AntoineH/smart_augmentation
1
0
Fork 0
mirror of https://github.com/AntoineHX/smart_augmentation.git synced 2025-05-04 12:10:45 +02:00
Code Issues Projects Releases Packages Wiki Activity

Modification du early stopping (sur test data...)

Browse source
This commit is contained in:
Harle, Antoine (Contracteur) 2019-11-13 12:03:54 -05:00
parent bc8e5f2817
commit 198fb06065
3 changed files with 15 additions and 19 deletions
Download patch file Download diff file Expand all files Collapse all files

4
higher/datasets.py
View file

@ -3,8 +3,8 @@ from torch.utils.data import SubsetRandomSampler
import torchvision import torchvision
BATCH_SIZE = 300 BATCH_SIZE = 300
#TEST_SIZE = 300 TEST_SIZE = 300
TEST_SIZE = 10000 #TEST_SIZE = 10000
#ATTENTION : Dataug (Kornia) Expect image in the range of [0, 1] #ATTENTION : Dataug (Kornia) Expect image in the range of [0, 1]
#transform_train = torchvision.transforms.Compose([ #transform_train = torchvision.transforms.Compose([

4
higher/test_dataug.py
View file

@ -37,8 +37,8 @@ else:
########################################## ##########################################
if __name__ == "__main__": if __name__ == "__main__":
n_inner_iter = 0 n_inner_iter = 10
epochs = 100 epochs = 200
dataug_epoch_start=0 dataug_epoch_start=0
#### Classic #### #### Classic ####

26
higher/train_utils.py
View file

@ -6,12 +6,6 @@ import higher
from datasets import * from datasets import *
from utils import * from utils import *
#Variables a definir
#dl_train = None
#dl_val = None
#dl_test = None
#device = torch.device('cuda')
def test(model): def test(model):
device = next(model.parameters()).device device = next(model.parameters()).device
model.eval() model.eval()
@ -21,13 +15,13 @@ def test(model):
pred = model.forward(features) pred = model.forward(features)
return pred.argmax(dim=1).eq(labels).sum().item() / dl_test.batch_size * 100 return pred.argmax(dim=1).eq(labels).sum().item() / dl_test.batch_size * 100
def compute_vaLoss(model, dl_val_it): def compute_loss(model, dl_it, dl):
device = next(model.parameters()).device device = next(model.parameters()).device
try: try:
xs_val, ys_val = next(dl_val_it) xs_val, ys_val = next(dl_it)
except StopIteration: #Fin epoch val except StopIteration: #Fin epoch val
dl_val_it = iter(dl_val) dl_val_it = iter(dl)
xs_val, ys_val = next(dl_val_it) xs_val, ys_val = next(dl_it)
xs_val, ys_val = xs_val.to(device), ys_val.to(device) xs_val, ys_val = xs_val.to(device), ys_val.to(device)
try: try:
@ -528,6 +522,7 @@ def run_dist_dataugV2(model, epochs=1, inner_it=0, dataug_epoch_start=0, print_f
countcopy=0 countcopy=0
val_loss=torch.tensor(0) #Necessaire si pas de metastep sur une epoch val_loss=torch.tensor(0) #Necessaire si pas de metastep sur une epoch
dl_val_it = iter(dl_val) dl_val_it = iter(dl_val)
dl_test_it = iter(dl_test) #ATTENTION A UTILISER SEULEMT POUR EARLY STOP
meta_opt = torch.optim.Adam(model['data_aug'].parameters(), lr=1e-2) meta_opt = torch.optim.Adam(model['data_aug'].parameters(), lr=1e-2)
inner_opt = torch.optim.SGD(model['model'].parameters(), lr=1e-2, momentum=0.9) inner_opt = torch.optim.SGD(model['model'].parameters(), lr=1e-2, momentum=0.9)
@ -542,7 +537,7 @@ def run_dist_dataugV2(model, epochs=1, inner_it=0, dataug_epoch_start=0, print_f
val_loss_monitor= None val_loss_monitor= None
if loss_patience != None : if loss_patience != None :
if dataug_epoch_start==-1: val_loss_monitor = loss_monitor(patience=loss_patience, end_train=2) #1st limit = dataug start if dataug_epoch_start==-1: val_loss_monitor = loss_monitor(patience=loss_patience, end_train=2) #1st limit = dataug start
else: val_loss_monitor = loss_monitor(patience=loss_patience) #Val loss monitor else: val_loss_monitor = loss_monitor(patience=loss_patience) #Val loss monitor (Not on val data : used by Dataug... => Test data)
model.train() model.train()
@ -594,7 +589,7 @@ def run_dist_dataugV2(model, epochs=1, inner_it=0, dataug_epoch_start=0, print_f
if(high_grad_track and i%inner_it==0): #Perform Meta step if(high_grad_track and i%inner_it==0): #Perform Meta step
#print("meta") #print("meta")
#Peu utile si high_grad_track = False #Peu utile si high_grad_track = False
val_loss = compute_vaLoss(model=fmodel, dl_val_it=dl_val_it) val_loss = compute_loss(model=fmodel, dl_it=dl_val_it, dl=dl_val)
#print_graph(val_loss) #print_graph(val_loss)
@ -619,7 +614,7 @@ def run_dist_dataugV2(model, epochs=1, inner_it=0, dataug_epoch_start=0, print_f
countcopy+=1 countcopy+=1
model_copy(src=fmodel, dst=model) model_copy(src=fmodel, dst=model)
optim_copy(dopt=diffopt, opt=inner_opt) optim_copy(dopt=diffopt, opt=inner_opt)
val_loss = compute_vaLoss(model=fmodel, dl_val_it=dl_val_it) val_loss = compute_loss(model=fmodel, dl_it=dl_val_it, dl=dl_val)
#Necessaire pour reset higher (Accumule les fast_param meme avec track_higher_grads = False) #Necessaire pour reset higher (Accumule les fast_param meme avec track_higher_grads = False)
fmodel = higher.patch.monkeypatch(model, device=None, copy_initial_weights=True) fmodel = higher.patch.monkeypatch(model, device=None, copy_initial_weights=True)
@ -652,9 +647,10 @@ def run_dist_dataugV2(model, epochs=1, inner_it=0, dataug_epoch_start=0, print_f
log.append(data) log.append(data)
############# #############
if val_loss_monitor : if val_loss_monitor :
val_loss_monitor.register(val_loss.item()) model.eval()
val_loss_monitor.register(compute_loss(model, dl_it=dl_test_it, dl=dl_test))#val_loss.item())
if val_loss_monitor.end_training(): break #Stop training if val_loss_monitor.end_training(): break #Stop training
model.train()
if not model.is_augmenting() and (epoch == dataug_epoch_start or (val_loss_monitor and val_loss_monitor.limit_reached()==1)): if not model.is_augmenting() and (epoch == dataug_epoch_start or (val_loss_monitor and val_loss_monitor.limit_reached()==1)):
print('Starting Data Augmention...') print('Starting Data Augmention...')