def get_random_faltness_radius_upper_lower_bs(self, precision=0.0001):
'''
The goal of this is for the easy of computation of the epsilon radius of
a network which is defined as follows:
r(dx,eps,W) = sup{r \in R : |I(W) - I(W+r*dx)|<=eps}
W_all = r*dx
dx = isotropic unit vector from the net
'''
''' record reference errors/losses '''
#stats_collector.record_errors_loss_reference_net(criterion,error_criterion,net,trainloader,testloader,device)
''' get isotropic direction '''
nb_params = nn_mdls.count_nb_params(self.net)
v = torch.normal(torch.zeros(nb_params),
torch.ones(nb_params)).to(self.device)
dx = v / v.norm(2)
''' set up '''
lb, ub = 0, 2 * self.r_initial
_, error_minima = evalaute_running_mdl_data_set(
self.criterion, self.error_criterion, self.net, self.trainloader,
self.device, self.iterations)
y_target = error_minima + self.epsilon # I(W) + eps = I(W+0*dx) + eps
def f(r):
net_rdx = produce_new_translated_net(self.net, r, dx)
Loss_rdx, Error_rdx = evalaute_running_mdl_data_set(
self.criterion, self.error_criterion, net_rdx,
self.trainloader, self.device, self.iterations)
return Error_rdx - y_target
''' start bisect method ''' # https://en.wikipedia.org/wiki/Bisection_method
f_lb = error_minima - y_target
#f_ub = f(ub)
while True:
r = (lb + ub) / 2
f_r = f(r) # note this value could be negative
# print('--')
# print(f'lb={lb}')
# print(f'ub={ub}')
# print(f'r = {r}')
# print(f'f(r) = f({r}) = {f_r}')
# print(f'f(lb) = f({lb}) = {f_lb}')
# print(f'abs(f_r - 0) < precision = {abs(f_r - 0) < precision}')
# print(f'abs(ub - lb) < precision = {abs(ub - lb) < precision}')
''' check if we reached epsilon jump '''
if abs(f_r - 0) < precision or abs(
ub -
lb) < precision: ## 10^-4.5 for half machine precision
''' compute I(W+r*dx) = I(W+W_all)'''
return r
elif np.sign(f_r) == np.sign(f_lb):
lb = r
f_lb = f_r
else:
ub = r
def get_radius_errors_loss_list_via_interpolation(dir_index, net, r_large,
interpolations, device,
stats_collector, criterion,
error_criterion, trainloader,
testloader, iterations):
'''
Computes I = [..., I(W+r*dx),...]. A sequence of errors/losses
from a starting minimum W to the final minum net r*dx.
The goal of this is for the easy of computation of the epsilon radius of
a network which is defined as follows:
r(dx,eps,W) = sup{r \in R : |I(W) - I(W+r*dx)|<=eps}
W_all = r*dx
dx = isotropic unit vector from the net
'''
''' record reference errors/losses '''
stats_collector.record_errors_loss_reference_net(criterion,
error_criterion, net,
trainloader, testloader,
device)
''' get isotropic direction '''
nb_params = nn_mdls.count_nb_params(net)
v = torch.normal(torch.zeros(nb_params), torch.ones(nb_params)).to(device)
dx = v / v.norm(2)
''' fill up I list '''
net_r = copy.deepcopy(net)
net_end = produce_new_translated_net(net, r_large, dx)
#print(f'||net1 - net2|| = {weight_diff_btw_nets(net_end,net)}')
for epoch, alpha in enumerate(interpolations):
''' compute I(W+r*dx) = I(W+W_all)'''
net_r = convex_interpolate_nets(
net_r, net1=net_end, net2=net,
alpha=alpha) # alpha*net_end+(1-alpha)*net
Er_train_loss, Er_train_error = evalaute_running_mdl_data_set(
criterion, error_criterion, net_r, trainloader, device, iterations)
Er_test_loss, Er_test_error = evalaute_running_mdl_data_set(
criterion, error_criterion, net_r, testloader, device, iterations)
''' record result '''
stats_collector.append_losses_errors_accs(Er_train_loss,
Er_train_error, Er_test_loss,
Er_test_error)
errors_losses = [
Er_train_loss, Er_train_error, Er_test_loss, Er_test_error
]
stats_collector.append_all_losses_errors_accs(dir_index, epoch,
errors_losses)
''' record current r '''
r = alpha * r_large
stats_collector.rs.append(r)
#stats_collector.random_dirs.append(dx.cpu().numpy())
return Er_train_loss, Er_train_error, Er_test_loss, Er_test_error, net_r
def get_random_faltness_radius_repeated_doubling(self, precision=0.0001):
'''
The goal of this is for the easy of computation of the epsilon radius of
a network which is defined as follows:
r(dx,eps,W) = sup{r \in R : |I(W) - I(W+r*dx)|<=eps}
W_all = r*dx
dx = isotropic unit vector from the net
'''
''' record reference errors/losses '''
#stats_collector.record_errors_loss_reference_net(criterion,error_criterion,net,trainloader,testloader,device)
''' get isotropic direction '''
nb_params = nn_mdls.count_nb_params(self.net)
v = torch.normal(torch.zeros(nb_params),
torch.ones(nb_params)).to(self.device)
dx = v / v.norm(2)
''' fill up I list '''
r = self.r_initial
Loss_minima, Error_minima = evalaute_running_mdl_data_set(
self.criterion, self.error_criterion, self.net, self.trainloader,
self.device, self.iterations)
while True:
net_rdx = produce_new_translated_net(self.net, r, dx)
Loss_rdx, Error_rdx = evalaute_running_mdl_data_set(
self.criterion, self.error_criterion, net_rdx,
self.trainloader, self.device, self.iterations)
diff = Error_rdx - Error_minima
print(f'\n--\nr = {r}')
print(f'Error_minima={Error_minima}')
print(f'Error_rdx={Error_rdx}')
print(f'diff = {diff}')
print(f'epsilon={self.epsilon}')
print(f'abs(abs(diff)-eps)={abs(abs(diff)-self.epsilon)}')
print(f'precision={precision}')
print(
f'abs(abs(diff)-eps) < precision={abs(abs(diff)-self.epsilon) < precision}'
)
print(
f'approx_equals(diff,self.epsilon,precision=precision)={approx_equals(diff,self.epsilon,precision=precision)}'
)
''' check if we reached epsilon jump '''
if approx_equals(diff, self.epsilon, precision=precision
): ## 10^-4.5 for half machine precision
''' compute I(W+r*dx) = I(W+W_all)'''
st()
return r
elif diff > self.epsilon: # I(w+rdx) - I(W) > eps, r is too large
r /= 2
else: # I(w+rdx) - I(W) < eps, r is too small
r *= 1.5
def get_radius_errors_loss_list(dir_index, net, r_large, rs, device,
stats_collector, criterion, error_criterion,
trainloader, testloader, iterations):
'''
Computes I = [..., I(W+r*dx),...]. A sequence of errors/losses
from a starting minimum W to the final minum net r*dx.
The goal of this is for the easy of computation of the epsilon radius of
a network which is defined as follows:
r(dx,eps,W) = sup{r \in R : |I(W) - I(W+r*dx)|<=eps}
W_all = r*dx
dx = isotropic unit vector from the net
'''
''' record reference errors/losses '''
stats_collector.record_errors_loss_reference_net(criterion,
error_criterion, net,
trainloader, testloader,
device)
''' get isotropic direction '''
nb_params = nn_mdls.count_nb_params(net)
#mvg_sampler = torch.distributions.multivariate_normal.MultivariateNormal(torch.zeros(nb_params), torch.eye(nb_params))
v = torch.normal(torch.zeros(nb_params), torch.ones(nb_params)).to(device)
dx = v / v.norm(2)
''' fill up I list '''
net_r = copy.deepcopy(net)
for epoch, r in enumerate(rs):
''' compute I(W+r*dx) = I(W+W_all)'''
net_r = translate_net_by_rdx(net, net_r, r, dx)
Er_train_loss, Er_train_error = evalaute_running_mdl_data_set(
criterion, error_criterion, net_r, trainloader, device)
Er_test_loss, Er_test_error = evalaute_running_mdl_data_set(
criterion, error_criterion, net_r, testloader, device)
''' record result '''
stats_collector.append_losses_errors_accs(Er_train_loss,
Er_train_error, Er_test_loss,
Er_test_error)
errors_losses = [
Er_train_loss, Er_train_error, Er_test_loss, Er_test_error
]
stats_collector.append_all_losses_errors_accs(dir_index, epoch,
errors_losses)
return Er_train_loss, Er_train_error, Er_test_loss, Er_test_error, net_r
def main(plot=True):
if args.means != '':
means = [float(x.strip()) for x in args.means.strip('[').strip(']').split(',')]
else:
means = []
if args.stds != '':
stds = [float(x.strip()) for x in args.stds.strip('[').strip(']').split(',')]
else:
stds = []
##
hostname = utils.get_hostname()
''' cuda '''
use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
print(f'device = {device}')
''' '''
store_net = True
other_stats = dict({'sj':sj,'satid':satid,'hostname':hostname,'label_corrupt_prob':args.label_corrupt_prob})
''' reproducibility setup/params'''
#num_workers = 2 # how many subprocesses to use for data loading. 0 means that the data will be loaded in the main process.
githash = subprocess.check_output(["git", "describe", "--always"]).strip()
seed = args.seed
if seed is None: # if seed is None it has not been set, so get a random seed, else use the seed that was set
seed = int.from_bytes(os.urandom(7), byteorder="big")
print(f'seed: {seed}')
## SET SEED/determinism
num_workers = 3
torch.manual_seed(seed)
#torch.backends.cudnn.deterministic=True
''' date parameters setup'''
today_obj = date.today() # contains datetime.date(year, month, day); accessible via .day etc
day = today_obj.day
month = calendar.month_name[today_obj.month]
setup_time = time.time()
''' filenames '''
## folder names
results_root = './test_runs_flatness5_ProperOriginalExpt'
expt_folder = f'flatness_{month}_label_corrupt_prob_{args.label_corrupt_prob}_exptlabel_{args.exptlabel}_' \
f'only_1st_layer_BIAS_{args.only_1st_layer_bias}_data_set_{args.data_set}_reg_param_{args.reg_param}'
## filenames
matlab_file_name = f'flatness_{day}_{month}_sj_{sj}_staid_{satid}_seed_{seed}_{hostname}'
net_file_name = f'net_{day}_{month}_sj_{sj}_staid_{satid}_seed_{seed}_{hostname}'
## folder to hold all nets
all_nets_folder = f'nets_folder_{day}_{month}_sj_{sj}_staid_{satid}_seed_{seed}_{hostname}'
## experiment path
expt_path = os.path.join(results_root,expt_folder)
''' data set '''
data_path = './data'
standardize = not args.dont_standardize_data # x - mu / std , [-1,+1]
trainset, testset, classes = data_class.get_data_processors(data_path,args.label_corrupt_prob,dataset_type=args.data_set,standardize=standardize,type_standardize=args.type_standardize)
''' experiment params '''
evalaute_mdl_data_set = get_function_evaluation_from_name(args.evalaute_mdl_data_set)
suffle_test = False
shuffle_train = True
nb_epochs = 4 if args.epochs is None else args.epochs
batch_size = 256
#batch_size_train,batch_size_test = batch_size,batch_size
batch_size_train = batch_size
batch_size_test = 256
''' get NN '''
nets = []
mdl = args.mdl
do_bn = args.use_bn
other_stats = dict({'mdl':mdl,'do_bn':do_bn, 'type_standardize':args.type_standardize},**other_stats)
print(f'model = {mdl}')
if mdl == 'cifar_10_tutorial_net':
suffle_test = False
net = nn_mdls.Net()
nets.append(net)
elif mdl == 'debug':
suffle_test = False
nb_conv_layers=1
## conv params
Fs = [3]*nb_conv_layers
Ks = [2]*nb_conv_layers
## fc params
FC = len(classes)
C,H,W = 3,32,32
net = nn_mdls.LiaoNet(C,H,W,Fs,Ks,FC,do_bn)
nets.append(net)
elif mdl == 'sequential':
batch_size_train = 256
batch_size_test = 256
##
batch_size = batch_size_train
suffle_test = False
##
FC = [10,10]
C,H,W = 3, 32, 32
# net = torch.nn.Sequential(OrderedDict([
# ('Flatten',Flatten()),
# ('FC1', torch.nn.Linear(C*H*W,FC[0])),
# ('FC2', torch.nn.Linear(FC[0],FC[1]))
# ]))
# net = torch.nn.Sequential(OrderedDict([
# ('Flatten',Flatten()),
# ('FC1', torch.nn.Linear(C*H*W,FC[0])),
# ('relu1', torch.nn.ReLU()),
# ('FC2', torch.nn.Linear(FC[0],FC[1]))
# ]))
net = torch.nn.Sequential(OrderedDict([
('conv0', torch.nn.Conv2d(3,420,5,bias=True)),
('relu0', torch.nn.ReLU()),
('conv1', torch.nn.Conv2d(420,50,5, bias=True)),
('relu1', torch.nn.ReLU()),
('Flatten',Flatten()),
('FC1', torch.nn.Linear(28800,50,bias=True)),
('relu2', torch.nn.ReLU()),
('FC2', torch.nn.Linear(50, 10, bias=True))
]))
##
nets.append(net)
elif mdl == 'BoixNet':
batch_size_train = 256
batch_size_test = 256
##
batch_size = batch_size_train
suffle_test = False
## conv params
nb_filters1,nb_filters2 = 32, 32
nb_filters1, nb_filters2 = 32, 32
kernel_size1,kernel_size2 = 5,5
## fc params
nb_units_fc1,nb_units_fc2,nb_units_fc3 = 512,256,len(classes)
C,H,W = 3,32,32
net = nn_mdls.BoixNet(C,H,W,nb_filters1,nb_filters2, kernel_size1,kernel_size2, nb_units_fc1,nb_units_fc2,nb_units_fc3,do_bn)
nets.append(net)
elif mdl == 'LiaoNet':
suffle_test = False
nb_conv_layers=5
## conv params
Fs = [32]*nb_conv_layers
Ks = [10]*nb_conv_layers
## fc params
FC = len(classes)
C,H,W = 3,32,32
net = nn_mdls.LiaoNet(C,H,W,Fs,Ks,FC,do_bn)
nets.append(net)
elif mdl == 'GBoixNet':
#batch_size_train = 16384 # 2**14
#batch_size_test = 16384
batch_size_train = 2**10
batch_size_test = 2**10
##
batch_size = batch_size_train
suffle_test = False
## conv params
nb_conv_layers=2
Fs = [34]*nb_conv_layers
Ks = [5]*nb_conv_layers
#nb_conv_layers = 4
#Fs = [60] * nb_conv_layers
#Ks = [5] * nb_conv_layers
## fc params
FCs = [len(classes)]
##
print(f'------> FCs = {FCs}')
if args.data_set == 'mnist':
CHW = (1, 28, 28)
else:
CHW = (3,32,32)
net = nn_mdls.GBoixNet(CHW,Fs,Ks,FCs,do_bn,only_1st_layer_bias=args.only_1st_layer_bias)
print(f'net = {net}')
##
if len(means) != 0 and len(stds) != 0:
params = net.named_parameters()
dict_params = dict(params)
i = 0
for name, param in dict_params.items():
if name in dict_params:
print(name)
if name != 'conv0.bias':
mu,s = means[i], stds[i]
param.data.normal_(mean=mu,std=s)
i+=1
##
expt_path = f'{expt_path}_means_{args.means}_stds_{args.stds}'
other_stats = dict({'means': means, 'stds': stds}, **other_stats)
##
nets.append(net)
other_stats = dict({'only_1st_layer_bias': args.only_1st_layer_bias}, **other_stats)
elif mdl == 'AllConvNetStefOe':
#batch_size_train = 16384 # 2**14
#batch_size_test = 16384
#batch_size_train = 2**10
# batch_size_train = 2**10
# batch_size_test = 2**10
batch_size_train = 32
batch_size_test = 124
##
batch_size = batch_size_train
suffle_test = False
## AllConvNet
only_1st_layer_bias = args.only_1st_layer_bias
CHW = (3,32,32)
dropout = args.use_dropout
net = nn_mdls.AllConvNetStefOe(nc=len(CHW),dropout=dropout,only_1st_layer_bias=only_1st_layer_bias)
##
nets.append(net)
other_stats = dict({'only_1st_layer_bias': args.only_1st_layer_bias,'dropout':dropout}, **other_stats)
expt_path = f'{expt_path}_dropout_{dropout}'
elif mdl == 'AndyNet':
#batch_size_train = 16384 # 2**14
#batch_size_test = 16384
#batch_size_train = 2**10
batch_size_train = 2**10
batch_size_test = 2**10
# batch_size_train = 32
# batch_size_test = 124
##
batch_size = batch_size_train
suffle_test = False
## AndyNet
#only_1st_layer_bias = args.only_1st_layer_bias ## TODO fix
only_1st_layer_bias = args.only_1st_layer_bias
CHW = (3,32,32)
net = nn_mdls.get_AndyNet()
##
nets.append(net)
other_stats = dict({'only_1st_layer_bias': args.only_1st_layer_bias}, **other_stats)
expt_path = f'{expt_path}'
elif mdl == 'interpolate':
suffle_test = True
batch_size = 2**10
batch_size_train, batch_size_test = batch_size, batch_size
iterations = inf # controls how many epochs to stop before returning the data set error
#iterations = 1 # controls how many epochs to stop before returning the data set error
''' '''
path_nl = os.path.join(results_root,'flatness_27_April_label_corrupt_prob_0.0_exptlabel_GB_24_24_10_2C1FC_momentum_NL_polestar/net_27_April_sj_343_staid_1_seed_56134200848018679')
path_rl_nl = os.path.join(results_root,'flatness_27_April_label_corrupt_prob_0.0_exptlabel_GB_24_24_10_2C1FC_momentum_RLNL_polestar/net_27_April_sj_345_staid_1_seed_57700439347820897')
''' restore nets'''
net_nl = utils.restore_entire_mdl(path_nl)
net_rlnl = utils.restore_entire_mdl(path_rl_nl)
nets.append(net_nl)
nets.append(net_rlnl)
elif mdl == 'radius_flatness':
suffle_test = True
batch_size = 2**10
batch_size_train, batch_size_test = batch_size, batch_size
iterations = 11 # controls how many epochs to stop before returning the data set error
#iterations = inf # controls how many epochs to stop before returning the data set error
other_stats = dict({'iterations':iterations},**other_stats)
''' load net '''
if args.net_name == 'NL':
#path = os.path.join(results_root,'flatness_28_March_label_corrupt_prob_0.0_exptlabel_BoixNet_polestar_300_stand_natural_labels/net_28_March_206')
path = os.path.join(results_root,'flatness_27_April_label_corrupt_prob_0.0_exptlabel_GB_24_24_10_2C1FC_momentum_NL_polestar/net_27_April_sj_343_staid_1_seed_56134200848018679')
else: # RLNL
#path = os.path.join(results_root,'flatness_28_March_label_corrupt_prob_0.0_exptlabel_re_train_RLBoixNet_noBN_polestar_150/net_28_March_18')
path = os.path.join(results_root,'flatness_27_April_label_corrupt_prob_0.0_exptlabel_GB_24_24_10_2C1FC_momentum_RLNL_polestar/net_27_April_sj_345_staid_1_seed_57700439347820897')
''' restore nets'''
net = utils.restore_entire_mdl(path)
nets.append(net)
store_net = False
elif mdl == 'sharpness':
suffle_test=False #doesn't matter
''' load net '''
if args.net_name == 'NL':
#path = os.path.join(results_root,'flatness_28_March_label_corrupt_prob_0.0_exptlabel_BoixNet_polestar_300_stand_natural_labels/net_28_March_206')
path = os.path.join(results_root,'flatness_27_April_label_corrupt_prob_0.0_exptlabel_GB_24_24_10_2C1FC_momentum_NL_polestar/net_27_April_sj_343_staid_1_seed_56134200848018679')
path_adverserial_data = os.path.join('./data/sharpness_data_NL/','sdata_NL_net_27_April_sj_343_staid_1_seed_56134200848018679.npz')
else: # RLNL
#path = os.path.join(results_root,'flatness_28_March_label_corrupt_prob_0.0_exptlabel_re_train_RLBoixNet_noBN_polestar_150/net_28_March_18')
path = os.path.join(results_root,'flatness_27_April_label_corrupt_prob_0.0_exptlabel_GB_24_24_10_2C1FC_momentum_RLNL_polestar/net_27_April_sj_345_staid_1_seed_57700439347820897')
path_adverserial_data = os.path.join('./data/sharpness_data_RLNL/','sdata_RLNL_net_27_April_sj_345_staid_1_seed_57700439347820897.npz')
''' restore nets'''
net = torch.load(path)
nets.append(net)
store_net = False
elif mdl == 'divide_constant':
''' ''' # both false because we want low variation on the output of the error
iterations = inf # controls how many epochs to stop before returning the data set error
#iterations = 11 # controls how many epochs to stop before returning the data set error
batch_size = 2**10
batch_size_train, batch_size_test = batch_size, batch_size
shuffle_train = True
suffle_test = False
''' load net '''
## NL
#path_nl = os.path.join(results_root,'flatness_27_April_label_corrupt_prob_0.0_exptlabel_GB_24_24_10_2C1FC_momentum_NL_polestar/net_27_April_sj_343_staid_1_seed_56134200848018679')
#path_nl = os.path.join(results_root,'flatness_May_label_corrupt_prob_0.0_exptlabel_SGD_ManyRuns_Momentum0.9/net_17_May_sj_641_staid_5_seed_31866864409272026_polestar-old')
path_nl = os.path.join(results_root,'flatness_May_label_corrupt_prob_0.0_exptlabel_MovieNL_lr_0.01_momentum_0.9/net_22_May_sj_1168_staid_1_seed_59937023958974481_polestar-old_epoch_173')
## RLNL
#path_rlnl = os.path.join(results_root,'flatness_27_April_label_corrupt_prob_0.0_exptlabel_GB_24_24_10_2C1FC_momentum_RLNL_polestar/net_27_April_sj_345_staid_1_seed_57700439347820897')
path_rlnl = os.path.join(results_root,'flatness_May_label_corrupt_prob_0.0_exptlabel_MovieRLNLmdls_label_corruption0.5_lr_0.01_momentum_0.9/net_22_May_sj_1172_staid_1_seed_38150714758131256_polestar-old_epoch_148')
##
net_nl = torch.load(path_nl)
net_rlnl = torch.load(path_rlnl)
''' '''
print('NL')
l2_norm_all_params(net_nl)
print('RLNL')
l2_norm_all_params(net_rlnl)
''' modify nets '''
W_nl = 1
W_rlnl = (get_norm(net_rlnl, l=2)/get_norm(net_nl, l=2)) # 2.284937620162964
W_rlnl = (10)**(1.0/3.0)
#W_rlnl = 1/0.57775
#W_rlnl = 1/0.7185
#W_rlnl = 1/0.85925
#W_rlnl = 1
print(f'W_rlnl = {W_rlnl}')
print(f'norm of weight BEFORE division: get_norm(net_nl,l=2)={get_norm(net_nl,l=2)}, get_norm(net_rlnl,l=2)={get_norm(net_rlnl,l=2)}')
#net_nl = divide_params_by(W_nl, net_nl)
#net_rlnl = divide_params_by(W_rlnl, net_rlnl)
net_rlnl = divide_params_by_taking_bias_into_account(W=W_rlnl,net=net_rlnl)
print(f'norm of weight AFTER division: get_norm(net_nl,l=2)={get_norm(net_nl,l=2)}, get_norm(net_rlnl,l=2)={get_norm(net_rlnl,l=2)}')
nets.append(net_nl)
nets.append(net_rlnl)
other_stats = dict({'W_rlnl':W_rlnl,'W_nl':W_nl})
elif mdl == 'load_nl_and_rlnl':
''' load net '''
# NL
#path = os.path.join(results_root,'flatness_27_April_label_corrupt_prob_0.0_exptlabel_GB_24_24_10_2C1FC_momentum_NL_polestar/net_27_April_sj_343_staid_1_seed_56134200848018679')
path = os.path.join(results_root,'flatness_May_label_corrupt_prob_0.0_exptlabel_MovieNL_lr_0.01_momentum_0.9/net_22_May_sj_1168_staid_1_seed_59937023958974481_polestar-old_epoch_173')
net = torch.load(path)
nets.append(net)
# RLNL
#path_rlnl = os.path.join(results_root,'flatness_27_April_label_corrupt_prob_0.0_exptlabel_GB_24_24_10_2C1FC_momentum_RLNL_polestar/net_27_April_sj_345_staid_1_seed_57700439347820897')
path_rlnl = os.path.join(results_root,'flatness_May_label_corrupt_prob_0.0_exptlabel_MovieRLNLmdls_label_corruption0.5_lr_0.01_momentum_0.9/net_22_May_sj_1172_staid_1_seed_38150714758131256_polestar-old_epoch_148')
net_rlnl = torch.load(path_rlnl)
nets.append(net_rlnl)
other_stats = dict({'path': path, 'path_rlnl': path_rlnl}, **other_stats)
elif mdl == 'load_one_net':
# path = os.path.join(results_root, '/')
''' load net '''
## 0.0001
path = os.path.join(results_root,'flatness_June_label_corrupt_prob_0.0001_exptlabel_RLInits_only_1st_layer_BIAS_True_batch_size_train_1024_lr_0.01_momentum_0.9_scheduler_milestones_200,250,300_gamma_1.0/net_21_June_sj_974_staid_1_seed_44940314088747654_polestar-old')
## 0.001
path = os.path.join(results_root,'flatness_June_label_corrupt_prob_0.001_exptlabel_RLInits_only_1st_layer_BIAS_True_batch_size_train_1024_lr_0.01_momentum_0.9_scheduler_milestones_200,250,300_gamma_1.0/net_21_June_sj_967_staid_1_seed_1986409594254668_polestar-old')
## 0.01
path = os.path.join(results_root, 'flatness_June_label_corrupt_prob_0.01_exptlabel_RLInits_only_1st_layer_BIAS_True_batch_size_train_1024_lr_0.01_momentum_0.9_scheduler_milestones_200,250,300_gamma_1.0/net_21_June_sj_976_staid_1_seed_34669758900780265_polestar-old')
## 0.1
path = os.path.join(results_root,'flatness_June_label_corrupt_prob_0.1_exptlabel_RLInits_only_1st_layer_BIAS_True_batch_size_train_1024_lr_0.01_momentum_0.9_scheduler_milestones_200,250,300_gamma_1.0/net_21_June_sj_977_staid_1_seed_57003505407221650_polestar-old')
## 0.2
path = os.path.join(results_root, 'flatness_June_label_corrupt_prob_0.2_exptlabel_RLInits_only_1st_layer_BIAS_True_batch_size_train_1024_lr_0.01_momentum_0.9_scheduler_milestones_200,250,300_gamma_1.0/net_21_June_sj_978_staid_1_seed_63479113068450657_polestar-old')
## 0.5
path = os.path.join(results_root, 'flatness_June_label_corrupt_prob_0.5_exptlabel_RLInits_only_1st_layer_BIAS_True_batch_size_train_1024_lr_0.01_momentum_0.9_scheduler_milestones_200,250,300_gamma_1.0/net_21_June_sj_979_staid_1_seed_51183371945505111_polestar-old')
## 0.75
path = os.path.join(results_root, 'flatness_June_label_corrupt_prob_0.75_exptlabel_RLInits_only_1st_layer_BIAS_True_batch_size_train_1024_lr_0.01_momentum_0.9_scheduler_milestones_200,250,300_gamma_1.0/net_21_June_sj_980_staid_1_seed_63292262317939652_polestar-old')
## 1.0
path = os.path.join(results_root, 'flatness_June_label_corrupt_prob_1.0_exptlabel_RLInits_only_1st_layer_BIAS_True_batch_size_train_1024_lr_0.01_momentum_0.9_scheduler_milestones_200,250,300_gamma_1.0/net_21_June_sj_981_staid_1_seed_34295360820373818_polestar-old')
''' load net '''
net = torch.load(path)
nets.append(net)
other_stats = dict({'path': path}, **other_stats)
elif mdl == 'l2_norm_all_params':
''' load net '''
# path = os.path.join(results_root,'flatness_June_label_corrupt_sqprob_0.0_exptlabel_WeightDecay_lambda100_lr_0.1_momentum_0.0/net_1_June_sj_2833_staid_2_seed_45828051420330772_polestar-old')
# path = os.path.join(results_root,'flatness_June_label_corrupt_prob_0.0_exptlabel_WeightDecay_lambda1_lr_0.1_momentum_0.0/net_1_June_sj_2830_staid_1_seed_53714812690274511_polestar-old')
# path = os.path.join(results_root,'flatness_June_label_corrupt_prob_0.0_exptlabel_WeightDecay_lambda0.1_lr_0.1_momentum_0.0/net_1_June_sj_2835_staid_2_seed_66755608399194708_polestar-old')
# path = os.path.join(results_root,'flatness_June_label_corrupt_prob_0.0_exptlabel_WeightDecay_lambda0.01_lr_0.1_momentum_0.0/net_1_June_sj_2832_staid_1_seed_47715620118836168_polestar-old')
#path = os.path.join(results_root,'flatness_May_label_corrupt_prob_0.0_exptlabel_WeightDecay_lambda0.1_lr_0.01_momentum_0.9/net_31_May_sj_2784_staid_1_seed_59165331201064855_polestar-old')
#path = os.path.join(results_root,'flatness_May_label_corrupt_prob_0.0_exptlabel_WeightDecay_lambda0.01_lr_0.01_momentum_0.9/net_31_May_sj_2792_staid_1_seed_42391375291583068_polestar-old')
#path = os.path.join(results_root,'flatness_May_label_corrupt_prob_0.0_exptlabel_WeightDecay_lambda0.001_lr_0.01_momentum_0.9/net_31_May_sj_2793_staid_2_seed_47559284752010338_polestar-old')
#path = os.path.join(results_root,'flatness_June_label_corrupt_prob_0.0_exptlabel_L2_squared_lambda1_lr_0.1_momentum_0.0/net_1_June_sj_2841_staid_2_seed_29441453139027048_polestar-old')
#path = os.path.join(results_root,'flatness_June_label_corrupt_prob_0.0_exptlabel_L2_squared_lambda0.1_lr_0.1_momentum_0.0/net_1_June_sj_2839_staid_2_seed_35447208985369634_polestar-old')
#path = os.path.join(results_root,'flatness_June_label_corrupt_prob_0.0_exptlabel_L2_squared_lambda0.01_lr_0.1_momentum_0.0/net_1_June_sj_2837_staid_2_seed_57556488720733908_polestar-old')
#path = os.path.join(results_root,'flatness_June_label_corrupt_prob_0.0_exptlabel_L2_squared_lambda0.001_lr_0.1_momentum_0.0/net_1_June_sj_2848_staid_1_seed_48943421305461120_polestar-old')
#path = os.path.join(results_root,'flatness_June_label_corrupt_prob_0.0_exptlabel_L2_squared_lambda0.0001_lr_0.1_momentum_0.0/net_1_June_sj_2850_staid_1_seed_2881772832480048_polestar-old')
#path = os.path.join(results_root,'flatness_June_label_corrupt_prob_0.0_exptlabel_L2_squared_lambda0.00001_lr_0.1_momentum_0.0/net_1_June_sj_2852_staid_1_seed_24293440492629928_polestar-old')
print(f'path = {path}')
net = torch.load(path)
''' l2_norm_all_params '''
l2_norm_all_params(net)
''' evaluate data set '''
standardize = not args.dont_standardize_data # x - mu / std , [-1,+1]
error_criterion = metrics.error_criterion
criterion = torch.nn.CrossEntropyLoss()
trainset, testset, classes = data_class.get_data_processors(data_path, args.label_corrupt_prob,dataset_type=args.data_set,standardize=standardize)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size_train, shuffle=shuffle_train,num_workers=num_workers)
testloader = torch.utils.data.DataLoader(testset, batch_size=batch_size_test, shuffle=suffle_test,num_workers=num_workers)
train_loss_epoch, train_error_epoch = evalaute_mdl_data_set(criterion, error_criterion, net,trainloader,device)
test_loss_epoch, test_error_epoch = evalaute_mdl_data_set(criterion, error_criterion, net,testloader,device)
print(f'[-1, -1], (train_loss: {train_loss_epoch}, train error: {train_error_epoch}) , (test loss: {test_loss_epoch}, test error: {test_error_epoch})')
''' end '''
nets.append(net)
sys.exit()
else:
print('RESTORED FROM PRE-TRAINED NET')
suffle_test = False
''' RESTORED PRE-TRAINED NET '''
# example name of file, os.path.join(results_root,expt_path,f'net_{day}_{month}_{seed}')
# args.net_path = 'flatness_27_March_label_corrupt_prob_0_exptlabel_BoixNet_stand_600_OM/net_27_Match_64'
path_to_mdl = args.mdl
path = os.path.join(results_root,path_to_mdl)
# net = utils.restore_entire_mdl(path)
net = torch.load(path)
nets.append(net)
print(f'nets = {nets}')
''' cuda/gpu '''
for net in nets:
net.to(device)
nb_params = nn_mdls.count_nb_params(net)
''' stats collector '''
stats_collector = StatsCollector(net)
''' get data set '''
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size_train, shuffle=shuffle_train, num_workers=num_workers)
testloader = torch.utils.data.DataLoader(testset, batch_size=batch_size_test, shuffle=suffle_test, num_workers=num_workers)
''' Cross Entropy + Optmizer '''
lr = args.lr
momentum = 0.9
## Error/Loss criterions
error_criterion = metrics.error_criterion
criterion = torch.nn.CrossEntropyLoss()
#criterion = torch.nn.MultiMarginLoss()
#criterion = torch.nn.MSELoss(size_average=True)
print(f'Training Algorithm = {args.train_alg}')
if args.train_alg == 'SGD':
optimizer = optim.SGD(net.parameters(), lr=lr, momentum=momentum)
elif args.train_alg == 'Adam':
optimizer = optim.Adam(net.parameters(), lr=lr)
else:
raise ValueError(f'Training alg not existent: {args.train_alg}')
other_stats = dict({'nb_epochs':nb_epochs,'batch_size':batch_size,'mdl':mdl,'lr':lr,'momentum':momentum, 'seed':seed,'githash':githash},**other_stats)
expt_path = f'{expt_path}_args.train_alg_{args.train_alg}_batch_train_{batch_size_train}_lr_{lr}_moment_{momentum}_epochs_{nb_epochs}'
''' scheduler '''
#milestones = [20, 30, 40]
milestones = [200, 250, 300]
#milestones = [700, 800, 900]
#milestones = [1700, 1800, 1900]
scheduler_gamma = args.decay_rate
scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=milestones, gamma=scheduler_gamma)
other_stats = dict({'milestones': milestones, 'scheduler_gamma': scheduler_gamma}, **other_stats)
milestones_str = ','.join(str(m) for m in milestones)
#expt_path = f'{expt_path}_scheduler_milestones_{milestones_str}_gamma_{gamma}'
expt_path = f'{expt_path}_scheduler_gamma_{scheduler_gamma}'
print(f'scheduler_gamma = {scheduler_gamma}')
''' Verify model you got has the right error'''
train_loss_epoch, train_error_epoch = evalaute_mdl_data_set(criterion, error_criterion, net, trainloader, device)
test_loss_epoch, test_error_epoch = evalaute_mdl_data_set(criterion, error_criterion, net, testloader, device)
print(f'train_loss_epoch, train_error_epoch = {train_loss_epoch}, {train_error_epoch} \n test_loss_epoch, test_error_epoch = {test_loss_epoch}, {test_error_epoch}')
''' Is it over parametrized?'''
overparametrized = len(trainset)<nb_params # N < W ?
print(f'Model overparametrized? N, W = {len(trainset)} vs {nb_params}')
print(f'Model overparametrized? N < W = {overparametrized}')
other_stats = dict({'overparametrized':overparametrized,'nb_params':nb_params}, **other_stats)
''' report time for setup'''
seconds_setup,minutes_setup,hours_setup = utils.report_times(setup_time,'setup')
other_stats = dict({'seconds_setup': seconds_setup, 'minutes_setup': minutes_setup, 'hours_setup': hours_setup}, **other_stats)
''' Start Training '''
training_time = time.time()
print(f'----\nSTART training: label_corrupt_prob={args.label_corrupt_prob},nb_epochs={nb_epochs},batch_size={batch_size},lr={lr},momentum={momentum},mdl={mdl},batch-norm={do_bn},nb_params={nb_params}')
## START TRAIN
if args.train_alg == 'SGD' or args.train_alg == 'Adam':
#iterations = 4 # the number of iterations to get a sense of test error, smaller faster larger more accurate. Grows as sqrt(n) though.
iterations = inf
''' set up Trainer '''
if args.save_every_epoch:
save_every_epoch = args.save_every_epoch
trainer = Trainer(trainloader, testloader, optimizer, scheduler, criterion, error_criterion, stats_collector,
device, expt_path,net_file_name,all_nets_folder,save_every_epoch,args.evalaute_mdl_data_set,
reg_param=args.reg_param,p=args.Lp_norm)
else:
trainer = Trainer(trainloader,testloader, optimizer, scheduler, criterion,error_criterion, stats_collector,
device,evalaute_mdl_data_set=args.evalaute_mdl_data_set,reg_param=args.reg_param,p=args.Lp_norm)
last_errors = trainer.train_and_track_stats(net, nb_epochs,iterations)
''' Test the Network on the test data '''
train_loss_epoch, train_error_epoch, test_loss_epoch, test_error_epoch = last_errors
print(f'train_loss_epoch={train_loss_epoch} \ntrain_error_epoch={train_error_epoch} \ntest_loss_epoch={test_loss_epoch} \ntest_error_epoch={test_error_epoch}')
elif args.train_alg == 'pert':
''' batch sizes '''
batch_size_train, batch_size_test = 50*10**3, 10*10**3
''' number of repetitions '''
nb_perturbation_trials = nb_epochs
''' noise level '''
nb_layers = len(list(net.parameters()))
noise_level = args.noise_level
perturbation_magnitudes = nb_layers*[noise_level]
print(f'noise_level={noise_level}')
''' locate where to save it '''
folder_name_noise = f'noise_{perturbation_magnitudes[0]}'
expt_path = os.path.join(expt_path,folder_name_noise)
matlab_file_name = f'noise_{perturbation_magnitudes}_{matlab_file_name}'
## TODO collect by perburbing current model X number of times with current perturbation_magnitudes
use_w_norm2 = args.not_pert_w_norm2
train_loss,train_error,test_loss,test_error = get_errors_for_all_perturbations(net,perturbation_magnitudes,use_w_norm2,device,nb_perturbation_trials,stats_collector,criterion,error_criterion,trainloader,testloader)
print(f'noise_level={noise_level},train_loss,train_error,test_loss,test_error={train_loss},{train_error},{test_loss},{test_error}')
other_stats = dict({'perturbation_magnitudes':perturbation_magnitudes}, **other_stats)
elif args.train_alg == 'interpolate':
''' prints stats before interpolation'''
print_evaluation_of_nets(net_nl, net_rlnl, criterion, error_criterion, trainloader, testloader, device, iterations)
''' do interpolation of nets'''
nb_interpolations = nb_epochs
interpolations = np.linspace(0,1,nb_interpolations)
get_landscapes_stats_between_nets(net_nl,net_rlnl,interpolations, device,stats_collector,criterion,error_criterion,trainloader,testloader,iterations)
''' print stats of the net '''
other_stats = dict({'interpolations':interpolations},**other_stats)
#print_evaluation_of_nets(net_nl, net_rlnl, criterion, error_criterion, trainloader, testloader, device, iterations)
elif args.train_alg == 'brando_chiyuan_radius_inter':
r_large = args.r_large ## check if this number is good
nb_radius_samples = nb_epochs
interpolations = np.linspace(0,1,nb_radius_samples)
expt_path = os.path.join(expt_path+f'_RLarge_{r_large}')
''' '''
nb_dirs = args.nb_dirs
stats_collector = StatsCollector(net,nb_dirs,nb_epochs)
get_all_radius_errors_loss_list_interpolate(nb_dirs,net,r_large,interpolations,device,stats_collector,criterion,error_criterion,trainloader,testloader,iterations)
other_stats = dict({'nb_dirs':nb_dirs,'interpolations':interpolations,'nb_radius_samples':nb_radius_samples,'r_large':r_large},**other_stats)
elif args.train_alg == 'sharpness':
''' load the data set '''
print('About to load the data set')
shuffle_train = True
#batch_size = 2**10
batch_size = 2**5
batch_size_train, batch_size_test = batch_size, batch_size
iterations = inf # controls how many epochs to stop before returning the data set error
#eps = 2500/50000
eps = 1 / 50000
other_stats = dict({'iterations':iterations,'eps':eps},**other_stats)
trainset,trainloader = data_class.load_only_train(path_adverserial_data,eps,batch_size_train,shuffle_train,num_workers)
''' three musketeers '''
print('Preparing the three musketeers')
net_pert = copy.deepcopy(net)
#nn_mdls.reset_parameters(net_pert)
net_original = dont_train(net)
#net_original = net
initialize_to_zero(net_original)
debug=False
if debug:
## conv params
nb_conv_layers=3
Fs = [24]*nb_conv_layers
Ks = [5]*nb_conv_layers
## fc params
FCs = [len(classes)]
CHW = (3,32,32)
net_pert = nn_mdls.GBoixNet(CHW,Fs,Ks,FCs,do_bn).to(device)
print('Musketeers are prepared')
''' optimizer + criterion stuff '''
optimizer = optim.SGD(net_pert.parameters(), lr=lr, momentum=momentum)
#optimizer = optim.Adam(net_pert.parameters(), lr=lr)
error_criterion = metrics.error_criterion
criterion = torch.nn.CrossEntropyLoss()
#criterion = torch.nn.MultiMarginLoss()
#criterion = torch.nn.MultiLabelMarginLoss()
''' Landscape Inspector '''
save_all_learning_curves = True
save_all_perts = False
nb_lambdas = 1
lambdas = np.linspace(1,10,nb_lambdas)
print('Do Sharpness expt!')
sharpness_inspector = LandscapeInspector(net_original,net_pert, nb_epochs,iterations, trainloader,testloader, optimizer,
criterion,error_criterion, device, lambdas,save_all_learning_curves=save_all_learning_curves,save_all_perts=save_all_perts)
sharpness_inspector.do_sharpness_experiment()
elif args.train_alg == 'flatness_bs':
''' BS params '''
r_initial = 50
epsilon = args.epsilon ## check if this number is good
# nb_radius_samples = nb_epochs could use this number as a cap of # iterations of BS
expt_path = os.path.join(expt_path+f'_BS')
''' Do BS '''
precision = 0.001
nb_dirs = args.nb_dirs
# stats_collector = StatsCollector(net,nb_dirs,nb_epochs) TODO
rand_inspector = RandLandscapeInspector(epsilon,net,r_initial,device,criterion,error_criterion,trainloader,testloader,iterations)
rand_inspector.get_faltness_radii_for_isotropic_directions(nb_dirs=nb_dirs,precision=precision)
other_stats = dict({'nb_dirs':nb_dirs,'flatness_radii':rand_inspector.flatness_radii},**other_stats)
elif args.train_alg == 'evaluate_nets':
plot = False
print('')
iterations = inf
print(f'W_nl = {W_nl}')
print(f'W_rlnl = {W_rlnl}')
''' train errors '''
loss_nl_train, error_nl_train = evalaute_mdl_data_set(criterion, error_criterion, net_nl, trainloader, device, iterations)
loss_rlnl_train, error_rlnl_train = evalaute_mdl_data_set(criterion,error_criterion,net_rlnl,trainloader,device,iterations)
print(f'loss_nl_train, error_nl_train = {loss_nl_train, error_nl_train}')
print(f'loss_rlnl_train, error_rlnl_train = {loss_rlnl_train, error_rlnl_train}')
''' test errors '''
loss_nl_test, error_nl_test = evalaute_mdl_data_set(criterion, error_criterion, net_nl, testloader, device, iterations)
loss_rlnl_test, error_rlnl_test = evalaute_mdl_data_set(criterion,error_criterion,net_rlnl,testloader,device,iterations)
print(f'loss_nl_test, error_nl_test = {loss_nl_test, error_nl_test}')
print(f'loss_rlnl_test, error_rlnl_test = {loss_rlnl_test, error_rlnl_test}')
''' '''
store_results = False
store_net = False
# elif args.train_alg == 'reach_target_loss':
# iterations = inf
# precision = 0.00001
# ''' set target loss '''
# loss_rlnl_train, error_rlnl_train = evalaute_mdl_data_set(criterion, error_criterion, net_rlnl, trainloader,device, iterations)
# target_train_loss = loss_rlnl_train
# ''' do SGD '''
# trainer = Trainer(trainloader,testloader, optimizer,criterion,error_criterion, stats_collector, device)
# last_errors = trainer.train_and_track_stats(net,nb_epochs,iterations=iterations,target_train_loss=target_train_loss,precision=precision)
# ''' Test the Network on the test data '''
# train_loss_epoch, train_error_epoch, test_loss_epoch, test_error_epoch = last_errors
# print(f'train_loss_epoch={train_loss_epoch} train_error_epoch={train_error_epoch}')
# print(f'test_loss_epoch={test_loss_epoch} test_error_epoch={test_error_epoch}')
# st()
elif args.train_alg == 'no_train':
print('NO TRAIN BRANCH')
print(f'expt_path={expt_path}')
utils.make_and_check_dir(expt_path)
''' save times '''
seconds_training, minutes_training, hours_training = utils.report_times(training_time,meta_str='training')
other_stats = dict({'seconds_training': seconds_training, 'minutes_training': minutes_training, 'hours_training': hours_training}, **other_stats)
seconds, minutes, hours = seconds_training+seconds_setup, minutes_training+minutes_setup, hours_training+hours_setup
other_stats = dict({'seconds':seconds,'minutes':minutes,'hours':hours}, **other_stats)
print(f'nb_epochs = {nb_epochs}')
print(f'Finished Training, hours={hours}')
print(f'seed = {seed}, githash = {githash}')
''' save results from experiment '''
store_results = not args.dont_save_expt_results
print(f'ALL other_stats={other_stats}')
if store_results:
print(f'storing results!')
matlab_path_to_filename = os.path.join(expt_path,matlab_file_name)
save2matlab.save2matlab_flatness_expt(matlab_path_to_filename, stats_collector,other_stats=other_stats)
''' save net model '''
if store_net:
print(f'saving final net mdl!')
net_path_to_filename = os.path.join(expt_path,net_file_name)
torch.save(net,net_path_to_filename)
''' check the error of net saved '''
loss_original, error_original = evalaute_mdl_data_set(criterion, error_criterion, net, trainloader,device)
restored_net = utils.restore_entire_mdl(net_path_to_filename)
loss_restored,error_restored = evalaute_mdl_data_set(criterion,error_criterion,restored_net,trainloader,device)
print()
print(f'net_path_to_filename = {net_path_to_filename}')
print(f'loss_original={loss_original},error_original={error_original}\a')
print(f'loss_restored={loss_restored},error_restored={error_restored}\a')
''' send e-mail '''
if hostname == 'polestar' or args.email:
message = f'SLURM Job_id=MANUAL Name=flatness_expts.py Ended, ' \
f'Total Run time hours:{hours},minutes:{minutes},seconds:{seconds} COMPLETED, ExitCode [0-0]'
utils.send_email(message,destination='*****@*****.**')
''' plot '''
if sj == 0 and plot:
#TODO
plot_utils.plot_loss_and_accuracies(stats_collector)
plt.show()
testloader = torch.utils.data.DataLoader(testset,
batch_size=1024,
shuffle=False,
num_workers=10)
''' Criterion '''
error_criterion = metrics.error_criterion
criterion = torch.nn.CrossEntropyLoss()
iterations = math.inf
''' Nets'''
net = utils.restore_entire_mdl(path).cuda()
#net2 = utils.restore_entire_mdl(path).cuda()
#net3 = utils.restore_entire_mdl(path).cuda()
''' stats about the nets '''
train_loss_epoch, train_error_epoch = evalaute_mdl_data_set(
criterion, error_criterion, net, trainloader, device)
test_loss_epoch, test_error_epoch = evalaute_mdl_data_set(
criterion, error_criterion, net, testloader, device)
nb_params = nn_mdls.count_nb_params(net)
''' print net stats '''
print(
f'train_loss_epoch, train_error_epoch = {train_loss_epoch}, {train_error_epoch}'
)
print(
f'test_loss_epoch, test_error_epoch = {test_loss_epoch}, {test_error_epoch}'
)
print(f'nb_params {nb_params}')
''' '''
st()
''' END '''
print('end no issue \a')
def main(plot=False):
''' date parameters setup'''
today_obj = date.today() # contains datetime.date(year, month, day); accessible via .day etc
day = today_obj.day
month = calendar.month_name[today_obj.month]
start_time = time.time()
''' '''
label_corrupt_prob = 0
results_root = './test_runs_flatness'
expt_path = f'flatness_label_corrupt_prob_{label_corrupt_prob}_debug2'
matlab_file_name = f'flatness_{day}_{month}'
''' '''
nb_epochs = 200
batch_size = 256
#batch_size_train,batch_size_test = batch_size,batch_size
batch_size_train = batch_size
batch_size_test = 256
data_path = './data'
num_workers = 2 # how many subprocesses to use for data loading. 0 means that the data will be loaded in the main process.
''' get data set '''
standardize = True
trainset,trainloader, testset,testloader, classes = data_class.get_cifer_data_processors(data_path,batch_size_train,batch_size_test,num_workers,label_corrupt_prob,standardize=standardize)
''' get NN '''
mdl = 'cifar_10_tutorial_net'
mdl = 'BoixNet'
mdl = 'LiaoNet'
##
print(f'model = {mdl}')
if mdl == 'cifar_10_tutorial_net':
do_bn = False
net = nn_mdls.Net()
elif mdl == 'BoixNet':
do_bn=False
## conv params
nb_filters1,nb_filters2 = 32, 32
kernel_size1,kernel_size2 = 5,5
## fc params
nb_units_fc1,nb_units_fc2,nb_units_fc3 = 512,256,len(classes)
C,H,W = 3,32,32
net = nn_mdls.BoixNet(C,H,W,nb_filters1,nb_filters2, kernel_size1,kernel_size2, nb_units_fc1,nb_units_fc2,nb_units_fc3,do_bn)
elif mdl == 'LiaoNet':
do_bn=False
nb_conv_layers=3
## conv params
Fs = [32]*nb_conv_layers
Ks = [5]*nb_conv_layers
## fc params
FC = len(classes)
C,H,W = 3,32,32
net = nn_mdls.LiaoNet(C,H,W,Fs,Ks,FC,do_bn)
# elif mdl == 'MMNISTNet':
# net = MMNISTNet()
if args.enable_cuda:
net.cuda()
nb_params = nn_mdls.count_nb_params(net)
''' Cross Entropy + Optmizer'''
lr = 0.01
momentum = 0.0
#error_criterion = metrics.error_criterion
error_criterion = metrics.error_criterion2
#criterion = torch.nn.CrossEntropyLoss()
criterion = torch.nn.MultiMarginLoss()
#loss = torch.nn.MSELoss(size_average=True)
optimizer = optim.SGD(net.parameters(), lr=lr, momentum=momentum)
''' stats collector '''
stats_collector = tr_alg.StatsCollector(net,None,None)
dynamic_stats = ['final_act',(np.zeros(10,nb_epochs),tr_alg.)]
''' Train the Network '''
print(f'----\nSTART training: label_corrupt_prob={label_corrupt_prob},nb_epochs={nb_epochs},batch_size={batch_size},lr={lr},mdl={mdl},batch-norm={do_bn},nb_params={nb_params}')
overparametrized = len(trainset)<nb_params # N < W ?
print(f'Model over parametrized? N, W = {len(trainset)} vs {nb_params}')
print(f'Model over parametrized? N < W = {overparametrized}')
# We simply have to loop over our data iterator, and feed the inputs to the network and optimize.
#tr_alg.train_cifar(args, nb_epochs, trainloader,testloader, net,optimizer,criterion)
train_loss_epoch, train_error_epoch, test_loss_epoch, test_error_epoch = tr_alg.train_and_track_stats2(args, nb_epochs, trainloader,testloader, net,optimizer,criterion,error_criterion, stats_collector)
seconds,minutes,hours = utils.report_times(start_time)
print(f'Finished Training, hours={hours}')
''' Test the Network on the test data '''
print(f'train_loss_epoch={train_loss_epoch} \ntrain_error_epoch={train_error_epoch} \ntest_loss_epoch={test_loss_epoch} \ntest_error_epoch={test_error_epoch}')
''' save results from experiment '''
other_stats = {'nb_epochs':nb_epochs,'batch_size':batch_size,'mdl':mdl,'lr':lr,'momentum':momentum}
save2matlab.save2matlab_flatness_expt(results_root,expt_path,matlab_file_name, stats_collector,other_stats=other_stats)
''' save net model '''
path = os.path.join(results_root,expt_path,f'net_{day}_{month}')
utils.save_entire_mdl(path,net)
restored_net = utils.restore_entire_mdl(path)
loss_restored,error_restored = tr_alg.evalaute_mdl_data_set(criterion,error_criterion,restored_net,testloader,args.enable_cuda)
print(f'\nloss_restored={loss_restored},error_restored={error_restored}\a')
''' plot '''
if plot:
#TODO
plot_utils.plot_loss_and_accuracies(stats_collector)
plt.show()