def _load_ckpt(self):
if self.args.EXPER.resume == "load_voc":
load_tf_weights(self.model, 'vocweights.pkl')
else: # iter or best
ckptfile = torch.load(os.path.join(self.save_path, 'checkpoint-{}.pth'.format(self.args.EXPER.resume)))
self.model.load_state_dict(ckptfile['state_dict'])
#load_checkpoint(self.model,ckptfile)
if not self.args.finetune and not self.args.do_test and not self.args.Prune.do_test:
self.optimizer.load_state_dict(ckptfile['opti_dict'])
self.global_epoch = ckptfile['epoch']
self.global_iter = ckptfile['iter']
self.best_mAP = ckptfile['metric']
print("successfully load checkpoint {}".format(self.args.EXPER.resume))
def _load_ckpt(self):
if self.args.EXPER.resume == "load_voc":
load_tf_weights(self.model, 'vocweights.pkl')
else: # iter or best
ckptfile = torch.load(
os.path.join(
self.save_path,
'checkpoint-{}.pth'.format(self.args.EXPER.resume)))
if next(iter(ckptfile['state_dict'])).startswith(
'module.'): #if checkpoint comes from parallelized model
self.model = nn.DataParallel(self.model)
self.model.load_state_dict(ckptfile['state_dict'])
#load_checkpoint(self.model,ckptfile)
if not self.args.finetune:
self.optimizer.load_state_dict(ckptfile['opti_dict'])
self.global_epoch = ckptfile['epoch']
self.global_iter = ckptfile['iter']
self.best_mAP = ckptfile['metric']
print("successfully load checkpoint {}".format(self.args.EXPER.resume))
def _load_ckpt_quantile_adjusted(self, name):
cfg2.merge_from_file("configs/strongerv3_asff.yaml")
net = eval(cfg2.MODEL.modeltype)(cfg=cfg2.MODEL).cuda()
if self.args.EXPER.resume == "load_voc":
load_tf_weights(self.model, 'vocweights.pkl')
else: # iter or best
ckptfile = torch.load(
os.path.join('./checkpoints/strongerv3_asff/',
'checkpoint-{}.pth'.format(name)))
state_dict = ckptfile['state_dict']
new_state_dict = {}
for k, v in state_dict.items():
name = k #remove 'module.' of DataParallel
#name = "module."+k
new_state_dict[name] = v
net.load_state_dict(new_state_dict)
backbone, headslarge, detlarge, mergelarge, headsmid, detmid, mergemid, headsmall, detsmall = net.get_info(
)
self.model.load_partial_state(backbone, headslarge, detlarge,
mergelarge, headsmid, detmid,
mergemid, headsmall, detsmall)
def _load_ckpt(self):
if self.args.EXPER.resume == "load_voc":
load_tf_weights(self.model, 'vocweights.pkl')
else: # iter or best
ckptfile = torch.load(
os.path.join(
self.save_path,
'checkpoint-{}.pth'.format(self.args.EXPER.resume)))
state_dict = ckptfile['state_dict']
new_state_dict = {}
for k, v in state_dict.items():
#name=k #remove 'module.' of DataParallel
name = "module." + k
new_state_dict[name] = v
self.model.load_state_dict(new_state_dict)
#load_checkpoint(self.model,ckptfile)
if not self.args.finetune and not self.args.do_test and not self.args.Prune.do_test:
self.optimizer.load_state_dict(ckptfile['opti_dict'])
self.global_epoch = ckptfile['epoch']
self.global_iter = ckptfile['iter']
self.best_mAP = ckptfile['metric']
print("successfully load checkpoint {}".format(self.args.EXPER.resume))
def _load_ckpt(self):
if self.args.EXPER.resume == "load_voc":
load_tf_weights(self.model, 'vocweights.pkl')
else: # iter or best
ckptfile = torch.load(os.path.join(self.save_path, 'checkpoint-{}.pth'.format(self.args.EXPER.resume)))
# take care of the distributed model
if 'module.' in list(self.model.state_dict().keys())[0]:
newdict = OrderedDict()
for k, v in ckptfile['state_dict'].items():
if 'module.' not in k:
newdict['module.' + k] = v
else:
newdict[k] = v
ckptfile['state_dict'] = newdict
else:
newdict = OrderedDict()
for k, v in ckptfile['state_dict'].items():
if 'module.' in k:
newdict[k[7:]] = v
else:
newdict[k] = v
ckptfile['state_dict'] = newdict
# check the consistency
for k,v in ckptfile['state_dict'].items():
if k not in self.model.state_dict():
ckptfile['state_dict'].pop(k)
for (k1,v1) in self.model.state_dict().items():
if k1 not in ckptfile['state_dict'] or ckptfile['state_dict'][k1].shape!=self.model.state_dict()[k1].shape:
print("weight {} will be initialized from scratch".format(k1))
ckptfile['state_dict'][k1]=self.model.state_dict()[k1]
# just ignore the bn_not_save parameters
self.model.load_state_dict(ckptfile['state_dict'], strict=True)
# load_checkpoint(self.model,ckptfile)
if not self.args.finetune and not self.args.do_test and not self.args.Prune.do_test:
self.optimizer.load_state_dict(ckptfile['opti_dict'])
self.global_epoch = ckptfile['epoch']
self.global_iter = ckptfile['iter']
self.best_mAP = ckptfile['metric']
print("successfully load checkpoint {}".format(self.args.EXPER.resume))
def _load_ckpt(self):
if self.args.EXPER.resume == "load_voc":
load_tf_weights(self.model, 'vocweights.pkl')
else: # iter or best
ckptfile = torch.load(os.path.join(self.save_path, 'checkpoint-{}.pth'.format(self.args.EXPER.resume)))
# take care of the distributed model
if 'module.' in list(self.model.state_dict().keys())[0]:
newdict=OrderedDict()
for k,v in ckptfile['state_dict'].items():
if 'module.' not in k:
newdict['module.'+k]=v
else:
newdict[k]=v
ckptfile['state_dict']=newdict
# just ignore the bn_not_save parameters
self.model.load_state_dict(ckptfile['state_dict'],strict=False)
# load_checkpoint(self.model,ckptfile)
if not self.args.finetune and not self.args.do_test and not self.args.Prune.do_test:
self.optimizer.load_state_dict(ckptfile['opti_dict'])
self.global_epoch = ckptfile['epoch']
self.global_iter = ckptfile['iter']
self.best_mAP = ckptfile['metric']
print("successfully load checkpoint {}".format(self.args.EXPER.resume))