from domainlab.algos.a_algo_builder import NodeAlgoBuilder
from domainlab.algos.msels.c_msel_oracle import MSelOracleVisitor
from domainlab.algos.msels.c_msel_val import MSelValPerf
from domainlab.algos.observers.b_obvisitor import ObVisitor
from domainlab.algos.trainers.zoo_trainer import TrainerChainNodeGetter
from domainlab.compos.zoo_nn import FeatExtractNNBuilderChainNodeGetter
from domainlab.utils.utils_cuda import get_device
from domainlab.compos.nn_zoo.nn import LayerId
[docs]
def make_basic_trainer(class_name_model):
"""make_basic_trainer.
:param class_name_model:
"""
class NodeAlgoBuilderCustom(NodeAlgoBuilder):
"""NodeAlgoBuilderCustom."""
def get_trainer(self, args):
"""
chain of responsibility pattern for fetching trainer from commandline parsed arguments
"""
trainer = TrainerChainNodeGetter(args.trainer)(default="basic")
return trainer
def _set_args(self, args, val_arg_na, prefix, argname):
"""_set_args.
Since we could not hard code all possible strings in the argparser,
we add attributes to args name space according to the user
specification from the custom python file they provide
the custom model the user wrote should have
model.dict_net_module_na2arg_na with something like
{"net1":"name1", "net2":"name2"}
val_arg_na below will be filled with "name1" for instance
python main_out.py –te_d=caltech –task=mini_vlcs –debug –bs=3
–apath=examples/algos/demo_custom_model.py
–model=custom –nname_argna2val net1 –nname_argna2val alexnet
:param args: the namespace of command line arguemnts
:param val_arg_na: the custom argument name the user specified
:param prefix: nname or npath to be consistent with the rest of
the package
:param argname: nname_argna2val or "npath_argna2val", hard coded
"""
if getattr(args, argname) is None:
setattr(args, prefix + val_arg_na, None)
return
list_args = getattr(args, argname)
ind = list_args.index(val_arg_na)
if ind + 1 >= len(list_args): # list of args always even length
raise RuntimeError(
f"\n nname_argna2val or npath_argna2val should \
\n always be specified in pairs instead of \
odd number:\
\n {str(list_args)}"
)
val = list_args[ind + 1]
# add attributes to namespaces args, the attributes are provided by
# user in the custom model file
setattr(args, prefix + val_arg_na, val)
def set_nets_from_dictionary(self, args, task, model):
"""set_nets_from_dictionary.
the custom model the user wrote should have
model.dict_net_module_na2arg_na with something like
{"net1":"name1", "net2":"name2"}
python main_out.py –te_d=caltech –task=mini_vlcs –debug –bs=3
–apath=examples/algos/demo_custom_model.py
–model=custom –nname_argna2val net1 –nname_argna2val alexnet
"""
for key_module_na, val_arg_na in model.dict_net_module_na2arg_na.items():
#
if args.nname_argna2val is None and args.npath_argna2val is None:
raise RuntimeError(
"either specify nname_argna2val or \
npath_argna2val"
)
self._set_args(args, val_arg_na, "nname", "nname_argna2val")
self._set_args(args, val_arg_na, "npath", "npath_argna2val")
#
builder = FeatExtractNNBuilderChainNodeGetter(
args,
arg_name_of_net="nname" + val_arg_na,
arg_path_of_net="npath" + val_arg_na,
)()
net = builder.init_business(
flag_pretrain=True,
dim_out=task.dim_y,
remove_last_layer=False,
args=args,
isize=(task.isize.i_c, task.isize.i_h, task.isize.i_w),
)
model.add_module("%s" % (key_module_na), net)
# FIXME: this is just a stub to let net_aux be invariant feature
model._net_invar_feat = model.net_aux
model._net_classifier = LayerId()
def init_business(self, exp):
"""
return trainer, model, observer
"""
task = exp.task
args = exp.args
device = get_device(args)
model_sel = MSelOracleVisitor(MSelValPerf(max_es=args.es),
val_threshold=args.val_threshold)
observer = ObVisitor(model_sel)
model = class_name_model(net_classifier=None, list_str_y=task.list_str_y)
model = self.init_next_model(model, exp)
self.set_nets_from_dictionary(args, task, model)
trainer = self.get_trainer(args)
trainer.init_business(model, task, observer, device, args)
return trainer, model, observer, device
return NodeAlgoBuilderCustom