def train():
args = cli_options()
fr = FlowerRecognizor(args.arch, args.hidden_units, args.learning_rate,
args.gpu)
train_loader, valid_loader, test_loader, class_to_idx = create_data_loaders(
args.data_directory)
fr.train(args.save_directory, train_loader, valid_loader, class_to_idx,
args.epochs)
fr.test(test_loader)
import torch
import numpy as np
import utils
from args import get_args
from data.transforms import complex_abs
import matplotlib.pyplot as plt
args = get_args()
train_loader, dev_loader = utils.create_data_loaders(args, if_shuffle=False)
prev_file = ""
tensor_dict = {'blurred_images': [], 'target_images': []}
print("All the outputs will be stored in the folder: ", args.out_dir)
for i, data in enumerate(train_loader):
original_kspace, masked_kspace, mask, target, fname, slice_index = data
fname = fname[0]
blurred_image = utils.kspaceto2dimage(masked_kspace,
cropping=True,
resolution=args.resolution)
if prev_file != fname and prev_file != "":
utils.save_tensors(tensor_dict, args.out_dir, prev_file)
tensor_dict['blurred_images'] = [blurred_image.squeeze()]
tensor_dict['target_images'] = [target.squeeze()]
prev_file = fname
elif prev_file == fname:
tensor_dict['blurred_images'].append(blurred_image.squeeze())
tensor_dict['target_images'].append(target.squeeze())
elif prev_file == "":
def main():
"""
Main call function for the script
:return:
"""
import os
source_path = os.path.abspath(__file__)
base_path = os.path.dirname(os.path.dirname(source_path))
args = get_args()
if "params/" not in args.config:
args.config = "../params/" + args.config
with open(os.path.join(base_path, args.config), 'r',
encoding='utf-8') as config_file:
json_string = config_file.read()
if args.split.lower() == 'cluster':
random_split = False
print("Using cluster split for train and validation")
else:
random_split = True
print("Using random split for train and validation")
import os
source_path = os.path.abspath(__file__)
os.chdir(os.path.dirname(source_path))
params = json.loads(json_string)
print('Neuraldecipher training with param settings:')
print(params)
if params['neuraldecipher'].get('norm_before') is None:
params['neuraldecipher']['norm_before'] = True
# instantiate neuraldecipher model
neuraldecipher = Neuraldecipher(**params['neuraldecipher'])
print("Neuraldecipher model:")
print(neuraldecipher)
# instantiate trainer object
trainer = Trainer(model=neuraldecipher, trainparams=params['training'])
earlystopping = EarlyStopping(mode='min',
patience=params['training']['patience'])
optimizer = torch.optim.Adam(
params=neuraldecipher.parameters(),
betas=(params['training']['b1'], params['training']['b2']),
lr=params['training']['lr'],
weight_decay=params['training']['weight_decay'])
scheduler = ReduceLROnPlateau(optimizer,
mode='min',
factor=0.70,
patience=10,
verbose=True)
if params['training']['loss'] == 'mse':
criterion = torch.nn.MSELoss()
elif params['training']['loss'] == 'x-sigmoid':
criterion = XSigmoidLoss()
elif params['training']['loss'] == 'x-tanh':
criterion = XTanhLoss()
elif params['training']['loss'] == 'log-cosh':
criterion = LogCoshLoss()
else:
criterion = torch.nn.MSELoss()
if str_to_bool(args.cosineloss):
criterion_2 = CosineSimLoss()
criteria = [criterion, criterion_2]
print("Using {} and cosine difference loss.".format(
params['training']['loss']))
else:
criteria = [criterion]
print("Using {} loss.".format(params['training']['loss']))
seed = params['training']['seed']
torch.manual_seed(seed)
np.random.seed(seed)
if 'cuda' in params['training']['device']:
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# obtain datasets for training and validation
train_data, test_data = create_train_and_test_set(
ecfp_path=params['training']['data_dir'],
test_group=7,
random_split=random_split)
# create dataloaders
train_loader, test_loader = create_data_loaders(
train_data,
test_data,
batch_size=params['training']['batch_size'],
num_workers=args.num_workers)
trainer._train(criteria,
earlystopping,
scheduler,
optimizer,
train_loader,
test_loader,
verbose=True)
import torch
import torch.cuda as cuda
import torch.nn as nn
from skimage.measure import compare_ssim as ssim
from tensorboardX import SummaryWriter
from torch.autograd import Variable
from torch.nn import functional as F
from torch.utils.data import Dataset
import utils
from data import transforms
from anet_model import AnetModel
from args import get_args
args = get_args()
train_loader, dev_loader = utils.create_data_loaders(args)
# ### Custom dataset class
def build_model(args):
model = AnetModel(in_chans=2,
out_chans=2,
chans=args.num_chans,
num_pool_layers=args.num_pools,
drop_prob=args.drop_prob).to(args.device)
return model
# def build_optim(args, params):
# optimizer = torch.optim.RMSprop(params, args.learning_rate, weight_decay=args.weight_decay)
def train():
# initiate command line arguments, configuration file and logging block
args = parse_args()
config = read_config()
try:
if args.overwrite:
shutil.rmtree(f"./logs/{args.name}", ignore_errors=True)
os.mkdir(f"./logs/{args.name}")
except:
print(f"log folder {args.name} already exits.")
init_logging(log_path=f"./logs/{args.name}")
# determine train model on which device, cuda or cpu
device = 'cuda' if torch.cuda.is_available() else 'cpu'
logger.info(f"running training on {device}")
device += f':{args.main_cuda}'
# prepare training and validation datasets
logger.info('creating dataset and data loaders')
dataset = args.dataset
train_dataset = AerialDataset("train", dataset,
config[dataset]["train"]["image_path"],
config[dataset]["train"]["mask_path"])
val_dataset = AerialDataset("val", dataset,
config[dataset]["val"]["image_path"],
config[dataset]["val"]["mask_path"])
train_loader, train_metrics_loader, val_metrics_loader = create_data_loaders(
train_dataset=train_dataset,
val_dataset=val_dataset,
num_workers=config["num_workers"],
batch_size=config["batchsize"],
)
# create model
logger.info(
f'creating BiseNetv2 and optimizer with initial lr of {config["learning_rate"]}'
)
model = BiSeNetV2(config["n_classes"])
model = nn.DataParallel(model,
device_ids=[x for x in range(args.main_cuda, 4)
]).to(device)
# initiate loss function and optimizer
optimizer_fn = init_optimizer(config)
optimizer = optimizer_fn(model.parameters(), lr=config["learning_rate"])
logger.info('creating trainer and evaluator engines')
_loss_fn = init_loss(config["loss_fn"])
loss_fn = LossWithAux(_loss_fn)
# create trainer and evaluator wiht ignite.engine
trainer = engine.create_supervised_trainer(
model=model,
optimizer=optimizer,
loss_fn=loss_fn,
device=device,
non_blocking=True,
)
evaluator = engine.create_supervised_evaluator(
model=model,
metrics={
'loss':
metrics.Loss(nn.CrossEntropyLoss()),
"[email protected]":
metrics.Accuracy(thresholded_transform(0.3)),
"[email protected]":
metrics.Accuracy(thresholded_transform(0.3)),
"IOU":
metrics.IoU(
metrics.ConfusionMatrix(num_classes=config["n_classes"])),
"mIOU":
metrics.mIoU(
metrics.ConfusionMatrix(num_classes=config["n_classes"])),
},
device=device,
non_blocking=True,
output_transform=lambda x, y, y_pred:
(torch.sigmoid(y_pred["out"]), y),
)
# attach event listener to do post process after each iteration and epoch
logger.info(f'creating summary writer with tag {config["model_tag"]}')
writer = tensorboard.SummaryWriter(log_dir=f'logs/{config["model_tag"]}')
# logger.info('attaching lr scheduler')
# lr_scheduler = optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.9)
# attach_lr_scheduler(trainer, lr_scheduler, writer)
logger.info('attaching event driven calls')
attach_model_checkpoint(trainer, {config["model_tag"]: model.module},
args.name)
attach_training_logger(trainer, writer=writer)
attach_metric_logger(trainer, evaluator, 'train', train_metrics_loader,
writer)
attach_metric_logger(trainer, evaluator, 'val', val_metrics_loader, writer)
# start training (evaluation is included too)
logger.info('training...')
trainer.run(train_loader, max_epochs=config["epochs"])
def main(exp_name="cifar_for_images", load=False):
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print(device)
seed_everything()
print(torch.cuda.get_device_name(0))
torch.seed()
if str(my_computer) == "False":
n_clusters = None
if os.environ["dataset_name"] == "imagenet":
n_clusters = 512
os.environ["batch_size"] = str(512)
elif os.environ["dataset_name"] == "cifar10":
n_clusters = 20
elif os.environ["dataset_name"] == "tiny_imagenet":
n_clusters = 200
os.environ["n_cluster"] = str(n_clusters)
else:
os.environ["n_cluster"] = "10"
print(f"n clustrs is {os.environ['n_cluster']}")
print(f"batch size is {os.environ['batch_size']}")
n_classes = None
if os.environ["dataset_name"] == "imagenet":
n_classes = 1000
elif os.environ["dataset_name"] == "cifar10":
n_classes = 10
elif os.environ["dataset_name"] == "tiny_imagenet":
n_classes = 200
if network_to_use == "DenseNet":
models = [
DenseNet(
200,
clustering_algorithm=clustering_algorithms.KmeanSklearnByBatch(
n_clusters=int(os.environ['n_cluster']))),
DenseNet(200)
]
# models = [DenseNet(200),resnet50(num_classes=200,pretrained=False)]
base_lrs = [0.0001, 0.00001, 0.00001, 0.000001]
max_lrs = [0.0006, 0.00006, 0.00006, 0.000006]
step_sizes_up = [4686, 4686, 3128, 1564]
ths = [0.52, 0.61, 0.62, 0.99]
optimizer1 = torch.optim.RMSprop(models[0].parameters(),
lr=0.0001,
eps=1e-08,
weight_decay=2e-4)
scheduler1 = chainedCyclicLr(optimizer=optimizer1,
base_lrs=base_lrs,
max_lrs=max_lrs,
step_sizes_up=step_sizes_up,
ths=ths)
optimizer2 = torch.optim.RMSprop(models[1].parameters(),
lr=0.0001,
eps=1e-08,
weight_decay=2e-4)
scheduler2 = chainedCyclicLr(optimizer=optimizer2,
base_lrs=base_lrs,
max_lrs=max_lrs,
step_sizes_up=step_sizes_up,
ths=ths)
loss_func = nn.NLLLoss
elif network_to_use == "ResNet50":
models = [
resnet50(
num_classes=n_classes,
clustering_algorithm=clustering_algorithms.KmeanSklearnByBatch(
n_clusters=int(os.environ['n_cluster'])),
pretrained=False),
resnet50(num_classes=n_classes, pretrained=False)
]
fake = torch.optim.SGD(models[0].parameters(),
lr=0.001,
momentum=0.9,
nesterov=True,
weight_decay=5e-4)
optimizer1 = torch.optim.Adam(models[0].parameters(),
lr=0.001,
betas=(0.9, 0.999),
eps=1e-08,
weight_decay=0,
amsgrad=False)
scheduler1 = torch.optim.lr_scheduler.CyclicLR(fake,
base_lr=0.00001,
max_lr=0.01,
step_size_up=5000,
mode="triangular2")
optimizer2 = torch.optim.Adam(models[1].parameters(),
lr=0.001,
betas=(0.9, 0.999),
eps=1e-08,
weight_decay=0,
amsgrad=False)
scheduler2 = torch.optim.lr_scheduler.CyclicLR(fake,
base_lr=0.00001,
max_lr=0.01,
step_size_up=5000,
mode="triangular2")
loss_func = nn.CrossEntropyLoss
else:
models = []
for model_idx, model in enumerate(models):
print(f"copy model {model_idx} to device")
# print(model)
# print(f"model {model_idx} total param is {sum(p.numel() for p in model.parameters())}")
# print(f"model {model_idx} traineble param is {sum(p.numel() for p in model.parameters() if p.requires_grad)}")
model.to(device=device)
# print(os.popen('nvidia-smi').read())
train_dls, eval_dls, test_dls = [], [], []
# create cluster resnet data
if os.environ["dataset_name"] == "imagenet":
data_root = "/home/ML_courses/datasets/imagenet/"
train_set_normal, test_set = utils.ImageNetDs(
data_root=data_root, max_index=500,
do_aug=True), utils.ImageNetDs(data_root=data_root,
is_train=False,
is_eval=False,
do_aug=False)
train_set_clustered, eval_set = utils.ImageNetDs(
data_root=data_root, max_index=400,
do_aug=False), utils.ImageNetDs(data_root=data_root,
is_eval=True,
is_train=False,
max_index=400,
do_aug=False)
# train_set_normal, test_set = utils.DS_by_batch(
# data_root=os.path.join(os.path.dirname(os.getcwd()), "data", "data_clustering", "imagenet"),
# max_index=10), utils.DS_by_batch(
# data_root=os.path.join(os.path.dirname(os.getcwd()), "data", "data_clustering", "imagenet"), is_train=False,
# is_eval=False)
# train_set_clustered, eval_set = utils.DS_by_batch(
# data_root=os.path.join(os.path.dirname(os.getcwd()), "data", "data_clustering", "imagenet"),
# max_index=9), utils.DS_by_batch(
# data_root=os.path.join(os.path.dirname(os.getcwd()), "data", "data_clustering", "imagenet"), is_eval=True,
# is_train=False)
elif os.environ["dataset_name"] == "cifar10":
train_set_normal, test_set = utils.Cifar10Ds(
data_root=os.path.join(os.path.dirname(os.getcwd()), "data",
"data_clustering"),
max_index=5), utils.Cifar10Ds(data_root=os.path.join(
os.path.dirname(os.getcwd()), "data", "data_clustering"),
is_train=False,
is_eval=False)
train_set_clustered, eval_set = utils.Cifar10Ds(
data_root=os.path.join(os.path.dirname(os.getcwd()), "data",
"data_clustering"),
max_index=4), utils.Cifar10Ds(data_root=os.path.join(
os.path.dirname(os.getcwd()), "data", "data_clustering"),
is_eval=True,
is_train=False)
elif os.environ["dataset_name"] == "tiny_imagenet":
data_root = "/content/tiny-imagenet-200" if my_computer == "False" else os.path.join(
os.path.dirname(os.getcwd()), "data", "data_clustering",
"tiny-imagenet-200")
train_set_normal, test_set = utils.TinyInDs(
data_root=data_root, max_index=500,
do_aug=True), utils.TinyInDs(data_root=data_root,
is_train=False,
is_eval=False,
do_aug=False)
train_set_clustered, eval_set = utils.TinyInDs(
data_root=data_root, max_index=400,
do_aug=False), utils.TinyInDs(data_root=data_root,
is_eval=True,
is_train=False,
max_index=400,
do_aug=False)
else:
raise Exception("1")
tb = utils.Tb(exp_name=exp_name)
print("clustreee")
if str(my_computer) == "True":
start_clustering = 6
else:
if os.environ["dataset_name"] == "imagenet":
start_clustering = 20000
elif os.environ["dataset_name"] == "cifar10":
start_clustering = 200
elif os.environ["dataset_name"] == "tiny_imagenet":
start_clustering = 3000
clustered_smapler = ClusteredSampler(train_set_normal, tb=tb)
train_dl, eval_dl, test_dl = utils.create_data_loaders(
[train_set_clustered, eval_set, test_set], [
RegularSampler(train_set_clustered),
RegularSampler(eval_set),
RegularSampler(test_set)
])
train_dls.append(train_dl)
eval_dls.append(eval_dl)
test_dls.append(test_dl)
# normal resnet data
train_dl, eval_dl, test_dl = utils.create_data_loaders(
[train_set_normal, [], test_set],
[RegularSampler(train_set_normal), None,
RegularSampler(test_set)])
train_dls.append(train_dl)
eval_dls.append(eval_dl)
test_dls.append(test_dl)
trainer = Trainer(models=models,
train_dls=train_dls,
eval_dls=eval_dls,
test_dls=test_dls,
loss_fn=loss_func(),
loss_fn_eval=loss_func(reduction="none"),
optimizers=[optimizer1, optimizer2],
schedulers=[scheduler1, scheduler2],
num_steps=300000,
tb=tb,
load=load,
clustered_sampler=clustered_smapler,
start_clustering=start_clustering)
trainer.train_models()
