def generality_experiment(
arch_params,
optimization_params ,
dataset,
original_filename_params,
filename_params_retrain,
retrain_params,
visual_params,
validate_after_epochs = 1,
n_epochs = 50,
ft_epochs = 200,
verbose = False
):
params_loaded, arch_params_loaded = load_network (original_filename_params ["network_save_name"] ,
data_params = False,
optimization_params = False)
# retrain is used to do the dataset some wierd experiments.
retrain_net = cnn_mlp( filename_params = filename_params_retrain,
arch_params = arch_params,
optimization_params = optimization_params,
retrain_params = retrain_params,
init_params = params_loaded,
verbose =verbose )
retrain_net.init_data ( dataset = dataset , outs = arch_params ["outs"], visual_params = visual_params, verbose = verbose )
retrain_net.build_network ( verbose = verbose )
retrain_net.train ( n_epochs = n_epochs,
ft_epochs = ft_epochs,
validate_after_epochs = validate_after_epochs,
verbose = verbose)
retrain_net.test ( verbose = verbose )
retrain_net.save_network ()
def run_cnn(
arch_params,
optimization_params ,
dataset,
filename_params,
visual_params,
n_epochs = 50,
ft_epochs = 200,
validate_after_epochs = 1,
verbose = False,
):
net = cnn_mlp( filename_params = filename_params,
arch_params = arch_params,
optimization_params = optimization_params,
retrain_params = None,
init_params = None,
verbose =verbose )
net.init_data ( dataset = dataset , outs = arch_params ["outs"], visual_params = visual_params, verbose = verbose )
net.build_network(verbose = verbose)
net.train( n_epochs = n_epochs,
ft_epochs = ft_epochs,
validate_after_epochs = validate_after_epochs,
verbose = verbose )
net.test( verbose = verbose )
net.save_network ()
def solo(
arch_params,
optimization_params ,
dataset,
filename_params,
visual_params,
n_epochs = 200,
ft_epochs = 100,
validate_after_epochs = 1,
verbose = False,
):
print "solo network"
"""
# use this if solo was trained and you want to fine tune..
params_loaded, arch_params_loaded = load_network ( filename_params ["network_save_name"] ,
data_params = False,
optimization_params = False)
retrain_params = {
"copy_from_old" : [ True ] * (len(arch_params_loaded["nkerns"]) + len(arch_params_loaded["num_nodes"])+1),
"freeze" : [ False ] * (len(arch_params_loaded["nkerns"]) + len(arch_params_loaded["num_nodes"])+1)
}
net = cnn_mlp( filename_params = filename_params,
arch_params = arch_params_loaded,
optimization_params = optimization_params,
retrain_params = retrain_params,
init_params = params_loaded,
verbose =verbose )
"""
net = cnn_mlp( filename_params = filename_params,
arch_params = arch_params,
optimization_params = optimization_params,
retrain_params = None,
init_params = None,
verbose =verbose )
net.init_data ( dataset = dataset , outs = arch_params ["outs"], visual_params = visual_params, verbose = verbose )
net.build_network(verbose = verbose)
net.build_learner(verbose =verbose)
net.train( n_epochs = n_epochs,
ft_epochs = ft_epochs,
validate_after_epochs = validate_after_epochs,
patience_epochs = 4,
verbose = verbose )
net.test( verbose = verbose )
net.save_network ()
pdb.set_trace()
def regularized_train(
parent_params,
child_params,
joint_params,
visual_params,
dataset,
verbose = False,
):
print "obi-wan network"
parent_filename_params = parent_params["parent_filename_params"]
parent_arch_params = parent_params["parent_arch_params"]
parent_optimization_params = parent_params["parent_optimization_params"]
parent_retrain_params = None
parent_init_params = None
"""
## use this if Parent is VGG ...
model = './dataset/vgg/vgg19.pkl' # if vgg being loaded
freeze_layer_params = True
obi_wan = load_vgg(
model = model,
dataset = dataset[0],
filename_params = parent_filename_params,
optimization_params = parent_optimization_params,
freeze_layer_params = freeze_layer_params,
visual_params = visual_params,
outs = parent_arch_params["outs"],
verbose = verbose
)
obi_wan.build_network(verbose = verbose)
"""
# Use these to load a pre-trained parent in samosa.
parent_init_params, parent_arch_params = load_network ( parent_filename_params ["network_save_name"] ,
data_params = False,
optimization_params = False)
parent_arch_params["outs"] = child_params["child_arch_params"]["outs"]
parent_retrain_params = {
"copy_from_old" : [ True ] * (len(parent_arch_params["nkerns"]) + len(parent_arch_params["num_nodes"])) + [True],
"freeze" : [ False ] * (len(parent_arch_params["nkerns"]) + len(parent_arch_params["num_nodes"])) + [False]
}
obi_wan = cnn_mlp(
filename_params = parent_filename_params,
arch_params = parent_arch_params,
optimization_params = parent_optimization_params,
retrain_params = parent_retrain_params,
init_params = parent_init_params,
verbose =verbose )
obi_wan.init_data ( dataset = dataset[1] , outs = parent_arch_params ["outs"], visual_params = visual_params, verbose = verbose )
obi_wan.build_network(verbose = verbose)
# The next three lines might be skipped if not retraining.
"""
## Create a new parent from scratch in samosa
obi_wan = cnn_mlp(
filename_params = parent_filename_params,
arch_params = parent_arch_params,
optimization_params = parent_optimization_params,
retrain_params = parent_retrain_params,
init_params = parent_init_params,
verbose =verbose )
obi_wan.init_data ( dataset = dataset[0] , outs = parent_arch_params ["outs"], visual_params = visual_params, verbose = verbose )
obi_wan.build_network(verbose = verbose)
obi_wan.build_learner(verbose =verbose)
obi_wan.train( n_epochs = parent_params["parent_n_epochs"],
ft_epochs = parent_params ["parent_ft_epochs"],
validate_after_epochs = parent_params ["parent_validate_after_epochs"],
patience_epochs = 4,
verbose = verbose )
obi_wan.save_network ()
obi_wan.test( verbose = verbose )
"""
"""
# use this to fine tune a pre-trained obi-wan on the skywalker dataset.
print "obi-wan fine tuning on target dataset"
#setting the network up for child dataset and fine tunining with the child dataset also..
parent_arch_params["outs"] = child_params["child_arch_params"]["outs"]
parent_init_params, parent_arch_params = load_network ( parent_filename_params ["network_save_name"] ,
data_params = False,
optimization_params = False)
parent_retrain_params = {
"copy_from_old" : [ True ] * (len(parent_arch_params["nkerns"]) + len(parent_arch_params["num_nodes"])) + [False] ,
"freeze" : [ False] * (len(parent_arch_params["nkerns"]) + len(parent_arch_params["num_nodes"]) + 1)
}
obi_wan = cnn_mlp(
filename_params = parent_filename_params,
arch_params = parent_arch_params,
optimization_params = parent_optimization_params,
retrain_params = parent_retrain_params,
init_params = parent_init_params,
verbose =verbose
)
obi_wan.init_data ( dataset = dataset[1] , outs = parent_arch_params ["outs"], visual_params = visual_params, verbose = verbose )
obi_wan.build_network(verbose = verbose)
obi_wan.build_learner(verbose =verbose)
obi_wan.train( n_epochs = parent_params["parent_n_epochs"],
ft_epochs = parent_params ["parent_ft_epochs"],
validate_after_epochs = parent_params ["parent_validate_after_epochs"],
patience_epochs = 4,
verbose = verbose )
obi_wan.save_network ()
obi_wan.test( verbose = verbose )
"""
# Make true for unsupervised
un_supervised = True
if un_supervised is True:
print "skywalker network"
# use this for unsupervised mentoring ...
print "unsupervised mentoring"
child_filename_params = child_params["child_filename_params"]
child_arch_params = child_params["child_arch_params"]
child_optimization_params = child_params["child_optimization_params"]
child_retrain_params = None
child_init_params = None
# make soft label and hard label OFF
unsupervised_filename_params = child_filename_params
unsupervised_filename_params ["network_save_name"] = "../results/skywalker_network_unsupervised.pkl"
joint_params_unsupervised = {
"learn_layers" : [ (0,0), (2,1), (3,2) ], # 0 is the first layer
"learn_layers_coeffs" : [
(1,1,75), # This is probes
(0,0,0), # This is soft outputs
(0,0,1) # this is for hard labels
],
# weights for each probes and soft outputs and labels
# first term is begining weight,
# second term is ending weight,
# third term is until which epoch.
"error" : rmse, # erros of the probes.
"print_probe_costs" : True
}
skywalker = regularizer_net(
filename_params = child_filename_params,
arch_params = child_arch_params,
optimization_params = child_optimization_params,
retrain_params = child_retrain_params,
init_params = child_init_params,
parent = obi_wan,
joint_params = joint_params_unsupervised,
verbose =verbose )
skywalker.init_data ( dataset = dataset[0] , outs = child_arch_params ["outs"], visual_params = visual_params, verbose = verbose )
skywalker.build_network(verbose = verbose)
# build probes also builds the learer Don't bother calling a build learner.
# Uncomment for learning the unsupervised network also.
skywalker.build_probes(verbose = verbose)
skywalker.train( n_epochs = 75,
ft_epochs = 0 ,
validate_after_epochs = child_params ["child_validate_after_epochs"],
patience_epochs = 75, # this will avoid early stopping.
verbose = verbose )
skywalker.save_network ()
skywalker.print_net(epoch = 'final')
pdb.set_trace()
# regular ...
print "supervised learning"
child_filename_params = child_params["child_filename_params"]
child_arch_params = child_params["child_arch_params"]
child_optimization_params = child_params["child_optimization_params"]
child_retrain_params = None
child_init_params = None
if un_supervised is True:
# use the next few lines if an unsupervised mentor is already ready or to pre-load a skywalker and fine tune ..
child_init_params, child_arch_params = load_network ( unsupervised_filename_params["network_save_name"] ,
data_params = False,
optimization_params = False)
child_retrain_params = {
"copy_from_old" : [ True ] * (len(parent_arch_params["nkerns"]) + len(parent_arch_params["num_nodes"])) + [True],
"freeze" : [ False ] * (len(parent_arch_params["nkerns"]) + len(parent_arch_params["num_nodes"])) + [False]
}
#####
skywalker = regularizer_net(
filename_params = child_filename_params,
arch_params = child_arch_params,
optimization_params = child_optimization_params,
retrain_params = child_retrain_params,
init_params = child_init_params,
parent = obi_wan,
joint_params = joint_params,
verbose =verbose )
skywalker.init_data ( dataset = dataset[1] , outs = child_arch_params ["outs"], visual_params = visual_params, verbose = verbose )
skywalker.build_network(verbose = verbose)
# build probes also builds the leayer Don't bother calling a build_learner.
skywalker.build_probes(verbose = verbose)
skywalker.train( n_epochs = child_params ["child_n_epochs"],
ft_epochs = child_params ["child_ft_epochs"] ,
validate_after_epochs = child_params ["child_validate_after_epochs"],
patience_epochs = 50,
verbose = verbose )
skywalker.save_network ()
skywalker.test( verbose = verbose )
pdb.set_trace()
