def object_2_attributes(vals):
if isinstance(vals, dict):
if 'type' in vals and 'par' in vals and len(vals) == 2:
v1 = ModelParPair(vals['type'])
update_config(v1, vals['par'])
return v1
else:
v2 = AttrDict()
for key, value in vals.items():
v2[key] = object_2_attributes(value)
return v2
elif isinstance(vals, list):
v3 = []
for val in vals:
v3.append(object_2_attributes(val))
return v3
else:
return vals
__C.training_parameters.lambda_grl_steps = 0
__C.training_parameters.lambda_grl_start = 0
__C.training_parameters.lambda_q = 1.0
__C.training_parameters.static_lr = True
# --------------------------------------------------------------------------- #
# loss options:
# --------------------------------------------------------------------------- #
__C.loss = 'logitBCE'
# --------------------------------------------------------------------------- #
# optimizer options:
# --------------------------------------------------------------------------- #
__C.optimizer = ModelParPair('Adamax')
# --------------------------------------------------------------------------- #
# adv_optimizer options:
# --------------------------------------------------------------------------- #
__C.adv_optimizer = ModelParPair('adv_opt')
# --------------------------------------------------------------------------- #
# model options: Note default is our
# --------------------------------------------------------------------------- #
__C.model = AttrDict()
__C.model.image_feat_dim = 2048
__C.model.question_embedding = [ModelParPair("att_que_embed")]
__C.model.image_feature_encoding = [ModelParPair('default_image')]
__C.training_parameters.wu_iters = 1000
__C.training_parameters.max_iter = 12000
__C.training_parameters.lr_steps = [5000, 7000, 9000, 11000]
__C.training_parameters.lr_ratio = 0.1
# --------------------------------------------------------------------------- #
# loss options:
# --------------------------------------------------------------------------- #
__C.loss = "logitBCE"
# --------------------------------------------------------------------------- #
# optimizer options:
# --------------------------------------------------------------------------- #
__C.optimizer = ModelParPair("Adamax")
# --------------------------------------------------------------------------- #
# model options: Note default is our
# --------------------------------------------------------------------------- #
__C.model = AttrDict()
__C.model.image_feat_dim = 2048
__C.model.question_embedding = [ModelParPair("att_que_embed")]
__C.model.image_feature_encoding = [ModelParPair("default_image")]
__C.model.image_embedding_models = []
__C.model.modal_combine = ModelParPair("non_linear_elmt_multiply")
__C.model.classifier = ModelParPair("logit_classifier")
top_down_bottom_up = AttrDict()
top_down_bottom_up.modal_combine = ModelParPair("non_linear_elmt_multiply")