def initialize(self,
pretrain_dict: Dict[str, nn.Parameter],
verbosity: int = 2,
**kwargs):
weight_dict = dict()
matched_dict, missed_keys = self.find_match_miss(pretrain_dict)
weight_dict.update(matched_dict)
if len(missed_keys) > 0:
print("Initial seg head from cls head")
strip_pretrain_dict = strip_named_params(pretrain_dict)
for model_param_name in missed_keys:
if 'fc8' in model_param_name:
# Don't init fc_8
continue
if re.search(r'fc\d_\d', model_param_name) is None:
continue
# Fuzzy Match
strip_model_param_name = strip_param_name(model_param_name)
fuzzy_strip_model_param_name = re.sub(
r'(fc\d_)\d', lambda match: match.group(1) + '0',
strip_model_param_name)
weight_dict[model_param_name] = strip_pretrain_dict[
fuzzy_strip_model_param_name]
self.update_model_state_dict(weight_dict, verbosity)
def initialize(self,
pretrain_dict: Dict[str, nn.Parameter],
verbosity: int = 1,
**kwargs):
strip_pretrain_dict = strip_named_params(pretrain_dict)
weight_dict = {}
for model_param_name in self.state_dict().keys():
strip_model_param_name = strip_param_name(model_param_name)
if strip_model_param_name in strip_pretrain_dict:
weight_dict[model_param_name] = strip_pretrain_dict[
strip_model_param_name]
update_model_state_dict(self, weight_dict, verbosity)
def initialize(self, pretrain_dict: Dict[str, nn.Parameter], verbosity: int=1, **kwargs):
strip_pretrain_dict = strip_named_params(pretrain_dict)
weight_dict = {}
for model_param_name in self.state_dict().keys():
strip_model_param_name = strip_param_name(model_param_name)
if strip_model_param_name in strip_pretrain_dict:
weight_dict[model_param_name] = strip_pretrain_dict[strip_model_param_name]
elif 'fc8' in strip_model_param_name:
continue
else:
# Using param from cls branch with same layer_ID
print(f"Init {strip_model_param_name} from cls head")
pattern = re.sub(r'fc\d_\d', lambda match: match.group(0)[:-1] + '0', strip_model_param_name)
weight = list(named_param_match_pattern(pattern,
strip_pretrain_dict,
full_match=True,
multi_match=False).values())[0]
weight_dict[model_param_name] = weight
update_model_state_dict(self, weight_dict, verbosity)
def initialize_from_deeplab_v1(self,
pretrain_dict: Dict[str, nn.Parameter],
verbosity: int = 1):
update_dict = {}
matched_dict, missed_keys = self.find_match_miss(pretrain_dict)
update_dict.update(matched_dict)
if len(missed_keys) > 0:
print("Init cls head from deeplab_v1")
strip_pretrain_dict = strip_named_params(pretrain_dict)
for model_param_name in missed_keys:
strip_model_param_name = strip_param_name(model_param_name)
if 'fc8' in strip_model_param_name:
continue
else:
param_name_in_pretrain = re.sub(
r'(fc\d)_\d', lambda match: match.group(1),
strip_model_param_name)
update_dict[model_param_name] = strip_pretrain_dict[
param_name_in_pretrain]
self.update_model_state_dict(update_dict, verbosity)
def initialize_from_deeplab_v2(self,
pretrain_dict: Dict[str, nn.Parameter],
verbosity: int = 1):
strip_pretrain_dict = strip_named_params(pretrain_dict)
weight_dict = {}
for model_param_name in self.state_dict().keys():
strip_model_param_name = strip_param_name(model_param_name)
if strip_model_param_name in strip_pretrain_dict:
# Exact Match
weight_dict[model_param_name] = strip_pretrain_dict[
strip_model_param_name]
else:
if 'fc8' in model_param_name:
# Don't init fc_8
continue
# Fuzzy Match
fuzzy_strip_model_param_name = re.sub(
r'fc\d_\d', lambda match: match.group(0)[:-1] + '1',
strip_model_param_name)
weight_dict[model_param_name] = strip_pretrain_dict[
fuzzy_strip_model_param_name]
update_model_state_dict(self, weight_dict, verbosity)