def main(path_continue_learning=None, total_epoch=0, new_name=None):
"""
:param path_continue_learning: Path were the network is already saved
(don t use if it is the beginning of the training)
:param total_epoch: Number of epoch needed don t use if it is the beginning of the training)
:param new_name: New name of the network, if we want to use again a network already train.
:return: Nothing but train the network and save CSV files for the error and also save the network regularly
"""
# Manual seed of the network to have reproducible experiment
torch.manual_seed(26542461)
# If the network was already train we import it
if path_continue_learning is not None:
# Load the trained Network
parameters, network = Save_import.load_from_checkpoint(path_checkpoint=path_continue_learning)
# Here we can change some parameters, the only one necessary is the total_epoch
parameters.epoch_total = total_epoch
# parameters.learning_rate_decay = - 4.5 * 10 ** (-5)
# parameters.batch_size = 4
# parameters.batch_size_val = 4
# parameters.learning_rate = 0.01
# parameters.momentum_IoU = 0
# parameters.loss = "IoU_Lovasz"
# Put weight to GPU
if torch.cuda.is_available():
parameters.weight_grad = parameters.weight_grad.cuda()
# If a new name is define, we create new CSV files associated and change the name of the network
if new_name is not None:
# Init the csv file that will store the error, this time we make a copy of the existing error
Save_import.duplicated_csv(path_CSV=parameters.path_CSV,
old_name_network=parameters.name_network,
new_name_network=new_name,
train_number=parameters.train_number)
parameters.name_network = new_name
with open(parameters.path_print, 'w') as txtfile:
txtfile.write('\n The program will continue \n')
# If the network was not train, we start from scratch
else:
# Define the weight
weight_grad = torch.FloatTensor([2.381681e+09, 3.856594e+08, 1.461642e+09, 4.291781e+07,
5.597591e+07, 8.135516e+07, 1.328548e+07, 3.654657e+07,
1.038652e+09, 7.157456e+07, 2.527450e+08, 7.923985e+07,
9.438758e+06, 4.460595e+08, 1.753254e+07, 1.655341e+07,
1.389560e+07, 6.178567e+06, 2.936571e+07])
sum_grad = weight_grad.sum()
# normalize and then take the invert
for i in range(weight_grad.size(0)):
weight_grad[i] = sum_grad / weight_grad[i]
# Normalize again and mult by the number of classes
weight_grad = (weight_grad / weight_grad.sum()) * weight_grad.size(0)
# if you want to keep the wiehgt, comment the next line
weight_grad = torch.FloatTensor([1 for i in range(19)])
# Define all the parameters
parameters = Parameters.Parameters(nColumns=8,
nFeatMaps=[16, 32, 64, 128, 256],
nFeatureMaps_init=3,
number_classes=20 - 1,
label_DF=Label.create_label(),
width_image_initial=2048, height_image_initial=1024,
size_image_crop=401,
dropFactor=0.1,
learning_rate=0.01,
learning_rate_decay=1 * (10 ** (-2)),
weight_decay=0,
beta1=0.9,
beta2=0.999,
epsilon=1 * 10 ** (-8),
batch_size=5,
batch_size_val=5,
epoch_total=400,
actual_epoch=0,
ratio=(1, 1),
weight_grad=weight_grad,
loss="focal_loss",
momentum_IoU=0,
path_save_net="./Model/",
name_network="focal_loss2",
train_number=0,
path_CSV="./CSV/",
path_data="/home_expes/collections/Cityscapes/",
path_print="./Python_print_focal_loss.txt",
path_result="./Result",
num_workers=2)
# Define the GridNet
network = GridNet_structure.gridNet(nInputs=parameters.nFeatureMaps_init,
nOutputs=parameters.number_classes,
nColumns=parameters.nColumns,
nFeatMaps=parameters.nFeatMaps,
dropFactor=parameters.dropFactor)
with open(parameters.path_print, 'w') as txtfile:
txtfile.write('\n Start of the program \n')
# Init the csv file that will store the error
Save_import.init_csv(name_network=parameters.name_network,
train_number=parameters.train_number,
path_CSV=parameters.path_CSV,
path_print=parameters.path_print)
# Import both DataSets with the transformation
train_dataset = Save_import.cityscapes_create_dataset(quality='fine',
mode='train',
transform=parameters.transforms_input,
transform_target=parameters.transforms_output,
parameters=parameters)
val_dataset = Save_import.cityscapes_create_dataset(quality='fine',
mode='val',
transform=parameters.transforms_input,
transform_target=parameters.transforms_output,
parameters=parameters)
# Create the DataSets for Pytorch used
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=parameters.batch_size,
shuffle=True,
num_workers=parameters.num_workers,
drop_last=False)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=parameters.batch_size_val,
shuffle=True,
num_workers=parameters.num_workers,
drop_last=False)
# If there is more than one GPU we can use them
if torch.cuda.device_count() > 1:
with open(parameters.path_print, 'a') as txtfile:
txtfile.write("\nLet's use " + str(torch.cuda.device_count()) + " GPUs! \n")
network = torch.nn.DataParallel(network)
else:
with open(parameters.path_print, 'a') as txtfile:
txtfile.write("\nWe don t have more than one GPU \n")
# ... But we still use it in this case ? ... TODO try without to check if it is working
network = torch.nn.DataParallel(network)
# Put the network on GPU if possible
if torch.cuda.is_available():
network.cuda()
else:
with open(parameters.path_print, 'a') as txtfile:
txtfile.write("\nAccording to torch Cuda is not available \n")
# Train the network
train(network=network,
parameters=parameters,
train_loader=train_loader,
val_loader=val_loader)
def main(path_continue_learning=None, total_epoch=0, new_name=None):
"""
:param path_continue_learning: Path were the network is already saved
(don t use if it is the beginning of the training)
:param total_epoch: Number of epoch needed don t use if it is the beginning of the training)
:param new_name: New name of the network, if we want to use again a network already train.
:return: Nothing but train the network and save CSV files for the error and also save the network regularly
"""
# Manual seed of the network to have reproducible experiment
torch.manual_seed(945682461)
# If the network was already train we import it
if path_continue_learning is not None:
# Load the trained Network
parameters, network = Save_import.load_from_checkpoint(
path_checkpoint=path_continue_learning)
# Here we can change some parameters, the only one necessary is the total_epoch
parameters.epoch_total = total_epoch
# parameters.learning_rate_decay = 0.5 * (10 ** (-2))
# parameters.batch_size = 5
# parameters.batch_size_val = 5
parameters.learning_rate = 0.001
# parameters.momentum_IoU = 0.9
# If a new name is define, we create new CSV files associated and change the name of the network
if new_name is not None:
# Init the csv file that will store the error, this time we make a copy of the existing error
Save_import.duplicated_csv(
path_CSV=parameters.path_CSV,
old_name_network=parameters.name_network,
new_name_network=new_name,
train_number=parameters.train_number)
parameters.name_network = new_name
with open(parameters.path_print, 'w') as txtfile:
txtfile.write('\n The program will continue \n')
# If the network was not train, we start from scratch
else:
# Define all the parameters
parameters = Parameters.Parameters(
nColumns=6,
nFeatMaps=[16, 32, 64, 128],
nFeatureMaps_init=3,
number_classes=20 - 1,
label_DF=Label.create_label(),
width_image_initial=2048,
height_image_initial=1024,
size_image_crop=401,
dropFactor=0.1,
learning_rate=0.0001,
learning_rate_decay=1 * (10**(-2)),
weight_decay=0,
beta1=0.9,
beta2=0.999,
epsilon=1 * 10**(-8),
batch_size=40,
batch_size_val=40,
epoch_total=100,
actual_epoch=0,
ratio=(1, 1),
weight_grad=torch.FloatTensor([1 for i in range(19)]),
loss="cross_entropy_pretrain",
momentum_IoU=0,
pretrain=True,
path_save_net="./Model/",
name_network="resnet18_1000classes",
train_number=0,
path_CSV="./CSV/",
# path_data="/home_expes/collections/Cityscapes/",
path_data="/home_expes/collections/imagenet_1000_classes/",
path_print="./Python_print_resnet18_1000classes.txt",
path_result="./Result",
num_workers=2)
# Define the GridNet
network = GridNet_structure.gridNet_imagenet(
nInputs=parameters.nFeatureMaps_init,
nOutputs=parameters.number_classes,
nColumns=parameters.nColumns,
nFeatMaps=parameters.nFeatMaps,
dropFactor=parameters.dropFactor)
network = GridNet_structure.ResNet18(
nOutputs=len(Label.create_imagenet_class()))
with open(parameters.path_print, 'w') as txtfile:
txtfile.write('\n Start of the program \n')
# Init the csv file that will store the error
Save_import.init_csv(name_network=parameters.name_network,
train_number=parameters.train_number,
path_CSV=parameters.path_CSV,
path_print=parameters.path_print)
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
# Import both DataSets with the transformation
train_dataset = Save_import.cityscapes_create_dataset_pretrain(
mode="train",
parameters=parameters,
sliding_crop=None,
transform=transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]))
val_dataset = Save_import.cityscapes_create_dataset_pretrain(
mode="val",
parameters=parameters,
sliding_crop=None,
transform=transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.ToTensor(),
normalize,
]))
# Create the DataSets for Pytorch used
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=parameters.batch_size,
shuffle=True,
num_workers=parameters.num_workers,
drop_last=False)
val_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=parameters.batch_size_val,
shuffle=True,
num_workers=parameters.num_workers,
drop_last=False)
# If there is more than one GPU we can sue them
if torch.cuda.device_count() > 1:
with open(parameters.path_print, 'a') as txtfile:
txtfile.write("\nLet's use " + str(torch.cuda.device_count()) +
" GPUs! \n")
network = torch.nn.DataParallel(network)
else:
with open(parameters.path_print, 'a') as txtfile:
txtfile.write("\nWe don t have more than one GPU \n")
# ... But we still use it in this case ? ... TODO try without to check if it is still working
# network = torch.nn.DataParallel(network)
# Put the network on GPU if possible
if torch.cuda.is_available():
network.cuda()
else:
with open(parameters.path_print, 'a') as txtfile:
txtfile.write("\nAccording to torch Cuda is not even available \n")
# Train the network
train(network=network,
parameters=parameters,
train_loader=train_loader,
val_loader=val_loader)