def test_img(net_g, datatest, logwriter,args, step):
net_g.eval()
# testing
test_loss = 0
correct = 0
data_loader = datatest
loss_func = additional_losses.CombinedLoss()
for batch_idx, (images, labels, w) in enumerate(data_loader):
images, labels, w = images.type(torch.FloatTensor), labels.type(torch.LongTensor), w.type(torch.FloatTensor)
if args.gpu != -1:
images, labels, w = images.cuda(), labels.cuda(), w.cuda()
log_probs = net_g(images)
if (batch_idx%200 ==0):
log_images(images, labels, log_probs,logwriter,step, batch_idx)
# sum up batch loss
loss = loss_func(log_probs, labels, w).item()
test_loss += loss
#logwriter.add_scalar('Test Loss', loss, batch_idx)
# get the index of the max log-probability
test_loss /= len(data_loader.dataset)
logwriter.add_scalar('Test Loss', test_loss, step)
def __init__(self,
model,
exp_name,
device,
num_class,
optim=torch.optim.SGD,
optim_args={},
loss_func=additional_losses.CombinedLoss(),
model_name='quicknat',
labels=None,
num_epochs=10,
log_nth=5,
lr_scheduler_step_size=5,
lr_scheduler_gamma=0.5,
use_last_checkpoint=True,
exp_dir='experiments',
log_dir='logs',
arch_file_path=None):
self.device = device
self.model = model
# self.swa_model = torch.optim.swa_utils.AveragedModel(self.model)
self.model_name = model_name
self.labels = labels
self.num_epochs = num_epochs
if torch.cuda.is_available():
self.loss_func = loss_func.cuda(device)
else:
self.loss_func = loss_func
self.optim = optim(model.parameters(), **optim_args)
# self.scheduler = lr_scheduler.StepLR(self.optim, step_size=lr_scheduler_step_size,
# gamma=lr_scheduler_gamma)
self.scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(self.optim, T_max=100)
# self.swa_start = -1 #int(np.round(self.num_epochs*0.75))
# print(self.swa_start)
# self.swa_scheduler = torch.optim.swa_utils.SWALR(self.optim, swa_lr=0.05)
exp_dir_path = os.path.join(exp_dir, exp_name)
common_utils.create_if_not(exp_dir_path)
common_utils.create_if_not(os.path.join(exp_dir_path, CHECKPOINT_DIR))
self.exp_dir_path = exp_dir_path
self.save_architectural_files(arch_file_path)
self.log_nth = log_nth
self.logWriter = LogWriter(num_class, log_dir, exp_name, use_last_checkpoint, labels)
# self.wandb = wandb
self.use_last_checkpoint = use_last_checkpoint
self.start_epoch = 1
self.start_iteration = 1
self.best_ds_mean = 0
self.best_ds_mean_epoch = 0
if use_last_checkpoint:
self.load_checkpoint()
print(self.best_ds_mean, self.best_ds_mean_epoch, self.start_epoch)
def __init__(self,
model,
exp_name,
device,
num_class,
optim=torch.optim.Adam,
optim_args={},
loss_func=additional_losses.CombinedLoss(),
model_name='quicknat',
labels=None,
num_epochs=10,
log_nth=5,
lr_scheduler_step_size=5,
lr_scheduler_gamma=0.5,
use_last_checkpoint=True,
exp_dir='experiments',
log_dir='logs'):
self.device = device
self.model = model
self.model_name = model_name
self.labels = labels
self.num_epochs = num_epochs
if torch.cuda.is_available():
self.loss_func = loss_func.cuda(device)
else:
self.loss_func = loss_func
self.optim = optim(model.parameters(), **optim_args)
self.scheduler = lr_scheduler.StepLR(self.optim,
step_size=lr_scheduler_step_size,
gamma=lr_scheduler_gamma)
exp_dir_path = os.path.join(exp_dir, exp_name)
common_utils.create_if_not(exp_dir_path)
common_utils.create_if_not(os.path.join(exp_dir_path, CHECKPOINT_DIR))
self.exp_dir_path = exp_dir_path
self.log_nth = log_nth
self.logWriter = LogWriter(num_class, log_dir, exp_name,
use_last_checkpoint, labels)
self.use_last_checkpoint = use_last_checkpoint
self.start_epoch = 1
self.start_iteration = 1
self.best_ds_mean = 0
self.best_ds_mean_epoch = 0
if use_last_checkpoint:
self.load_checkpoint()
def test_img(self, net_g, dataLoader, logwriter, args, user, step):
net_g.eval()
# testing
test_loss = 0
data_loader = dataLoader
loss_func = additional_losses.CombinedLoss()
for batch_idx, (images, labels, w) in enumerate(data_loader):
images, labels, w = images.type(torch.FloatTensor), labels.type(
torch.LongTensor), w.type(torch.FloatTensor)
if args.gpu != -1:
images, labels, w = images.cuda(), labels.cuda(), w.cuda()
log_probs = net_g(images)
#if (batch_idx%500 ==0 and (step - self.epoch)%3 == 0):
#log_images(images, labels, log_probs,logwriter, step)
# sum up batch loss
loss = loss_func(log_probs, labels, w).item()
test_loss += loss
test_loss /= len(data_loader.dataset)
logwriter.add_scalar('Test_Loss_User {:3d}'.format(user), test_loss,
step)
def __init__(self,
args,
dataset=None,
idxs=None,
logwriter=None,
user_id=None,
testLoader=None,
epoch=None):
self.args = args
self.logwriter = logwriter
self.user_id = user_id
self.test_loader = testLoader
self.epoch = epoch
#weights = [1.80434783, 0.34583333, 1.25757576, 0.84693878, 1.06410256, 1.09210526, 0.25, 3.45833333, 0.32170543, 0.94318182]
#class_weights = torch.FloatTensor(weights).cuda()
self.loss_func = additional_losses.CombinedLoss()
self.selected_clients = []
#self.ldr_train = DataLoader(DatasetSplit(dataset, idxs), batch_size=self.args.local_bs, shuffle=True)
self.ldr_train = DataLoader(dataset,
batch_size=self.args.local_bs,
shuffle=True)
