def update_learning_rate(self, step):
# Warm up
if step < self.cfg.SOLVER.WARM_UP_ITERS:
method = self.cfg.SOLVER.WARM_UP_METHOD
if method == 'constant':
warmup_factor = self.cfg.SOLVER.WARM_UP_FACTOR
elif method == 'linear':
alpha = step / self.cfg.SOLVER.WARM_UP_ITERS
warmup_factor = self.cfg.SOLVER.WARM_UP_FACTOR * (1 - alpha) + alpha
else:
raise KeyError('Unknown SOLVER.WARM_UP_METHOD: {}'.format(method))
lr_new = self.cfg.SOLVER.BASE_LR * warmup_factor
net_utils.update_learning_rate(self.optimizer, self.lr, lr_new)
self.lr = self.optimizer.param_groups[0]['lr']
assert self.lr == lr_new
elif step == self.cfg.SOLVER.WARM_UP_ITERS:
net_utils.update_learning_rate(self.optimizer, self.lr, self.cfg.SOLVER.BASE_LR)
self.lr = self.optimizer.param_groups[0]['lr']
assert self.lr == self.cfg.SOLVER.BASE_LR
# Learning rate decay
if self.decay_steps_ind < len(self.cfg.SOLVER.STEPS) and step == self.cfg.SOLVER.STEPS[self.decay_steps_ind]:
logger.info('Decay the learning on step %d', step)
lr_new = self.lr * self.cfg.SOLVER.GAMMA
net_utils.update_learning_rate(self.optimizer, self.lr, lr_new)
self.lr = self.optimizer.param_groups[0]['lr']
assert self.lr == lr_new
self.decay_steps_ind += 1
def main():
"""Main function"""
args = parse_args()
print('Called with args:')
print(args)
if not torch.cuda.is_available():
sys.exit("Need a CUDA device to run the code.")
if args.cuda or cfg.NUM_GPUS > 0:
cfg.CUDA = True
else:
raise ValueError("Need Cuda device to run !")
if args.dataset == "coco2017":
cfg.TRAIN.DATASETS = ('coco_2017_train',)
cfg.MODEL.NUM_CLASSES = 81
elif args.dataset == "keypoints_coco2017":
cfg.TRAIN.DATASETS = ('keypoints_coco_2017_train',)
cfg.MODEL.NUM_CLASSES = 2
else:
raise ValueError("Unexpected args.dataset: {}".format(args.dataset))
cfg_from_file(args.cfg_file)
if args.set_cfgs is not None:
cfg_from_list(args.set_cfgs)
### Adaptively adjust some configs ###
original_batch_size = cfg.NUM_GPUS * cfg.TRAIN.IMS_PER_BATCH
original_ims_per_batch = cfg.TRAIN.IMS_PER_BATCH
original_num_gpus = cfg.NUM_GPUS
if args.batch_size is None:
args.batch_size = original_batch_size
cfg.NUM_GPUS = torch.cuda.device_count()
assert (args.batch_size % cfg.NUM_GPUS) == 0, \
'batch_size: %d, NUM_GPUS: %d' % (args.batch_size, cfg.NUM_GPUS)
cfg.TRAIN.IMS_PER_BATCH = args.batch_size // cfg.NUM_GPUS
effective_batch_size = args.iter_size * args.batch_size
print('effective_batch_size = batch_size * iter_size = %d * %d' % (args.batch_size, args.iter_size))
print('Adaptive config changes:')
print(' effective_batch_size: %d --> %d' % (original_batch_size, effective_batch_size))
print(' NUM_GPUS: %d --> %d' % (original_num_gpus, cfg.NUM_GPUS))
print(' IMS_PER_BATCH: %d --> %d' % (original_ims_per_batch, cfg.TRAIN.IMS_PER_BATCH))
### Adjust learning based on batch size change linearly
# For iter_size > 1, gradients are `accumulated`, so lr is scaled based
# on batch_size instead of effective_batch_size
old_base_lr = cfg.SOLVER.BASE_LR
cfg.SOLVER.BASE_LR *= args.batch_size / original_batch_size
print('Adjust BASE_LR linearly according to batch_size change:\n'
' BASE_LR: {} --> {}'.format(old_base_lr, cfg.SOLVER.BASE_LR))
### Adjust solver steps
step_scale = original_batch_size / effective_batch_size
old_solver_steps = cfg.SOLVER.STEPS
old_max_iter = cfg.SOLVER.MAX_ITER
cfg.SOLVER.STEPS = list(map(lambda x: int(x * step_scale + 0.5), cfg.SOLVER.STEPS))
cfg.SOLVER.MAX_ITER = int(cfg.SOLVER.MAX_ITER * step_scale + 0.5)
print('Adjust SOLVER.STEPS and SOLVER.MAX_ITER linearly based on effective_batch_size change:\n'
' SOLVER.STEPS: {} --> {}\n'
' SOLVER.MAX_ITER: {} --> {}'.format(old_solver_steps, cfg.SOLVER.STEPS,
old_max_iter, cfg.SOLVER.MAX_ITER))
# Scale FPN rpn_proposals collect size (post_nms_topN) in `collect` function
# of `collect_and_distribute_fpn_rpn_proposals.py`
#
# post_nms_topN = int(cfg[cfg_key].RPN_POST_NMS_TOP_N * cfg.FPN.RPN_COLLECT_SCALE + 0.5)
if cfg.FPN.FPN_ON and cfg.MODEL.FASTER_RCNN:
cfg.FPN.RPN_COLLECT_SCALE = cfg.TRAIN.IMS_PER_BATCH / original_ims_per_batch
print('Scale FPN rpn_proposals collect size directly propotional to the change of IMS_PER_BATCH:\n'
' cfg.FPN.RPN_COLLECT_SCALE: {}'.format(cfg.FPN.RPN_COLLECT_SCALE))
if args.num_workers is not None:
cfg.DATA_LOADER.NUM_THREADS = args.num_workers
print('Number of data loading threads: %d' % cfg.DATA_LOADER.NUM_THREADS)
### Overwrite some solver settings from command line arguments
if args.optimizer is not None:
cfg.SOLVER.TYPE = args.optimizer
if args.lr is not None:
cfg.SOLVER.BASE_LR = args.lr
if args.lr_decay_gamma is not None:
cfg.SOLVER.GAMMA = args.lr_decay_gamma
assert_and_infer_cfg()
timers = defaultdict(Timer)
### Dataset ###
timers['roidb'].tic()
roidb, ratio_list, ratio_index = combined_roidb_for_training(
cfg.TRAIN.DATASETS, cfg.TRAIN.PROPOSAL_FILES)
timers['roidb'].toc()
roidb_size = len(roidb)
logger.info('{:d} roidb entries'.format(roidb_size))
logger.info('Takes %.2f sec(s) to construct roidb', timers['roidb'].average_time)
# Effective training sample size for one epoch
train_size = roidb_size // args.batch_size * args.batch_size
batchSampler = BatchSampler(
sampler=MinibatchSampler(ratio_list, ratio_index),
batch_size=args.batch_size,
drop_last=True
)
dataset = RoiDataLoader(
roidb,
cfg.MODEL.NUM_CLASSES,
training=True)
dataloader = torch.utils.data.DataLoader(
dataset,
batch_sampler=batchSampler,
num_workers=cfg.DATA_LOADER.NUM_THREADS,
collate_fn=collate_minibatch)
dataiterator = iter(dataloader)
### Model ###
maskRCNN = Generalized_RCNN()
if cfg.CUDA:
maskRCNN.cuda()
### Optimizer ###
gn_param_nameset = set()
for name, module in maskRCNN.named_modules():
if isinstance(module, nn.GroupNorm):
gn_param_nameset.add(name+'.weight')
gn_param_nameset.add(name+'.bias')
gn_params = []
gn_param_names = []
bias_params = []
bias_param_names = []
nonbias_params = []
nonbias_param_names = []
nograd_param_names = []
for key, value in maskRCNN.named_parameters():
if value.requires_grad:
if 'bias' in key:
bias_params.append(value)
bias_param_names.append(key)
elif key in gn_param_nameset:
gn_params.append(value)
gn_param_names.append(key)
else:
nonbias_params.append(value)
nonbias_param_names.append(key)
else:
nograd_param_names.append(key)
assert (gn_param_nameset - set(nograd_param_names) - set(bias_param_names)) == set(gn_param_names)
# Learning rate of 0 is a dummy value to be set properly at the start of training
params = [
{'params': nonbias_params,
'lr': 0,
'weight_decay': cfg.SOLVER.WEIGHT_DECAY},
{'params': bias_params,
'lr': 0 * (cfg.SOLVER.BIAS_DOUBLE_LR + 1),
'weight_decay': cfg.SOLVER.WEIGHT_DECAY if cfg.SOLVER.BIAS_WEIGHT_DECAY else 0},
{'params': gn_params,
'lr': 0,
'weight_decay': cfg.SOLVER.WEIGHT_DECAY_GN}
]
# names of paramerters for each paramter
param_names = [nonbias_param_names, bias_param_names, gn_param_names]
if cfg.SOLVER.TYPE == "SGD":
optimizer = torch.optim.SGD(params, momentum=cfg.SOLVER.MOMENTUM)
elif cfg.SOLVER.TYPE == "Adam":
optimizer = torch.optim.Adam(params)
### Load checkpoint
if args.load_ckpt:
load_name = args.load_ckpt
logging.info("loading checkpoint %s", load_name)
checkpoint = torch.load(load_name, map_location=lambda storage, loc: storage)
net_utils.load_ckpt(maskRCNN, checkpoint['model'])
if args.resume:
args.start_step = checkpoint['step'] + 1
if 'train_size' in checkpoint: # For backward compatibility
if checkpoint['train_size'] != train_size:
print('train_size value: %d different from the one in checkpoint: %d'
% (train_size, checkpoint['train_size']))
# reorder the params in optimizer checkpoint's params_groups if needed
# misc_utils.ensure_optimizer_ckpt_params_order(param_names, checkpoint)
# There is a bug in optimizer.load_state_dict on Pytorch 0.3.1.
# However it's fixed on master.
optimizer.load_state_dict(checkpoint['optimizer'])
# misc_utils.load_optimizer_state_dict(optimizer, checkpoint['optimizer'])
del checkpoint
torch.cuda.empty_cache()
if args.load_detectron: #TODO resume for detectron weights (load sgd momentum values)
logging.info("loading Detectron weights %s", args.load_detectron)
load_detectron_weight(maskRCNN, args.load_detectron)
lr = optimizer.param_groups[0]['lr'] # lr of non-bias parameters, for commmand line outputs.
maskRCNN = mynn.DataParallel(maskRCNN, cpu_keywords=['im_info', 'roidb'],
minibatch=True)
### Training Setups ###
args.run_name = misc_utils.get_run_name() + '_step'
output_dir = misc_utils.get_output_dir(args, args.run_name)
args.cfg_filename = os.path.basename(args.cfg_file)
if not args.no_save:
if not os.path.exists(output_dir):
os.makedirs(output_dir)
blob = {'cfg': yaml.dump(cfg), 'args': args}
with open(os.path.join(output_dir, 'config_and_args.pkl'), 'wb') as f:
pickle.dump(blob, f, pickle.HIGHEST_PROTOCOL)
if args.use_tfboard:
from tensorboardX import SummaryWriter
# Set the Tensorboard logger
tblogger = SummaryWriter(output_dir)
### Training Loop ###
maskRCNN.train()
CHECKPOINT_PERIOD = int(cfg.TRAIN.SNAPSHOT_ITERS / cfg.NUM_GPUS)
# Set index for decay steps
decay_steps_ind = None
for i in range(1, len(cfg.SOLVER.STEPS)):
if cfg.SOLVER.STEPS[i] >= args.start_step:
decay_steps_ind = i
break
if decay_steps_ind is None:
decay_steps_ind = len(cfg.SOLVER.STEPS)
training_stats = TrainingStats(
args,
args.disp_interval,
tblogger if args.use_tfboard and not args.no_save else None)
try:
logger.info('Training starts !')
step = args.start_step
for step in range(args.start_step, cfg.SOLVER.MAX_ITER):
# Warm up
if step < cfg.SOLVER.WARM_UP_ITERS:
method = cfg.SOLVER.WARM_UP_METHOD
if method == 'constant':
warmup_factor = cfg.SOLVER.WARM_UP_FACTOR
elif method == 'linear':
alpha = step / cfg.SOLVER.WARM_UP_ITERS
warmup_factor = cfg.SOLVER.WARM_UP_FACTOR * (1 - alpha) + alpha
else:
raise KeyError('Unknown SOLVER.WARM_UP_METHOD: {}'.format(method))
lr_new = cfg.SOLVER.BASE_LR * warmup_factor
net_utils.update_learning_rate(optimizer, lr, lr_new)
lr = optimizer.param_groups[0]['lr']
assert lr == lr_new
elif step == cfg.SOLVER.WARM_UP_ITERS:
net_utils.update_learning_rate(optimizer, lr, cfg.SOLVER.BASE_LR)
lr = optimizer.param_groups[0]['lr']
assert lr == cfg.SOLVER.BASE_LR
# Learning rate decay
if decay_steps_ind < len(cfg.SOLVER.STEPS) and \
step == cfg.SOLVER.STEPS[decay_steps_ind]:
logger.info('Decay the learning on step %d', step)
lr_new = lr * cfg.SOLVER.GAMMA
net_utils.update_learning_rate(optimizer, lr, lr_new)
lr = optimizer.param_groups[0]['lr']
assert lr == lr_new
decay_steps_ind += 1
training_stats.IterTic()
optimizer.zero_grad()
for inner_iter in range(args.iter_size):
try:
input_data = next(dataiterator)
except StopIteration:
dataiterator = iter(dataloader)
input_data = next(dataiterator)
for key in input_data:
if key != 'roidb': # roidb is a list of ndarrays with inconsistent length
input_data[key] = list(map(Variable, input_data[key]))
net_outputs = maskRCNN(**input_data)
training_stats.UpdateIterStats(net_outputs, inner_iter)
loss = net_outputs['total_loss']
loss.backward()
optimizer.step()
training_stats.IterToc()
training_stats.LogIterStats(step, lr)
if (step+1) % CHECKPOINT_PERIOD == 0:
save_ckpt(output_dir, args, step, train_size, maskRCNN, optimizer)
# ---- Training ends ----
# Save last checkpoint
save_ckpt(output_dir, args, step, train_size, maskRCNN, optimizer)
except (RuntimeError, KeyboardInterrupt):
del dataiterator
logger.info('Save ckpt on exception ...')
save_ckpt(output_dir, args, step, train_size, maskRCNN, optimizer)
logger.info('Save ckpt done.')
stack_trace = traceback.format_exc()
print(stack_trace)
finally:
if args.use_tfboard and not args.no_save:
tblogger.close()
def main():
"""Main function"""
args = parse_args()
print('Called with args:')
print(args)
if not torch.cuda.is_available():
sys.exit("Need a CUDA device to run the code.")
if args.cuda or cfg.NUM_GPUS > 0:
cfg.CUDA = True
else:
raise ValueError("Need Cuda device to run !")
if args.dataset == "coco2017":
cfg.TRAIN.DATASETS = ('coco_2017_train', )
cfg.MODEL.NUM_CLASSES = 81
elif args.dataset == "keypoints_coco2017":
cfg.TRAIN.DATASETS = ('keypoints_coco_2017_train', )
cfg.MODEL.NUM_CLASSES = 2
else:
raise ValueError("Unexpected args.dataset: {}".format(args.dataset))
cfg_from_file(args.cfg_file)
if args.set_cfgs is not None:
cfg_from_list(args.set_cfgs)
### Adaptively adjust some configs ###
original_batch_size = cfg.NUM_GPUS * cfg.TRAIN.IMS_PER_BATCH
if args.batch_size is None:
args.batch_size = original_batch_size
cfg.NUM_GPUS = torch.cuda.device_count()
assert (args.batch_size % cfg.NUM_GPUS) == 0, \
'batch_size: %d, NUM_GPUS: %d' % (args.batch_size, cfg.NUM_GPUS)
cfg.TRAIN.IMS_PER_BATCH = args.batch_size // cfg.NUM_GPUS
print('Batch size change from {} (in config file) to {}'.format(
original_batch_size, args.batch_size))
print('NUM_GPUs: %d, TRAIN.IMS_PER_BATCH: %d' %
(cfg.NUM_GPUS, cfg.TRAIN.IMS_PER_BATCH))
if args.num_workers is not None:
cfg.DATA_LOADER.NUM_THREADS = args.num_workers
print('Number of data loading threads: %d' % cfg.DATA_LOADER.NUM_THREADS)
### Adjust learning based on batch size change linearly
old_base_lr = cfg.SOLVER.BASE_LR
cfg.SOLVER.BASE_LR *= args.batch_size / original_batch_size
print('Adjust BASE_LR linearly according to batch size change: {} --> {}'.
format(old_base_lr, cfg.SOLVER.BASE_LR))
### Overwrite some solver settings from command line arguments
if args.optimizer is not None:
cfg.SOLVER.TYPE = args.optimizer
if args.lr is not None:
cfg.SOLVER.BASE_LR = args.lr
if args.lr_decay_gamma is not None:
cfg.SOLVER.GAMMA = args.lr_decay_gamma
assert_and_infer_cfg()
timers = defaultdict(Timer)
### Dataset ###
timers['roidb'].tic()
roidb, ratio_list, ratio_index = combined_roidb_for_training(
cfg.TRAIN.DATASETS, cfg.TRAIN.PROPOSAL_FILES)
timers['roidb'].toc()
roidb_size = len(roidb)
logger.info('{:d} roidb entries'.format(roidb_size))
logger.info('Takes %.2f sec(s) to construct roidb',
timers['roidb'].average_time)
# Effective training sample size for one epoch
train_size = roidb_size // args.batch_size * args.batch_size
sampler = MinibatchSampler(ratio_list, ratio_index)
dataset = RoiDataLoader(roidb, cfg.MODEL.NUM_CLASSES, training=True)
dataloader = torch.utils.data.DataLoader(
dataset,
batch_size=args.batch_size,
drop_last=True,
sampler=sampler,
num_workers=cfg.DATA_LOADER.NUM_THREADS,
collate_fn=collate_minibatch)
dataiterator = iter(dataloader)
### Model ###
maskRCNN = Generalized_RCNN()
if cfg.CUDA:
maskRCNN.cuda()
### Optimizer ###
bias_params = []
nonbias_params = []
for key, value in dict(maskRCNN.named_parameters()).items():
if value.requires_grad:
if 'bias' in key:
bias_params.append(value)
else:
nonbias_params.append(value)
# Learning rate of 0 is a dummy value to be set properly at the start of training
params = [{
'params': nonbias_params,
'lr': 0,
'weight_decay': cfg.SOLVER.WEIGHT_DECAY
}, {
'params':
bias_params,
'lr':
0 * (cfg.SOLVER.BIAS_DOUBLE_LR + 1),
'weight_decay':
cfg.SOLVER.WEIGHT_DECAY if cfg.SOLVER.BIAS_WEIGHT_DECAY else 0
}]
if cfg.SOLVER.TYPE == "SGD":
optimizer = torch.optim.SGD(params, momentum=cfg.SOLVER.MOMENTUM)
elif cfg.SOLVER.TYPE == "Adam":
optimizer = torch.optim.Adam(params)
### Load checkpoint
if args.load_ckpt:
load_name = args.load_ckpt
logging.info("loading checkpoint %s", load_name)
checkpoint = torch.load(load_name,
map_location=lambda storage, loc: storage)
net_utils.load_ckpt(maskRCNN, checkpoint['model'])
if args.resume:
args.start_step = checkpoint['step'] + 1
assert checkpoint['train_size'] == train_size
# There is a bug in optimizer.load_state_dict on Pytorch 0.3.1.
# However it's fixed on master.
# optimizer.load_state_dict(checkpoint['optimizer'])
misc_utils.load_optimizer_state_dict(optimizer,
checkpoint['optimizer'])
del checkpoint
torch.cuda.empty_cache()
if args.load_detectron: #TODO resume for detectron weights (load sgd momentum values)
logging.info("loading Detectron weights %s", args.load_detectron)
load_detectron_weight(maskRCNN, args.load_detectron)
lr = optimizer.param_groups[0][
'lr'] # lr of non-bias parameters, for commmand line outputs.
maskRCNN = mynn.DataParallel(maskRCNN,
cpu_keywords=['im_info', 'roidb'],
minibatch=True)
### Training Setups ###
run_name = misc_utils.get_run_name()
output_dir = misc_utils.get_output_dir(args, run_name)
if not args.no_save:
if not os.path.exists(output_dir):
os.makedirs(output_dir)
blob = {'cfg': yaml.dump(cfg), 'args': args}
with open(os.path.join(output_dir, 'config_and_args.pkl'), 'wb') as f:
pickle.dump(blob, f, pickle.HIGHEST_PROTOCOL)
if args.use_tfboard:
from tensorboardX import SummaryWriter
# Set the Tensorboard logger
tblogger = SummaryWriter(output_dir)
### Training Loop ###
maskRCNN.train()
CHECKPOINT_PERIOD = int(cfg.TRAIN.SNAPSHOT_ITERS / cfg.NUM_GPUS)
# Set index for decay steps
decay_steps_ind = None
for i in range(1, len(cfg.SOLVER.STEPS)):
if cfg.SOLVER.STEPS[i] > args.start_step:
decay_steps_ind = i
if decay_steps_ind is None:
decay_steps_ind = len(cfg.SOLVER.STEPS)
logger.info('Training starts !')
loss_avg = 0
try:
timers['train_loop'].tic()
for step in range(args.start_step, cfg.SOLVER.MAX_ITER):
# Warm up
if step < cfg.SOLVER.WARM_UP_ITERS:
method = cfg.SOLVER.WARM_UP_METHOD
if method == 'constant':
warmup_factor = cfg.SOLVER.WARM_UP_FACTOR
elif method == 'linear':
alpha = step / cfg.SOLVER.WARM_UP_ITERS
warmup_factor = cfg.SOLVER.WARM_UP_FACTOR * (1 -
alpha) + alpha
else:
raise KeyError(
'Unknown SOLVER.WARM_UP_METHOD: {}'.format(method))
lr_new = cfg.SOLVER.BASE_LR * warmup_factor
net_utils.update_learning_rate(optimizer, lr, lr_new)
lr = lr_new
elif step == cfg.SOLVER.WARM_UP_ITERS:
lr = cfg.SOLVER.BASE_LR
# Learning rate decay
if decay_steps_ind < len(cfg.SOLVER.STEPS) and \
step == cfg.SOLVER.STEPS[decay_steps_ind]:
logger.info('Decay the learning on step %d', step)
lr_new = lr * cfg.SOLVER.GAMMA
net_utils.update_learning_rate(optimizer, lr, lr_new)
lr = lr_new
decay_steps_ind += 1
try:
input_data = next(dataiterator)
except StopIteration:
dataiterator = iter(dataloader)
input_data = next(dataiterator)
for key in input_data:
if key != 'roidb': # roidb is a list of ndarrays with inconsistent length
input_data[key] = list(map(Variable, input_data[key]))
outputs = maskRCNN(**input_data)
rois_label = outputs['rois_label']
cls_score = outputs['cls_score']
bbox_pred = outputs['bbox_pred']
loss_rpn_cls = outputs['loss_rpn_cls'].mean()
loss_rpn_bbox = outputs['loss_rpn_bbox'].mean()
loss_rcnn_cls = outputs['loss_rcnn_cls'].mean()
loss_rcnn_bbox = outputs['loss_rcnn_bbox'].mean()
loss = loss_rpn_cls + loss_rpn_bbox + loss_rcnn_cls + loss_rcnn_bbox
if cfg.MODEL.MASK_ON:
loss_rcnn_mask = outputs['loss_rcnn_mask'].mean()
loss += loss_rcnn_mask
if cfg.MODEL.KEYPOINTS_ON:
loss_rcnn_keypoints = outputs['loss_rcnn_keypoints'].mean()
loss += loss_rcnn_keypoints
loss_avg += loss.data.cpu().numpy()[0]
optimizer.zero_grad()
loss.backward()
optimizer.step()
if (step + 1) % CHECKPOINT_PERIOD == 0:
save_ckpt(output_dir, args, step, train_size, maskRCNN,
optimizer)
if ((step % args.disp_interval == 0 and
(step - args.start_step >= args.disp_interval))
or step == cfg.SOLVER.MAX_ITER - 1):
diff = timers['train_loop'].toc(average=False)
loss_avg /= args.disp_interval
loss_rpn_cls = loss_rpn_cls.data[0]
loss_rpn_bbox = loss_rpn_bbox.data[0]
loss_rcnn_cls = loss_rcnn_cls.data[0]
loss_rcnn_bbox = loss_rcnn_bbox.data[0]
fg_cnt = torch.sum(rois_label.data.ne(0))
bg_cnt = rois_label.data.numel() - fg_cnt
print("[ %s ][ step %d ]" % (run_name, step))
print("\t\tloss: %.4f, lr: %.2e" % (loss_avg, lr))
print("\t\tfg/bg=(%d/%d), time cost: %f" %
(fg_cnt, bg_cnt, diff))
print(
"\t\trpn_cls: %.4f, rpn_bbox: %.4f, rcnn_cls: %.4f, rcnn_bbox %.4f"
% (loss_rpn_cls, loss_rpn_bbox, loss_rcnn_cls,
loss_rcnn_bbox))
print_prefix = "\t\t"
if cfg.MODEL.MASK_ON:
loss_rcnn_mask = loss_rcnn_mask.data[0]
print("%srcnn_mask %.4f" % (print_prefix, loss_rcnn_mask))
print_prefix = ", "
if cfg.MODEL.KEYPOINTS_ON:
loss_rcnn_keypoints = loss_rcnn_keypoints.data[0]
print("%srcnn_keypoints %.4f" %
(print_prefix, loss_rcnn_keypoints))
if args.use_tfboard and not args.no_save:
info = {
'lr': lr,
'loss': loss_avg,
'loss_rpn_cls': loss_rpn_cls,
'loss_rpn_box': loss_rpn_bbox,
'loss_rcnn_cls': loss_rcnn_cls,
'loss_rcnn_box': loss_rcnn_bbox,
}
if cfg.MODEL.MASK_ON:
info['loss_rcnn_mask'] = loss_rcnn_mask
if cfg.MODEL.KEYPOINTS_ON:
info['loss_rcnn_keypoints'] = loss_rcnn_keypoints
for tag, value in info.items():
tblogger.add_scalar(tag, value, step)
loss_avg = 0
timers['train_loop'].tic()
# Save last checkpoint
save_ckpt(output_dir, args, step, train_size, maskRCNN, optimizer)
except (RuntimeError, KeyboardInterrupt) as e:
print('Save on exception:', e)
save_ckpt(output_dir, args, step, train_size, maskRCNN, optimizer)
stack_trace = traceback.format_exc()
print(stack_trace)
finally:
# ---- Training ends ----
if args.use_tfboard and not args.no_save:
tblogger.close()
def main():
args = parse_args()
use_gpu = torch.cuda.is_available()
if args.use_cpu: use_gpu = False
pin_memory = True if use_gpu else False
#args.labelsmooth = True
if not args.evaluate:
sys.stdout = Logger(osp.join(args.save_dir, 'log_train.txt'))
else:
sys.stdout = Logger(osp.join(args.save_dir, 'log_test.txt'))
print("==========\nArgs:{}\n==========".format(args))
cfg_from_file(args.cfg_file)
if args.set_cfgs is not None:
cfg_from_list(args.set_cfgs)
if args.deterministic:
# Experiment reproducibility is sometimes important. Pete Warden expounded about this
# in his blog: https://petewarden.com/2018/03/19/the-machine-learning-reproducibility-crisis/
# In Pytorch, support for deterministic execution is still a bit clunky.
# Use a well-known seed, for repeatability of experiments
torch.manual_seed(args.seed)
random.seed(args.seed)
np.random.seed(args.seed)
cudnn.deterministic = True
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
cudnn.benchmark = True
torch.cuda.manual_seed_all(args.seed)
else:
print("Currently using CPU (GPU is highly recommended)")
# if cfg.RESNETS.SSN:
# rank, world_size = init_dist(backend='nccl', port=29500)
# print("The world_size is : {}".format(world_size))
# summary_writer = SummaryWriter(osp.join(args.save_dir, 'tensorboard_log'))
print("Initializing dataset {}".format(args.dataset))
dataset = data_manager.init_img_dataset(
root=args.root,
name=args.dataset,
split_id=args.split_id,
cuhk03_labeled=args.cuhk03_labeled,
cuhk03_classic_split=args.cuhk03_classic_split,
WEIGHT_TEST=cfg.REID.WEIGHT_TEST,
)
trainloader = DataLoader(
ImageDataset(dataset.train,
transform=TrainTransform(cfg.REID.HEIGHT, cfg.REID.WIDTH,
cfg.REID.PRE_PRO_TYPE)),
sampler=RandomIdentitySampler(
dataset.train, num_instances=cfg.REID.TRI_NUM_INSTANCES),
batch_size=cfg.REID.TRAIN_BATCH,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=True,
)
if cfg.REID.WEIGHT_TEST:
queryloader_1 = DataLoader(
ImageDataset(dataset.query_1,
transform=TestTransform(cfg.REID.HEIGHT,
cfg.REID.WIDTH,
cfg.REID.PRE_PRO_TYPE)),
batch_size=cfg.REID.TEST_BATCH,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
galleryloader_1 = DataLoader(
ImageDataset(dataset.gallery_1,
transform=TestTransform(cfg.REID.HEIGHT,
cfg.REID.WIDTH,
cfg.REID.PRE_PRO_TYPE)),
batch_size=cfg.REID.TEST_BATCH,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
queryloader_2 = DataLoader(
ImageDataset(dataset.query_2,
transform=TestTransform(cfg.REID.HEIGHT,
cfg.REID.WIDTH,
cfg.REID.PRE_PRO_TYPE)),
batch_size=cfg.REID.TEST_BATCH,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
galleryloader_2 = DataLoader(
ImageDataset(dataset.gallery_2,
transform=TestTransform(cfg.REID.HEIGHT,
cfg.REID.WIDTH,
cfg.REID.PRE_PRO_TYPE)),
batch_size=cfg.REID.TEST_BATCH,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
else:
queryloader = DataLoader(
ImageDataset(dataset.query,
transform=TestTransform(cfg.REID.HEIGHT,
cfg.REID.WIDTH,
cfg.REID.PRE_PRO_TYPE)),
batch_size=cfg.REID.TEST_BATCH,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
galleryloader = DataLoader(
ImageDataset(dataset.gallery,
transform=TestTransform(cfg.REID.HEIGHT,
cfg.REID.WIDTH,
cfg.REID.PRE_PRO_TYPE)),
batch_size=cfg.REID.TEST_BATCH,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
print("Initializing model: {}".format(
cfg.MODEL.CONV_BODY.split('.')[-1].split('_')[1]))
model = Generalized_FPN(num_classes=dataset.num_train_pids,
loss=cfg.REID.LOSS,
aligned=cfg.REID.ALIGNED,
strong_baseline=cfg.REID.STRONG_BASELINE)
print("Model size: {:.5f}M".format(
sum(p.numel() for p in model.parameters()) / 1000000.0))
if cfg.REID.FOCALLOSS:
criterion_class = FocalLoss(gamma=2, alpha=0.25, \
labelsmooth=cfg.REID.LABElSMOOTH,
num_classes=dataset.num_train_pids,
epsilon=0.1)
elif cfg.REID.LABLESMOOTH:
criterion_class = CrossEntropyLabelSmooth(
num_classes=dataset.num_train_pids, use_gpu=use_gpu)
else:
criterion_class = CrossEntropyLoss(use_gpu=use_gpu)
criterion_class_oimwarmup = None
criterion_metric = TripletLossAlignedReID(margin=cfg.REID.TRI_MARGIN,
aligned=cfg.REID.ALIGNED)
criterion_center = CenterLoss(num_classes=dataset.num_train_pids,
feat_dim=2048)
if cfg.REID.CAMIDCLASS:
criterion_class_cam = nn.BCEWithLogitsLoss().cuda()
else:
criterion_class_cam = None
criterion_class_mixup = None #CrossEntropyLoss(use_gpu=use_gpu)
def load_ckpt(model, ckpt):
"""Load checkpoint"""
mapping, _ = model.detectron_weight_mapping
model_state_dict = model.state_dict()
state_dict = {}
for name in ckpt:
try:
if mapping[name] and name.split('.')[1] != 'posthoc_modules'\
and name.split('.')[1] !='topdown_lateral_modules' \
and name.split('.')[1] !='conv_top':
state_dict[name] = ckpt[name]
except:
if name.split('.')[0] != 'bottleneck_sub2' and name.split('.')[0] != 'classifier' \
and name.split('.')[0].split('_')[0] != 'classifier':
state_dict[name] = ckpt[name]
print('parameters: {}'.format(name))
model.load_state_dict(state_dict, strict=False)
if args.load_ckpt:
load_name = args.load_ckpt
print("loading checkpoint %s", load_name)
checkpoint = torch.load(load_name,
map_location=lambda storage, loc: storage)
load_ckpt(model, checkpoint['state_dict'])
start_epoch = args.start_epoch
if args.resume:
print("Loading checkpoint from '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
# print(checkpoint['state_dict'].keys())
model.load_state_dict(checkpoint['state_dict'])
start_epoch = checkpoint['epoch']
### Optimizer ###
if cfg.REID.DETECTRON_OPTIMIZER:
gn_params = []
bias_params = []
bias_param_names = []
nonbias_params = []
nonbias_param_names = []
ssn_params = []
ssn_param_names = []
for key, value in dict(model.named_parameters()).items():
if value.requires_grad:
if 'gn' in key:
gn_params.append(value)
elif 'bias' in key:
bias_params.append(value)
bias_param_names.append(key)
elif key.endswith('_weight') and cfg.RESNETS.SSN:
# print('********{}'.format(key))
ssn_params.append(value)
ssn_param_names.append(key)
else:
nonbias_params.append(value)
nonbias_param_names.append(key)
else:
# print('NO!')
print('FREEZE para: {}'.format(key))
# Learning rate of 0 is a dummy value to be set properly at the start of training
params = [
{
'params': nonbias_params,
'lr': cfg.SOLVER.BASE_LR, #0,
'weight_decay': cfg.SOLVER.WEIGHT_DECAY
},
{
'params':
bias_params,
'lr':
cfg.SOLVER.BASE_LR, #0 * (cfg.SOLVER.BIAS_DOUBLE_LR + 1),
'weight_decay':
cfg.SOLVER.WEIGHT_DECAY if cfg.SOLVER.BIAS_WEIGHT_DECAY else 0
},
{
'params': gn_params,
'lr': cfg.SOLVER.BASE_LR, #0,
'weight_decay': cfg.SOLVER.WEIGHT_DECAY_GN
},
{
'params': ssn_params,
'lr': cfg.SOLVER.BASE_LR / 10, # 0, / 10
'weight_decay': 0
} #cfg.SOLVER.WEIGHT_DECAY
]
# names of paramerters for each paramter
param_names = [nonbias_param_names, bias_param_names, ssn_param_names]
if cfg.SOLVER.TYPE == "SGD":
optimizer = torch.optim.SGD(params,
momentum=cfg.SOLVER.MOMENTUM,
weight_decay=cfg.SOLVER.WEIGHT_DECAY)
elif cfg.SOLVER.TYPE == "Adam":
optimizer = torch.optim.Adam(params,
weight_decay=cfg.SOLVER.WEIGHT_DECAY)
elif cfg.SOLVER.TYPE == 'Rmsprop':
optimizer = torch.optim.RMSprop(
params,
momentum=cfg.SOLVER.MOMENTUM,
weight_decay=cfg.SOLVER.WEIGHT_DECAY)
else:
raise KeyError("Unsupported optim: {}".format(cfg.SOLVER.TYPE))
lr = optimizer.param_groups[0][
'lr'] # lr of non-bias parameters, for commmand line outputs.
if cfg.SOLVER.LR_POLICY == 'cosine_annealing':
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, float(400 - 18), eta_min=0.000001)
if cfg.REID.SWA:
# SWA: initialize SWA optimizer wrapper
print('SWA training')
# steps_per_epoch = len(loaders['train'].dataset) / args.batch_size
# steps_per_epoch = int(steps_per_epoch)
# print("Steps per epoch:", steps_per_epoch)
optimizer = SWA(optimizer,
swa_start=cfg.REID.SWA_START,
swa_freq=cfg.REID.SWA_FREQ,
swa_lr=cfg.REID.SWA_LR)
# Set index for decay steps
decay_steps_ind = None
for i in range(1, len(cfg.SOLVER.STEPS)):
if cfg.SOLVER.STEPS[i] >= args.start_epoch:
decay_steps_ind = i
break
if decay_steps_ind is None:
decay_steps_ind = len(cfg.SOLVER.STEPS)
else:
raise KeyError("Unimplemented cfg.REID.DETECTRON_OPTIMIZER: {}".format(
cfg.REID.DETECTRON_OPTIMIZER))
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
if cfg.REID.WEIGHT_TEST:
rank1 = wwn_test_2(model, queryloader_1, galleryloader_1,
queryloader_2, galleryloader_2, use_gpu, args)
else:
wwn_test(model, queryloader, galleryloader, use_gpu, args)
return 0
start_time = time.time()
train_time = 0
best_rank1 = -np.inf
best_epoch = 0
print("==> Start training")
max_epoch = cfg.SOLVER.MAX_ITER
for epoch in range(start_epoch, max_epoch):
if cfg.REID.DETECTRON_OPTIMIZER:
# Warm up
if epoch < cfg.SOLVER.WARM_UP_ITERS:
method = cfg.SOLVER.WARM_UP_METHOD
if method == 'constant':
warmup_factor = cfg.SOLVER.WARM_UP_FACTOR
elif method == 'linear':
# alpha = (epoch+1) / cfg.SOLVER.WARM_UP_ITERS
# warmup_factor = cfg.SOLVER.WARM_UP_FACTOR * (1 - alpha) + alpha
alpha = (cfg.SOLVER.BASE_LR - cfg.SOLVER.WARM_UP_FACTOR
) / cfg.SOLVER.WARM_UP_ITERS
warmup_factor = epoch * alpha + cfg.SOLVER.WARM_UP_FACTOR
# print(warmup_factor)
else:
raise KeyError(
'Unknown SOLVER.WARM_UP_METHOD: {}'.format(method))
lr_new = warmup_factor
net_utils.update_learning_rate(optimizer, lr, lr_new)
lr = optimizer.param_groups[0]['lr']
assert lr == lr_new
elif epoch == cfg.SOLVER.WARM_UP_ITERS:
net_utils.update_learning_rate(optimizer, lr,
cfg.SOLVER.BASE_LR)
lr = optimizer.param_groups[0]['lr']
assert lr == cfg.SOLVER.BASE_LR
# Learning rate decay
if cfg.SOLVER.LR_POLICY == 'steps_with_decay':
if decay_steps_ind < len(cfg.SOLVER.STEPS) and \
epoch == cfg.SOLVER.STEPS[decay_steps_ind]:
print('Decay the learning on step %d', epoch)
lr_new = lr * cfg.SOLVER.GAMMA
net_utils.update_learning_rate(optimizer, lr, lr_new)
lr = optimizer.param_groups[0]['lr']
assert lr == lr_new
decay_steps_ind += 1
elif cfg.SOLVER.LR_POLICY == 'cosine_annealing' and epoch >= cfg.SOLVER.WARM_UP_ITERS:
scheduler.step()
lr_new = scheduler.get_lr()[0]
net_utils.update_learning_rate(optimizer, lr, lr_new)
lr = optimizer.param_groups[0]['lr']
assert lr == lr_new
else:
if args.strong_baseline:
adjust_lr(optimizer, epoch + 1)
else:
if args.stepsize > 0:
scheduler.step()
start_train_time = time.time()
train(epoch, model, criterion_class, criterion_metric, criterion_center,optimizer, trainloader, use_gpu, \
criterion_class_oimwarmup,criterion_class_cam, \
criterion_class_mixup, args)
train_time += round(time.time() - start_train_time)
# if cfg.REID.SWA and (epoch + 1) >= cfg.REID.SWA_START:
# # utils.moving_average(swa_model, model, 1.0 / (swa_n + 1))
# # swa_n += 1
# if epoch == 0 or epoch % args.eval_step == args.eval_step - 1 or epoch == max_epoch - 1:
# # Batchnorm update
# optimizer.swap_swa_sgd()
# # print('WWN')
# optimizer.bn_update(trainloader, model, device='cuda')
# # swa_res = utils.eval(loaders['test'], model, criterion)
if (epoch + 1) > args.start_eval and args.eval_step > 0 and (
epoch + 1) % args.eval_step == 0 or (epoch + 1) == max_epoch:
print("==> Test")
if cfg.REID.SWA and (epoch + 1) >= cfg.REID.SWA_START:
optimizer.swap_swa_sgd()
# print('WWN')
optimizer.bn_update(trainloader, model, device='cuda')
if cfg.REID.WEIGHT_TEST:
rank1 = wwn_test_2(model, queryloader_1, galleryloader_1,
queryloader_2, galleryloader_2, use_gpu,
args)
else:
rank1 = wwn_test(model, queryloader, galleryloader,
use_gpu, args)
optimizer.swap_swa_sgd()
else:
if cfg.REID.WEIGHT_TEST:
rank1 = wwn_test_2(model, queryloader_1, galleryloader_1,
queryloader_2, galleryloader_2, use_gpu,
args)
else:
rank1 = wwn_test(model, queryloader, galleryloader,
use_gpu, args)
is_best = rank1 > best_rank1
if is_best:
best_rank1 = rank1
best_epoch = epoch + 1
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint_best(
{
'state_dict': state_dict,
'rank1': rank1,
'epoch': epoch,
}, is_best,
osp.join(args.save_dir,
'checkpoint_ep' + str(epoch + 1) + '.pth.tar'))
print("==> Best Rank-1 {:.1%}, achieved at epoch {}".format(
best_rank1, best_epoch))
elapsed = round(time.time() - start_time)
elapsed = str(datetime.timedelta(seconds=elapsed))
train_time = str(datetime.timedelta(seconds=train_time))
print(
"Finished. Total elapsed time (h:m:s): {}. Training time (h:m:s): {}.".
format(elapsed, train_time))
def main():
"""Main function"""
args = parse_cfg()
timers = defaultdict(Timer)
### Dataset ###
timers['roidb'].tic()
roidb, ratio_list, ratio_index = combined_roidb_for_training(
cfg.TRAIN.DATASETS, cfg.TRAIN.PROPOSAL_FILES)
timers['roidb'].toc()
roidb_size = len(roidb)
logger.info('{:d} roidb entries'.format(roidb_size))
logger.info('Takes %.2f sec(s) to construct roidb',
timers['roidb'].average_time)
# Effective training sample size for one epoch
train_size = roidb_size // args.batch_size * args.batch_size
batchSampler = BatchSampler(sampler=MinibatchSampler(
ratio_list, ratio_index),
batch_size=args.batch_size,
drop_last=True)
dataset = RoiDataLoader(roidb, cfg.MODEL.NUM_CLASSES, training=True)
dataloader = torch.utils.data.DataLoader(
dataset,
batch_sampler=batchSampler,
num_workers=cfg.DATA_LOADER.NUM_THREADS,
collate_fn=collate_minibatch)
dataiterator = iter(dataloader)
### Model ###
maskRCNN = GetRCNNModel()
if cfg.CUDA:
maskRCNN.cuda()
### Optimizer ###
gn_param_nameset = set()
for name, module in maskRCNN.named_modules():
if isinstance(module, nn.GroupNorm):
gn_param_nameset.add(name + '.weight')
gn_param_nameset.add(name + '.bias')
gn_params = []
gn_param_names = []
bias_params = []
bias_param_names = []
nonbias_params = []
nonbias_param_names = []
nograd_param_names = []
for key, value in dict(maskRCNN.named_parameters()).items():
if value.requires_grad:
if 'bias' in key:
bias_params.append(value)
bias_param_names.append(key)
elif key in gn_param_nameset:
gn_params.append(value)
gn_param_names.append(key)
else:
nonbias_params.append(value)
nonbias_param_names.append(key)
else:
nograd_param_names.append(key)
assert (gn_param_nameset - set(nograd_param_names) -
set(bias_param_names)) == set(gn_param_names)
# Learning rate of 0 is a dummy value to be set properly at the start of training
params = [{
'params': nonbias_params,
'lr': 0,
'weight_decay': cfg.SOLVER.WEIGHT_DECAY
}, {
'params':
bias_params,
'lr':
0 * (cfg.SOLVER.BIAS_DOUBLE_LR + 1),
'weight_decay':
cfg.SOLVER.WEIGHT_DECAY if cfg.SOLVER.BIAS_WEIGHT_DECAY else 0
}, {
'params': gn_params,
'lr': 0,
'weight_decay': cfg.SOLVER.WEIGHT_DECAY_GN
}]
# names of paramerters for each paramter
param_names = [nonbias_param_names, bias_param_names, gn_param_names]
if cfg.SOLVER.TYPE == "SGD":
optimizer = torch.optim.SGD(params, momentum=cfg.SOLVER.MOMENTUM)
elif cfg.SOLVER.TYPE == "Adam":
optimizer = torch.optim.Adam(params)
### Load checkpoint
if args.load_ckpt:
load_name = args.load_ckpt
logging.info("loading checkpoint %s", load_name)
checkpoint = torch.load(load_name,
map_location=lambda storage, loc: storage)
net_utils.load_ckpt(maskRCNN, checkpoint['model'])
if args.resume:
args.start_step = checkpoint['step'] + 1
if 'train_size' in checkpoint: # For backward compatibility
if checkpoint['train_size'] != train_size:
print(
'train_size value: %d different from the one in checkpoint: %d'
% (train_size, checkpoint['train_size']))
# reorder the params in optimizer checkpoint's params_groups if needed
# misc_utils.ensure_optimizer_ckpt_params_order(param_names, checkpoint)
# There is a bug in optimizer.load_state_dict on Pytorch 0.3.1.
# However it's fixed on master.
# optimizer.load_state_dict(checkpoint['optimizer'])
misc_utils.load_optimizer_state_dict(optimizer,
checkpoint['optimizer'])
del checkpoint
torch.cuda.empty_cache()
if args.load_detectron: #TODO resume for detectron weights (load sgd momentum values)
logging.info("loading Detectron weights %s", args.load_detectron)
load_detectron_weight(maskRCNN, args.load_detectron)
lr = optimizer.param_groups[0][
'lr'] # lr of non-bias parameters, for commmand line outputs.
maskRCNN = mynn.DataParallel(maskRCNN,
cpu_keywords=['im_info', 'roidb'],
minibatch=True)
if cfg.TRAIN_SYNC_BN:
# Shu:For synchorinized BN
patch_replication_callback(maskRCNN)
### Training Setups ###
args.run_name = misc_utils.get_run_name() + '_step'
output_dir = misc_utils.get_output_dir(args, args.run_name)
args.cfg_filename = os.path.basename(args.cfg_file)
if not args.no_save:
if not os.path.exists(output_dir):
os.makedirs(output_dir)
blob = {'cfg': yaml.dump(cfg), 'args': args}
with open(os.path.join(output_dir, 'config_and_args.pkl'), 'wb') as f:
pickle.dump(blob, f, pickle.HIGHEST_PROTOCOL)
if args.use_tfboard:
from tensorboardX import SummaryWriter
# Set the Tensorboard logger
tblogger = SummaryWriter(output_dir)
### Training Loop ###
maskRCNN.train()
CHECKPOINT_PERIOD = int(cfg.TRAIN.SNAPSHOT_ITERS / cfg.NUM_GPUS)
# Set index for decay steps
decay_steps_ind = None
for i in range(1, len(cfg.SOLVER.STEPS)):
if cfg.SOLVER.STEPS[i] >= args.start_step:
decay_steps_ind = i
break
if decay_steps_ind is None:
decay_steps_ind = len(cfg.SOLVER.STEPS)
training_stats = TrainingStats(
args, args.disp_interval,
tblogger if args.use_tfboard and not args.no_save else None)
try:
logger.info('Training starts !')
step = args.start_step
for step in range(args.start_step, cfg.SOLVER.MAX_ITER):
# Warm up
if step < cfg.SOLVER.WARM_UP_ITERS:
method = cfg.SOLVER.WARM_UP_METHOD
if method == 'constant':
warmup_factor = cfg.SOLVER.WARM_UP_FACTOR
elif method == 'linear':
alpha = step / cfg.SOLVER.WARM_UP_ITERS
warmup_factor = cfg.SOLVER.WARM_UP_FACTOR * (1 -
alpha) + alpha
else:
raise KeyError(
'Unknown SOLVER.WARM_UP_METHOD: {}'.format(method))
lr_new = cfg.SOLVER.BASE_LR * warmup_factor
net_utils.update_learning_rate(optimizer, lr, lr_new)
lr = optimizer.param_groups[0]['lr']
assert lr == lr_new
elif step == cfg.SOLVER.WARM_UP_ITERS:
net_utils.update_learning_rate(optimizer, lr,
cfg.SOLVER.BASE_LR)
lr = optimizer.param_groups[0]['lr']
assert lr == cfg.SOLVER.BASE_LR
# Learning rate decay
if decay_steps_ind < len(cfg.SOLVER.STEPS) and \
step == cfg.SOLVER.STEPS[decay_steps_ind]:
logger.info('Decay the learning on step %d', step)
lr_new = lr * cfg.SOLVER.GAMMA
net_utils.update_learning_rate(optimizer, lr, lr_new)
lr = optimizer.param_groups[0]['lr']
assert lr == lr_new
decay_steps_ind += 1
training_stats.IterTic()
optimizer.zero_grad()
for inner_iter in range(args.iter_size):
try:
input_data = next(dataiterator)
except StopIteration:
dataiterator = iter(dataloader)
input_data = next(dataiterator)
for key in input_data:
if cfg.MODEL.LR_VIEW_ON or cfg.MODEL.GIF_ON or cfg.MODEL.LRASY_MAHA_ON:
if key != 'roidb' and key != 'data': # roidb is a list of ndarrays with inconsistent length
input_data[key] = list(
map(Variable, input_data[key]))
if key == 'data':
input_data[key] = [
torch.squeeze(item) for item in input_data[key]
]
input_data[key] = list(
map(Variable, input_data[key]))
else:
if key != 'roidb': # roidb is a list of ndarrays with inconsistent length
input_data[key] = list(
map(Variable, input_data[key]))
net_outputs = maskRCNN(**input_data)
training_stats.UpdateIterStats(net_outputs, inner_iter)
loss = net_outputs['total_loss']
loss.backward()
optimizer.step()
training_stats.IterToc()
training_stats.LogIterStats(step, lr)
if (step + 1) % CHECKPOINT_PERIOD == 0:
net_utils.train_save_ckpt(output_dir, args, step, train_size,
maskRCNN, optimizer)
# ---- Training ends ----
# Save last checkpoint
net_utils.train_save_ckpt(output_dir, args, step, train_size, maskRCNN,
optimizer)
except (RuntimeError, KeyboardInterrupt):
del dataiterator
logger.info('Save ckpt on exception ...')
net_utils.train_save_ckpt(output_dir, args, step, train_size, maskRCNN,
optimizer)
logger.info('Save ckpt done.')
stack_trace = traceback.format_exc()
print(stack_trace)
finally:
if args.use_tfboard and not args.no_save:
tblogger.close()
def main():
"""Main function"""
args = parse_args()
print('Called with args:')
print(args)
if not torch.cuda.is_available():
sys.exit("Need a CUDA device to run the code.")
if args.cuda or cfg.NUM_GPUS > 0:
cfg.CUDA = True
else:
raise ValueError("Need Cuda device to run !")
if args.dataset == "coco2017":
cfg.TRAIN.DATASETS = ('coco_2017_train', )
cfg.MODEL.NUM_CLASSES = 81
elif args.dataset == "keypoints_coco2017":
cfg.TRAIN.DATASETS = ('keypoints_coco_2017_train', )
cfg.MODEL.NUM_CLASSES = 2
# ADE20k as a detection dataset
elif args.dataset == "ade_train":
cfg.TRAIN.DATASETS = ('ade_train', )
cfg.MODEL.NUM_CLASSES = 446
# Noisy CS6+WIDER datasets
elif args.dataset == 'cs6_noise020+WIDER':
cfg.TRAIN.DATASETS = ('cs6_noise020', 'wider_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'cs6_noise050+WIDER':
cfg.TRAIN.DATASETS = ('cs6_noise050', 'wider_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'cs6_noise080+WIDER':
cfg.TRAIN.DATASETS = ('cs6_noise080', 'wider_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'cs6_noise085+WIDER':
cfg.TRAIN.DATASETS = ('cs6_noise085', 'wider_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'cs6_noise090+WIDER':
cfg.TRAIN.DATASETS = ('cs6_noise090', 'wider_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'cs6_noise095+WIDER':
cfg.TRAIN.DATASETS = ('cs6_noise095', 'wider_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'cs6_noise100+WIDER':
cfg.TRAIN.DATASETS = ('cs6_noise100', 'wider_train')
cfg.MODEL.NUM_CLASSES = 2
# Just Noisy CS6 datasets
elif args.dataset == 'cs6_noise020':
cfg.TRAIN.DATASETS = ('cs6_noise020', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'cs6_noise030':
cfg.TRAIN.DATASETS = ('cs6_noise030', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'cs6_noise040':
cfg.TRAIN.DATASETS = ('cs6_noise040', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'cs6_noise050':
cfg.TRAIN.DATASETS = ('cs6_noise050', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'cs6_noise060':
cfg.TRAIN.DATASETS = ('cs6_noise060', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'cs6_noise070':
cfg.TRAIN.DATASETS = ('cs6_noise070', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'cs6_noise080':
cfg.TRAIN.DATASETS = ('cs6_noise080', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'cs6_noise085':
cfg.TRAIN.DATASETS = ('cs6_noise085', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'cs6_noise090':
cfg.TRAIN.DATASETS = ('cs6_noise090', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'cs6_noise095':
cfg.TRAIN.DATASETS = ('cs6_noise095', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'cs6_noise100':
cfg.TRAIN.DATASETS = ('cs6_noise100', )
cfg.MODEL.NUM_CLASSES = 2
# Cityscapes 7 classes
elif args.dataset == "cityscapes":
cfg.TRAIN.DATASETS = ('cityscapes_train', )
cfg.MODEL.NUM_CLASSES = 8
# BDD 7 classes
elif args.dataset == "bdd_any_any_any":
cfg.TRAIN.DATASETS = ('bdd_any_any_any_train', )
cfg.MODEL.NUM_CLASSES = 8
elif args.dataset == "bdd_any_any_daytime":
cfg.TRAIN.DATASETS = ('bdd_any_any_daytime_train', )
cfg.MODEL.NUM_CLASSES = 8
elif args.dataset == "bdd_clear_any_daytime":
cfg.TRAIN.DATASETS = ('bdd_clear_any_daytime_train', )
cfg.MODEL.NUM_CLASSES = 8
# Cistyscapes Pedestrian sets
elif args.dataset == "cityscapes_peds":
cfg.TRAIN.DATASETS = ('cityscapes_peds_train', )
cfg.MODEL.NUM_CLASSES = 2
# Cityscapes Car sets
elif args.dataset == "cityscapes_cars_HPlen3+kitti_car_train":
cfg.TRAIN.DATASETS = ('cityscapes_cars_HPlen3', 'kitti_car_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cityscapes_cars_HPlen5+kitti_car_train":
cfg.TRAIN.DATASETS = ('cityscapes_cars_HPlen5', 'kitti_car_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cityscapes_car_train+kitti_car_train":
cfg.TRAIN.DATASETS = ('cityscapes_car_train', 'kitti_car_train')
cfg.MODEL.NUM_CLASSES = 2
# KITTI Car set
elif args.dataset == "kitti_car_train":
cfg.TRAIN.DATASETS = ('kitti_car_train', )
cfg.MODEL.NUM_CLASSES = 2
# BDD pedestrians sets
elif args.dataset == "bdd_peds":
cfg.TRAIN.DATASETS = ('bdd_peds_train',
) # bdd peds: clear_any_daytime
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "bdd_peds_full":
cfg.TRAIN.DATASETS = ('bdd_peds_full_train', ) # bdd peds: any_any_any
cfg.MODEL.NUM_CLASSES = 2
# Pedestrians with constraints
elif args.dataset == "bdd_peds_not_clear_any_daytime":
cfg.TRAIN.DATASETS = ('bdd_peds_not_clear_any_daytime_train', )
cfg.MODEL.NUM_CLASSES = 2
# Ashish's 20k samples videos
elif args.dataset == "bdd_peds_not_clear_any_daytime_20k":
cfg.TRAIN.DATASETS = ('bdd_peds_not_clear_any_daytime_20k_train', )
cfg.MODEL.NUM_CLASSES = 2
# Source domain + Target domain detections
elif args.dataset == "bdd_peds+DETS_20k":
cfg.TRAIN.DATASETS = ('bdd_peds_dets_20k_target_domain',
'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
# Source domain + Target domain detections -- same 18k images as HP18k
elif args.dataset == "bdd_peds+DETS18k":
cfg.TRAIN.DATASETS = ('bdd_peds_dets18k_target_domain',
'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
# Only Dets
elif args.dataset == "DETS20k":
cfg.TRAIN.DATASETS = ('bdd_peds_dets_20k_target_domain', )
cfg.MODEL.NUM_CLASSES = 2
# Only Dets18k - same images as HP18k
elif args.dataset == 'DETS18k':
cfg.TRAIN.DATASETS = ('bdd_peds_dets18k_target_domain', )
cfg.MODEL.NUM_CLASSES = 2
# Only HP
elif args.dataset == 'HP':
cfg.TRAIN.DATASETS = ('bdd_peds_HP_target_domain', )
cfg.MODEL.NUM_CLASSES = 2
# Only HP 18k videos
elif args.dataset == 'HP18k':
cfg.TRAIN.DATASETS = ('bdd_peds_HP18k_target_domain', )
cfg.MODEL.NUM_CLASSES = 2
# Source domain + Target domain HP
elif args.dataset == 'bdd_peds+HP':
cfg.TRAIN.DATASETS = ('bdd_peds_train', 'bdd_peds_HP_target_domain')
cfg.MODEL.NUM_CLASSES = 2
# Source domain + Target domain HP 18k videos
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+HP18k':
cfg.TRAIN.DATASETS = ('bdd_peds_HP18k_target_domain', 'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
#### Source domain + Target domain with different conf threshold theta ####
elif args.dataset == 'bdd_peds+HP18k_thresh-050':
cfg.TRAIN.DATASETS = ('bdd_HP18k_thresh-050', 'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+HP18k_thresh-060':
cfg.TRAIN.DATASETS = ('bdd_HP18k_thresh-060', 'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+HP18k_thresh-070':
cfg.TRAIN.DATASETS = ('bdd_HP18k_thresh-070', 'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+HP18k_thresh-090':
cfg.TRAIN.DATASETS = ('bdd_HP18k_thresh-090', 'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
##############################
#### Data distillation on BDD -- for rebuttal
elif args.dataset == 'bdd_peds+bdd_data_dist_small':
cfg.TRAIN.DATASETS = ('bdd_data_dist_small', 'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+bdd_data_dist_mid':
cfg.TRAIN.DATASETS = ('bdd_data_dist_mid', 'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+bdd_data_dist':
cfg.TRAIN.DATASETS = ('bdd_data_dist', 'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
##############################
#### Source domain + **Labeled** Target domain with varying number of images
elif args.dataset == 'bdd_peds+labeled_100':
cfg.TRAIN.DATASETS = ('bdd_peds_not_clear_any_daytime_train_100',
'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+labeled_075':
cfg.TRAIN.DATASETS = ('bdd_peds_not_clear_any_daytime_train_075',
'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+labeled_050':
cfg.TRAIN.DATASETS = ('bdd_peds_not_clear_any_daytime_train_050',
'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+labeled_025':
cfg.TRAIN.DATASETS = ('bdd_peds_not_clear_any_daytime_train_025',
'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+labeled_010':
cfg.TRAIN.DATASETS = ('bdd_peds_not_clear_any_daytime_train_010',
'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+labeled_005':
cfg.TRAIN.DATASETS = ('bdd_peds_not_clear_any_daytime_train_005',
'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+labeled_001':
cfg.TRAIN.DATASETS = ('bdd_peds_not_clear_any_daytime_train_001',
'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
##############################
# Source domain + Target domain HP tracker bboxes only
elif args.dataset == 'bdd_peds+HP18k_track_only':
cfg.TRAIN.DATASETS = ('bdd_HP18k_track_only', 'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
##### subsets of bdd_HP18k with different constraints
# Source domain + HP tracker images at NIGHT
elif args.dataset == 'bdd_peds+HP18k_any_any_night':
cfg.TRAIN.DATASETS = ('bdd_HP18k_any_any_night', 'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+HP18k_rainy_any_daytime':
cfg.TRAIN.DATASETS = ('bdd_HP18k_rainy_any_daytime', 'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+HP18k_rainy_any_night':
cfg.TRAIN.DATASETS = ('bdd_HP18k_rainy_any_night', 'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+HP18k_overcast,rainy_any_daytime':
cfg.TRAIN.DATASETS = ('bdd_HP18k_overcast,rainy_any_daytime',
'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+HP18k_overcast,rainy_any_night':
cfg.TRAIN.DATASETS = ('bdd_HP18k_overcast,rainy_any_night',
'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+HP18k_overcast,rainy,snowy_any_daytime':
cfg.TRAIN.DATASETS = ('bdd_HP18k_overcast,rainy,snowy_any_daytime',
'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
############# end of bdd constraned subsets #####################
# Source domain + Target domain HP18k -- after histogram matching
elif args.dataset == 'bdd_peds+HP18k_remap_hist':
cfg.TRAIN.DATASETS = ('bdd_peds_HP18k_target_domain_remap_hist',
'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+HP18k_remap_cityscape_hist':
cfg.TRAIN.DATASETS = (
'bdd_peds_HP18k_target_domain_remap_cityscape_hist',
'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
# Source domain + Target domain HP18k -- after histogram matching
elif args.dataset == 'bdd_peds+HP18k_remap_random':
cfg.TRAIN.DATASETS = ('bdd_peds_HP18k_target_domain_remap_random',
'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
# Source+Noisy Target domain -- prevent domain adv from using HP roi info
elif args.dataset == 'bdd_peds+bdd_HP18k_noisy_100k':
cfg.TRAIN.DATASETS = ('bdd_HP18k_noisy_100k', 'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+bdd_HP18k_noisy_080':
cfg.TRAIN.DATASETS = ('bdd_HP18k_noisy_080', 'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+bdd_HP18k_noisy_060':
cfg.TRAIN.DATASETS = ('bdd_HP18k_noisy_060', 'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == 'bdd_peds+bdd_HP18k_noisy_070':
cfg.TRAIN.DATASETS = ('bdd_HP18k_noisy_070', 'bdd_peds_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "wider_train":
cfg.TRAIN.DATASETS = ('wider_train', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-subset":
cfg.TRAIN.DATASETS = ('cs6-subset', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-subset-score":
cfg.TRAIN.DATASETS = ('cs6-subset-score', )
elif args.dataset == "cs6-subset-gt":
cfg.TRAIN.DATASETS = ('cs6-subset-gt', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-3013-gt":
cfg.TRAIN.DATASETS = ('cs6-3013-gt',
) # DEBUG: overfit on one video annots
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-subset-gt+WIDER":
cfg.TRAIN.DATASETS = ('cs6-subset-gt', 'wider_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-subset+WIDER":
cfg.TRAIN.DATASETS = ('cs6-subset', 'wider_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-train-gt":
cfg.TRAIN.DATASETS = ('cs6-train-gt', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-train-gt-noisy-0.3":
cfg.TRAIN.DATASETS = ('cs6-train-gt-noisy-0.3', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-train-gt-noisy-0.5":
cfg.TRAIN.DATASETS = ('cs6-train-gt-noisy-0.5', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-train-det-score":
cfg.TRAIN.DATASETS = ('cs6-train-det-score', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-train-det-score-0.5":
cfg.TRAIN.DATASETS = ('cs6-train-det-score-0.5', )
elif args.dataset == "cs6-train-det":
cfg.TRAIN.DATASETS = ('cs6-train-det', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-train-det-0.5":
cfg.TRAIN.DATASETS = ('cs6-train-det-0.5', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-train-hp":
cfg.TRAIN.DATASETS = ('cs6-train-hp', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-train-easy-gt":
cfg.TRAIN.DATASETS = ('cs6-train-easy-gt', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-train-easy-gt-sub":
cfg.TRAIN.DATASETS = ('cs6-train-easy-gt-sub', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-train-easy-hp":
cfg.TRAIN.DATASETS = ('cs6-train-easy-hp', )
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-train-easy-det":
cfg.TRAIN.DATASETS = ('cs6-train-easy-det', )
cfg.MODEL.NUM_CLASSES = 2
# Joint training with CS6 and WIDER
elif args.dataset == "cs6-train-easy-gt-sub+WIDER":
cfg.TRAIN.DATASETS = ('cs6-train-easy-gt-sub', 'wider_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-train-gt+WIDER":
cfg.TRAIN.DATASETS = ('cs6-train-gt', 'wider_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-train-hp+WIDER":
cfg.TRAIN.DATASETS = ('cs6-train-hp', 'wider_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-train-dummy+WIDER":
cfg.TRAIN.DATASETS = ('cs6-train-dummy', 'wider_train')
cfg.MODEL.NUM_CLASSES = 2
elif args.dataset == "cs6-train-det+WIDER":
cfg.TRAIN.DATASETS = ('cs6-train-det', 'wider_train')
cfg.MODEL.NUM_CLASSES = 2
# Dets dataset created by removing tracker results from the HP json
elif args.dataset == "cs6_train_det_from_hp+WIDER":
cfg.TRAIN.DATASETS = ('cs6_train_det_from_hp', 'wider_train')
cfg.MODEL.NUM_CLASSES = 2
# Dataset created by removing det results from the HP json -- HP tracker only
elif args.dataset == "cs6_train_hp_tracker_only+WIDER":
cfg.TRAIN.DATASETS = ('cs6_train_hp_tracker_only', 'wider_train')
cfg.MODEL.NUM_CLASSES = 2
# HP dataset with noisy labels: used to prevent DA from getting any info from HP
elif args.dataset == "cs6_train_hp_noisy_100+WIDER":
cfg.TRAIN.DATASETS = ('cs6_train_hp_noisy_100', 'wider_train')
cfg.MODEL.NUM_CLASSES = 2
else:
raise ValueError("Unexpected args.dataset: {}".format(args.dataset))
cfg_from_file(args.cfg_file)
if args.set_cfgs is not None:
cfg_from_list(args.set_cfgs)
### Adaptively adjust some configs ###
original_batch_size = cfg.NUM_GPUS * cfg.TRAIN.IMS_PER_BATCH
original_ims_per_batch = cfg.TRAIN.IMS_PER_BATCH
original_num_gpus = cfg.NUM_GPUS
if args.batch_size is None:
args.batch_size = original_batch_size
cfg.NUM_GPUS = torch.cuda.device_count()
assert (args.batch_size % cfg.NUM_GPUS) == 0, \
'batch_size: %d, NUM_GPUS: %d' % (args.batch_size, cfg.NUM_GPUS)
cfg.TRAIN.IMS_PER_BATCH = args.batch_size // cfg.NUM_GPUS
effective_batch_size = args.iter_size * args.batch_size
print('effective_batch_size = batch_size * iter_size = %d * %d' %
(args.batch_size, args.iter_size))
print('Adaptive config changes:')
print(' effective_batch_size: %d --> %d' %
(original_batch_size, effective_batch_size))
print(' NUM_GPUS: %d --> %d' %
(original_num_gpus, cfg.NUM_GPUS))
print(' IMS_PER_BATCH: %d --> %d' %
(original_ims_per_batch, cfg.TRAIN.IMS_PER_BATCH))
### Adjust learning based on batch size change linearly
# For iter_size > 1, gradients are `accumulated`, so lr is scaled based
# on batch_size instead of effective_batch_size
old_base_lr = cfg.SOLVER.BASE_LR
cfg.SOLVER.BASE_LR *= args.batch_size / original_batch_size
print('Adjust BASE_LR linearly according to batch_size change:\n'
' BASE_LR: {} --> {}'.format(old_base_lr, cfg.SOLVER.BASE_LR))
### Adjust solver steps
step_scale = original_batch_size / effective_batch_size
old_solver_steps = cfg.SOLVER.STEPS
old_max_iter = cfg.SOLVER.MAX_ITER
cfg.SOLVER.STEPS = list(
map(lambda x: int(x * step_scale + 0.5), cfg.SOLVER.STEPS))
cfg.SOLVER.MAX_ITER = int(cfg.SOLVER.MAX_ITER * step_scale + 0.5)
print(
'Adjust SOLVER.STEPS and SOLVER.MAX_ITER linearly based on effective_batch_size change:\n'
' SOLVER.STEPS: {} --> {}\n'
' SOLVER.MAX_ITER: {} --> {}'.format(old_solver_steps,
cfg.SOLVER.STEPS, old_max_iter,
cfg.SOLVER.MAX_ITER))
# Scale FPN rpn_proposals collect size (post_nms_topN) in `collect` function
# of `collect_and_distribute_fpn_rpn_proposals.py`
#
# post_nms_topN = int(cfg[cfg_key].RPN_POST_NMS_TOP_N * cfg.FPN.RPN_COLLECT_SCALE + 0.5)
if cfg.FPN.FPN_ON and cfg.MODEL.FASTER_RCNN:
cfg.FPN.RPN_COLLECT_SCALE = cfg.TRAIN.IMS_PER_BATCH / original_ims_per_batch
print(
'Scale FPN rpn_proposals collect size directly propotional to the change of IMS_PER_BATCH:\n'
' cfg.FPN.RPN_COLLECT_SCALE: {}'.format(
cfg.FPN.RPN_COLLECT_SCALE))
if args.num_workers is not None:
cfg.DATA_LOADER.NUM_THREADS = args.num_workers
print('Number of data loading threads: %d' % cfg.DATA_LOADER.NUM_THREADS)
### Overwrite some solver settings from command line arguments
if args.optimizer is not None:
cfg.SOLVER.TYPE = args.optimizer
if args.lr is not None:
cfg.SOLVER.BASE_LR = args.lr
if args.lr_decay_gamma is not None:
cfg.SOLVER.GAMMA = args.lr_decay_gamma
assert_and_infer_cfg()
timers = defaultdict(Timer)
"""def _init_fn(worker_id):
random.seed(999)
np.random.seed(999)
torch.cuda.manual_seed(999)
torch.cuda.manual_seed_all(999)
torch.manual_seed(999)
torch.backends.cudnn.deterministic = True
"""
### Dataset ###
timers['roidb'].tic()
roidb, ratio_list, ratio_index = combined_roidb_for_training(
cfg.TRAIN.DATASETS, cfg.TRAIN.PROPOSAL_FILES)
timers['roidb'].toc()
roidb_size = len(roidb)
logger.info('{:d} roidb entries'.format(roidb_size))
logger.info('Takes %.2f sec(s) to construct roidb',
timers['roidb'].average_time)
if cfg.TRAIN.JOINT_TRAINING:
if len(cfg.TRAIN.DATASETS) == 2:
print('Joint training on two datasets')
else:
raise NotImplementedError
joint_training_roidb = []
for i, dataset_name in enumerate(cfg.TRAIN.DATASETS):
# ROIDB construction
timers['roidb'].tic()
roidb, ratio_list, ratio_index = combined_roidb_for_training(
(dataset_name), cfg.TRAIN.PROPOSAL_FILES)
timers['roidb'].toc()
roidb_size = len(roidb)
logger.info('{:d} roidb entries'.format(roidb_size))
logger.info('Takes %.2f sec(s) to construct roidb',
timers['roidb'].average_time)
if i == 0:
roidb_size = len(roidb)
batchSampler = BatchSampler(sampler=MinibatchSampler(
ratio_list, ratio_index),
batch_size=args.batch_size,
drop_last=True)
dataset = RoiDataLoader(roidb,
cfg.MODEL.NUM_CLASSES,
training=True)
dataloader = torch.utils.data.DataLoader(
dataset,
batch_sampler=batchSampler,
num_workers=cfg.DATA_LOADER.NUM_THREADS,
collate_fn=collate_minibatch)
#worker_init_fn=_init_fn)
# decrease num-threads when using two dataloaders
dataiterator = iter(dataloader)
joint_training_roidb.append({
'dataloader': dataloader,
'dataiterator': dataiterator,
'dataset_name': dataset_name
})
else:
roidb, ratio_list, ratio_index = combined_roidb_for_training(
cfg.TRAIN.DATASETS, cfg.TRAIN.PROPOSAL_FILES)
timers['roidb'].toc()
roidb_size = len(roidb)
logger.info('{:d} roidb entries'.format(roidb_size))
logger.info('Takes %.2f sec(s) to construct roidb',
timers['roidb'].average_time)
batchSampler = BatchSampler(sampler=MinibatchSampler(
ratio_list, ratio_index),
batch_size=args.batch_size,
drop_last=True)
dataset = RoiDataLoader(roidb, cfg.MODEL.NUM_CLASSES, training=True)
dataloader = torch.utils.data.DataLoader(
dataset,
batch_sampler=batchSampler,
num_workers=cfg.DATA_LOADER.NUM_THREADS,
collate_fn=collate_minibatch)
#worker_init_fn=init_fn)
dataiterator = iter(dataloader)
# Effective training sample size for one epoch
train_size = roidb_size // args.batch_size * args.batch_size
if cfg.TRAIN.JOINT_SELECTIVE_FG:
orig_fg_batch_ratio = cfg.TRAIN.FG_FRACTION
### Model ###
maskRCNN = Generalized_RCNN()
if cfg.CUDA:
maskRCNN.cuda()
### Optimizer ###
gn_param_nameset = set()
for name, module in maskRCNN.named_modules():
if isinstance(module, nn.GroupNorm):
gn_param_nameset.add(name + '.weight')
gn_param_nameset.add(name + '.bias')
gn_params = []
gn_param_names = []
bias_params = []
bias_param_names = []
nonbias_params = []
nonbias_param_names = []
nograd_param_names = []
for key, value in dict(maskRCNN.named_parameters()).items():
if value.requires_grad:
if 'bias' in key:
bias_params.append(value)
bias_param_names.append(key)
elif key in gn_param_nameset:
gn_params.append(value)
gn_param_names.append(key)
else:
nonbias_params.append(value)
nonbias_param_names.append(key)
else:
nograd_param_names.append(key)
assert (gn_param_nameset - set(nograd_param_names) -
set(bias_param_names)) == set(gn_param_names)
# Learning rate of 0 is a dummy value to be set properly at the start of training
params = [{
'params': nonbias_params,
'lr': 0,
'weight_decay': cfg.SOLVER.WEIGHT_DECAY
}, {
'params':
bias_params,
'lr':
0 * (cfg.SOLVER.BIAS_DOUBLE_LR + 1),
'weight_decay':
cfg.SOLVER.WEIGHT_DECAY if cfg.SOLVER.BIAS_WEIGHT_DECAY else 0
}, {
'params': gn_params,
'lr': 0,
'weight_decay': cfg.SOLVER.WEIGHT_DECAY_GN
}]
# names of paramerters for each paramter
param_names = [nonbias_param_names, bias_param_names, gn_param_names]
if cfg.SOLVER.TYPE == "SGD":
optimizer = torch.optim.SGD(params, momentum=cfg.SOLVER.MOMENTUM)
elif cfg.SOLVER.TYPE == "Adam":
optimizer = torch.optim.Adam(params)
### Load checkpoint
if args.load_ckpt:
load_name = args.load_ckpt
logging.info("loading checkpoint %s", load_name)
checkpoint = torch.load(load_name,
map_location=lambda storage, loc: storage)
net_utils.load_ckpt(maskRCNN, checkpoint['model'])
if args.resume:
args.start_step = checkpoint['step'] + 1
if 'train_size' in checkpoint: # For backward compatibility
if checkpoint['train_size'] != train_size:
print(
'train_size value: %d different from the one in checkpoint: %d'
% (train_size, checkpoint['train_size']))
# reorder the params in optimizer checkpoint's params_groups if needed
# misc_utils.ensure_optimizer_ckpt_params_order(param_names, checkpoint)
# There is a bug in optimizer.load_state_dict on Pytorch 0.3.1.
# However it's fixed on master.
# optimizer.load_state_dict(checkpoint['optimizer'])
misc_utils.load_optimizer_state_dict(optimizer,
checkpoint['optimizer'])
del checkpoint
torch.cuda.empty_cache()
if args.load_detectron: #TODO resume for detectron weights (load sgd momentum values)
logging.info("loading Detectron weights %s", args.load_detectron)
load_detectron_weight(maskRCNN, args.load_detectron)
lr = optimizer.param_groups[0][
'lr'] # lr of non-bias parameters, for commmand line outputs.
maskRCNN = mynn.DataParallel(maskRCNN,
cpu_keywords=['im_info', 'roidb'],
minibatch=True)
### Training Setups ###
args.run_name = misc_utils.get_run_name() + '_' + str(
cfg.TRAIN.DATASETS) + '_step'
output_dir = misc_utils.get_output_dir(args, args.run_name)
args.cfg_filename = os.path.basename(args.cfg_file)
if not args.no_save:
if not os.path.exists(output_dir):
os.makedirs(output_dir)
blob = {'cfg': yaml.dump(cfg), 'args': args}
with open(os.path.join(output_dir, 'config_and_args.pkl'), 'wb') as f:
pickle.dump(blob, f, pickle.HIGHEST_PROTOCOL)
if args.use_tfboard:
from tensorboardX import SummaryWriter
# Set the Tensorboard logger
tblogger = SummaryWriter(output_dir)
### Training Loop ###
maskRCNN.train()
CHECKPOINT_PERIOD = int(cfg.TRAIN.SNAPSHOT_ITERS / cfg.NUM_GPUS)
# Set index for decay steps
decay_steps_ind = None
for i in range(1, len(cfg.SOLVER.STEPS)):
if cfg.SOLVER.STEPS[i] >= args.start_step:
decay_steps_ind = i
break
if decay_steps_ind is None:
decay_steps_ind = len(cfg.SOLVER.STEPS)
training_stats = TrainingStats(
args, args.disp_interval,
tblogger if args.use_tfboard and not args.no_save else None)
try:
logger.info('Training starts !')
step = args.start_step
for step in range(args.start_step, cfg.SOLVER.MAX_ITER):
"""
random.seed(cfg.RNG_SEED)
np.random.seed(cfg.RNG_SEED)
torch.cuda.manual_seed(cfg.RNG_SEED)
torch.cuda.manual_seed_all(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
torch.backends.cudnn.deterministic = True
"""
# Warm up
if step < cfg.SOLVER.WARM_UP_ITERS:
method = cfg.SOLVER.WARM_UP_METHOD
if method == 'constant':
warmup_factor = cfg.SOLVER.WARM_UP_FACTOR
elif method == 'linear':
alpha = step / cfg.SOLVER.WARM_UP_ITERS
warmup_factor = cfg.SOLVER.WARM_UP_FACTOR * (1 -
alpha) + alpha
else:
raise KeyError(
'Unknown SOLVER.WARM_UP_METHOD: {}'.format(method))
lr_new = cfg.SOLVER.BASE_LR * warmup_factor
net_utils.update_learning_rate(optimizer, lr, lr_new)
lr = optimizer.param_groups[0]['lr']
assert lr == lr_new
elif step == cfg.SOLVER.WARM_UP_ITERS:
net_utils.update_learning_rate(optimizer, lr,
cfg.SOLVER.BASE_LR)
lr = optimizer.param_groups[0]['lr']
assert lr == cfg.SOLVER.BASE_LR
# Learning rate decay
if decay_steps_ind < len(cfg.SOLVER.STEPS) and \
step == cfg.SOLVER.STEPS[decay_steps_ind]:
logger.info('Decay the learning on step %d', step)
lr_new = lr * cfg.SOLVER.GAMMA
net_utils.update_learning_rate(optimizer, lr, lr_new)
lr = optimizer.param_groups[0]['lr']
assert lr == lr_new
decay_steps_ind += 1
training_stats.IterTic()
optimizer.zero_grad()
for inner_iter in range(args.iter_size):
# use a iter counter for optional alternating batches
if args.iter_size == 1:
iter_counter = step
else:
iter_counter = inner_iter
if cfg.TRAIN.JOINT_TRAINING:
# alternate batches between dataset[0] and dataset[1]
if iter_counter % 2 == 0:
if True: #DEBUG:
print('Dataset: %s' %
joint_training_roidb[0]['dataset_name'])
dataloader = joint_training_roidb[0]['dataloader']
dataiterator = joint_training_roidb[0]['dataiterator']
# NOTE: if available FG samples cannot fill minibatch
# then batchsize will be smaller than cfg.TRAIN.BATCH_SIZE_PER_IM.
else:
if True: #DEBUG:
print('Dataset: %s' %
joint_training_roidb[1]['dataset_name'])
dataloader = joint_training_roidb[1]['dataloader']
dataiterator = joint_training_roidb[1]['dataiterator']
try:
input_data = next(dataiterator)
except StopIteration:
# end of epoch for dataloader
dataiterator = iter(dataloader)
input_data = next(dataiterator)
if cfg.TRAIN.JOINT_TRAINING:
if iter_counter % 2 == 0:
joint_training_roidb[0][
'dataiterator'] = dataiterator
else:
joint_training_roidb[1][
'dataiterator'] = dataiterator
for key in input_data:
if key != 'roidb': # roidb is a list of ndarrays with inconsistent length
input_data[key] = list(map(Variable, input_data[key]))
net_outputs = maskRCNN(**input_data)
training_stats.UpdateIterStats(net_outputs, inner_iter)
loss = net_outputs['total_loss']
# [p.data.get_device() for p in maskRCNN.parameters()]
# [(name, p.data.get_device()) for name, p in maskRCNN.named_parameters()]
loss.backward()
optimizer.step()
training_stats.IterToc()
training_stats.LogIterStats(step, lr)
if (step + 1) % CHECKPOINT_PERIOD == 0:
save_ckpt(output_dir, args, step, train_size, maskRCNN,
optimizer)
# ---- Training ends ----
# Save last checkpoint
save_ckpt(output_dir, args, step, train_size, maskRCNN, optimizer)
except (RuntimeError, KeyboardInterrupt):
del dataiterator
logger.info('Save ckpt on exception ...')
save_ckpt(output_dir, args, step, train_size, maskRCNN, optimizer)
logger.info('Save ckpt done.')
stack_trace = traceback.format_exc()
print(stack_trace)
finally:
if args.use_tfboard and not args.no_save:
tblogger.close()
def main():
"""Main function"""
args = parse_args()
print('Called with args:')
print(args)
if not torch.cuda.is_available():
sys.exit("Need a CUDA device to run the code.")
if args.cuda or cfg.NUM_GPUS > 0:
cfg.CUDA = True
else:
raise ValueError("Need Cuda device to run !")
if args.dataset == "davis2017":
cfg.TRAIN.DATASETS = ('davis_train', )
#For davis, coco category is used.
cfg.MODEL.NUM_CLASSES = 81 #80 foreground + 1 background
else:
raise ValueError("Unexpected args.dataset: {}".format(args.dataset))
cfg_from_file(args.cfg_file)
if args.set_cfgs is not None:
cfg_from_list(args.set_cfgs)
if cfg.MODEL.IDENTITY_TRAINING and cfg.MODEL.IDENTITY_REPLACE_CLASS:
cfg.MODEL.NUM_CLASSES = 145
cfg.MODEL.IDENTITY_TRAINING = False
cfg.MODEL.ADD_UNKNOWN_CLASS = False
#Add unknow class type if necessary.
if cfg.MODEL.ADD_UNKNOWN_CLASS is True:
cfg.MODEL.NUM_CLASSES += 1
### Adaptively adjust some configs ###
original_batch_size = cfg.NUM_GPUS * cfg.TRAIN.IMS_PER_BATCH
original_ims_per_batch = cfg.TRAIN.IMS_PER_BATCH
original_num_gpus = cfg.NUM_GPUS
if args.batch_size is None:
args.batch_size = original_batch_size
cfg.NUM_GPUS = torch.cuda.device_count()
assert (args.batch_size % cfg.NUM_GPUS) == 0, \
'batch_size: %d, NUM_GPUS: %d' % (args.batch_size, cfg.NUM_GPUS)
cfg.TRAIN.IMS_PER_BATCH = args.batch_size // cfg.NUM_GPUS
effective_batch_size = args.iter_size * args.batch_size
print('effective_batch_size = batch_size * iter_size = %d * %d' %
(args.batch_size, args.iter_size))
print('Adaptive config changes:')
print(' effective_batch_size: %d --> %d' %
(original_batch_size, effective_batch_size))
print(' NUM_GPUS: %d --> %d' %
(original_num_gpus, cfg.NUM_GPUS))
print(' IMS_PER_BATCH: %d --> %d' %
(original_ims_per_batch, cfg.TRAIN.IMS_PER_BATCH))
### Adjust learning based on batch size change linearly
# For iter_size > 1, gradients are `accumulated`, so lr is scaled based
# on batch_size instead of effective_batch_size
old_base_lr = cfg.SOLVER.BASE_LR
cfg.SOLVER.BASE_LR *= args.batch_size / original_batch_size
cfg.SOLVER.BASE_LR *= 1.0 / cfg.MODEL.SEQUENCE_LENGTH
print(
'Adjust BASE_LR linearly according to batch_size change and sequence length change:\n'
' BASE_LR: {} --> {}'.format(old_base_lr, cfg.SOLVER.BASE_LR))
### Adjust solver steps
step_scale = original_batch_size / effective_batch_size
old_solver_steps = cfg.SOLVER.STEPS
old_max_iter = cfg.SOLVER.MAX_ITER
cfg.SOLVER.STEPS = list(
map(lambda x: int(x * step_scale + 0.5), cfg.SOLVER.STEPS))
cfg.SOLVER.MAX_ITER = int(cfg.SOLVER.MAX_ITER * step_scale + 0.5)
print(
'Adjust SOLVER.STEPS and SOLVER.MAX_ITER linearly based on effective_batch_size change:\n'
' SOLVER.STEPS: {} --> {}\n'
' SOLVER.MAX_ITER: {} --> {}'.format(old_solver_steps,
cfg.SOLVER.STEPS, old_max_iter,
cfg.SOLVER.MAX_ITER))
# Scale FPN rpn_proposals collect size (post_nms_topN) in `collect` function
# of `collect_and_distribute_fpn_rpn_proposals.py`
#
# post_nms_topN = int(cfg[cfg_key].RPN_POST_NMS_TOP_N * cfg.FPN.RPN_COLLECT_SCALE + 0.5)
if cfg.FPN.FPN_ON and cfg.MODEL.FASTER_RCNN:
cfg.FPN.RPN_COLLECT_SCALE = cfg.TRAIN.IMS_PER_BATCH / original_ims_per_batch
print(
'Scale FPN rpn_proposals collect size directly propotional to the change of IMS_PER_BATCH:\n'
' cfg.FPN.RPN_COLLECT_SCALE: {}'.format(
cfg.FPN.RPN_COLLECT_SCALE))
if args.num_workers is not None:
cfg.DATA_LOADER.NUM_THREADS = args.num_workers
print('Number of data loading threads: %d' % cfg.DATA_LOADER.NUM_THREADS)
### Overwrite some solver settings from command line arguments
if args.optimizer is not None:
cfg.SOLVER.TYPE = args.optimizer
if args.lr is not None:
cfg.SOLVER.BASE_LR = args.lr
if args.lr_decay_gamma is not None:
cfg.SOLVER.GAMMA = args.lr_decay_gamma
assert_and_infer_cfg()
timers = defaultdict(Timer)
### Dataset ###
timers['roidb'].tic()
merged_roidb, seq_num, seq_start_end = sequenced_roidb_for_training(
cfg.TRAIN.DATASETS, cfg.TRAIN.PROPOSAL_FILES, load_inv_db=True)
timers['roidb'].toc()
roidb_size = len(merged_roidb)
logger.info('{:d} roidbs sequences.'.format(roidb_size))
logger.info('Takes %.2f sec(s) to construct roidbs',
timers['roidb'].average_time)
# Effective training sample size for one epoch, number of sequences.
train_size = roidb_size // args.batch_size * args.batch_size
### Model ###
maskRCNN = vos_model_builder.Generalized_VOS_RCNN()
if cfg.CUDA:
maskRCNN.cuda()
### Optimizer ###
gn_param_nameset = set()
for name, module in maskRCNN.named_modules():
if isinstance(module, nn.GroupNorm):
gn_param_nameset.add(name + '.weight')
gn_param_nameset.add(name + '.bias')
gn_params = []
gn_param_names = []
bias_params = []
bias_param_names = []
nonbias_params = []
nonbias_param_names = []
nograd_param_names = []
for key, value in maskRCNN.named_parameters():
if value.requires_grad:
if 'bias' in key:
bias_params.append(value)
bias_param_names.append(key)
elif key in gn_param_nameset:
gn_params.append(value)
gn_param_names.append(key)
else:
nonbias_params.append(value)
nonbias_param_names.append(key)
else:
nograd_param_names.append(key)
assert (gn_param_nameset - set(nograd_param_names) -
set(bias_param_names)) == set(gn_param_names)
# Learning rate of 0 is a dummy value to be set properly at the start of training
params = [{
'params': nonbias_params,
'lr': 0,
'weight_decay': cfg.SOLVER.WEIGHT_DECAY
}, {
'params':
bias_params,
'lr':
0 * (cfg.SOLVER.BIAS_DOUBLE_LR + 1),
'weight_decay':
cfg.SOLVER.WEIGHT_DECAY if cfg.SOLVER.BIAS_WEIGHT_DECAY else 0
}, {
'params': gn_params,
'lr': 0,
'weight_decay': cfg.SOLVER.WEIGHT_DECAY_GN
}]
# names of paramerters for each paramter
param_names = [nonbias_param_names, bias_param_names, gn_param_names]
if cfg.SOLVER.TYPE == "SGD":
optimizer = torch.optim.SGD(params, momentum=cfg.SOLVER.MOMENTUM)
elif cfg.SOLVER.TYPE == "Adam":
optimizer = torch.optim.Adam(params)
### Load checkpoint
if args.load_ckpt:
load_name = args.load_ckpt
logging.info("loading checkpoint %s", load_name)
checkpoint = torch.load(load_name,
map_location=lambda storage, loc: storage)
net_utils.load_ckpt_no_mapping(maskRCNN, checkpoint['model'])
if args.resume:
args.start_step = checkpoint['step'] + 1
if 'train_size' in checkpoint: # For backward compatibility
if checkpoint['train_size'] != train_size:
print(
'train_size value: %d different from the one in checkpoint: %d'
% (train_size, checkpoint['train_size']))
# reorder the params in optimizer checkpoint's params_groups if needed
# misc_utils.ensure_optimizer_ckpt_params_order(param_names, checkpoint)
# There is a bug in optimizer.load_state_dict on Pytorch 0.3.1.
# However it's fixed on master.
optimizer.load_state_dict(checkpoint['optimizer'])
# misc_utils.load_optimizer_state_dict(optimizer, checkpoint['optimizer'])
del checkpoint
torch.cuda.empty_cache()
if args.load_detectron: #TODO resume for detectron weights (load sgd momentum values)
logging.info("loading Detectron weights %s", args.load_detectron)
load_detectron_weight(maskRCNN,
args.load_detectron,
force_load_all=False)
lr = optimizer.param_groups[0][
'lr'] # lr of non-bias parameters, for commmand line outputs.
maskRCNN = mynn.DataParallel(maskRCNN,
cpu_keywords=['im_info', 'roidb'],
minibatch=True)
### Training Setups ###
args.run_name = misc_utils.get_run_name() + '_step'
output_dir = misc_utils.get_output_dir(args, args.run_name)
args.cfg_filename = os.path.basename(args.cfg_file)
if not args.no_save:
if not os.path.exists(output_dir):
os.makedirs(output_dir)
blob = {'cfg': yaml.dump(cfg), 'args': args}
with open(os.path.join(output_dir, 'config_and_args.pkl'), 'wb') as f:
pickle.dump(blob, f, pickle.HIGHEST_PROTOCOL)
if args.use_tfboard:
from tensorboardX import SummaryWriter
# Set the Tensorboard logger
tblogger = SummaryWriter(output_dir)
### Training Loop ###
maskRCNN.train()
CHECKPOINT_PERIOD = int(cfg.TRAIN.SNAPSHOT_ITERS / cfg.NUM_GPUS)
# Set index for decay steps
decay_steps_ind = None
for i in range(1, len(cfg.SOLVER.STEPS)):
if cfg.SOLVER.STEPS[i] >= args.start_step:
decay_steps_ind = i
break
if decay_steps_ind is None:
decay_steps_ind = len(cfg.SOLVER.STEPS)
training_stats = TrainingStats(
args, args.disp_interval,
tblogger if args.use_tfboard and not args.no_save else None)
dataloader, dataiterator, warmup_length = gen_sequence_data_sampler(
merged_roidb, seq_num, seq_start_end, use_seq_warmup=False)
try:
logger.info('Training starts !')
step = args.start_step
for step in range(args.start_step, cfg.SOLVER.MAX_ITER):
# Warm up
if step < cfg.SOLVER.WARM_UP_ITERS:
method = cfg.SOLVER.WARM_UP_METHOD
if method == 'constant':
warmup_factor = cfg.SOLVER.WARM_UP_FACTOR
elif method == 'linear':
alpha = step / cfg.SOLVER.WARM_UP_ITERS
warmup_factor = cfg.SOLVER.WARM_UP_FACTOR * (1 -
alpha) + alpha
else:
raise KeyError(
'Unknown SOLVER.WARM_UP_METHOD: {}'.format(method))
lr_new = cfg.SOLVER.BASE_LR * warmup_factor
net_utils.update_learning_rate(optimizer, lr, lr_new)
lr = optimizer.param_groups[0]['lr']
assert lr == lr_new
elif step == cfg.SOLVER.WARM_UP_ITERS:
net_utils.update_learning_rate(optimizer, lr,
cfg.SOLVER.BASE_LR)
lr = optimizer.param_groups[0]['lr']
assert lr == cfg.SOLVER.BASE_LR
# Learning rate decay
if decay_steps_ind < len(cfg.SOLVER.STEPS) and \
step == cfg.SOLVER.STEPS[decay_steps_ind]:
logger.info('Decay the learning on step %d', step)
lr_new = lr * cfg.SOLVER.GAMMA
net_utils.update_learning_rate(optimizer, lr, lr_new)
lr = optimizer.param_groups[0]['lr']
assert lr == lr_new
decay_steps_ind += 1
training_stats.IterTic()
optimizer.zero_grad()
if cfg.TRAIN.USE_SEQ_WARMUP is True and step >= cfg.TRAIN.ITER_BEFORE_USE_SEQ_WARMUP and (
(step - cfg.TRAIN.ITER_BEFORE_USE_SEQ_WARMUP) %
cfg.TRAIN.WARMUP_LENGTH_CHANGE_STEP) == 0:
#TODO better dataiterator creator.
raise NotImplementedError('not able to delete.')
dataloader, dataiterator, warmup_length = gen_sequence_data_sampler(
merged_roidb, seq_num, seq_start_end, use_seq_warmup=True)
print('update warmup length:', warmup_length)
try:
input_data_sequence = next(dataiterator)
except StopIteration:
dataiterator = iter(dataloader)
input_data_sequence = next(dataiterator)
# clean hidden states before training.
maskRCNN.module.clean_hidden_states()
maskRCNN.module.clean_flow_features()
assert len(input_data_sequence['data']
) == cfg.MODEL.SEQUENCE_LENGTH + warmup_length, print(
len(input_data_sequence['data']), '!=',
cfg.MODEL.SEQUENCE_LENGTH + warmup_length)
# if train_part == 0: train backbone.
# else train_part == 1: train heads.
train_part = 1
if cfg.TRAIN.ALTERNATE_TRAINING:
if step % 2 == 0:
maskRCNN.module.freeze_conv_body_only()
train_part = 1
else:
maskRCNN.module.train_conv_body_only()
train_part = 0
# this is used for longer sequence training.
# when reach maximum trainable length, detach the hidden states.
cnter_for_detach_hidden_states = 0
for inner_iter in range(cfg.MODEL.SEQUENCE_LENGTH + warmup_length):
#get input_data
input_data = {}
for key in input_data_sequence.keys():
input_data[key] = input_data_sequence[key][
inner_iter:inner_iter + 1]
for key in input_data:
if key != 'roidb' and key != 'data_flow': # roidb is a list of ndarrays with inconsistent length
input_data[key] = list(map(Variable, input_data[key]))
if key == 'data_flow':
if inner_iter != 0 and input_data[key][0][0][
0] is not None: # flow is not None.
input_data[key] = [
Variable(
torch.tensor(
np.expand_dims(
np.squeeze(
np.array(input_data[key],
dtype=np.float32)),
0),
device=input_data['data'][0].device))
]
assert input_data['data'][0].shape[
-2:] == input_data[key][0].shape[
-2:], "Spatial shape of image and flow are not equal."
else:
input_data[key] = [None]
if cfg.TRAIN.USE_SEQ_WARMUP and inner_iter < warmup_length:
maskRCNN.module.set_stop_after_hidden_states(stop=True)
net_outputs = maskRCNN(**input_data)
assert net_outputs is None
maskRCNN.module.detach_hidden_states()
maskRCNN.module.set_stop_after_hidden_states(stop=False)
continue
net_outputs = maskRCNN(**input_data)
training_stats.UpdateIterStats(net_outputs, inner_iter)
loss = net_outputs['total_loss']
if train_part == 0 or cnter_for_detach_hidden_states >= cfg.TRAIN.MAX_TRAINABLE_SEQ_LENGTH or inner_iter == cfg.MODEL.SEQUENCE_LENGTH + warmup_length - 1:
# if reach the max trainable length or end of the sequence. free the graph and detach the hidden states.
loss.backward()
maskRCNN.module.detach_hidden_states()
cnter_for_detach_hidden_states = 0
else:
loss.backward(retain_graph=True)
cnter_for_detach_hidden_states += 1
#TODO step every time?
if cnter_for_detach_hidden_states == 0:
optimizer.step()
optimizer.zero_grad()
training_stats.IterToc()
training_stats.LogIterStats(step, lr)
if (step + 1) % CHECKPOINT_PERIOD == 0:
save_ckpt(output_dir, args, step, train_size, maskRCNN,
optimizer)
# ---- Training ends ----
# Save last checkpoint
save_ckpt(output_dir, args, step, train_size, maskRCNN, optimizer)
except (RuntimeError):
del dataiterator
logger.info('Save ckpt on exception ...')
save_ckpt(output_dir, args, step, train_size, maskRCNN, optimizer)
logger.info('Save ckpt done.')
stack_trace = traceback.format_exc()
print(stack_trace)
except (KeyboardInterrupt):
del dataiterator
logger.info('Save ckpt on Keyboard Interrupt ...')
save_ckpt(output_dir, args, step, train_size, maskRCNN, optimizer)
logger.info('Save ckpt done.')
stack_trace = traceback.format_exc()
print(stack_trace)
finally:
if args.use_tfboard and not args.no_save:
tblogger.close()
def main():
"""Main function"""
args = parse_args()
print('Called with args:')
print(args)
if not torch.cuda.is_available():
sys.exit("Need a CUDA device to run the code.")
if args.cuda or cfg.NUM_GPUS > 0:
cfg.CUDA = True
else:
raise ValueError("Need Cuda device to run !")
if args.dataset == "coco2017":
cfg.TRAIN.DATASETS = ('coco_2017_train',)
cfg.MODEL.NUM_CLASSES = 81
elif args.dataset == "keypoints_coco2017":
cfg.TRAIN.DATASETS = ('keypoints_coco_2017_train',)
cfg.MODEL.NUM_CLASSES = 2
else:
raise ValueError("Unexpected args.dataset: {}".format(args.dataset))
cfg_from_file(args.cfg_file)
if args.set_cfgs is not None:
cfg_from_list(args.set_cfgs)
### Adaptively adjust some configs ###
original_batch_size = cfg.NUM_GPUS * cfg.TRAIN.IMS_PER_BATCH
original_ims_per_batch = cfg.TRAIN.IMS_PER_BATCH
original_num_gpus = cfg.NUM_GPUS
if args.batch_size is None:
args.batch_size = original_batch_size
cfg.NUM_GPUS = torch.cuda.device_count()
assert (args.batch_size % cfg.NUM_GPUS) == 0, \
'batch_size: %d, NUM_GPUS: %d' % (args.batch_size, cfg.NUM_GPUS)
cfg.TRAIN.IMS_PER_BATCH = args.batch_size // cfg.NUM_GPUS
effective_batch_size = args.iter_size * args.batch_size
print('effective_batch_size = batch_size * iter_size = %d * %d' % (args.batch_size, args.iter_size))
print('Adaptive config changes:')
print(' effective_batch_size: %d --> %d' % (original_batch_size, effective_batch_size))
print(' NUM_GPUS: %d --> %d' % (original_num_gpus, cfg.NUM_GPUS))
print(' IMS_PER_BATCH: %d --> %d' % (original_ims_per_batch, cfg.TRAIN.IMS_PER_BATCH))
### Adjust learning based on batch size change linearly
# For iter_size > 1, gradients are `accumulated`, so lr is scaled based
# on batch_size instead of effective_batch_size
old_base_lr = cfg.SOLVER.BASE_LR
cfg.SOLVER.BASE_LR *= args.batch_size / original_batch_size
print('Adjust BASE_LR linearly according to batch_size change:\n'
' BASE_LR: {} --> {}'.format(old_base_lr, cfg.SOLVER.BASE_LR))
### Adjust solver steps
step_scale = original_batch_size / effective_batch_size
old_solver_steps = cfg.SOLVER.STEPS
old_max_iter = cfg.SOLVER.MAX_ITER
cfg.SOLVER.STEPS = list(map(lambda x: int(x * step_scale + 0.5), cfg.SOLVER.STEPS))
cfg.SOLVER.MAX_ITER = int(cfg.SOLVER.MAX_ITER * step_scale + 0.5)
print('Adjust SOLVER.STEPS and SOLVER.MAX_ITER linearly based on effective_batch_size change:\n'
' SOLVER.STEPS: {} --> {}\n'
' SOLVER.MAX_ITER: {} --> {}'.format(old_solver_steps, cfg.SOLVER.STEPS,
old_max_iter, cfg.SOLVER.MAX_ITER))
# Scale FPN rpn_proposals collect size (post_nms_topN) in `collect` function
# of `collect_and_distribute_fpn_rpn_proposals.py`
#
# post_nms_topN = int(cfg[cfg_key].RPN_POST_NMS_TOP_N * cfg.FPN.RPN_COLLECT_SCALE + 0.5)
if cfg.FPN.FPN_ON and cfg.MODEL.FASTER_RCNN:
cfg.FPN.RPN_COLLECT_SCALE = cfg.TRAIN.IMS_PER_BATCH / original_ims_per_batch
print('Scale FPN rpn_proposals collect size directly propotional to the change of IMS_PER_BATCH:\n'
' cfg.FPN.RPN_COLLECT_SCALE: {}'.format(cfg.FPN.RPN_COLLECT_SCALE))
if args.num_workers is not None:
cfg.DATA_LOADER.NUM_THREADS = args.num_workers
print('Number of data loading threads: %d' % cfg.DATA_LOADER.NUM_THREADS)
### Overwrite some solver settings from command line arguments
if args.optimizer is not None:
cfg.SOLVER.TYPE = args.optimizer
if args.lr is not None:
cfg.SOLVER.BASE_LR = args.lr
if args.lr_decay_gamma is not None:
cfg.SOLVER.GAMMA = args.lr_decay_gamma
assert_and_infer_cfg()
timers = defaultdict(Timer)
### Dataset ###
timers['roidb'].tic()
roidb, ratio_list, ratio_index = combined_roidb_for_training(
cfg.TRAIN.DATASETS, cfg.TRAIN.PROPOSAL_FILES)
timers['roidb'].toc()
roidb_size = len(roidb)
logger.info('{:d} roidb entries'.format(roidb_size))
logger.info('Takes %.2f sec(s) to construct roidb', timers['roidb'].average_time)
# Effective training sample size for one epoch
train_size = roidb_size // args.batch_size * args.batch_size
batchSampler = BatchSampler(
sampler=MinibatchSampler(ratio_list, ratio_index),
batch_size=args.batch_size,
drop_last=True
)
dataset = RoiDataLoader(
roidb,
cfg.MODEL.NUM_CLASSES,
training=True)
dataloader = torch.utils.data.DataLoader(
dataset,
batch_sampler=batchSampler,
num_workers=cfg.DATA_LOADER.NUM_THREADS,
collate_fn=collate_minibatch)
dataiterator = iter(dataloader)
### Model ###
maskRCNN = Generalized_RCNN()
if cfg.CUDA:
maskRCNN.cuda()
### Optimizer ###
gn_param_nameset = set()
for name, module in maskRCNN.named_modules():
if isinstance(module, nn.GroupNorm):
gn_param_nameset.add(name+'.weight')
gn_param_nameset.add(name+'.bias')
gn_params = []
gn_param_names = []
bias_params = []
bias_param_names = []
nonbias_params = []
nonbias_param_names = []
nograd_param_names = []
for key, value in dict(maskRCNN.named_parameters()).items():
if value.requires_grad:
if 'bias' in key:
bias_params.append(value)
bias_param_names.append(key)
elif key in gn_param_nameset:
gn_params.append(value)
gn_param_names.append(key)
else:
nonbias_params.append(value)
nonbias_param_names.append(key)
else:
nograd_param_names.append(key)
assert (gn_param_nameset - set(nograd_param_names) - set(bias_param_names)) == set(gn_param_names)
# Learning rate of 0 is a dummy value to be set properly at the start of training
params = [
{'params': nonbias_params,
'lr': 0,
'weight_decay': cfg.SOLVER.WEIGHT_DECAY},
{'params': bias_params,
'lr': 0 * (cfg.SOLVER.BIAS_DOUBLE_LR + 1),
'weight_decay': cfg.SOLVER.WEIGHT_DECAY if cfg.SOLVER.BIAS_WEIGHT_DECAY else 0},
{'params': gn_params,
'lr': 0,
'weight_decay': cfg.SOLVER.WEIGHT_DECAY_GN}
]
# names of paramerters for each paramter
param_names = [nonbias_param_names, bias_param_names, gn_param_names]
if cfg.SOLVER.TYPE == "SGD":
optimizer = torch.optim.SGD(params, momentum=cfg.SOLVER.MOMENTUM)
elif cfg.SOLVER.TYPE == "Adam":
optimizer = torch.optim.Adam(params)
### Load checkpoint
if args.load_ckpt:
load_name = args.load_ckpt
logging.info("loading checkpoint %s", load_name)
checkpoint = torch.load(load_name, map_location=lambda storage, loc: storage)
net_utils.load_ckpt(maskRCNN, checkpoint['model'])
if args.resume:
args.start_step = checkpoint['step'] + 1
if 'train_size' in checkpoint: # For backward compatibility
if checkpoint['train_size'] != train_size:
print('train_size value: %d different from the one in checkpoint: %d'
% (train_size, checkpoint['train_size']))
# reorder the params in optimizer checkpoint's params_groups if needed
# misc_utils.ensure_optimizer_ckpt_params_order(param_names, checkpoint)
# There is a bug in optimizer.load_state_dict on Pytorch 0.3.1.
# However it's fixed on master.
# optimizer.load_state_dict(checkpoint['optimizer'])
misc_utils.load_optimizer_state_dict(optimizer, checkpoint['optimizer'])
del checkpoint
torch.cuda.empty_cache()
if args.load_detectron: #TODO resume for detectron weights (load sgd momentum values)
logging.info("loading Detectron weights %s", args.load_detectron)
load_detectron_weight(maskRCNN, args.load_detectron)
lr = optimizer.param_groups[0]['lr'] # lr of non-bias parameters, for commmand line outputs.
maskRCNN = mynn.DataParallel(maskRCNN, cpu_keywords=['im_info', 'roidb'],
minibatch=True)
### Training Setups ###
args.run_name = misc_utils.get_run_name() + '_step'
output_dir = misc_utils.get_output_dir(args, args.run_name)
args.cfg_filename = os.path.basename(args.cfg_file)
if not args.no_save:
if not os.path.exists(output_dir):
os.makedirs(output_dir)
blob = {'cfg': yaml.dump(cfg), 'args': args}
with open(os.path.join(output_dir, 'config_and_args.pkl'), 'wb') as f:
pickle.dump(blob, f, pickle.HIGHEST_PROTOCOL)
if args.use_tfboard:
from tensorboardX import SummaryWriter
# Set the Tensorboard logger
tblogger = SummaryWriter(output_dir)
### Training Loop ###
maskRCNN.train()
CHECKPOINT_PERIOD = int(cfg.TRAIN.SNAPSHOT_ITERS / cfg.NUM_GPUS)
# Set index for decay steps
decay_steps_ind = None
for i in range(1, len(cfg.SOLVER.STEPS)):
if cfg.SOLVER.STEPS[i] >= args.start_step:
decay_steps_ind = i
break
if decay_steps_ind is None:
decay_steps_ind = len(cfg.SOLVER.STEPS)
training_stats = TrainingStats(
args,
args.disp_interval,
tblogger if args.use_tfboard and not args.no_save else None)
try:
logger.info('Training starts !')
step = args.start_step
for step in range(args.start_step, cfg.SOLVER.MAX_ITER):
# Warm up
if step < cfg.SOLVER.WARM_UP_ITERS:
method = cfg.SOLVER.WARM_UP_METHOD
if method == 'constant':
warmup_factor = cfg.SOLVER.WARM_UP_FACTOR
elif method == 'linear':
alpha = step / cfg.SOLVER.WARM_UP_ITERS
warmup_factor = cfg.SOLVER.WARM_UP_FACTOR * (1 - alpha) + alpha
else:
raise KeyError('Unknown SOLVER.WARM_UP_METHOD: {}'.format(method))
lr_new = cfg.SOLVER.BASE_LR * warmup_factor
net_utils.update_learning_rate(optimizer, lr, lr_new)
lr = optimizer.param_groups[0]['lr']
assert lr == lr_new
elif step == cfg.SOLVER.WARM_UP_ITERS:
net_utils.update_learning_rate(optimizer, lr, cfg.SOLVER.BASE_LR)
lr = optimizer.param_groups[0]['lr']
assert lr == cfg.SOLVER.BASE_LR
# Learning rate decay
if decay_steps_ind < len(cfg.SOLVER.STEPS) and \
step == cfg.SOLVER.STEPS[decay_steps_ind]:
logger.info('Decay the learning on step %d', step)
lr_new = lr * cfg.SOLVER.GAMMA
net_utils.update_learning_rate(optimizer, lr, lr_new)
lr = optimizer.param_groups[0]['lr']
assert lr == lr_new
decay_steps_ind += 1
training_stats.IterTic()
optimizer.zero_grad()
for inner_iter in range(args.iter_size):
try:
input_data = next(dataiterator)
except StopIteration:
dataiterator = iter(dataloader)
input_data = next(dataiterator)
for key in input_data:
if key != 'roidb': # roidb is a list of ndarrays with inconsistent length
input_data[key] = list(map(Variable, input_data[key]))
net_outputs = maskRCNN(**input_data)
training_stats.UpdateIterStats(net_outputs, inner_iter)
loss = net_outputs['total_loss']
loss.backward()
optimizer.step()
training_stats.IterToc()
training_stats.LogIterStats(step, lr)
if (step+1) % CHECKPOINT_PERIOD == 0:
save_ckpt(output_dir, args, step, train_size, maskRCNN, optimizer)
# ---- Training ends ----
# Save last checkpoint
save_ckpt(output_dir, args, step, train_size, maskRCNN, optimizer)
except (RuntimeError, KeyboardInterrupt):
del dataiterator
logger.info('Save ckpt on exception ...')
save_ckpt(output_dir, args, step, train_size, maskRCNN, optimizer)
logger.info('Save ckpt done.')
stack_trace = traceback.format_exc()
print(stack_trace)
finally:
if args.use_tfboard and not args.no_save:
tblogger.close()
def train_val(model,
args,
optimizer,
lr,
dataloader,
train_size,
output_dir,
tblogger=None):
dataiterator = iter(dataloader)
model.train()
CHECKPOINT_PERIOD = cfg.TRAIN.SNAPSHOT_ITERS
# Set index for decay steps
decay_steps_ind = None
for i in range(1, len(cfg.SOLVER.STEPS)):
if cfg.SOLVER.STEPS[i] >= args.start_step:
decay_steps_ind = i
break
if decay_steps_ind is None:
decay_steps_ind = len(cfg.SOLVER.STEPS)
training_stats = TrainingStats(
args, args.disp_interval,
tblogger if args.use_tfboard and not args.no_save else None)
try:
logger.info('Training starts !')
step = args.start_step
best_ap = 0
best_step = 0
running_tr_loss = 0.
DRAW_STEP = args.start_step
for step in range(args.start_step, cfg.SOLVER.MAX_ITER):
# print('stepppp: ', step)
# print(cfg.SOLVER.WARM_UP_ITERS)
# Warm up
if step < cfg.SOLVER.WARM_UP_ITERS:
method = cfg.SOLVER.WARM_UP_METHOD
if method == 'constant':
warmup_factor = cfg.SOLVER.WARM_UP_FACTOR
elif method == 'linear':
alpha = step / cfg.SOLVER.WARM_UP_ITERS
warmup_factor = cfg.SOLVER.WARM_UP_FACTOR * (1 -
alpha) + alpha
else:
raise KeyError(
'Unknown SOLVER.WARM_UP_METHOD: {}'.format(method))
lr_new = cfg.SOLVER.BASE_LR * warmup_factor
net_utils.update_learning_rate(optimizer, lr, lr_new)
lr = optimizer.param_groups[0]['lr']
assert lr == lr_new
elif step == cfg.SOLVER.WARM_UP_ITERS:
net_utils.update_learning_rate(optimizer, lr,
cfg.SOLVER.BASE_LR)
lr = optimizer.param_groups[0]['lr']
assert lr == cfg.SOLVER.BASE_LR
# Learning rate decay
if decay_steps_ind < len(cfg.SOLVER.STEPS) and \
step == cfg.SOLVER.STEPS[decay_steps_ind]:
logger.info('Decay the learning on step %d', step)
lr_new = lr * cfg.SOLVER.GAMMA
net_utils.update_learning_rate(optimizer, lr, lr_new)
lr = optimizer.param_groups[0]['lr']
assert lr == lr_new
decay_steps_ind += 1
training_stats.IterTic()
optimizer.zero_grad()
for inner_iter in range(args.iter_size):
try:
input_data = next(dataiterator)
except StopIteration:
dataiterator = iter(dataloader)
input_data = next(dataiterator)
for key in input_data:
if key != 'roidb': # roidb is a list of ndarrays with inconsistent length
input_data[key] = list(map(Variable, input_data[key]))
net_outputs = model(**input_data)
training_stats.UpdateIterStats(net_outputs, inner_iter)
loss = net_outputs['total_loss']
#print('555')
running_tr_loss += loss.item()
if loss.requires_grad:
loss.backward()
optimizer.step()
training_stats.IterToc()
training_stats.LogIterStats(step, lr)
# print('CHECKPOINT_PERIOD: ', CHECKPOINT_PERIOD)
# if (step+1) % 800==0 or step==cfg.SOLVER.MAX_ITER-1:
# print('\tAverage Training Runing loss of step {}: {:.8f}'.format(step+1, running_tr_loss/(step+1-DRAW_STEP)))
# tblogger.add_scalar('Runing_Training_loss', running_tr_loss/(step+1-DRAW_STEP), step+1)
# DRAW_STEP = step+1
# running_tr_loss = 0.
CHECKPOINT_PERIOD = 2000
if ((step + 1) % CHECKPOINT_PERIOD
== 0) or step == cfg.SOLVER.MAX_ITER - 1:
if (step + 1) > 15000:
save_ckpt(output_dir, args, step, train_size, model,
optimizer)
# ---- Training ends ----
# Save last checkpoint
#save_ckpt(output_dir, args, step, train_size, model, optimizer)
except (RuntimeError, KeyboardInterrupt):
del dataiterator
logger.info('Save ckpt on exception ...')
save_ckpt(output_dir, args, step, train_size, model, optimizer)
logger.info('Save ckpt done.')
stack_trace = traceback.format_exc()
print(stack_trace)
finally:
if args.use_tfboard and not args.no_save:
tblogger.close()
def main():
"""Main function"""
args = parse_args()
print('Called with args:')
print(args)
if not torch.cuda.is_available():
sys.exit("Need a CUDA device to run the code.")
if args.cuda or cfg.NUM_GPUS > 0:
cfg.CUDA = True
else:
raise ValueError("Need Cuda device to run !")
merge_cfg_from_file(args.cfg_file)
# Some manual adjustment for the ApolloScape dataset parameters here
cfg.OUTPUT_DIR = args.output_dir
cfg.TRAIN.DATASETS = 'Car3D'
cfg.MODEL.NUM_CLASSES = 8
if cfg.CAR_CLS.SIM_MAT_LOSS:
cfg.MODEL.NUMBER_CARS = 79
else:
# Loss is only cross entropy, hence, we detect only car categories in the training set.
cfg.MODEL.NUMBER_CARS = 34
cfg.TRAIN.MIN_AREA = 196 # 14*14
cfg.TRAIN.USE_FLIPPED = False # Currently I don't know how to handle the flipped case
cfg.TRAIN.IMS_PER_BATCH = 1
cfg.NUM_GPUS = torch.cuda.device_count()
effective_batch_size = cfg.TRAIN.IMS_PER_BATCH * cfg.NUM_GPUS * args.iter_size
### Adaptively adjust some configs ###
original_batch_size = cfg.NUM_GPUS * cfg.TRAIN.IMS_PER_BATCH
original_ims_per_batch = cfg.TRAIN.IMS_PER_BATCH
original_num_gpus = cfg.NUM_GPUS
if args.batch_size is None:
args.batch_size = original_batch_size
assert (args.batch_size %
cfg.NUM_GPUS) == 0, 'batch_size: %d, NUM_GPUS: %d' % (
args.batch_size, cfg.NUM_GPUS)
print('effective_batch_size = batch_size * iter_size = %d * %d' %
(args.batch_size, args.iter_size))
print('Adaptive config changes:')
print(' effective_batch_size: %d --> %d' %
(original_batch_size, effective_batch_size))
print(' NUM_GPUS: %d --> %d' %
(original_num_gpus, cfg.NUM_GPUS))
print(' IMS_PER_BATCH: %d --> %d' %
(original_ims_per_batch, cfg.TRAIN.IMS_PER_BATCH))
### Adjust learning based on batch size change linearly
# For iter_size > 1, gradients are `accumulated`, so lr is scaled based
# on batch_size instead of effective_batch_size
old_base_lr = cfg.SOLVER.BASE_LR
cfg.SOLVER.BASE_LR *= args.batch_size / original_batch_size
print(
'Adjust BASE_LR linearly according to batch_size change:\n BASE_LR: {} --> {}'
.format(old_base_lr, cfg.SOLVER.BASE_LR))
### Adjust solver steps
step_scale = original_batch_size / effective_batch_size
old_solver_steps = cfg.SOLVER.STEPS
old_max_iter = cfg.SOLVER.MAX_ITER
cfg.SOLVER.STEPS = list(
map(lambda x: int(x * step_scale + 0.5), cfg.SOLVER.STEPS))
cfg.SOLVER.MAX_ITER = int(cfg.SOLVER.MAX_ITER * step_scale + 0.5)
print(
'Adjust SOLVER.STEPS and SOLVER.MAX_ITER linearly based on effective_batch_size change:\n'
' SOLVER.STEPS: {} --> {}\n'
' SOLVER.MAX_ITER: {} --> {}'.format(old_solver_steps,
cfg.SOLVER.STEPS, old_max_iter,
cfg.SOLVER.MAX_ITER))
# Scale FPN rpn_proposals collect size (post_nms_topN) in `collect` function
# of `collect_and_distribute_fpn_rpn_proposals.py`
#
# post_nms_topN = int(cfg[cfg_key].RPN_POST_NMS_TOP_N * cfg.FPN.RPN_COLLECT_SCALE + 0.5)
if cfg.FPN.FPN_ON and cfg.MODEL.FASTER_RCNN:
cfg.FPN.RPN_COLLECT_SCALE = cfg.TRAIN.IMS_PER_BATCH / original_ims_per_batch
print(
'Scale FPN rpn_proposals collect size directly propotional to the change of IMS_PER_BATCH:\n'
' cfg.FPN.RPN_COLLECT_SCALE: {}'.format(
cfg.FPN.RPN_COLLECT_SCALE))
if args.num_workers is not None:
cfg.DATA_LOADER.NUM_THREADS = args.num_workers
print('Number of data loading threads: %d' % cfg.DATA_LOADER.NUM_THREADS)
### Overwrite some solver settings from command line arguments
if args.optimizer is not None:
cfg.SOLVER.TYPE = args.optimizer
if args.lr is not None:
cfg.SOLVER.BASE_LR = args.lr
if args.lr_decay_gamma is not None:
cfg.SOLVER.GAMMA = args.lr_decay_gamma
assert_and_infer_cfg()
timers = defaultdict(Timer)
### Dataset ###
timers['roidb'].tic()
if cfg.MODEL.LOSS_3D_2D_ON:
roidb, ratio_list, ratio_index, ds = combined_roidb_for_training(
cfg.TRAIN.DATASETS, args.dataset_dir)
else:
roidb, ratio_list, ratio_index = combined_roidb_for_training(
cfg.TRAIN.DATASETS, args.dataset_dir)
timers['roidb'].toc()
roidb_size = len(roidb)
logger.info('{:d} roidb entries'.format(roidb_size))
logger.info('Takes %.2f sec(s) to construct roidb',
timers['roidb'].average_time)
# Effective training sample size for one epoch
train_size = roidb_size // args.batch_size * args.batch_size
sampler = MinibatchSampler(ratio_list, ratio_index)
dataset = RoiDataLoader(roidb,
cfg.MODEL.NUM_CLASSES,
training=True,
valid_keys=[
'has_visible_keypoints', 'boxes', 'seg_areas',
'gt_classes', 'gt_overlaps',
'box_to_gt_ind_map', 'is_crowd',
'car_cat_classes', 'poses', 'quaternions',
'im_info'
])
dataloader = torch.utils.data.DataLoader(
dataset,
batch_size=args.batch_size,
drop_last=True,
sampler=sampler,
num_workers=cfg.DATA_LOADER.NUM_THREADS,
collate_fn=collate_minibatch)
dataiterator = iter(dataloader)
### Model ###
if cfg.MODEL.LOSS_3D_2D_ON:
maskRCNN = Generalized_RCNN(ds.Car3D)
else:
maskRCNN = Generalized_RCNN()
if cfg.CUDA:
maskRCNN.cuda()
### Optimizer ###
bias_params = []
bias_param_names = []
nonbias_params = []
nonbias_param_names = []
for key, value in dict(maskRCNN.named_parameters()).items():
if value.requires_grad:
if 'bias' in key:
bias_params.append(value)
bias_param_names.append(key)
else:
nonbias_params.append(value)
nonbias_param_names.append(key)
# Learning rate of 0 is a dummy value to be set properly at the start of training
params = [{
'params': nonbias_params,
'lr': 0,
'weight_decay': cfg.SOLVER.WEIGHT_DECAY
}, {
'params':
bias_params,
'lr':
0 * (cfg.SOLVER.BIAS_DOUBLE_LR + 1),
'weight_decay':
cfg.SOLVER.WEIGHT_DECAY if cfg.SOLVER.BIAS_WEIGHT_DECAY else 0
}]
if cfg.SOLVER.TYPE == "SGD":
optimizer = torch.optim.SGD(params, momentum=cfg.SOLVER.MOMENTUM)
elif cfg.SOLVER.TYPE == "Adam":
optimizer = torch.optim.Adam(params)
### Load checkpoint
if args.load_ckpt:
load_name = args.load_ckpt
logging.info("loading checkpoint %s", load_name)
checkpoint = torch.load(load_name,
map_location=lambda storage, loc: storage)
net_utils.load_ckpt(maskRCNN,
checkpoint['model'],
ignore_list=args.ckpt_ignore_head)
if args.resume:
args.start_step = checkpoint['step'] + 1
if 'train_size' in checkpoint: # For backward compatibility
if checkpoint['train_size'] != train_size:
print(
'train_size value: %d different from the one in checkpoint: %d'
% (train_size, checkpoint['train_size']))
# reorder the params in optimizer checkpoint's params_groups if needed
# misc_utils.ensure_optimizer_ckpt_params_order(param_names, checkpoint)
# There is a bug in optimizer.load_state_dict on Pytorch 0.3.1.
# However it's fixed on master.
# optimizer.load_state_dict(checkpoint['optimizer'])
misc_utils.load_optimizer_state_dict(optimizer,
checkpoint['optimizer'])
del checkpoint
torch.cuda.empty_cache()
if args.load_detectron: # TODO resume for detectron weights (load sgd momentum values)
logging.info("loading Detectron weights %s", args.load_detectron)
load_detectron_weight(maskRCNN, args.load_detectron)
lr = optimizer.param_groups[0][
'lr'] # lr of non-bias parameters, for commmand line outputs.
maskRCNN = mynn.DataParallel(maskRCNN,
cpu_keywords=['im_info', 'roidb'],
minibatch=True)
### Training Setups ###
args.run_name = misc_utils.get_run_name() + '_step'
# output_dir = os.path.join('/media/SSD_1TB/zzy/ApolloScape/ECCV2018_apollo/train', args.run_name)
output_dir = misc_utils.get_output_dir(args, args.run_name)
args.cfg_filename = os.path.basename(args.cfg_file)
if not args.no_save:
if not os.path.exists(output_dir):
os.makedirs(output_dir)
blob = {'cfg': yaml.dump(cfg), 'args': args}
with open(os.path.join(output_dir, 'config_and_args.pkl'), 'wb') as f:
pickle.dump(blob, f, pickle.HIGHEST_PROTOCOL)
if args.use_tfboard:
from tensorboardX import SummaryWriter
# Set the Tensorboard logger
tblogger = SummaryWriter(output_dir)
### Training Loop ###
maskRCNN.train()
CHECKPOINT_PERIOD = int(cfg.TRAIN.SNAPSHOT_ITERS / cfg.NUM_GPUS)
# Set index for decay steps
decay_steps_ind = None
for i in range(1, len(cfg.SOLVER.STEPS)):
if cfg.SOLVER.STEPS[i] >= args.start_step:
decay_steps_ind = i
break
if decay_steps_ind is None:
decay_steps_ind = len(cfg.SOLVER.STEPS)
training_stats = TrainingStats(
args, args.disp_interval,
tblogger if args.use_tfboard and not args.no_save else None)
# warmup_factor_trans = 1.0
try:
logger.info('Training starts !')
step = args.start_step
for step in range(args.start_step, cfg.SOLVER.MAX_ITER):
# Warm up
if step < cfg.SOLVER.WARM_UP_ITERS:
method = cfg.SOLVER.WARM_UP_METHOD
if method == 'constant':
warmup_factor = cfg.SOLVER.WARM_UP_FACTOR
elif method == 'linear':
alpha = step / cfg.SOLVER.WARM_UP_ITERS
warmup_factor = cfg.SOLVER.WARM_UP_FACTOR * (1 -
alpha) + alpha
# warmup_factor_trans = cfg.SOLVER.WARM_UP_FACTOR_TRANS * (1 - alpha) + alpha
# warmup_factor_trans *= cfg.TRANS_HEAD.LOSS_BETA
warmup_factor_trans = 1.0
else:
raise KeyError(
'Unknown SOLVER.WARM_UP_METHOD: {}'.format(method))
lr_new = cfg.SOLVER.BASE_LR * warmup_factor
net_utils.update_learning_rate(optimizer, lr, lr_new)
lr = optimizer.param_groups[0]['lr']
assert lr == lr_new
elif step == cfg.SOLVER.WARM_UP_ITERS:
net_utils.update_learning_rate(optimizer, lr,
cfg.SOLVER.BASE_LR)
lr = optimizer.param_groups[0]['lr']
assert lr == cfg.SOLVER.BASE_LR
# Learning rate decay
if decay_steps_ind < len(
cfg.SOLVER.STEPS
) and step == cfg.SOLVER.STEPS[decay_steps_ind]:
logger.info('Decay the learning on step %d', step)
lr_new = lr * cfg.SOLVER.GAMMA
net_utils.update_learning_rate(optimizer, lr, lr_new)
lr = optimizer.param_groups[0]['lr']
assert lr == lr_new
decay_steps_ind += 1
training_stats.IterTic()
optimizer.zero_grad()
for inner_iter in range(args.iter_size):
try:
input_data = next(dataiterator)
except StopIteration:
dataiterator = iter(dataloader)
input_data = next(dataiterator)
for key in input_data:
if key != 'roidb': # roidb is a list of ndarrays with inconsistent length
input_data[key] = list(map(Variable, input_data[key]))
net_outputs = maskRCNN(**input_data)
net_outputs['losses'][
'loss_car_cls'] *= cfg.CAR_CLS.CAR_CLS_LOSS_BETA
net_outputs['losses']['loss_rot'] *= cfg.CAR_CLS.ROT_LOSS_BETA
if cfg.MODEL.TRANS_HEAD_ON:
net_outputs['losses'][
'loss_trans'] *= cfg.TRANS_HEAD.TRANS_LOSS_BETA
training_stats.UpdateIterStats_car_3d(net_outputs)
# start training
# loss_car_cls: 2.233790, loss_rot: 0.296853, loss_trans: ~100
loss = net_outputs['losses']['loss_car_cls'] + net_outputs[
'losses']['loss_rot']
if cfg.MODEL.TRANS_HEAD_ON:
loss += net_outputs['losses']['loss_trans']
if cfg.MODEL.LOSS_3D_2D_ON:
loss += net_outputs['losses']['UV_projection_loss']
if not cfg.TRAIN.FREEZE_CONV_BODY and not cfg.TRAIN.FREEZE_RPN and not cfg.TRAIN.FREEZE_FPN:
loss += net_outputs['total_loss_conv']
loss.backward()
optimizer.step()
training_stats.IterToc()
training_stats.LogIterStats(step, lr, warmup_factor_trans)
if (step + 1) % CHECKPOINT_PERIOD == 0:
save_ckpt(output_dir, args, step, train_size, maskRCNN,
optimizer)
# ---- Training ends ----
# Save last checkpoint
save_ckpt(output_dir, args, step, train_size, maskRCNN, optimizer)
except (RuntimeError, KeyboardInterrupt):
del dataiterator
logger.info('Save ckpt on exception ...')
save_ckpt(output_dir, args, step, train_size, maskRCNN, optimizer)
logger.info('Save ckpt done.')
stack_trace = traceback.format_exc()
print(stack_trace)
finally:
if args.use_tfboard and not args.no_save:
tblogger.close()