def loss(self, datum, output):
score = utils.precision_at_one(output['pred'],
datum['semantic_target'])
return {
'score': torch.FloatTensor([score])[0].to(output['outputs'].F),
'loss': self.criterion(output['outputs'].F,
datum['semantic_target'])
}
def test(model, data_loader, config, transform_data_fn=None, has_gt=True):
device = get_torch_device(config.is_cuda)
dataset = data_loader.dataset
num_labels = dataset.NUM_LABELS
global_timer, data_timer, iter_timer = Timer(), Timer(), Timer()
criterion = nn.CrossEntropyLoss(ignore_index=config.ignore_label)
losses, scores, ious = AverageMeter(), AverageMeter(), 0
aps = np.zeros((0, num_labels))
hist = np.zeros((num_labels, num_labels))
logging.info('===> Start testing')
global_timer.tic()
data_iter = data_loader.__iter__()
max_iter = len(data_loader)
max_iter_unique = max_iter
# Fix batch normalization running mean and std
model.eval()
# Clear cache (when run in val mode, cleanup training cache)
torch.cuda.empty_cache()
if config.save_prediction or config.test_original_pointcloud:
if config.save_prediction:
save_pred_dir = config.save_pred_dir
os.makedirs(save_pred_dir, exist_ok=True)
else:
save_pred_dir = tempfile.mkdtemp()
if os.listdir(save_pred_dir):
raise ValueError(f'Directory {save_pred_dir} not empty. '
'Please remove the existing prediction.')
with torch.no_grad():
for iteration in range(max_iter):
data_timer.tic()
if config.return_transformation:
coords, input, target, transformation = data_iter.next()
else:
coords, input, target = data_iter.next()
transformation = None
data_time = data_timer.toc(False)
# Preprocess input
iter_timer.tic()
if config.wrapper_type != 'None':
color = input[:, :3].int()
if config.normalize_color:
input[:, :3] = input[:, :3] / 255. - 0.5
sinput = SparseTensor(input, coords).to(device)
# Feed forward
inputs = (sinput, ) if config.wrapper_type == 'None' else (sinput,
coords,
color)
soutput = model(*inputs)
output = soutput.F
pred = get_prediction(dataset, output, target).int()
iter_time = iter_timer.toc(False)
if config.save_prediction or config.test_original_pointcloud:
save_predictions(coords, pred, transformation, dataset, config,
iteration, save_pred_dir)
if has_gt:
if config.evaluate_original_pointcloud:
raise NotImplementedError('pointcloud')
output, pred, target = permute_pointcloud(
coords, pointcloud, transformation, dataset.label_map,
output, pred)
target_np = target.numpy()
num_sample = target_np.shape[0]
target = target.to(device)
cross_ent = criterion(output, target.long())
losses.update(float(cross_ent), num_sample)
scores.update(precision_at_one(pred, target), num_sample)
hist += fast_hist(pred.cpu().numpy().flatten(),
target_np.flatten(), num_labels)
ious = per_class_iu(hist) * 100
prob = torch.nn.functional.softmax(output, dim=1)
ap = average_precision(prob.cpu().detach().numpy(), target_np)
aps = np.vstack((aps, ap))
# Due to heavy bias in class, there exists class with no test label at all
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=RuntimeWarning)
ap_class = np.nanmean(aps, 0) * 100.
if iteration % config.test_stat_freq == 0 and iteration > 0:
reordered_ious = dataset.reorder_result(ious)
reordered_ap_class = dataset.reorder_result(ap_class)
class_names = dataset.get_classnames()
print_info(iteration,
max_iter_unique,
data_time,
iter_time,
has_gt,
losses,
scores,
reordered_ious,
hist,
reordered_ap_class,
class_names=class_names)
if iteration % config.empty_cache_freq == 0:
# Clear cache
torch.cuda.empty_cache()
global_time = global_timer.toc(False)
reordered_ious = dataset.reorder_result(ious)
reordered_ap_class = dataset.reorder_result(ap_class)
class_names = dataset.get_classnames()
print_info(iteration,
max_iter_unique,
data_time,
iter_time,
has_gt,
losses,
scores,
reordered_ious,
hist,
reordered_ap_class,
class_names=class_names)
if config.test_original_pointcloud:
logging.info('===> Start testing on original pointcloud space.')
dataset.test_pointcloud(save_pred_dir)
logging.info("Finished test. Elapsed time: {:.4f}".format(global_time))
return losses.avg, scores.avg, np.nanmean(ap_class), np.nanmean(
per_class_iu(hist)) * 100
def __call__(self, inferred_values, gt_data, phase='train'):
# Initialize the dictionary
metrics = {}
if (self.args['method']['flow'] and self.args['metrics']['flow']):
assert (('refined_flow' in inferred_values) &
('flow_eval' in gt_data)), "Flow metrics selected \
but either est or gt flow not provided"
gt_flow = gt_data['flow'] if phase == 'train' else gt_data[
'flow_eval']
# Compute the end point error of the flow vectors
# If bg/fg labels are available use them to also compute f-EPE and b-EPE
if 'fg_labels_eval_s' in gt_data and self.args['data'][
'dataset'] not in ["FlyingThings3D_ME", "StereoKITTI_ME"]:
gt_label = gt_data[
'fg_labels_s'] if phase == 'train' else gt_data[
'fg_labels_eval_s']
ego_metrics = compute_epe(
inferred_values['refined_rigid_flow'],
gt_flow,
sem_label=gt_label,
eval_stats=True)
else:
ego_metrics = compute_epe(
inferred_values['refined_rigid_flow'],
gt_flow,
eval_stats=True)
for key, value in ego_metrics.items():
metrics[key] = value
# Compute the ego-motion metric
if self.args['method']['ego_motion'] and self.args['metrics'][
'ego_motion']:
assert (('R_est' in inferred_values) &
('R_ego' in gt_data)), "Ego motion metric selected \
but either est or gt ego motion not provided"
r_error = rotation_error(inferred_values['R_est'],
gt_data['R_ego'])
metrics['mean_r_error'] = torch.mean(r_error).item()
metrics['max_r_error'] = torch.max(r_error).item()
metrics['min_r_error'] = torch.min(r_error).item()
t_error = translation_error(inferred_values['t_est'],
gt_data['t_ego'])
metrics['mean_t_error'] = torch.mean(t_error).item()
metrics['max_t_error'] = torch.max(t_error).item()
metrics['min_t_error'] = torch.min(t_error).item()
# Compute the background segmentation metric
if self.args['method']['semantic'] and self.args['metrics']['semantic']:
assert (('semantic_logits_s_all' in inferred_values) &
('fg_labels_eval_s'
in gt_data)), "Background segmentation metric selected \
but either est or gt labels not provided"
gt_label = gt_data['fg_labels_s'] if phase == 'train' else gt_data[
'fg_labels_eval_s']
pred_label = inferred_values['semantic_logits_s_all'].max(1)[1]
pre_f, pre_b, rec_f, rec_b = precision_at_one(pred_label, gt_label)
metrics['precision_f'] = pre_f.item()
metrics['recall_f'] = rec_f.item()
metrics['precision_b'] = pre_b.item()
metrics['recall_b'] = rec_b.item()
true_p, true_n, false_p, false_n = evaluate_binary_class(
pred_label, gt_label)
metrics['true_p'] = true_p.item()
metrics['true_n'] = true_n.item()
metrics['false_p'] = false_p.item()
metrics['false_n'] = false_n.item()
return metrics
def test(model,
data_loader,
config,
transform_data_fn=None,
has_gt=True,
validation=None,
epoch=None):
device = get_torch_device(config.is_cuda)
dataset = data_loader.dataset
num_labels = dataset.NUM_LABELS
global_timer, data_timer, iter_timer = Timer(), Timer(), Timer()
criterion = nn.CrossEntropyLoss(ignore_index=config.ignore_label)
alpha, gamma, eps = 1, 2, 1e-6 # Focal Loss parameters
losses, scores, ious = AverageMeter(), AverageMeter(), 0
aps = np.zeros((0, num_labels))
hist = np.zeros((num_labels, num_labels))
if not config.is_train:
checkpoint_fn = config.resume + '/weights.pth'
if osp.isfile(checkpoint_fn):
logging.info("=> loading checkpoint '{}'".format(checkpoint_fn))
state = torch.load(checkpoint_fn)
model.load_state_dict(state['state_dict'])
logging.info("=> loaded checkpoint '{}' (epoch {})".format(
checkpoint_fn, state['epoch']))
else:
raise ValueError(
"=> no checkpoint found at '{}'".format(checkpoint_fn))
if validation:
logging.info('===> Start validating')
else:
logging.info('===> Start testing')
global_timer.tic()
data_iter = data_loader.__iter__()
max_iter = len(data_loader)
max_iter_unique = max_iter
all_preds, all_labels, batch_losses, batch_loss = [], [], {}, 0
# Fix batch normalization running mean and std
model.eval()
# Clear cache (when run in val mode, cleanup training cache)
torch.cuda.empty_cache()
if config.save_prediction or config.test_original_pointcloud:
if config.save_prediction:
save_pred_dir = config.save_pred_dir
os.makedirs(save_pred_dir, exist_ok=True)
else:
save_pred_dir = tempfile.mkdtemp()
if os.listdir(save_pred_dir):
raise ValueError(f'Directory {save_pred_dir} not empty. '
'Please remove the existing prediction.')
with torch.no_grad():
for iteration in range(max_iter):
data_timer.tic()
if config.return_transformation:
coords, input, target, transformation = data_iter.next()
else:
coords, input, target = data_iter.next()
transformation = None
data_time = data_timer.toc(False)
# Preprocess input
iter_timer.tic()
if config.wrapper_type != 'None':
color = input[:, :3].int()
if config.normalize_color:
input[:, :3] = input[:, :3] / 255. - 0.5
sinput = SparseTensor(input, coords).to(device)
# Feed forward
inputs = (sinput, ) if config.wrapper_type == 'None' else (sinput,
coords,
color)
soutput = model(*inputs)
output = soutput.F
pred = get_prediction(dataset, output, target).int()
iter_time = iter_timer.toc(False)
all_preds.append(pred.cpu().detach().numpy())
all_labels.append(target.cpu().detach().numpy())
if config.save_prediction or config.test_original_pointcloud:
save_predictions(coords, pred, transformation, dataset, config,
iteration, save_pred_dir)
if has_gt:
if config.evaluate_original_pointcloud:
raise NotImplementedError('pointcloud')
output, pred, target = permute_pointcloud(
coords, pointcloud, transformation, dataset.label_map,
output, pred)
target_np = target.numpy()
num_sample = target_np.shape[0]
target = target.to(device)
"""# focal loss
input_soft = nn.functional.softmax(output, dim=1) + eps
focal_weight = torch.pow(-input_soft + 1., gamma)
loss = (-alpha * focal_weight * torch.log(input_soft)).mean()"""
loss = criterion(output, target.long())
batch_loss += loss
losses.update(float(loss), num_sample)
scores.update(precision_at_one(pred, target), num_sample)
hist += fast_hist(pred.cpu().numpy().flatten(),
target_np.flatten(), num_labels)
ious = per_class_iu(hist) * 100
prob = torch.nn.functional.softmax(output, dim=1)
ap = average_precision(prob.cpu().detach().numpy(), target_np)
aps = np.vstack((aps, ap))
# Due to heavy bias in class, there exists class with no test label at all
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=RuntimeWarning)
ap_class = np.nanmean(aps, 0) * 100.
if iteration % config.test_stat_freq == 0 and iteration > 0:
preds = np.concatenate(all_preds)
targets = np.concatenate(all_labels)
to_ignore = [
i for i in range(len(targets)) if targets[i] == 255
]
preds_trunc = [
preds[i] for i in range(len(preds)) if i not in to_ignore
]
targets_trunc = [
targets[i] for i in range(len(targets))
if i not in to_ignore
]
cm = confusion_matrix(targets_trunc,
preds_trunc,
normalize='true')
np.savetxt(config.log_dir + '/cm_epoch_{0}.txt'.format(epoch),
cm)
reordered_ious = dataset.reorder_result(ious)
reordered_ap_class = dataset.reorder_result(ap_class)
class_names = dataset.get_classnames()
print_info(iteration,
max_iter_unique,
data_time,
iter_time,
has_gt,
losses,
scores,
reordered_ious,
hist,
reordered_ap_class,
class_names=class_names)
if iteration % config.empty_cache_freq == 0:
# Clear cache
torch.cuda.empty_cache()
batch_losses[epoch] = batch_loss
global_time = global_timer.toc(False)
reordered_ious = dataset.reorder_result(ious)
reordered_ap_class = dataset.reorder_result(ap_class)
class_names = dataset.get_classnames()
print_info(iteration,
max_iter_unique,
data_time,
iter_time,
has_gt,
losses,
scores,
reordered_ious,
hist,
reordered_ap_class,
class_names=class_names)
if not config.is_train:
preds = np.concatenate(all_preds)
targets = np.concatenate(all_labels)
to_ignore = [i for i in range(len(targets)) if targets[i] == 255]
preds_trunc = [
preds[i] for i in range(len(preds)) if i not in to_ignore
]
targets_trunc = [
targets[i] for i in range(len(targets)) if i not in to_ignore
]
cm = confusion_matrix(targets_trunc, preds_trunc, normalize='true')
np.savetxt(config.log_dir + '/cm.txt', cm)
if config.test_original_pointcloud:
logging.info('===> Start testing on original pointcloud space.')
dataset.test_pointcloud(save_pred_dir)
logging.info("Finished test. Elapsed time: {:.4f}".format(global_time))
if validation:
loss_file_name = "/val_loss.txt"
with open(config.log_dir + loss_file_name, 'a') as val_loss_file:
for key in batch_losses:
val_loss_file.writelines('{0}, {1}\n'.format(
batch_losses[key], key))
val_loss_file.close()
return losses.avg, scores.avg, np.nanmean(ap_class), np.nanmean(
per_class_iu(hist)) * 100, batch_losses
else:
return losses.avg, scores.avg, np.nanmean(ap_class), np.nanmean(
per_class_iu(hist)) * 100
def train(model, data_loader, val_data_loader, config, transform_data_fn=None):
device = config.device_id
distributed = get_world_size() > 1
# Set up the train flag for batch normalization
model.train()
# Configuration
writer = SummaryWriter(log_dir=config.log_dir)
data_timer, iter_timer = Timer(), Timer()
fw_timer, bw_timer, ddp_timer = Timer(), Timer(), Timer()
data_time_avg, iter_time_avg = AverageMeter(), AverageMeter()
fw_time_avg, bw_time_avg, ddp_time_avg = AverageMeter(), AverageMeter(
), AverageMeter()
losses, scores = AverageMeter(), AverageMeter()
optimizer = initialize_optimizer(model.parameters(), config)
scheduler = initialize_scheduler(optimizer, config)
criterion = nn.CrossEntropyLoss(ignore_index=config.ignore_label)
writer = SummaryWriter(log_dir=config.log_dir)
# Train the network
logging.info('===> Start training on {} GPUs, batch-size={}'.format(
get_world_size(), config.batch_size * get_world_size()))
best_val_miou, best_val_iter, curr_iter, epoch, is_training = 0, 0, 1, 1, True
if config.resume:
checkpoint_fn = config.resume + '/weights.pth'
if osp.isfile(checkpoint_fn):
logging.info("=> loading checkpoint '{}'".format(checkpoint_fn))
state = torch.load(
checkpoint_fn,
map_location=lambda s, l: default_restore_location(s, 'cpu'))
curr_iter = state['iteration'] + 1
epoch = state['epoch']
load_state(model, state['state_dict'])
if config.resume_optimizer:
scheduler = initialize_scheduler(optimizer,
config,
last_step=curr_iter)
optimizer.load_state_dict(state['optimizer'])
if 'best_val' in state:
best_val_miou = state['best_val']
best_val_iter = state['best_val_iter']
logging.info("=> loaded checkpoint '{}' (epoch {})".format(
checkpoint_fn, state['epoch']))
else:
raise ValueError(
"=> no checkpoint found at '{}'".format(checkpoint_fn))
data_iter = data_loader.__iter__() # (distributed) infinite sampler
while is_training:
for iteration in range(len(data_loader) // config.iter_size):
optimizer.zero_grad()
data_time, batch_loss, batch_score = 0, 0, 0
iter_timer.tic()
# set random seed for every iteration for trackability
_set_seed(config, curr_iter)
for sub_iter in range(config.iter_size):
# Get training data
data_timer.tic()
coords, input, target = data_iter.next()
# For some networks, making the network invariant to even, odd coords is important
coords[:, :3] += (torch.rand(3) * 100).type_as(coords)
# Preprocess input
color = input[:, :3].int()
if config.normalize_color:
input[:, :3] = input[:, :3] / 255. - 0.5
sinput = SparseTensor(input, coords).to(device)
data_time += data_timer.toc(False)
# Feed forward
fw_timer.tic()
inputs = (sinput, ) if config.wrapper_type == 'None' else (
sinput, coords, color)
# model.initialize_coords(*init_args)
soutput = model(*inputs)
# The output of the network is not sorted
target = target.long().to(device)
loss = criterion(soutput.F, target.long())
# Compute and accumulate gradient
loss /= config.iter_size
pred = get_prediction(data_loader.dataset, soutput.F, target)
score = precision_at_one(pred, target)
fw_timer.toc(False)
bw_timer.tic()
# bp the loss
loss.backward()
bw_timer.toc(False)
# gather information
logging_output = {
'loss': loss.item(),
'score': score / config.iter_size
}
ddp_timer.tic()
if distributed:
logging_output = all_gather_list(logging_output)
logging_output = {
w: np.mean([a[w] for a in logging_output])
for w in logging_output[0]
}
batch_loss += logging_output['loss']
batch_score += logging_output['score']
ddp_timer.toc(False)
# Update number of steps
optimizer.step()
scheduler.step()
data_time_avg.update(data_time)
iter_time_avg.update(iter_timer.toc(False))
fw_time_avg.update(fw_timer.diff)
bw_time_avg.update(bw_timer.diff)
ddp_time_avg.update(ddp_timer.diff)
losses.update(batch_loss, target.size(0))
scores.update(batch_score, target.size(0))
if curr_iter >= config.max_iter:
is_training = False
break
if curr_iter % config.stat_freq == 0 or curr_iter == 1:
lrs = ', '.join(
['{:.3e}'.format(x) for x in scheduler.get_lr()])
debug_str = "===> Epoch[{}]({}/{}): Loss {:.4f}\tLR: {}\t".format(
epoch, curr_iter,
len(data_loader) // config.iter_size, losses.avg, lrs)
debug_str += "Score {:.3f}\tData time: {:.4f}, Forward time: {:.4f}, Backward time: {:.4f}, DDP time: {:.4f}, Total iter time: {:.4f}".format(
scores.avg, data_time_avg.avg, fw_time_avg.avg,
bw_time_avg.avg, ddp_time_avg.avg, iter_time_avg.avg)
logging.info(debug_str)
# Reset timers
data_time_avg.reset()
iter_time_avg.reset()
# Write logs
writer.add_scalar('training/loss', losses.avg, curr_iter)
writer.add_scalar('training/precision_at_1', scores.avg,
curr_iter)
writer.add_scalar('training/learning_rate',
scheduler.get_lr()[0], curr_iter)
losses.reset()
scores.reset()
# Save current status, save before val to prevent occational mem overflow
if curr_iter % config.save_freq == 0:
checkpoint(model, optimizer, epoch, curr_iter, config,
best_val_miou, best_val_iter)
# Validation
if curr_iter % config.val_freq == 0:
val_miou = validate(model, val_data_loader, writer, curr_iter,
config, transform_data_fn)
if val_miou > best_val_miou:
best_val_miou = val_miou
best_val_iter = curr_iter
checkpoint(model, optimizer, epoch, curr_iter, config,
best_val_miou, best_val_iter, "best_val")
logging.info("Current best mIoU: {:.3f} at iter {}".format(
best_val_miou, best_val_iter))
# Recover back
model.train()
if curr_iter % config.empty_cache_freq == 0:
# Clear cache
torch.cuda.empty_cache()
# End of iteration
curr_iter += 1
epoch += 1
# Explicit memory cleanup
if hasattr(data_iter, 'cleanup'):
data_iter.cleanup()
# Save the final model
checkpoint(model, optimizer, epoch, curr_iter, config, best_val_miou,
best_val_iter)
val_miou = validate(model, val_data_loader, writer, curr_iter, config,
transform_data_fn)
if val_miou > best_val_miou:
best_val_miou = val_miou
best_val_iter = curr_iter
checkpoint(model, optimizer, epoch, curr_iter, config, best_val_miou,
best_val_iter, "best_val")
logging.info("Current best mIoU: {:.3f} at iter {}".format(
best_val_miou, best_val_iter))
def train(model, data_loader, val_data_loader, config, transform_data_fn=None):
device = get_torch_device(config.is_cuda)
# Set up the train flag for batch normalization
model.train()
# Configuration
data_timer, iter_timer = Timer(), Timer()
data_time_avg, iter_time_avg = AverageMeter(), AverageMeter()
regs, losses, scores = AverageMeter(), AverageMeter(), AverageMeter()
optimizer = initialize_optimizer(model.parameters(), config)
scheduler = initialize_scheduler(optimizer, config)
criterion = nn.CrossEntropyLoss(ignore_index=config.ignore_label)
# Train the network
logging.info('===> Start training')
best_val_miou, best_val_iter, curr_iter, epoch, is_training = 0, 0, 1, 1, True
if config.resume:
# Test loaded ckpt first
v_loss, v_score, v_mAP, v_mIoU = test(model, val_data_loader, config)
checkpoint_fn = config.resume + '/weights.pth'
if osp.isfile(checkpoint_fn):
logging.info("=> loading checkpoint '{}'".format(checkpoint_fn))
state = torch.load(checkpoint_fn)
curr_iter = state['iteration'] + 1
epoch = state['epoch']
# we skip attention maps because the shape won't match because voxel number is different
# e.g. copyting a param with shape (23385, 8, 4) to (43529, 8, 4)
d = {
k: v
for k, v in state['state_dict'].items() if 'map' not in k
}
# handle those attn maps we don't load from saved dict
for k in model.state_dict().keys():
if k in d.keys(): continue
d[k] = model.state_dict()[k]
model.load_state_dict(d)
if config.resume_optimizer:
scheduler = initialize_scheduler(optimizer,
config,
last_step=curr_iter)
optimizer.load_state_dict(state['optimizer'])
if 'best_val' in state:
best_val_miou = state['best_val']
best_val_iter = state['best_val_iter']
logging.info("=> loaded checkpoint '{}' (epoch {})".format(
checkpoint_fn, state['epoch']))
else:
raise ValueError(
"=> no checkpoint found at '{}'".format(checkpoint_fn))
data_iter = data_loader.__iter__()
if config.dataset == "SemanticKITTI":
num_class = 19
config.normalize_color = False
config.xyz_input = False
val_freq_ = config.val_freq
config.val_freq = config.val_freq * 10
elif config.dataset == "S3DIS":
num_class = 13
config.normalize_color = False
config.xyz_input = False
val_freq_ = config.val_freq
config.val_freq = config.val_freq
elif config.dataset == "Nuscenes":
num_class = 16
config.normalize_color = False
config.xyz_input = False
val_freq_ = config.val_freq
config.val_freq = config.val_freq * 50
else:
num_class = 20
val_freq_ = config.val_freq
while is_training:
total_correct_class = torch.zeros(num_class, device=device)
total_iou_deno_class = torch.zeros(num_class, device=device)
for iteration in range(len(data_loader) // config.iter_size):
optimizer.zero_grad()
data_time, batch_loss = 0, 0
iter_timer.tic()
if curr_iter >= config.max_iter:
# if curr_iter >= max(config.max_iter, config.epochs*(len(data_loader) // config.iter_size):
is_training = False
break
elif curr_iter >= config.max_iter * (2 / 3):
config.val_freq = val_freq_ * 2 # valid more freq on lower half
for sub_iter in range(config.iter_size):
# Get training data
data_timer.tic()
pointcloud = None
if config.return_transformation:
coords, input, target, _, _, pointcloud, transformation, _ = data_iter.next(
)
else:
coords, input, target, _, _, _ = data_iter.next(
) # ignore unique_map and inverse_map
if config.use_aux:
assert target.shape[1] == 2
aux = target[:, 1]
target = target[:, 0]
else:
aux = None
# For some networks, making the network invariant to even, odd coords is important
coords[:, 1:] += (torch.rand(3) * 100).type_as(coords)
# Preprocess input
if config.normalize_color:
input[:, :3] = input[:, :3] / input[:, :3].max() - 0.5
coords_norm = coords[:, 1:] / coords[:, 1:].max() - 0.5
# cat xyz into the rgb feature
if config.xyz_input:
input = torch.cat([coords_norm, input], dim=1)
sinput = SparseTensor(input, coords, device=device)
starget = SparseTensor(
target.unsqueeze(-1).float(),
coordinate_map_key=sinput.coordinate_map_key,
coordinate_manager=sinput.coordinate_manager,
device=device
) # must share the same coord-manager to align for sinput
data_time += data_timer.toc(False)
# model.initialize_coords(*init_args)
# d = {}
# d['c'] = sinput.C
# d['l'] = starget.F
# torch.save('./plot/test-label.pth')
# import ipdb; ipdb.set_trace()
# Set up profiler
# memory_profiler = CUDAMemoryProfiler(
# [model, criterion],
# filename="cuda_memory.profile"
# )
# sys.settrace(memory_profiler)
# threading.settrace(memory_profiler)
# with torch.autograd.profiler.profile(enabled=True, use_cuda=True, record_shapes=False, profile_memory=True) as prof0:
if aux is not None:
soutput = model(sinput, aux)
elif config.enable_point_branch:
soutput = model(sinput,
iter_=curr_iter / config.max_iter,
enable_point_branch=True)
else:
# label-aux, feed it in as additional reg
soutput = model(
sinput, iter_=curr_iter / config.max_iter, aux=starget
) # feed in the progress of training for annealing inside the model
# The output of the network is not sorted
target = target.view(-1).long().to(device)
loss = criterion(soutput.F, target.long())
# ====== other loss regs =====
if hasattr(model, 'block1'):
cur_loss = torch.tensor([0.], device=device)
if hasattr(model.block1[0], 'vq_loss'):
if model.block1[0].vq_loss is not None:
cur_loss = torch.tensor([0.], device=device)
for n, m in model.named_children():
if 'block' in n:
cur_loss += m[
0].vq_loss # m is the nn.Sequential obj, m[0] is the TRBlock
logging.info(
'Cur Loss: {}, Cur vq_loss: {}'.format(
loss, cur_loss))
loss += cur_loss
if hasattr(model.block1[0], 'diverse_loss'):
if model.block1[0].diverse_loss is not None:
cur_loss = torch.tensor([0.], device=device)
for n, m in model.named_children():
if 'block' in n:
cur_loss += m[
0].diverse_loss # m is the nn.Sequential obj, m[0] is the TRBlock
logging.info(
'Cur Loss: {}, Cur diverse _loss: {}'.format(
loss, cur_loss))
loss += cur_loss
if hasattr(model.block1[0], 'label_reg'):
if model.block1[0].label_reg is not None:
cur_loss = torch.tensor([0.], device=device)
for n, m in model.named_children():
if 'block' in n:
cur_loss += m[
0].label_reg # m is the nn.Sequential obj, m[0] is the TRBlock
# logging.info('Cur Loss: {}, Cur diverse _loss: {}'.format(loss, cur_loss))
loss += cur_loss
# Compute and accumulate gradient
loss /= config.iter_size
batch_loss += loss.item()
loss.backward()
# soutput = model(sinput)
# Update number of steps
if not config.use_sam:
optimizer.step()
else:
optimizer.first_step(zero_grad=True)
soutput = model(sinput,
iter_=curr_iter / config.max_iter,
aux=starget)
criterion(soutput.F, target.long()).backward()
optimizer.second_step(zero_grad=True)
if config.lr_warmup is None:
scheduler.step()
else:
if curr_iter >= config.lr_warmup:
scheduler.step()
for g in optimizer.param_groups:
g['lr'] = config.lr * (iteration + 1) / config.lr_warmup
# CLEAR CACHE!
torch.cuda.empty_cache()
data_time_avg.update(data_time)
iter_time_avg.update(iter_timer.toc(False))
pred = get_prediction(data_loader.dataset, soutput.F, target)
score = precision_at_one(pred, target, ignore_label=-1)
regs.update(cur_loss.item(), target.size(0))
losses.update(batch_loss, target.size(0))
scores.update(score, target.size(0))
# calc the train-iou
for l in range(num_class):
total_correct_class[l] += ((pred == l) & (target == l)).sum()
total_iou_deno_class[l] += (((pred == l) & (target != -1)) |
(target == l)).sum()
if curr_iter % config.stat_freq == 0 or curr_iter == 1:
lrs = ', '.join(
['{:.3e}'.format(x) for x in scheduler.get_lr()])
IoU = ((total_correct_class) /
(total_iou_deno_class + 1e-6)).mean() * 100.
debug_str = "[{}] ===> Epoch[{}]({}/{}): Loss {:.4f}\tLR: {}\t".format(
config.log_dir.split('/')[-2], epoch, curr_iter,
len(data_loader) // config.iter_size, losses.avg, lrs)
debug_str += "Score {:.3f}\tIoU {:.3f}\tData time: {:.4f}, Iter time: {:.4f}".format(
scores.avg, IoU.item(), data_time_avg.avg,
iter_time_avg.avg)
if regs.avg > 0:
debug_str += "\n Additional Reg Loss {:.3f}".format(
regs.avg)
# print(debug_str)
logging.info(debug_str)
# Reset timers
data_time_avg.reset()
iter_time_avg.reset()
# Write logs
losses.reset()
scores.reset()
# Save current status, save before val to prevent occational mem overflow
if curr_iter % config.save_freq == 0:
checkpoint(model,
optimizer,
epoch,
curr_iter,
config,
best_val_miou,
best_val_iter,
save_inter=True)
# Validation
if curr_iter % config.val_freq == 0:
val_miou = validate(model, val_data_loader, None, curr_iter,
config, transform_data_fn)
if val_miou > best_val_miou:
best_val_miou = val_miou
best_val_iter = curr_iter
checkpoint(model,
optimizer,
epoch,
curr_iter,
config,
best_val_miou,
best_val_iter,
"best_val",
save_inter=True)
logging.info("Current best mIoU: {:.3f} at iter {}".format(
best_val_miou, best_val_iter))
# print("Current best mIoU: {:.3f} at iter {}".format(best_val_miou, best_val_iter))
# Recover back
model.train()
# End of iteration
curr_iter += 1
IoU = (total_correct_class) / (total_iou_deno_class + 1e-6)
logging.info('train point avg class IoU: %f' % ((IoU).mean() * 100.))
epoch += 1
# Explicit memory cleanup
if hasattr(data_iter, 'cleanup'):
data_iter.cleanup()
# Save the final model
checkpoint(model, optimizer, epoch, curr_iter, config, best_val_miou,
best_val_iter)
v_loss, v_score, v_mAP, val_miou = test(model, val_data_loader, config)
if val_miou > best_val_miou:
best_val_miou = val_miou
best_val_iter = curr_iter
checkpoint(model, optimizer, epoch, curr_iter, config, best_val_miou,
best_val_iter, "best_val")
logging.info("Current best mIoU: {:.3f} at iter {}".format(
best_val_miou, best_val_iter))
def train_distill(model,
data_loader,
val_data_loader,
config,
transform_data_fn=None):
'''
the distillation training
some cfgs here
'''
# distill_lambda = 1
# distill_lambda = 0.33
distill_lambda = 0.67
# TWO_STAGE=True: Transformer is first trained with L2 loss to match ResNet's activation, and then it fintunes like normal training on the second stage.
# TWO_STAGE=False: Transformer trains with combined loss
TWO_STAGE = False
# STAGE_PERCENTAGE = 0.7
device = get_torch_device(config.is_cuda)
# Set up the train flag for batch normalization
model.train()
# Configuration
data_timer, iter_timer = Timer(), Timer()
data_time_avg, iter_time_avg = AverageMeter(), AverageMeter()
losses, scores = AverageMeter(), AverageMeter()
optimizer = initialize_optimizer(model.parameters(), config)
scheduler = initialize_scheduler(optimizer, config)
criterion = nn.CrossEntropyLoss(ignore_index=config.ignore_label)
# Train the network
logging.info('===> Start training')
best_val_miou, best_val_iter, curr_iter, epoch, is_training = 0, 0, 1, 1, True
# TODO:
# load the sub-model only
# FIXME: some dirty hard-written stuff, only supporting current state
tch_model_cls = load_model('Res16UNet18A')
tch_model = tch_model_cls(3, 20, config).to(device)
# checkpoint_fn = "/home/zhaotianchen/project/point-transformer/SpatioTemporalSegmentation-ScanNet/outputs/ScannetSparseVoxelizationDataset/Res16UNet18A/resnet_base/weights.pth"
checkpoint_fn = "/home/zhaotianchen/project/point-transformer/SpatioTemporalSegmentation-ScanNet/outputs/ScannetSparseVoxelizationDataset/Res16UNet18A/Res18A/weights.pth" # voxel-size: 0.05
assert osp.isfile(checkpoint_fn)
logging.info("=> loading checkpoint '{}'".format(checkpoint_fn))
state = torch.load(checkpoint_fn)
d = {k: v for k, v in state['state_dict'].items() if 'map' not in k}
tch_model.load_state_dict(d)
if 'best_val' in state:
best_val_miou = state['best_val']
best_val_iter = state['best_val_iter']
logging.info("=> loaded checkpoint '{}' (epoch {})".format(
checkpoint_fn, state['epoch']))
if config.resume:
raise NotImplementedError
# Test loaded ckpt first
# checkpoint_fn = config.resume + '/weights.pth'
# if osp.isfile(checkpoint_fn):
# logging.info("=> loading checkpoint '{}'".format(checkpoint_fn))
# state = torch.load(checkpoint_fn)
# curr_iter = state['iteration'] + 1
# epoch = state['epoch']
# d = {k:v for k,v in state['state_dict'].items() if 'map' not in k }
# model.load_state_dict(d)
# if config.resume_optimizer:
# scheduler = initialize_scheduler(optimizer, config, last_step=curr_iter)
# optimizer.load_state_dict(state['optimizer'])
# if 'best_val' in state:
# best_val_miou = state['best_val']
# best_val_iter = state['best_val_iter']
# logging.info("=> loaded checkpoint '{}' (epoch {})".format(checkpoint_fn, state['epoch']))
# else:
# raise ValueError("=> no checkpoint found at '{}'".format(checkpoint_fn))
# test after loading the ckpt
v_loss, v_score, v_mAP, v_mIoU = test(tch_model, val_data_loader, config)
logging.info('Tch model tested, bes_miou: {}'.format(v_mIoU))
data_iter = data_loader.__iter__()
while is_training:
num_class = 20
total_correct_class = torch.zeros(num_class, device=device)
total_iou_deno_class = torch.zeros(num_class, device=device)
total_iteration = len(data_loader) // config.iter_size
for iteration in range(total_iteration):
# NOTE: for single stage distillation, L2 loss might be too large at first
# so we added a warmup training that don't use L2 loss
if iteration < 0:
use_distill = False
else:
use_distill = True
# Stage 1 / Stage 2 boundary
if TWO_STAGE:
stage_boundary = int(total_iteration * STAGE_PERCENTAGE)
optimizer.zero_grad()
data_time, batch_loss = 0, 0
iter_timer.tic()
for sub_iter in range(config.iter_size):
# Get training data
data_timer.tic()
if config.return_transformation:
coords, input, target, _, _, pointcloud, transformation = data_iter.next(
)
else:
coords, input, target, _, _ = data_iter.next(
) # ignore unique_map and inverse_map
if config.use_aux:
assert target.shape[1] == 2
aux = target[:, 1]
target = target[:, 0]
else:
aux = None
# For some networks, making the network invariant to even, odd coords is important
coords[:, 1:] += (torch.rand(3) * 100).type_as(coords)
# Preprocess input
if config.normalize_color:
input[:, :3] = input[:, :3] / 255. - 0.5
coords_norm = coords[:, 1:] / coords[:, 1:].max() - 0.5
# cat xyz into the rgb feature
if config.xyz_input:
input = torch.cat([coords_norm, input], dim=1)
sinput = SparseTensor(input, coords, device=device)
# TODO: return both-models
# in order to not breaking the valid interface, use a get_loss to get the regsitered loss
data_time += data_timer.toc(False)
# model.initialize_coords(*init_args)
if aux is not None:
raise NotImplementedError
# flatten ground truth tensor
target = target.view(-1).long().to(device)
if TWO_STAGE:
if iteration < stage_boundary:
# Stage 1: train transformer on L2 loss
soutput, anchor = model(sinput, save_anchor=True)
# Make sure gradient don't flow to teacher model
with torch.no_grad():
_, tch_anchor = tch_model(sinput, save_anchor=True)
loss = DistillLoss(tch_anchor, anchor)
else:
# Stage 2: finetune transformer on Cross-Entropy
soutput = model(sinput)
loss = criterion(soutput.F, target.long())
else:
if use_distill: # after warm up
soutput, anchor = model(sinput, save_anchor=True)
# if pretrained teacher, do not let the grad flow to teacher to update its params
with torch.no_grad():
tch_soutput, tch_anchor = tch_model(
sinput, save_anchor=True)
else: # warming up
soutput = model(sinput)
# The output of the network is not sorted
loss = criterion(soutput.F, target.long())
# Add L2 loss if use distillation
if use_distill:
distill_loss = DistillLoss(tch_anchor,
anchor) * distill_lambda
loss += distill_loss
# Compute and accumulate gradient
loss /= config.iter_size
batch_loss += loss.item()
loss.backward()
# Update number of steps
optimizer.step()
scheduler.step()
# CLEAR CACHE!
torch.cuda.empty_cache()
data_time_avg.update(data_time)
iter_time_avg.update(iter_timer.toc(False))
pred = get_prediction(data_loader.dataset, soutput.F, target)
score = precision_at_one(pred, target, ignore_label=-1)
losses.update(batch_loss, target.size(0))
scores.update(score, target.size(0))
# calc the train-iou
for l in range(num_class):
total_correct_class[l] += ((pred == l) & (target == l)).sum()
total_iou_deno_class[l] += (((pred == l) & (target != -1)) |
(target == l)).sum()
if curr_iter >= config.max_iter:
is_training = False
break
if curr_iter % config.stat_freq == 0 or curr_iter == 1:
lrs = ', '.join(
['{:.3e}'.format(x) for x in scheduler.get_lr()])
debug_str = "[{}] ===> Epoch[{}]({}/{}): Loss {:.4f}\tLR: {}\t".format(
config.log_dir, epoch, curr_iter,
len(data_loader) // config.iter_size, losses.avg, lrs)
debug_str += "Score {:.3f}\tData time: {:.4f}, Iter time: {:.4f}".format(
scores.avg, data_time_avg.avg, iter_time_avg.avg)
logging.info(debug_str)
if use_distill and not TWO_STAGE:
logging.info('Loss {} Distill Loss:{}'.format(
loss, distill_loss))
# Reset timers
data_time_avg.reset()
iter_time_avg.reset()
losses.reset()
scores.reset()
# Save current status, save before val to prevent occational mem overflow
if curr_iter % config.save_freq == 0:
checkpoint(model,
optimizer,
epoch,
curr_iter,
config,
best_val_miou,
best_val_iter,
save_inter=True)
# Validation
if curr_iter % config.val_freq == 0:
val_miou = validate(model, val_data_loader, None, curr_iter,
config, transform_data_fn)
if val_miou > best_val_miou:
best_val_miou = val_miou
best_val_iter = curr_iter
checkpoint(model,
optimizer,
epoch,
curr_iter,
config,
best_val_miou,
best_val_iter,
"best_val",
save_inter=True)
logging.info("Current best mIoU: {:.3f} at iter {}".format(
best_val_miou, best_val_iter))
# Recover back
model.train()
# End of iteration
curr_iter += 1
IoU = (total_correct_class) / (total_iou_deno_class + 1e-6)
logging.info('train point avg class IoU: %f' % ((IoU).mean() * 100.))
epoch += 1
# Explicit memory cleanup
if hasattr(data_iter, 'cleanup'):
data_iter.cleanup()
# Save the final model
checkpoint(model, optimizer, epoch, curr_iter, config, best_val_miou,
best_val_iter)
v_loss, v_score, v_mAP, val_miou = test(model, val_data_loader, config)
if val_miou > best_val_miou:
best_val_miou = val_miou
best_val_iter = curr_iter
checkpoint(model, optimizer, epoch, curr_iter, config, best_val_miou,
best_val_iter, "best_val")
logging.info("Current best mIoU: {:.3f} at iter {}".format(
best_val_miou, best_val_iter))
def train_point(model,
data_loader,
val_data_loader,
config,
transform_data_fn=None):
device = get_torch_device(config.is_cuda)
# Set up the train flag for batch normalization
model.train()
# Configuration
data_timer, iter_timer = Timer(), Timer()
data_time_avg, iter_time_avg = AverageMeter(), AverageMeter()
losses, scores = AverageMeter(), AverageMeter()
optimizer = initialize_optimizer(model.parameters(), config)
scheduler = initialize_scheduler(optimizer, config)
criterion = nn.CrossEntropyLoss(ignore_index=-1)
# Train the network
logging.info('===> Start training')
best_val_miou, best_val_iter, curr_iter, epoch, is_training = 0, 0, 1, 1, True
if config.resume:
checkpoint_fn = config.resume + '/weights.pth'
if osp.isfile(checkpoint_fn):
logging.info("=> loading checkpoint '{}'".format(checkpoint_fn))
state = torch.load(checkpoint_fn)
curr_iter = state['iteration'] + 1
epoch = state['epoch']
d = {
k: v
for k, v in state['state_dict'].items() if 'map' not in k
}
model.load_state_dict(d)
if config.resume_optimizer:
scheduler = initialize_scheduler(optimizer,
config,
last_step=curr_iter)
optimizer.load_state_dict(state['optimizer'])
if 'best_val' in state:
best_val_miou = state['best_val']
best_val_iter = state['best_val_iter']
logging.info("=> loaded checkpoint '{}' (epoch {})".format(
checkpoint_fn, state['epoch']))
else:
raise ValueError(
"=> no checkpoint found at '{}'".format(checkpoint_fn))
data_iter = data_loader.__iter__()
while is_training:
num_class = 20
total_correct_class = torch.zeros(num_class, device=device)
total_iou_deno_class = torch.zeros(num_class, device=device)
for iteration in range(len(data_loader) // config.iter_size):
optimizer.zero_grad()
data_time, batch_loss = 0, 0
iter_timer.tic()
for sub_iter in range(config.iter_size):
# Get training data
data = data_iter.next()
points, target, sample_weight = data
if config.pure_point:
sinput = points.transpose(1, 2).cuda().float()
# DEBUG: use the discrete coord for point-based
'''
feats = torch.unbind(points[:,:,:], dim=0)
voxel_size = config.voxel_size
coords = torch.unbind(points[:,:,:3]/voxel_size, dim=0) # 0.05 is the voxel-size
coords, feats= ME.utils.sparse_collate(coords, feats)
# assert feats.reshape([16, 4096, -1]) == points[:,:,3:]
points_ = ME.TensorField(features=feats.float(), coordinates=coords, device=device)
tmp_voxel = points_.sparse()
sinput_ = tmp_voxel.slice(points_)
sinput = torch.cat([sinput_.C[:,1:]*config.voxel_size, sinput_.F[:,3:]],dim=1).reshape([config.batch_size, config.num_points, 6])
# sinput = sinput_.F.reshape([config.batch_size, config.num_points, 6])
sinput = sinput.transpose(1,2).cuda().float()
# sinput = torch.cat([coords[:,1:], feats],dim=1).reshape([config.batch_size, config.num_points, 6])
# sinput = sinput.transpose(1,2).cuda().float()
'''
# For some networks, making the network invariant to even, odd coords is important
# coords[:, 1:] += (torch.rand(3) * 100).type_as(coords)
# Preprocess input
# if config.normalize_color:
# feats = feats / 255. - 0.5
# torch.save(points[:,:,:3], './sandbox/tensorfield-c.pth')
# torch.save(points_.C, './sandbox/points-c.pth')
else:
# feats = torch.unbind(points[:,:,3:], dim=0) # WRONG: should also feed in xyz as inupt feature
voxel_size = config.voxel_size
coords = torch.unbind(points[:, :, :3] / voxel_size,
dim=0) # 0.05 is the voxel-size
# Normalize the xyz in feature
# points[:,:,:3] = points[:,:,:3] / points[:,:,:3].mean()
feats = torch.unbind(points[:, :, :], dim=0)
coords, feats = ME.utils.sparse_collate(coords, feats)
# For some networks, making the network invariant to even, odd coords is important
coords[:, 1:] += (torch.rand(3) * 100).type_as(coords)
# Preprocess input
# if config.normalize_color:
# feats = feats / 255. - 0.5
# they are the same
points_ = ME.TensorField(features=feats.float(),
coordinates=coords,
device=device)
# points_1 = ME.TensorField(features=feats.float(), coordinates=coords, device=device, quantization_mode=ME.SparseTensorQuantizationMode.UNWEIGHTED_AVERAGE)
# points_2 = ME.TensorField(features=feats.float(), coordinates=coords, device=device, quantization_mode=ME.SparseTensorQuantizationMode.RANDOM_SUBSAMPLE)
sinput = points_.sparse()
data_time += data_timer.toc(False)
B, npoint = target.shape
# model.initialize_coords(*init_args)
soutput = model(sinput)
if config.pure_point:
soutput = soutput.reshape([B * npoint, -1])
else:
soutput = soutput.slice(points_).F
# s1 = soutput.slice(points_)
# print(soutput.quantization_mode)
# soutput.quantization_mode = ME.SparseTensorQuantizationMode.RANDOM_SUBSAMPLE
# s2 = soutput.slice(points_)
# The output of the network is not sorted
target = (target - 1).view(-1).long().to(device)
# catch NAN
if torch.isnan(soutput).sum() > 0:
import ipdb
ipdb.set_trace()
loss = criterion(soutput, target)
if torch.isnan(loss).sum() > 0:
import ipdb
ipdb.set_trace()
loss = (loss * sample_weight.to(device)).mean()
# Compute and accumulate gradient
loss /= config.iter_size
batch_loss += loss.item()
loss.backward()
# print(model.input_mlp[0].weight.max())
# print(model.input_mlp[0].weight.grad.max())
# Update number of steps
optimizer.step()
scheduler.step()
# CLEAR CACHE!
torch.cuda.empty_cache()
data_time_avg.update(data_time)
iter_time_avg.update(iter_timer.toc(False))
pred = get_prediction(data_loader.dataset, soutput, target)
score = precision_at_one(pred, target, ignore_label=-1)
losses.update(batch_loss, target.size(0))
scores.update(score, target.size(0))
# Calc the iou
for l in range(num_class):
total_correct_class[l] += ((pred == l) & (target == l)).sum()
total_iou_deno_class[l] += (((pred == l) & (target >= 0)) |
(target == l)).sum()
if curr_iter >= config.max_iter:
is_training = False
break
if curr_iter % config.stat_freq == 0 or curr_iter == 1:
lrs = ', '.join(
['{:.3e}'.format(x) for x in scheduler.get_lr()])
debug_str = "===> Epoch[{}]({}/{}): Loss {:.4f}\tLR: {}\t".format(
epoch, curr_iter,
len(data_loader) // config.iter_size, losses.avg, lrs)
debug_str += "Score {:.3f}\tData time: {:.4f}, Iter time: {:.4f}".format(
scores.avg, data_time_avg.avg, iter_time_avg.avg)
logging.info(debug_str)
# Reset timers
data_time_avg.reset()
iter_time_avg.reset()
# Write logs
losses.reset()
scores.reset()
# Save current status, save before val to prevent occational mem overflow
if curr_iter % config.save_freq == 0:
checkpoint(model,
optimizer,
epoch,
curr_iter,
config,
best_val_miou,
best_val_iter,
save_inter=True)
# Validation:
# for point-based should use alternate dataloader for eval
# if curr_iter % config.val_freq == 0:
# val_miou = test_points(model, val_data_loader, None, curr_iter, config, transform_data_fn)
# if val_miou > best_val_miou:
# best_val_miou = val_miou
# best_val_iter = curr_iter
# checkpoint(model, optimizer, epoch, curr_iter, config, best_val_miou, best_val_iter,
# "best_val")
# logging.info("Current best mIoU: {:.3f} at iter {}".format(best_val_miou, best_val_iter))
# # Recover back
# model.train()
# End of iteration
curr_iter += 1
IoU = (total_correct_class) / (total_iou_deno_class + 1e-6)
logging.info('train point avg class IoU: %f' % ((IoU).mean() * 100.))
epoch += 1
# Explicit memory cleanup
if hasattr(data_iter, 'cleanup'):
data_iter.cleanup()
# Save the final model
checkpoint(model, optimizer, epoch, curr_iter, config, best_val_miou,
best_val_iter)
test_points(model, val_data_loader, config)
if val_miou > best_val_miou:
best_val_miou = val_miou
best_val_iter = curr_iter
checkpoint(model, optimizer, epoch, curr_iter, config, best_val_miou,
best_val_iter, "best_val")
logging.info("Current best mIoU: {:.3f} at iter {}".format(
best_val_miou, best_val_iter))
def train(model, data_loader, val_data_loader, config, transform_data_fn=None):
all_losses = []
device = get_torch_device(config.is_cuda)
# Set up the train flag for batch normalization
model.train()
# Configuration
writer = SummaryWriter(log_dir=config.log_dir)
data_timer, iter_timer = Timer(), Timer()
data_time_avg, iter_time_avg = AverageMeter(), AverageMeter()
losses, scores, batch_losses = AverageMeter(), AverageMeter(), {}
optimizer = initialize_optimizer(model.parameters(), config)
scheduler = initialize_scheduler(optimizer, config)
criterion = nn.CrossEntropyLoss(ignore_index=config.ignore_label)
alpha, gamma, eps = 1, 2, 1e-6
writer = SummaryWriter(log_dir=config.log_dir)
# Train the network
logging.info('===> Start training')
best_val_miou, best_val_iter, curr_iter, epoch, is_training = 0, 0, 1, 1, True
if config.resume:
checkpoint_fn = config.resume + '/weights.pth'
if osp.isfile(checkpoint_fn):
logging.info("=> loading checkpoint '{}'".format(checkpoint_fn))
state = torch.load(checkpoint_fn)
curr_iter = state['iteration'] + 1
epoch = state['epoch']
model.load_state_dict(state['state_dict'])
if config.resume_optimizer:
scheduler = initialize_scheduler(optimizer,
config,
last_step=curr_iter)
optimizer.load_state_dict(state['optimizer'])
if 'best_val' in state:
best_val_miou = state['best_val']
best_val_iter = state['best_val_iter']
logging.info("=> loaded checkpoint '{}' (epoch {})".format(
checkpoint_fn, state['epoch']))
else:
raise ValueError(
"=> no checkpoint found at '{}'".format(checkpoint_fn))
data_iter = data_loader.__iter__()
while is_training:
print(
"********************************** epoch N° {0} ************************"
.format(epoch))
for iteration in range(len(data_loader) // config.iter_size):
print("####### Iteration N° {0}".format(iteration))
optimizer.zero_grad()
data_time, batch_loss = 0, 0
iter_timer.tic()
for sub_iter in range(config.iter_size):
print("------------- Sub_iteration N° {0}".format(sub_iter))
# Get training data
data_timer.tic()
coords, input, target = data_iter.next()
print("len of coords : {0}".format(len(coords)))
# For some networks, making the network invariant to even, odd coords is important
coords[:, :3] += (torch.rand(3) * 100).type_as(coords)
# Preprocess input
color = input[:, :3].int()
if config.normalize_color:
input[:, :3] = input[:, :3] / 255. - 0.5
sinput = SparseTensor(input, coords).to(device)
data_time += data_timer.toc(False)
# Feed forward
inputs = (sinput, ) if config.wrapper_type == 'None' else (
sinput, coords, color)
# model.initialize_coords(*init_args)
soutput = model(*inputs)
# The output of the network is not sorted
target = target.long().to(device)
print("count of classes : {0}".format(
np.unique(target.cpu().numpy(), return_counts=True)))
print("target : {0}\ntarget_len : {1}".format(
target, len(target)))
print("target [0]: {0}".format(target[0]))
input_soft = nn.functional.softmax(soutput.F, dim=1) + eps
print("input_soft[0] : {0}".format(input_soft[0]))
focal_weight = torch.pow(-input_soft + 1., gamma)
print("focal_weight : {0}\nweight[0] : {1}".format(
focal_weight, focal_weight[0]))
focal_loss = (-alpha * focal_weight *
torch.log(input_soft)).mean()
loss = criterion(soutput.F, target.long())
print("focal_loss :{0}\nloss : {1}".format(focal_loss, loss))
# Compute and accumulate gradient
loss /= config.iter_size
#batch_loss += loss
batch_loss += loss.item()
print("batch_loss : {0}".format(batch_loss))
loss.backward()
# Update number of steps
optimizer.step()
scheduler.step()
data_time_avg.update(data_time)
iter_time_avg.update(iter_timer.toc(False))
pred = get_prediction(data_loader.dataset, soutput.F, target)
score = precision_at_one(pred, target)
losses.update(batch_loss, target.size(0))
scores.update(score, target.size(0))
if curr_iter >= config.max_iter:
is_training = False
break
if curr_iter % config.stat_freq == 0 or curr_iter == 1:
lrs = ', '.join(
['{:.3e}'.format(x) for x in scheduler.get_lr()])
debug_str = "===> Epoch[{}]({}/{}): Loss {:.4f}\tLR: {}\t".format(
epoch, curr_iter,
len(data_loader) // config.iter_size, losses.avg, lrs)
debug_str += "Score {:.3f}\tData time: {:.4f}, Total iter time: {:.4f}".format(
scores.avg, data_time_avg.avg, iter_time_avg.avg)
logging.info(debug_str)
# Reset timers
data_time_avg.reset()
iter_time_avg.reset()
# Write logs
writer.add_scalar('training/loss', losses.avg, curr_iter)
writer.add_scalar('training/precision_at_1', scores.avg,
curr_iter)
writer.add_scalar('training/learning_rate',
scheduler.get_lr()[0], curr_iter)
losses.reset()
scores.reset()
# Save current status, save before val to prevent occational mem overflow
if curr_iter % config.save_freq == 0:
checkpoint(model, optimizer, epoch, curr_iter, config,
best_val_miou, best_val_iter)
# Validation
if curr_iter % config.val_freq == 0:
val_miou, val_losses = validate(model, val_data_loader, writer,
curr_iter, config,
transform_data_fn, epoch)
if val_miou > best_val_miou:
best_val_miou = val_miou
best_val_iter = curr_iter
checkpoint(model, optimizer, epoch, curr_iter, config,
best_val_miou, best_val_iter, "best_val")
logging.info("Current best mIoU: {:.3f} at iter {}".format(
best_val_miou, best_val_iter))
# Recover back
model.train()
if curr_iter % config.empty_cache_freq == 0:
# Clear cache
torch.cuda.empty_cache()
batch_losses[epoch] = batch_loss
# End of iteration
curr_iter += 1
with open(config.log_dir + "/train_loss.txt", 'a') as train_loss_log:
train_loss_log.writelines('{0}, {1}\n'.format(
batch_losses[epoch], epoch))
train_loss_log.close()
epoch += 1
# Explicit memory cleanup
if hasattr(data_iter, 'cleanup'):
data_iter.cleanup()
# Save the final model
checkpoint(model, optimizer, epoch, curr_iter, config, best_val_miou,
best_val_iter)
val_miou = validate(model, val_data_loader, writer, curr_iter, config,
transform_data_fn, epoch)[0]
if val_miou > best_val_miou:
best_val_miou = val_miou
best_val_iter = curr_iter
checkpoint(model, optimizer, epoch, curr_iter, config, best_val_miou,
best_val_iter, "best_val")
logging.info("Current best mIoU: {:.3f} at iter {}".format(
best_val_miou, best_val_iter))
def train_worker(gpu,
num_devices,
NetClass,
data_loader,
val_data_loader,
config,
transform_data_fn=None):
if gpu is not None:
print("Use GPU: {} for training".format(gpu))
rank = gpu
addr = 23491
dist.init_process_group(backend="nccl",
init_method="tcp://127.0.0.1:{}".format(addr),
world_size=num_devices,
rank=rank)
# replace with DistributedSampler
if config.multiprocess:
from lib.dataloader_dist import InfSampler
sampler = InfSampler(data_loader.dataset)
data_loader = DataLoader(dataset=data_loader.dataset,
num_workers=data_loader.num_workers,
batch_size=data_loader.batch_size,
collate_fn=data_loader.collate_fn,
worker_init_fn=data_loader.worker_init_fn,
sampler=sampler)
if data_loader.dataset.NUM_IN_CHANNEL is not None:
num_in_channel = data_loader.dataset.NUM_IN_CHANNEL
else:
num_in_channel = 3
num_labels = data_loader.dataset.NUM_LABELS
# load model
if config.pure_point:
model = NetClass(num_class=config.num_labels,
N=config.num_points,
normal_channel=config.num_in_channel)
else:
if config.model == 'MixedTransformer':
model = NetClass(config,
num_class=num_labels,
N=config.num_points,
normal_channel=num_in_channel)
elif config.model == 'MinkowskiVoxelTransformer':
model = NetClass(config, num_in_channel, num_labels)
elif config.model == 'MinkowskiTransformerNet':
model = NetClass(config, num_in_channel, num_labels)
elif "Res" in config.model:
model = NetClass(num_in_channel, num_labels, config)
else:
model = NetClass(num_in_channel, num_labels, config)
if config.weights == 'modelzoo':
model.preload_modelzoo()
elif config.weights.lower() != 'none':
state = torch.load(config.weights)
# delete the keys containing the attn since it raises size mismatch
d = {k: v for k, v in state['state' '_dict'].items() if 'map' not in k}
if config.weights_for_inner_model:
model.model.load_state_dict(d)
else:
if config.lenient_weight_loading:
matched_weights = load_state_with_same_shape(
model, state['state_dict'])
model_dict = model.state_dict()
model_dict.update(matched_weights)
model.load_state_dict(model_dict)
else:
model.load_state_dict(d, strict=False)
torch.cuda.set_device(gpu)
model.cuda(gpu)
# use model with DDP
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[gpu], find_unused_parameters=False)
# Synchronized batch norm
model = ME.MinkowskiSyncBatchNorm.convert_sync_batchnorm(model)
# Set up the train flag for batch normalization
model.train()
# Configuration
data_timer, iter_timer = Timer(), Timer()
data_time_avg, iter_time_avg = AverageMeter(), AverageMeter()
regs, losses, scores = AverageMeter(), AverageMeter(), AverageMeter()
optimizer = initialize_optimizer(model.parameters(), config)
scheduler = initialize_scheduler(optimizer, config)
criterion = nn.CrossEntropyLoss(ignore_index=config.ignore_label)
# Train the network
if rank == 0:
setup_logger(config)
logging.info('===> Start training')
best_val_miou, best_val_iter, curr_iter, epoch, is_training = 0, 0, 1, 1, True
if config.resume:
# Test loaded ckpt first
v_loss, v_score, v_mAP, v_mIoU = test(model, val_data_loader, config)
checkpoint_fn = config.resume + '/weights.pth'
if osp.isfile(checkpoint_fn):
logging.info("=> loading checkpoint '{}'".format(checkpoint_fn))
state = torch.load(checkpoint_fn)
curr_iter = state['iteration'] + 1
epoch = state['epoch']
# we skip attention maps because the shape won't match because voxel number is different
# e.g. copyting a param with shape (23385, 8, 4) to (43529, 8, 4)
d = {
k: v
for k, v in state['state_dict'].items() if 'map' not in k
}
# handle those attn maps we don't load from saved dict
for k in model.state_dict().keys():
if k in d.keys(): continue
d[k] = model.state_dict()[k]
model.load_state_dict(d)
if config.resume_optimizer:
scheduler = initialize_scheduler(optimizer,
config,
last_step=curr_iter)
optimizer.load_state_dict(state['optimizer'])
if 'best_val' in state:
best_val_miou = state['best_val']
best_val_iter = state['best_val_iter']
logging.info("=> loaded checkpoint '{}' (epoch {})".format(
checkpoint_fn, state['epoch']))
else:
raise ValueError(
"=> no checkpoint found at '{}'".format(checkpoint_fn))
data_iter = data_loader.__iter__()
device = gpu # multitrain fed in the device
if config.dataset == "SemanticKITTI":
num_class = 19
config.normalize_color = False
config.xyz_input = False
val_freq_ = config.val_freq
config.val_freq = config.val_freq * 10 # origianl val_freq_
elif config.dataset == 'S3DIS':
num_class = 13
config.normalize_color = False
config.xyz_input = False
val_freq_ = config.val_freq
elif config.dataset == "Nuscenes":
num_class = 16
config.normalize_color = False
config.xyz_input = False
val_freq_ = config.val_freq
config.val_freq = config.val_freq * 50
else:
val_freq_ = config.val_freq
num_class = 20
while is_training:
total_correct_class = torch.zeros(num_class, device=device)
total_iou_deno_class = torch.zeros(num_class, device=device)
for iteration in range(len(data_loader) // config.iter_size):
optimizer.zero_grad()
data_time, batch_loss = 0, 0
iter_timer.tic()
if curr_iter >= config.max_iter:
# if curr_iter >= max(config.max_iter, config.epochs*(len(data_loader) // config.iter_size):
is_training = False
break
elif curr_iter >= config.max_iter * (2 / 3):
config.val_freq = val_freq_ * 2 # valid more freq on lower half
for sub_iter in range(config.iter_size):
# Get training data
data_timer.tic()
if config.return_transformation:
coords, input, target, _, _, pointcloud, transformation = data_iter.next(
)
else:
coords, input, target, _, _ = data_iter.next(
) # ignore unique_map and inverse_map
if config.use_aux:
assert target.shape[1] == 2
aux = target[:, 1]
target = target[:, 0]
else:
aux = None
# For some networks, making the network invariant to even, odd coords is important
coords[:, 1:] += (torch.rand(3) * 100).type_as(coords)
# Preprocess input
if config.normalize_color:
input[:, :3] = input[:, :3] / input[:, :3].max() - 0.5
coords_norm = coords[:, 1:] / coords[:, 1:].max() - 0.5
# cat xyz into the rgb feature
if config.xyz_input:
input = torch.cat([coords_norm, input], dim=1)
# print(device)
sinput = SparseTensor(input, coords, device=device)
# d = {}
# d['coord'] = sinput.C
# d['feat'] = sinput.F
# torch.save(d, 'voxel.pth')
# import ipdb; ipdb.set_trace()
data_time += data_timer.toc(False)
# model.initialize_coords(*init_args)
if aux is not None:
soutput = model(sinput, aux)
elif config.enable_point_branch:
soutput = model(sinput,
iter_=curr_iter / config.max_iter,
enable_point_branch=True)
else:
soutput = model(
sinput, iter_=curr_iter / config.max_iter
) # feed in the progress of training for annealing inside the model
# soutput = model(sinput)
# The output of the network is not sorted
target = target.view(-1).long().to(device)
loss = criterion(soutput.F, target.long())
# ====== other loss regs =====
cur_loss = torch.tensor([0.], device=device)
if hasattr(model, 'module.block1'):
cur_loss = torch.tensor([0.], device=device)
if hasattr(model.module.block1[0], 'vq_loss'):
if model.block1[0].vq_loss is not None:
cur_loss = torch.tensor([0.], device=device)
for n, m in model.named_children():
if 'block' in n:
cur_loss += m[
0].vq_loss # m is the nn.Sequential obj, m[0] is the TRBlock
logging.info(
'Cur Loss: {}, Cur vq_loss: {}'.format(
loss, cur_loss))
loss += cur_loss
if hasattr(model.module.block1[0], 'diverse_loss'):
if model.block1[0].diverse_loss is not None:
cur_loss = torch.tensor([0.], device=device)
for n, m in model.named_children():
if 'block' in n:
cur_loss += m[
0].diverse_loss # m is the nn.Sequential obj, m[0] is the TRBlock
logging.info(
'Cur Loss: {}, Cur diverse _loss: {}'.format(
loss, cur_loss))
loss += cur_loss
if hasattr(model.module.block1[0], 'label_reg'):
if model.block1[0].label_reg is not None:
cur_loss = torch.tensor([0.], device=device)
for n, m in model.named_children():
if 'block' in n:
cur_loss += m[
0].label_reg # m is the nn.Sequential obj, m[0] is the TRBlock
# logging.info('Cur Loss: {}, Cur diverse _loss: {}'.format(loss, cur_loss))
loss += cur_loss
# Compute and accumulate gradient
loss /= config.iter_size
batch_loss += loss.item()
if not config.use_sam:
loss.backward()
else:
with model.no_sync():
loss.backward()
# Update number of steps
if not config.use_sam:
optimizer.step()
else:
optimizer.first_step(zero_grad=True)
soutput = model(sinput,
iter_=curr_iter / config.max_iter,
aux=starget)
criterion(soutput.F, target.long()).backward()
optimizer.second_step(zero_grad=True)
if config.lr_warmup is None:
scheduler.step()
else:
if curr_iter >= config.lr_warmup:
scheduler.step()
else:
for g in optimizer.param_groups:
g['lr'] = config.lr * (iteration +
1) / config.lr_warmup
# CLEAR CACHE!
torch.cuda.empty_cache()
data_time_avg.update(data_time)
iter_time_avg.update(iter_timer.toc(False))
pred = get_prediction(data_loader.dataset, soutput.F, target)
score = precision_at_one(pred, target, ignore_label=-1)
regs.update(cur_loss.item(), target.size(0))
losses.update(batch_loss, target.size(0))
scores.update(score, target.size(0))
# calc the train-iou
for l in range(num_class):
total_correct_class[l] += ((pred == l) & (target == l)).sum()
total_iou_deno_class[l] += (((pred == l) & (target != -1)) |
(target == l)).sum()
if curr_iter % config.stat_freq == 0 or curr_iter == 1:
lrs = ', '.join(
['{:.3e}'.format(g['lr']) for g in optimizer.param_groups])
IoU = ((total_correct_class) /
(total_iou_deno_class + 1e-6)).mean() * 100.
debug_str = "===> Epoch[{}]({}/{}): Loss {:.4f}\tLR: {}\t".format(
epoch, curr_iter,
len(data_loader) // config.iter_size, losses.avg, lrs)
debug_str += "Score {:.3f}\tIoU {:.3f}\tData time: {:.4f}, Iter time: {:.4f}".format(
scores.avg, IoU.item(), data_time_avg.avg,
iter_time_avg.avg)
if regs.avg > 0:
debug_str += "\n Additional Reg Loss {:.3f}".format(
regs.avg)
if rank == 0:
logging.info(debug_str)
# Reset timers
data_time_avg.reset()
iter_time_avg.reset()
# Write logs
losses.reset()
scores.reset()
# only save status on the 1st gpu
if rank == 0:
# Save current status, save before val to prevent occational mem overflow
if curr_iter % config.save_freq == 0:
checkpoint(model,
optimizer,
epoch,
curr_iter,
config,
best_val_miou,
best_val_iter,
save_inter=True)
# Validation
if curr_iter % config.val_freq == 0:
val_miou = validate(model, val_data_loader, None,
curr_iter, config, transform_data_fn
) # feedin None for SummaryWriter args
if val_miou > best_val_miou:
best_val_miou = val_miou
best_val_iter = curr_iter
checkpoint(model,
optimizer,
epoch,
curr_iter,
config,
best_val_miou,
best_val_iter,
"best_val",
save_inter=True)
if rank == 0:
logging.info(
"Current best mIoU: {:.3f} at iter {}".format(
best_val_miou, best_val_iter))
# Recover back
model.train()
# End of iteration
curr_iter += 1
IoU = (total_correct_class) / (total_iou_deno_class + 1e-6)
if rank == 0:
logging.info('train point avg class IoU: %f' %
((IoU).mean() * 100.))
epoch += 1
# Explicit memory cleanup
if hasattr(data_iter, 'cleanup'):
data_iter.cleanup()
# Save the final model
if rank == 0:
checkpoint(model, optimizer, epoch, curr_iter, config, best_val_miou,
best_val_iter)
v_loss, v_score, v_mAP, val_mIoU = test(model, val_data_loader, config)
if val_miou > best_val_miou and rank == 0:
best_val_miou = val_miou
best_val_iter = curr_iter
logging.info("Final best miou: {} at iter {} ".format(
val_miou, curr_iter))
checkpoint(model, optimizer, epoch, curr_iter, config,
best_val_miou, best_val_iter, "best_val")
logging.info("Current best mIoU: {:.3f} at iter {}".format(
best_val_miou, best_val_iter))
def test(model, data_loader, config, transform_data_fn=None, has_gt=True, save_pred=False, split=None, submit_dir=None):
device = get_torch_device(config.is_cuda)
dataset = data_loader.dataset
num_labels = dataset.NUM_LABELS
global_timer, data_timer, iter_timer = Timer(), Timer(), Timer()
criterion = nn.CrossEntropyLoss(ignore_index=config.ignore_label)
losses, scores, ious = AverageMeter(), AverageMeter(), 0
aps = np.zeros((0, num_labels))
hist = np.zeros((num_labels, num_labels))
# some cfgs concerning the usage of instance-level information
config.save_pred = save_pred
if split is not None:
assert save_pred
if config.save_pred:
save_dict = {}
save_dict['pred'] = []
save_dict['coord'] = []
logging.info('===> Start testing')
global_timer.tic()
data_iter = data_loader.__iter__()
max_iter = len(data_loader)
max_iter_unique = max_iter
# Fix batch normalization running mean and std
model.eval()
# Clear cache (when run in val mode, cleanup training cache)
torch.cuda.empty_cache()
# semantic kitti label inverse mapping
if config.submit:
remap_lut = Remap().getRemapLUT()
with torch.no_grad():
# Calc of the iou
total_correct = np.zeros(num_labels)
total_seen = np.zeros(num_labels)
total_positive = np.zeros(num_labels)
point_nums = np.zeros([19])
for iteration in range(max_iter):
data_timer.tic()
if config.return_transformation:
coords, input, target, unique_map_list, inverse_map_list, pointcloud, transformation, filename = data_iter.next()
else:
coords, input, target, unique_map_list, inverse_map_list, filename = data_iter.next()
data_time = data_timer.toc(False)
if config.use_aux:
assert target.shape[1] == 2
aux = target[:,1]
target = target[:,0]
else:
aux = None
# Preprocess input
iter_timer.tic()
if config.normalize_color:
input[:, :3] = input[:, :3] / input[:,:3].max() - 0.5
coords_norm = coords[:,1:] / coords[:,1:].max() - 0.5
XYZ_INPUT = config.xyz_input
# cat xyz into the rgb feature
if XYZ_INPUT:
input = torch.cat([coords_norm, input], dim=1)
sinput = ME.SparseTensor(input, coords, device=device)
# Feed forward
if aux is not None:
soutput = model(sinput)
else:
soutput = model(sinput, iter_ = iteration / max_iter, enable_point_branch=config.enable_point_branch)
output = soutput.F
if torch.isnan(output).sum() > 0:
import ipdb; ipdb.set_trace()
pred = get_prediction(dataset, output, target).int()
assert sum([int(t.shape[0]) for t in unique_map_list]) == len(pred), "number of points in unique_map doesn't match predition, do not enable preprocessing"
iter_time = iter_timer.toc(False)
if config.save_pred or config.submit:
# troublesome processing for splitting each batch's data, and export
batch_ids = sinput.C[:,0]
splits_at = torch.stack([torch.where(batch_ids == i)[0][-1] for i in torch.unique(batch_ids)]).int()
splits_at = splits_at + 1
splits_at_leftshift_one = splits_at.roll(shifts=1)
splits_at_leftshift_one[0] = 0
# len_per_batch = splits_at - splits_at_leftshift_one
len_sum = 0
batch_id = 0
for start, end in zip(splits_at_leftshift_one, splits_at):
len_sum += len(pred[int(start):int(end)])
pred_this_batch = pred[int(start):int(end)]
coord_this_batch = pred[int(start):int(end)]
if config.save_pred:
save_dict['pred'].append(pred_this_batch[inverse_map_list[batch_id]])
else: # save submit result
submission_path = filename[batch_id].replace(config.semantic_kitti_path, submit_dir).replace('velodyne', 'predictions').replace('.bin', '.label')
parent_dir = Path(submission_path).parent.absolute()
if not os.path.exists(parent_dir):
os.makedirs(parent_dir)
label_pred = pred_this_batch[inverse_map_list[batch_id]].cpu().numpy()
label_pred = remap_lut[label_pred].astype(np.uint32)
label_pred.tofile(submission_path)
print(submission_path)
batch_id += 1
assert len_sum == len(pred)
# Unpack it to original length
REVERT_WHOLE_POINTCLOUD = True
# print('{}/{}'.format(iteration, max_iter))
if REVERT_WHOLE_POINTCLOUD:
whole_pred = []
whole_target = []
for batch_ in range(config.batch_size):
batch_mask_ = (soutput.C[:,0] == batch_).cpu().numpy()
if batch_mask_.sum() == 0: # for empty batch, skip em
continue
try:
whole_pred_ = soutput.F[batch_mask_][inverse_map_list[batch_]]
except:
import ipdb; ipdb.set_trace()
whole_target_ = target[batch_mask_][inverse_map_list[batch_]]
whole_pred.append(whole_pred_)
whole_target.append(whole_target_)
whole_pred = torch.cat(whole_pred, dim=0)
whole_target = torch.cat(whole_target, dim=0)
pred = get_prediction(dataset, whole_pred, whole_target).int()
output = whole_pred
target = whole_target
if has_gt:
target_np = target.numpy()
num_sample = target_np.shape[0]
target = target.to(device)
output = output.to(device)
cross_ent = criterion(output, target.long())
losses.update(float(cross_ent), num_sample)
scores.update(precision_at_one(pred, target), num_sample)
hist += fast_hist(pred.cpu().numpy().flatten(), target_np.flatten(), num_labels) # within fast hist, mark label should >=0 & < num_label to filter out 255 / -1
ious = per_class_iu(hist) * 100
prob = torch.nn.functional.softmax(output, dim=-1)
pred = pred[target != -1]
target = target[target != -1]
# for _ in range(num_labels): # debug for SemKITTI: spvnas way of calc miou
# total_seen[_] += torch.sum(target == _)
# total_correct[_] += torch.sum((pred == target) & (target == _))
# total_positive[_] += torch.sum(pred == _)
# ious_ = []
# for _ in range(num_labels):
# if total_seen[_] == 0:
# ious_.append(1)
# else:
# ious_.append(total_correct[_]/(total_seen[_] + total_positive[_] - total_correct[_]))
# ious_ = torch.stack(ious_, dim=-1).cpu().numpy()*100
# print(np.nanmean(per_class_iu(hist)), np.nanmean(ious_))
# ious = np.array(ious_)*100
# calc the ratio of total points
# for i_ in range(19):
# point_nums[i_] += (target == i_).sum().detach()
# skip calculating aps
ap = average_precision(prob.cpu().detach().numpy(), target_np)
aps = np.vstack((aps, ap))
# Due to heavy bias in class, there exists class with no test label at all
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=RuntimeWarning)
ap_class = np.nanmean(aps, 0) * 100.
if iteration % config.test_stat_freq == 0 and iteration > 0 and not config.submit:
reordered_ious = dataset.reorder_result(ious)
reordered_ap_class = dataset.reorder_result(ap_class)
# dirty fix for semnaticcKITTI has no getclassnames
if hasattr(dataset, "class_names"):
class_names = dataset.get_classnames()
else: # semnantic KITTI
class_names = None
print_info(
iteration,
max_iter_unique,
data_time,
iter_time,
has_gt,
losses,
scores,
reordered_ious,
hist,
reordered_ap_class,
class_names=class_names)
if iteration % 5 == 0:
# Clear cache
torch.cuda.empty_cache()
if config.save_pred:
# torch.save(save_dict, os.path.join(config.log_dir, 'preds_{}_with_coord.pth'.format(split)))
torch.save(save_dict, os.path.join(config.log_dir, 'preds_{}.pth'.format(split)))
print("===> saved prediction result")
global_time = global_timer.toc(False)
save_map(model, config)
reordered_ious = dataset.reorder_result(ious)
reordered_ap_class = dataset.reorder_result(ap_class)
if hasattr(dataset, "class_names"):
class_names = dataset.get_classnames()
else:
class_names = None
print_info(
iteration,
max_iter_unique,
data_time,
iter_time,
has_gt,
losses,
scores,
reordered_ious,
hist,
reordered_ap_class,
class_names=class_names)
logging.info("Finished test. Elapsed time: {:.4f}".format(global_time))
# Explicit memory cleanup
if hasattr(data_iter, 'cleanup'):
data_iter.cleanup()
return losses.avg, scores.avg, np.nanmean(ap_class), np.nanmean(per_class_iu(hist)) * 100
def train(model, data_loader, val_data_loader, config, transform_data_fn=None):
device = get_torch_device(config.is_cuda)
# Set up the train flag for batch normalization
model.train()
# Configuration
writer = SummaryWriter(log_dir=config.log_dir)
data_timer, iter_timer = Timer(), Timer()
data_time_avg, iter_time_avg = AverageMeter(), AverageMeter()
losses, scores = AverageMeter(), AverageMeter()
optimizer = initialize_optimizer(model.parameters(), config)
scheduler = initialize_scheduler(optimizer, config)
criterion = nn.CrossEntropyLoss(ignore_index=config.ignore_label)
writer = SummaryWriter(log_dir=config.log_dir)
# Train the network
logging.info('===> Start training')
best_val_miou, best_val_iter, curr_iter, epoch, is_training = 0, 0, 1, 1, True
if config.resume:
checkpoint_fn = config.resume + '/weights.pth'
if osp.isfile(checkpoint_fn):
logging.info("=> loading checkpoint '{}'".format(checkpoint_fn))
state = torch.load(checkpoint_fn)
curr_iter = state['iteration'] + 1
epoch = state['epoch']
model.load_state_dict(state['state_dict'])
if config.resume_optimizer:
scheduler = initialize_scheduler(optimizer,
config,
last_step=curr_iter)
optimizer.load_state_dict(state['optimizer'])
if 'best_val' in state:
best_val_miou = state['best_val']
best_val_iter = state['best_val_iter']
logging.info("=> loaded checkpoint '{}' (epoch {})".format(
checkpoint_fn, state['epoch']))
else:
raise ValueError(
"=> no checkpoint found at '{}'".format(checkpoint_fn))
data_iter = data_loader.__iter__()
while is_training:
for iteration in range(len(data_loader) // config.iter_size):
optimizer.zero_grad()
data_time, batch_loss = 0, 0
iter_timer.tic()
for sub_iter in range(config.iter_size):
# Get training data
data_timer.tic()
if config.return_transformation:
coords, input, target, pointcloud, transformation = data_iter.next(
)
else:
coords, input, target = data_iter.next()
# For some networks, making the network invariant to even, odd coords is important
coords[:, 1:] += (torch.rand(3) * 100).type_as(coords)
# Preprocess input
if config.normalize_color:
input[:, :3] = input[:, :3] / 255. - 0.5
sinput = SparseTensor(input, coords).to(device)
data_time += data_timer.toc(False)
# model.initialize_coords(*init_args)
soutput = model(sinput)
# The output of the network is not sorted
target = target.long().to(device)
loss = criterion(soutput.F, target.long())
# Compute and accumulate gradient
loss /= config.iter_size
batch_loss += loss.item()
loss.backward()
# Update number of steps
optimizer.step()
scheduler.step()
data_time_avg.update(data_time)
iter_time_avg.update(iter_timer.toc(False))
pred = get_prediction(data_loader.dataset, soutput.F, target)
score = precision_at_one(pred, target)
losses.update(batch_loss, target.size(0))
scores.update(score, target.size(0))
if curr_iter >= config.max_iter:
is_training = False
break
if curr_iter % config.stat_freq == 0 or curr_iter == 1:
lrs = ', '.join(
['{:.3e}'.format(x) for x in scheduler.get_lr()])
debug_str = "===> Epoch[{}]({}/{}): Loss {:.4f}\tLR: {}\t".format(
epoch, curr_iter,
len(data_loader) // config.iter_size, losses.avg, lrs)
debug_str += "Score {:.3f}\tData time: {:.4f}, Iter time: {:.4f}".format(
scores.avg, data_time_avg.avg, iter_time_avg.avg)
logging.info(debug_str)
# Reset timers
data_time_avg.reset()
iter_time_avg.reset()
# Write logs
writer.add_scalar('training/loss', losses.avg, curr_iter)
writer.add_scalar('training/precision_at_1', scores.avg,
curr_iter)
writer.add_scalar('training/learning_rate',
scheduler.get_lr()[0], curr_iter)
losses.reset()
scores.reset()
# Save current status, save before val to prevent occational mem overflow
if curr_iter % config.save_freq == 0:
checkpoint(model, optimizer, epoch, curr_iter, config,
best_val_miou, best_val_iter)
# Validation
if curr_iter % config.val_freq == 0:
val_miou = validate(model, val_data_loader, writer, curr_iter,
config, transform_data_fn)
if val_miou > best_val_miou:
best_val_miou = val_miou
best_val_iter = curr_iter
checkpoint(model, optimizer, epoch, curr_iter, config,
best_val_miou, best_val_iter, "best_val")
logging.info("Current best mIoU: {:.3f} at iter {}".format(
best_val_miou, best_val_iter))
# Recover back
model.train()
# End of iteration
curr_iter += 1
epoch += 1
# Explicit memory cleanup
if hasattr(data_iter, 'cleanup'):
data_iter.cleanup()
# Save the final model
checkpoint(model, optimizer, epoch, curr_iter, config, best_val_miou,
best_val_iter)
val_miou = validate(model, val_data_loader, writer, curr_iter, config,
transform_data_fn)
if val_miou > best_val_miou:
best_val_miou = val_miou
best_val_iter = curr_iter
checkpoint(model, optimizer, epoch, curr_iter, config, best_val_miou,
best_val_iter, "best_val")
logging.info("Current best mIoU: {:.3f} at iter {}".format(
best_val_miou, best_val_iter))
