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 == "coco2014":
cfg.TRAIN.DATASETS = ('coco_2014_train', )
cfg.MODEL.NUM_CLASSES = 81
elif args.dataset == 'vcoco_trainval':
cfg.TRAIN.DATASETS = ('vcoco_trainval', )
cfg.MODEL.NUM_CLASSES = 81
elif args.dataset == 'vcoco_train':
cfg.TRAIN.DATASETS = ('vcoco_train', )
cfg.MODEL.NUM_CLASSES = 81
elif args.dataset == 'vcoco_val':
cfg.TRAIN.DATASETS = ('vcoco_val', )
cfg.MODEL.NUM_CLASSES = 81
elif args.dataset == 'keypoints_coco2014':
cfg.TRAIN.DATASETS = ('keypoints_coco_2014_train', )
cfg.MODEL.NUM_CLASSES = 2
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)
if args.vcoco_kp_on:
cfg.VCOCO.KEYPOINTS_ON = True
cfg.NETWORK_NAME = args.net_name # network name
print('Network name:', args.net_name)
cfg.MODEL.CONV_BODY = args.conv_body # backbone network name
print('Conv_body name:', args.conv_body)
cfg.TRAIN.FG_THRESH = args.fg_thresh
print('Train fg thresh:', args.fg_thresh)
cfg.RESNETS.FREEZE_AT = args.freeze_at
print('Freeze at: ', args.freeze_at)
cfg.VCOCO.MLP_HEAD_DIM = args.mlp_head_dim
print('MLP head dim: ', args.mlp_head_dim)
cfg.SOLVER.MAX_ITER = args.max_iter
print('MAX iter: ', args.max_iter)
cfg.TRAIN.SNAPSHOT_ITERS = args.snapshot
print('Snapshot Iters: ', args.snapshot)
if args.solver_steps is not None:
cfg.SOLVER.STEPS = args.solver_steps
print('Solver_steps: ', cfg.SOLVER.STEPS)
cfg.VCOCO.TRIPLETS_NUM_PER_IM = args.triplets_num_per_im
print('triplets_num_per_im: ', cfg.VCOCO.TRIPLETS_NUM_PER_IM)
cfg.VCOCO.HEATMAP_KERNEL_SIZE = args.heatmap_kernel_size
print('heatmap_kernel_size: ', cfg.VCOCO.HEATMAP_KERNEL_SIZE)
cfg.VCOCO.PART_CROP_SIZE = args.part_crop_size
print('part_crop_size: ', cfg.VCOCO.PART_CROP_SIZE)
print('use use_kps17 for part Align: ', args.use_kps17)
if args.use_kps17:
cfg.VCOCO.USE_KPS17 = True
else:
cfg.VCOCO.USE_KPS17 = False
print('MULTILEVEL_ROIS: ', cfg.FPN.MULTILEVEL_ROIS)
if args.vcoco_use_spatial:
cfg.VCOCO.USE_SPATIAL = True
if args.vcoco_use_union_feat:
cfg.VCOCO.USE_UNION_FEAT = True
if args.use_precomp_box:
cfg.VCOCO.USE_PRECOMP_BOX = True
cfg.DEBUG_TEST_WITH_GT = True
if args.lr is not None:
cfg.SOLVER.BASE_LR = args.lr
### Adaptively adjust some configs ###
original_batch_size = cfg.NUM_GPUS * cfg.TRAIN.IMS_PER_BATCH # 16
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))
print(' FG_THRESH: ', cfg.TRAIN.FG_THRESH)
### 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)
cfg.SOLVER.VAL_ITER = int(cfg.SOLVER.VAL_ITER * step_scale + 0.5)
cfg.TRAIN.SNAPSHOT_ITERS = int(cfg.TRAIN.SNAPSHOT_ITERS * 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)
# ipdb.set_trace()
### Overwrite some solver settings from command line arguments
if args.optimizer is not None:
cfg.SOLVER.TYPE = args.optimizer
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
# ToDo: shuffle?
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 ###
from modeling.model_builder import Generalized_RCNN
maskRCNN = Generalized_RCNN()
if cfg.CUDA:
maskRCNN.cuda()
### Optimizer ###
bias_hoi_params = []
bias_hoi_param_names = []
bias_faster_params = []
bias_faster_param_names = []
nobias_hoi_params = []
nobias_hoi_param_names = []
nobias_faster_params = []
nobias_faster_param_names = []
# bias_params = []
# bias_param_names = []
# nonbias_params = []
# nonbias_param_names = []
#base_model = torch.load('Outputs/baseline/baseline_512_32_nogt_1o3/ckpt/model_step47999.pth')
nograd_param_names = []
for key, value in maskRCNN.named_parameters():
#if key in base_model['model'].keys():
# value.requires_grad = False
#print('the key xxx:', key)
# Fix RPN module same as the paper
# ToDo: or key.startswith('Box')
# if 'affinity' not in key:
# value.requires_grad = False
print(key, value.size(), value.requires_grad)
if value.requires_grad:
if 'bias' in key:
if 'HOI_Head' in key:
bias_hoi_params.append(value)
bias_hoi_param_names.append(key)
else:
bias_faster_params.append(value)
bias_faster_param_names.append(key)
else:
if 'HOI_Head' in key:
nobias_hoi_params.append(value)
nobias_hoi_param_names.append(key)
else:
nobias_faster_params.append(value)
nobias_faster_param_names.append(key)
else:
nograd_param_names.append(key)
#del base_model
#ipdb.set_trace()
# Learning rate of 0 is a dummy value to be set properly at the start of training
params = [
{
'params': nobias_hoi_params,
'lr': 0,
'weight_decay': cfg.SOLVER.WEIGHT_DECAY
},
{
'params': nobias_faster_params,
'lr': 0 * cfg.SOLVER.FASTER_RCNN_WEIGHT,
'weight_decay': cfg.SOLVER.WEIGHT_DECAY
},
{
'params':
bias_hoi_params,
'lr':
0 * (cfg.SOLVER.BIAS_DOUBLE_LR + 1),
'weight_decay':
cfg.SOLVER.WEIGHT_DECAY if cfg.SOLVER.BIAS_WEIGHT_DECAY else 0
},
{
'params':
bias_faster_params,
'lr':
0 * (cfg.SOLVER.BIAS_DOUBLE_LR + 1) *
cfg.SOLVER.FASTER_RCNN_WEIGHT,
'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)
if args.krcnn_from_faster:
net_utils.load_krcnn_from_faster(maskRCNN, checkpoint['model'])
else:
net_utils.load_ckpt(maskRCNN, checkpoint['model'])
print('Original model loaded....')
if args.resume:
print('Resume, loaded step\n\n\n: ', checkpoint['step'])
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)
output_dir = os.path.join('Outputs', args.expDir, args.expID)
os.makedirs(output_dir, exist_ok=True)
args.cfg_filename = os.path.basename(args.cfg_file)
tblogger = None
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 ###
train_val(maskRCNN, args, optimizer, lr, dataloader, train_size,
output_dir, tblogger)
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
)
# ADD-------------------
from pycocotools.coco import COCO
coco = COCO('data/coco/annotations/instances_train2017.json')
dataset = RoiDataLoader(
roidb,
cfg.MODEL.NUM_CLASSES,
coco,
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)
CHECKPOINT_PERIOD = 5000
# 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 !")
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()
