def train_fully_supervised(model,n_epochs,train_loader,val_loader,criterion,optimizer,scheduler,auto_lr,\
save_folder,model_name,benchmark=False,save_all_ep=True, save_best=False, device='cpu',num_classes=21):
"""
A complete training of fully supervised model.
save_folder : Path to save the model, the courb of losses,metric...
benchmark : enable or disable backends.cudnn
save_all_ep : if True, the model is saved at each epoch in save_folder
scheduler : if True, the model will apply a lr scheduler during training
auto_lr : Auto lr finder
"""
torch.backends.cudnn.benchmark = benchmark
if auto_lr:
print('Auto finder for the Learning rate')
lr_finder = LRFinder(model,
optimizer,
criterion,
memory_cache=False,
cache_dir='/tmp',
device=device)
lr_finder.range_test(train_loader,
start_lr=10e-5,
end_lr=10,
num_iter=100)
if scheduler:
lr_scheduler = torch.optim.lr_scheduler.LambdaLR(
optimizer, lambda x: (1 - x / (len(train_loader) * n_epochs))**0.9)
loss_test = []
loss_train = []
iou_train = []
iou_test = []
accuracy_train = []
accuracy_test = []
model.to(device)
for ep in range(n_epochs):
print("EPOCH", ep)
model.train()
state = step_train_supervised(model,train_loader=train_loader,criterion=criterion,\
optimizer=optimizer,device=device,num_classes=num_classes)
iou = state.metrics['mean IoU']
acc = state.metrics['accuracy']
loss = state.metrics['CE Loss']
loss_train.append(loss)
iou_train.append(iou)
accuracy_train.append(acc)
print('TRAIN - EP:', ep, 'iou:', iou, 'Accuracy:', acc, 'Loss CE',
loss)
if scheduler:
lr_scheduler.step()
#Eval model
model.eval()
with torch.no_grad():
state = eval_model(model,
val_loader,
device=device,
num_classes=num_classes)
iou = state.metrics['mean IoU']
acc = state.metrics['accuracy']
loss = state.metrics['CE Loss']
loss_test.append(loss)
iou_test.append(iou)
accuracy_test.append(acc)
print('TEST - EP:', ep, 'iou:', iou, 'Accuracy:', acc, 'Loss CE',
loss)
## Save model
U.save_model(model,
save_all_ep,
save_best,
save_folder,
model_name,
ep=ep,
iou=iou,
iou_test=iou_test)
U.save_curves(path=save_folder,loss_train=loss_train,iou_train=iou_train,accuracy_train=accuracy_train\
,loss_test=loss_test,iou_test=iou_test,accuracy_test=accuracy_test)
def train_rot_equiv(model,n_epochs,train_loader_sup,train_dataset_unsup,val_loader,criterion_supervised,optimizer,scheduler,\
Loss,gamma,batch_size,iter_every,save_folder,model_name,benchmark=False,angle_max=30,size_img=520,\
eval_every=5,save_all_ep=True,dataroot_voc='~/data/voc2012',save_best=False,rot_cpu=False, device='cpu',num_classes=21):
"""
A complete training of rotation equivariance supervised model.
save_folder : Path to save the model, the courb of losses,metric...
benchmark : enable or disable backends.cudnn
Loss : Loss for unsupervised training 'KL' 'CE' 'L1' or 'MSE'
gamma : float btwn [0,1] -> Balancing two losses loss_sup*gamma + (1-gamma)*loss_unsup
save_all_ep : if True, the model is saved at each epoch in save_folder
scheduler : if True, the model will apply a lr scheduler during training
eval_every : Eval Model with different input image angle every n step
size_img : size of image during evaluation
angle_max : max angle rotation for input images
"""
torch.backends.cudnn.benchmark = benchmark
if scheduler:
lr_scheduler = torch.optim.lr_scheduler.LambdaLR(
optimizer, lambda x: (1 - x /
(len(train_loader_sup) * n_epochs))**0.9)
criterion_unsupervised = U.get_criterion(Loss)
iou_train = []
iou_test = []
combine_loss_train = []
combine_loss_test = []
loss_train_unsup = []
loss_train_sup = []
loss_test_unsup = []
loss_test_sup = []
equiv_accuracy_train = []
equiv_accuracy_test = []
accuracy_test = []
accuracy_train = []
for ep in range(n_epochs):
train_loader_equiv = torch.utils.data.DataLoader(train_dataset_unsup,batch_size=batch_size,\
shuffle=True,drop_last=True)
print("EPOCH", ep)
# TRAINING
d = train_step_rot_equiv(model,train_loader_sup,train_loader_equiv,criterion_supervised,criterion_unsupervised,\
optimizer,gamma,Loss,rot_cpu=rot_cpu,device=device,angle_max=angle_max,num_classes=num_classes,iter_every=iter_every)
if scheduler:
lr_scheduler.step()
combine_loss_train.append(d['loss'])
loss_train_unsup.append(d['loss_equiv'])
loss_train_sup.append(d['loss_sup'])
equiv_accuracy_train.append(d['equiv_acc'])
iou_train.append(d['iou_train'])
accuracy_train.append(d['accuracy_train'])
print('TRAIN - EP:',ep,'iou:',d['iou_train'],'Accuracy:',d['accuracy_train'],'Loss sup:',d['loss_sup'],\
'Loss equiv:',d['loss_equiv'],'Combine Loss:',d['loss'],'Equivariance Accuracy:',d['equiv_acc'],)
# EVALUATION
model.eval()
with torch.no_grad():
state = eval_model(model,
val_loader,
device=device,
num_classes=num_classes)
iou = state.metrics['mean IoU']
acc = state.metrics['accuracy']
loss = state.metrics['CE Loss']
loss_test_sup.append(loss)
iou_test.append(iou)
accuracy_test.append(acc)
print('TEST - EP:', ep, 'iou:', iou, 'Accuracy:', acc, 'Loss CE',
loss)
# SAVING MODEL
U.save_model(model,
save_all_ep,
save_best,
save_folder,
model_name,
ep=ep,
iou=iou,
iou_test=iou_test)
if ep % eval_every == 0: # Eval loss equiv and equivariance accuracy for the validation dataset
equiv_acc, m_loss_equiv = U.eval_accuracy_equiv(model,val_loader,criterion=criterion_unsupervised,\
nclass=21,device=device,Loss=Loss,plot=False,angle_max=angle_max,random_angle=False)
loss_test_unsup.append(m_loss_equiv)
equiv_accuracy_test.append(equiv_acc)
"""
print('VOC Dataset Train')
_ = eval_model_all_angle(model,size_img,dataroot_voc,train=True,device=device,num_classes=num_classes)
print('VOC Dataset Val')
_ = eval_model_all_angle(model,size_img,dataroot_voc,train=False,device=device,num_classes=num_classes)
## Save model"""
U.save_curves(path=save_folder,combine_loss_train=combine_loss_train,loss_train_sup=loss_train_sup,\
loss_train_unsup=loss_train_unsup,iou_train=iou_train,accuracy_train=accuracy_train,equiv_accuracy_train=equiv_accuracy_train,\
combine_loss_test=combine_loss_test,loss_test_unsup=loss_test_unsup,equiv_accuracy_test=equiv_accuracy_test,\
loss_test_sup= loss_test_sup,iou_test=iou_test,accuracy_test=accuracy_test)
def train_scale_equiv(model,n_epochs,train_loader_sup,train_dataset_unsup,val_loader,criterion_supervised,optimizer,scheduler,\
Loss,gamma,batch_size,save_folder,model_name,benchmark=False,angle_max=30,size_img=520,scale_factor=(0.5,1.2),\
save_all_ep=True,dataroot_voc='~/data/voc2012',save_best=False, device='cpu',num_classes=21):
"""
A complete training of rotation equivariance supervised model.
save_folder : Path to save the model, the courb of losses,metric...
benchmark : enable or disable backends.cudnn
Loss : Loss for unsupervised training 'KL' 'CE' 'L1' or 'MSE'
gamma : float btwn [0,1] -> Balancing two losses loss_sup*gamma + (1-gamma)*loss_unsup
save_all_ep : if True, the model is saved at each epoch in save_folder
scheduler : if True, the model will apply a lr scheduler during training
eval_every : Eval Model with different input image angle every n step
size_img : size of image during evaluation
scale_factor : scale between min*size_img and max*size_img
"""
torch.backends.cudnn.benchmark = benchmark
if scheduler:
lr_scheduler = torch.optim.lr_scheduler.LambdaLR(
optimizer, lambda x: (1 - x /
(len(train_loader_sup) * n_epochs))**0.9)
criterion_unsupervised = U.get_criterion(Loss)
print('Criterion Unsupervised', criterion_unsupervised)
iou_train = []
iou_test = []
combine_loss_train = []
combine_loss_test = []
loss_train_unsup = []
loss_train_sup = []
loss_test_unsup = []
loss_test_sup = []
equiv_accuracy_train = []
equiv_accuracy_test = []
accuracy_test = []
accuracy_train = []
torch.autograd.set_detect_anomaly(True)
for ep in range(n_epochs):
train_loader_equiv = torch.utils.data.DataLoader(train_dataset_unsup,batch_size=batch_size,\
shuffle=True,drop_last=True)
print("EPOCH", ep)
# TRAINING
d = train_step_scale_equiv(model,train_loader_sup,train_loader_equiv,criterion_supervised,criterion_unsupervised,\
optimizer,gamma,Loss,device,size_img=size_img,scale_factor=scale_factor)
if scheduler:
lr_scheduler.step()
combine_loss_train.append(d['loss'])
loss_train_unsup.append(d['loss_equiv'])
loss_train_sup.append(d['loss_sup'])
equiv_accuracy_train.append(d['equiv_acc'])
iou_train.append(d['iou_train'])
accuracy_train.append(d['accuracy_train'])
print('TRAIN - EP:',ep,'iou:',d['iou_train'],'Accuracy:',d['accuracy_train'],'Loss sup:',d['loss_sup'],\
'Loss equiv:',d['loss_equiv'],'Combine Loss:',d['loss'],'Equivariance Accuracy:',d['equiv_acc'],)
# EVALUATION
model.eval()
with torch.no_grad():
state = eval_model(model,
val_loader,
device=device,
num_classes=num_classes)
iou = state.metrics['mean IoU']
acc = state.metrics['accuracy']
loss = state.metrics['CE Loss']
loss_test_sup.append(loss)
iou_test.append(iou)
accuracy_test.append(acc)
print('TEST - EP:', ep, 'iou:', iou, 'Accuracy:', acc, 'Loss CE',
loss)
U.save_curves(path=save_folder,combine_loss_train=combine_loss_train,loss_train_sup=loss_train_sup,\
loss_train_unsup=loss_train_unsup,iou_train=iou_train,accuracy_train=accuracy_train,equiv_accuracy_train=equiv_accuracy_train,\
combine_loss_test=combine_loss_test,loss_test_unsup=loss_test_unsup,equiv_accuracy_test=equiv_accuracy_test,\
loss_test_sup= loss_test_sup,iou_test=iou_test,accuracy_test=accuracy_test)
from torchvision import transforms, datasets
import numpy as np
from matplotlib import pyplot as plt
import time
import sys
sys.path.append('../python_scripts/')
from utils import save_curves_no_smooth as save_curves
from sys import argv
# the path of the log is passed after the command
exp = argv[1]
ep = argv[2]
# define a CNN
sys.path.append('../python_scripts/')
from networks import alexnet_pytorch as Net
epoch = '_epoch_' + ep
model_path = '../saving_model/exp' + exp + '/exp' + exp + epoch + '.pth.tar'
checkpoint = torch.load(model_path)
print 'Model loaded from the following path:'
print model_path
print 'Model keys:'
print checkpoint.keys()
save_curves(checkpoint, './')