args['corr_set'] == 'both':
corr_set_train1 = correspondences.Correspondences(
corr_set_config1.correspondence_path,
corr_set_config1.correspondence_im_path,
input_size=(713, 713),
input_transform=input_transform,
joint_transform=train_joint_transform_corr,
listfile=corr_set_config1.correspondence_train_list_file)
corr_set_train2 = correspondences.Correspondences(
corr_set_config2.correspondence_path,
corr_set_config2.correspondence_im_path,
input_size=(713, 713),
input_transform=input_transform,
joint_transform=train_joint_transform_corr,
listfile=corr_set_config2.correspondence_train_list_file)
corr_set_train = merged.Merged([corr_set_train1, corr_set_train2])
else:
corr_set_train = correspondences.Correspondences(
corr_set_config.correspondence_path,
corr_set_config.correspondence_im_path,
input_size=(713, 713),
input_transform=input_transform,
joint_transform=train_joint_transform_corr,
listfile=corr_set_config.correspondence_train_list_file)
corr_loader_train = DataLoader(corr_set_train,
batch_size=1,
num_workers=args['n_workers'],
shuffle=True)
# Correspondences for validation
if args['corr_set'] == 'both':
corr_set_val1 = correspondences.Correspondences(
corr_set_config1.correspondence_path,
corr_set_config1.correspondence_im_path,
input_size=(713, 713),
input_transform=input_transform,
joint_transform=train_joint_transform_corr,
listfile=corr_set_config1.correspondence_val_list_file)
corr_set_val2 = correspondences.Correspondences(
corr_set_config2.correspondence_path,
corr_set_config2.correspondence_im_path,
input_size=(713, 713),
input_transform=input_transform,
joint_transform=train_joint_transform_corr,
listfile=corr_set_config2.correspondence_val_list_file)
corr_set_val = merged.Merged([corr_set_val1, corr_set_val2])
else:
corr_set_val = correspondences.Correspondences(
corr_set_config.correspondence_path,
corr_set_config.correspondence_im_path,
input_size=(713, 713),
input_transform=input_transform,
joint_transform=train_joint_transform_corr,
listfile=corr_set_config.correspondence_val_list_file)
corr_loader_val = DataLoader(corr_set_val,
batch_size=1,
num_workers=args['n_workers'],
shuffle=False)
# Loss setup
val_corr_loss_fct_feat = FeatureLoss(
input_size=[713, 713],
loss_type=args['feature_distance_measure'],
feat_dist_threshold_match=0.8,
feat_dist_threshold_nomatch=0.2,
n_not_matching=0)
val_corr_loss_fct_out = FeatureLoss(input_size=[713, 713],
loss_type='KL',
feat_dist_threshold_match=0.8,
feat_dist_threshold_nomatch=0.2,
n_not_matching=0)
loss_fct = ClusterCorrespondenceLoss(input_size=[713, 713],
size_average=True).to(device)
seg_loss_fct = torch.nn.CrossEntropyLoss(reduction='elementwise_mean')
if args['feature_hinge_loss_weight'] > 0:
feature_loss_fct = FeatureLoss(input_size=[713, 713],
loss_type='hingeF',
feat_dist_threshold_match=0.8,
feat_dist_threshold_nomatch=0.2,
n_not_matching=0)
# Validator
corr_validator = CorrValidator(corr_loader_val,
val_corr_loss_fct_feat,
val_corr_loss_fct_out,
loss_fct,
save_snapshot=True,
extra_name_str='Corr')
# Optimizer setup
optimizer = optim.SGD([{
'params': [
param for name, param in net.named_parameters()
if name[-4:] == 'bias' and param.requires_grad
],
'lr':
2 * args['lr']
}, {
'params': [
param for name, param in net.named_parameters()
if name[-4:] != 'bias' and param.requires_grad
],
'lr':
args['lr'],
'weight_decay':
args['weight_decay']
}],
momentum=args['momentum'],
nesterov=True)
# Clustering
deepcluster = clustering.Kmeans(args['n_clusters'])
if skip_clustering:
deepcluster.set_index(cluster_centroids)
if len(args['snapshot']) > 0:
optimizer.load_state_dict(
torch.load(os.path.join(save_folder, 'opt_' + args['snapshot'])))
optimizer.param_groups[0]['lr'] = 2 * args['lr']
optimizer.param_groups[1]['lr'] = args['lr']
open(os.path.join(save_folder,
str(datetime.datetime.now()) + '.txt'),
'w').write(str(args) + '\n\n')
if len(args['snapshot']) == 0:
f_handle = open(os.path.join(save_folder, 'log.log'), 'w', buffering=1)
else:
clean_log_before_continuing(os.path.join(save_folder, 'log.log'),
start_iter)
f_handle = open(os.path.join(save_folder, 'log.log'), 'a', buffering=1)
val_iter = 0
curr_iter = start_iter
while curr_iter <= args['max_iter']:
if not skip_clustering:
# Extract image features from reference images
net.eval()
net.output_features = True
features, _ = extract_features_for_reference(
net,
model_config,
ref_image_lists,
corr_im_paths,
ref_featurs_pos,
max_num_features_per_image=args['max_features_per_image'],
fraction_correspondeces=0.5)
cluster_features = np.vstack(features)
del features
# cluster the features
cluster_indices, clustering_loss, cluster_centroids, pca_info = deepcluster.cluster_imfeatures(
cluster_features, verbose=True, use_gpu=False)
# save cluster centroids
h5f = h5py.File(
os.path.join(save_folder, 'centroids_%d.h5' % curr_iter), 'w')
h5f.create_dataset('cluster_centroids', data=cluster_centroids)
h5f.create_dataset('pca_transform_Amat', data=pca_info[0])
h5f.create_dataset('pca_transform_bvec', data=pca_info[1])
h5f.close()
# Print distribution of clusters
cluster_distribution, _ = np.histogram(
cluster_indices,
bins=np.arange(args['n_clusters'] + 1),
density=True)
str2write = 'cluster distribution ' + \
np.array2string(cluster_distribution, formatter={'float_kind': '{0:.8f}'.format}).replace('\n', ' ')
print(str2write)
f_handle.write(str2write + "\n")
reinit_last_layers(
net) # set last layer weight to a normal distribution
# make a copy of current network state to do cluster assignment
net_for_clustering = copy.deepcopy(net)
else:
skip_clustering = False
optimizer.param_groups[0]['lr'] = 2 * args['lr'] * (
1 - float(curr_iter) / args['max_iter'])**args['lr_decay']
optimizer.param_groups[1]['lr'] = args['lr'] * (
1 - float(curr_iter) / args['max_iter'])**args['lr_decay']
net.train()
freeze_bn(net)
net.output_features = False
cluster_training_count = 0
# Train using the training correspondence set
corr_train_loss = AverageMeter()
seg_train_loss = AverageMeter()
feature_train_loss = AverageMeter()
while cluster_training_count < args[
'cluster_interval'] and curr_iter <= args['max_iter']:
# First extract cluster labels using saved network checkpoint
net.to("cpu")
net_for_clustering.to(device)
net_for_clustering.eval()
net_for_clustering.output_features = True
if args['feature_hinge_loss_weight'] > 0:
net_for_clustering.output_all = False
data_samples = []
extract_label_count = 0
while (extract_label_count < args['chunk_size']) and (
cluster_training_count + extract_label_count <
args['cluster_interval']
) and (val_iter + extract_label_count < args['val_interval']) and (
extract_label_count + curr_iter <= args['max_iter']):
img_ref, img_other, pts_ref, pts_other, _ = next(
iter(corr_loader_train))
# Transfer data to device
img_ref = img_ref.to(device)
img_other = img_other.to(device)
pts_ref = [p.to(device) for p in pts_ref]
pts_other = [p.to(device) for p in pts_other]
with torch.no_grad():
features = net_for_clustering(img_ref)
# assign feature to clusters for entire patch
output = features.cpu().numpy()
output_flat = output.reshape(
(output.shape[0], output.shape[1], -1))
cluster_image = np.zeros(
(output.shape[0], output.shape[2], output.shape[3]),
dtype=np.int64)
for b in range(output_flat.shape[0]):
out_f = output_flat[b]
out_f2, _ = preprocess_features(np.swapaxes(out_f, 0, 1),
pca_info=pca_info)
cluster_labels = deepcluster.assign(out_f2)
cluster_image[b] = cluster_labels.reshape(
(output.shape[2], output.shape[3]))
cluster_image = torch.from_numpy(cluster_image).to(device)
# assign cluster to correspondence positions
cluster_labels = assign_cluster_ids_to_correspondence_points(
features,
pts_ref, (deepcluster, pca_info),
inds_other=pts_other,
orig_im_size=(713, 713))
# Transfer data to cpu
img_ref = img_ref.cpu()
img_other = img_other.cpu()
pts_ref = [p.cpu() for p in pts_ref]
pts_other = [p.cpu() for p in pts_other]
cluster_labels = [p.cpu() for p in cluster_labels]
cluster_image = cluster_image.cpu()
data_samples.append((img_ref, img_other, pts_ref, pts_other,
cluster_labels, cluster_image))
extract_label_count += 1
net_for_clustering.to("cpu")
net.to(device)
for data_sample in data_samples:
img_ref, img_other, pts_ref, pts_other, cluster_labels, cluster_image = data_sample
# Transfer data to device
img_ref = img_ref.to(device)
img_other = img_other.to(device)
pts_ref = [p.to(device) for p in pts_ref]
pts_other = [p.to(device) for p in pts_other]
cluster_labels = [p.to(device) for p in cluster_labels]
cluster_image = cluster_image.to(device)
optimizer.zero_grad()
if args['feature_hinge_loss_weight'] > 0:
net.output_all = True
# Randomization to decide if reference or target image should be used for training
if args['fraction_reference_bp'] is None: # use both
if args['feature_hinge_loss_weight'] > 0:
out_feat_ref, aux_feat_ref, outputs_ref, aux_ref = net(
img_ref)
else:
outputs_ref, aux_ref = net(img_ref)
seg_main_loss = seg_loss_fct(outputs_ref, cluster_image)
seg_aux_loss = seg_loss_fct(aux_ref, cluster_image)
if args['feature_hinge_loss_weight'] > 0:
out_feat_other, aux_feat_other, outputs_other, aux_other = net(
img_other)
else:
outputs_other, aux_other = net(img_other)
elif np.random.rand(
1)[0] < args['fraction_reference_bp']: # use reference
if args['feature_hinge_loss_weight'] > 0:
out_feat_ref, aux_feat_ref, outputs_ref, aux_ref = net(
img_ref)
else:
outputs_ref, aux_ref = net(img_ref)
seg_main_loss = seg_loss_fct(outputs_ref, cluster_image)
seg_aux_loss = seg_loss_fct(aux_ref, cluster_image)
with torch.no_grad():
if args['feature_hinge_loss_weight'] > 0:
out_feat_other, aux_feat_other, outputs_other, aux_other = net(
img_other)
else:
outputs_other, aux_other = net(img_other)
else: # use target
with torch.no_grad():
if args['feature_hinge_loss_weight'] > 0:
out_feat_ref, aux_feat_ref, outputs_ref, aux_ref = net(
img_ref)
else:
outputs_ref, aux_ref = net(img_ref)
if args['feature_hinge_loss_weight'] > 0:
out_feat_other, aux_feat_other, outputs_other, aux_other = net(
img_other)
else:
outputs_other, aux_other = net(img_other)
seg_main_loss = 0.
seg_aux_loss = 0.
if args['feature_hinge_loss_weight'] > 0:
net.output_all = False
main_loss, _ = loss_fct(outputs_ref,
outputs_other,
None,
pts_ref,
pts_other,
cluster_labels=cluster_labels)
aux_loss, _ = loss_fct(aux_ref,
aux_other,
None,
pts_ref,
pts_other,
cluster_labels=cluster_labels)
if args['feature_hinge_loss_weight'] > 0:
feature_loss = feature_loss_fct(out_feat_ref,
out_feat_other, pts_ref,
pts_other, None)
feature_loss_aux = feature_loss_fct(
aux_feat_ref, aux_feat_other, pts_ref, pts_other, None)
loss = args['corr_loss_weight']*(main_loss + 0.4 * aux_loss) + args['seg_loss_weight']*(seg_main_loss + 0.4 * seg_aux_loss) \
+ args['feature_hinge_loss_weight']*(feature_loss + 0.4 * feature_loss_aux)
else:
feature_loss = 0.
loss = args['corr_loss_weight']*(main_loss + 0.4 * aux_loss) + \
args['seg_loss_weight']*(seg_main_loss + 0.4 * seg_aux_loss)
loss.backward()
optimizer.step()
cluster_training_count += 1
corr_train_loss.update(main_loss.item(), 1)
if type(seg_main_loss) == torch.Tensor:
seg_train_loss.update(seg_main_loss.item(), 1)
if type(feature_loss) == torch.Tensor:
feature_train_loss.update(feature_loss.item(), 1)
####################################################################################################
# LOGGING ETC
####################################################################################################
curr_iter += 1
val_iter += 1
writer.add_scalar('train_corr_loss', corr_train_loss.avg,
curr_iter)
writer.add_scalar('train_seg_loss', seg_train_loss.avg,
curr_iter)
writer.add_scalar('train_feature_loss', feature_train_loss.avg,
curr_iter)
writer.add_scalar('lr', optimizer.param_groups[1]['lr'],
curr_iter)
if (curr_iter + 1) % args['print_freq'] == 0:
str2write = '[iter %d / %d], [train seg loss %.5f], [train corr loss %.5f], [train feature loss %.5f]. [lr %.10f]' % (
curr_iter + 1, args['max_iter'], seg_train_loss.avg,
corr_train_loss.avg, feature_train_loss.avg,
optimizer.param_groups[1]['lr'])
print(str2write)
f_handle.write(str2write + "\n")
if val_iter >= args['val_interval']:
val_iter = 0
net_for_clustering.to(device)
corr_validator.run(net,
net_for_clustering,
(deepcluster, pca_info),
optimizer,
args,
curr_iter,
save_folder,
f_handle,
writer=writer)
net_for_clustering.to("cpu")
if curr_iter > args['max_iter']:
break
# Post training
f_handle.close()
writer.close()
def run_and_save(
self,
img_path,
save_path,
pre_sliding_crop_transform=None,
sliding_crop=joint_transforms.SlidingCropImageOnly(713, 2 / 3.),
input_transform=standard_transforms.ToTensor(),
verbose=False,
skip_if_seg_exists=False,
use_gpu=True,
):
"""
img - Path of input image
save_path - Path of output image (feature map)
sliding_crop - Transform that returns set of image slices
input_transform - Transform to apply to image before inputting to network
skip_if_seg_exists - Whether to overwrite or skip if segmentation exists already
"""
if save_path is not None:
if os.path.exists(save_path):
if skip_if_seg_exists:
if verbose:
print(
"Segmentation already exists, skipping: {}".format(
save_path))
return
else:
if verbose:
print("Segmentation already exists, overwriting: {}".
format(save_path))
if isinstance(img_path, str):
try:
img = Image.open(img_path).convert('RGB')
except OSError:
print(
"Error reading input image, skipping: {}".format(img_path))
else:
img = img_path
# creating sliding crop windows and transform them
img_size_orig = img.size
if pre_sliding_crop_transform is not None: # might reshape image
img = pre_sliding_crop_transform(img)
img_slices, slices_info = sliding_crop(img)
img_slices = [input_transform(e) for e in img_slices]
img_slices = torch.stack(img_slices, 0)
slices_info = torch.LongTensor(slices_info)
slices_info.squeeze_(0)
of_pre, oa_pre = self.net.output_features, self.net.output_all
self.net.output_features, self.net.output_all = True, False
feature_map = self.run_on_slices(
img_slices,
slices_info,
sliding_transform_step=sliding_crop.stride_rate,
use_gpu=use_gpu)
# restore previous settings
self.net.output_features, self.net.output_all = of_pre, oa_pre
if save_path is not None:
check_mkdir(os.path.dirname(save_path))
ext = save_path.split('.')[-1]
if ext == 'mat':
matdict = {
"features": np.transpose(feature_map, [1, 2, 0]),
"original_image_size": (img_size_orig[1], img_size_orig[0])
}
sio.savemat(save_path, matdict, appendmat=False)
elif ext == 'npy':
np.save(save_path, feature_map)
else:
raise ValueError(
'invalid file extension for save_path, only mat and np supported'
)
return feature_map
def run_and_save(
self,
img_path,
seg_path,
pre_sliding_crop_transform=None,
sliding_crop=joint_transforms.SlidingCropImageOnly(713, 2 / 3.),
input_transform=standard_transforms.ToTensor(),
verbose=False,
skip_if_seg_exists=False,
use_gpu=True,
):
"""
img - Path of input image
seg_path - Path of output image (segmentation)
sliding_crop - Transform that returns set of image slices
input_transform - Transform to apply to image before inputting to network
skip_if_seg_exists - Whether to overwrite or skip if segmentation exists already
"""
if seg_path is not None:
if os.path.exists(seg_path):
if skip_if_seg_exists:
if verbose:
print(
"Segmentation already exists, skipping: {}".format(
seg_path))
return
else:
if verbose:
print("Segmentation already exists, overwriting: {}".
format(seg_path))
if isinstance(img_path, str):
try:
img = Image.open(img_path).convert('RGB')
except OSError:
print(
"Error reading input image, skipping: {}".format(img_path))
return
else:
img = img_path
# creating sliding crop windows and transform them
img_size_orig = img.size
if pre_sliding_crop_transform is not None: # might reshape image
img = pre_sliding_crop_transform(img)
img_slices, slices_info = sliding_crop(img)
img_slices = [input_transform(e) for e in img_slices]
img_slices = torch.stack(img_slices, 0)
slices_info = torch.LongTensor(slices_info)
slices_info.squeeze_(0)
prediction_logits = self.run_on_slices(
img_slices,
slices_info,
sliding_transform_step=sliding_crop.stride_rate,
use_gpu=use_gpu,
return_logits=True)
prediction_orig = prediction_logits.max(0)[1].squeeze_(0).numpy()
prediction_logits = prediction_logits.numpy()
if self.colorize_fcn:
prediction_colorized = self.colorize_fcn(prediction_orig)
else:
prediction_colorized = Image.fromarray(
prediction_orig.astype(np.int32)).convert('I')
if prediction_colorized.size != img_size_orig:
prediction_colorized = F.resize(prediction_colorized,
img_size_orig[::-1],
interpolation=Image.NEAREST)
if seg_path is not None:
check_mkdir(os.path.dirname(seg_path))
prediction_colorized.save(seg_path)
return prediction_colorized