def run():
print 'Inside run!'
global args
args = parser.parse_args()
print args
# add checkpoint resume option
# load datasets
train_dataset = SketchData(root=path,
train=True,
transform=None,
target_transform=None,
)
val_dataset = SketchData(root=path,
train=False,
transform=None,
target_transform=None,
)
train_loader = torch.utils.data.DataLoader(dataset=train_dataset,
batch_size=args.b,
shuffle=True)
val_loader = torch.utils.data.DataLoader(dataset=val_dataset,
batch_size=args.b,
shuffle=False)
model = AlexNet()
optimizer = torch.optim.SGD(model.parameters(), args.lr,
momentum=args.momentum,
weight_decay=args.wd)
criterion = nn.CrossEntropyLoss()
best_prec = 0
for epoch in range(args.epochs):
print 'Epoch: ' + str(epoch)
adjust_learning_rate(optimizer, epoch)
print 'Adjusted learning rate'
train(train_loader, model, criterion, optimizer, epoch)
print 'Trained!'
precision = validate(val_loader, model, criterion)
print 'Got precision!'
best_prec = max(precision.data[0], best_prec)
print 'Updated best precision!'
save_checkpoint({
'epoch': epoch + 1,
'arch': args.arch,
'best_prec1': best_prec,
'state_dict': model.state_dict(),
'optimizer' : optimizer.state_dict(),
}, (precision.data[0] > best_prec))
def __init__(self, context: DeepSpeedTrialContext) -> None:
self.context = context
self.args = AttrDict(self.context.get_hparams())
model = AlexNet(10)
model = PipelineModule(
layers=join_layers(model),
loss_fn=torch.nn.CrossEntropyLoss(),
num_stages=self.args.pipe_parallel_size,
partition_method=self.args.part,
activation_checkpoint_interval=0,
)
ds_config = overwrite_deepspeed_config(
self.args.deepspeed_config,
self.args.get("overwrite_deepspeed_args", {}))
model_engine, optimizer, _, _ = deepspeed.initialize(
args=self.args,
model=model,
model_parameters=[
p for p in model.parameters() if p.requires_grad
],
config=ds_config,
)
self.model_engine = self.context.wrap_model_engine(model_engine)
class TrainNetwork():
def __init__(self, dataset, batch_size, epochs, lr, lr_decay_epoch,
momentum):
assert (dataset == 'letters' or dataset == 'mnist')
self.dataset = dataset
self.batch_size = batch_size
self.epochs = epochs
self.lr = lr
self.lr_decay_epoch = lr_decay_epoch
self.momentum = momentum
# letters contains 27 classes, digits contains 10 classes
num_classes = 27 if dataset == 'letters' else 10
# Load pre learned AlexNet with changed number of output classes
state_dict = torch.load('./trained_models/alexnet.pth')
state_dict['classifier.6.weight'] = torch.zeros(num_classes, 4096)
state_dict['classifier.6.bias'] = torch.zeros(num_classes)
self.model = AlexNet(num_classes)
self.model.load_state_dict(state_dict)
# Use cuda if available
if torch.cuda.is_available():
self.model.cuda()
# Load training dataset
kwargs = {
'num_workers': 1,
'pin_memory': True
} if torch.cuda.is_available() else {}
self.train_loader = torch.utils.data.DataLoader(
EMNIST('./data',
dataset,
download=True,
transform=transforms.Compose([
transforms.Lambda(correct_rotation),
transforms.Lambda(random_transform),
transforms.Resize((224, 224)),
transforms.RandomResizedCrop(224, (0.9, 1.1),
ratio=(0.9, 1.1)),
transforms.Grayscale(3),
transforms.ToTensor(),
])),
batch_size=batch_size,
shuffle=True,
**kwargs)
# Optimizer and loss function
self.optimizer = optim.SGD(self.model.parameters(),
lr=self.lr,
momentum=self.momentum)
self.loss_fn = nn.CrossEntropyLoss()
def reduce_learning_rate(self, epoch):
"""
Reduce the learning rate by factor 0.1 every lr_decay_epoch
:param optimizer: Optimizer containing the learning rate
:param epoch: Current epoch
:param init_lr: Initial learning rate
:param lr_decay_epoch: Number of epochs until learning rate gets reduced
:return: None
"""
lr = self.lr * (0.1**(epoch // self.lr_decay_epoch))
if epoch % self.lr_decay_epoch == 0:
print('LR is set to {}'.format(lr))
for param_group in self.optimizer.param_groups:
param_group['lr'] = lr
def train(self, epoch):
"""
Train the model for one epoch and save the result as a .pth file
:param epoch: Current epoch
:return: None
"""
self.model.train()
train_loss = 0
train_correct = 0
progress = None
for batch_idx, (data, target) in enumerate(self.train_loader):
# Get data and label
if torch.cuda.is_available():
data, target = data.cuda(), target.cuda()
data, target = Variable(data), Variable(target)
# Optimize using backpropagation
self.optimizer.zero_grad()
output = self.model(data)
loss = self.loss_fn(output, target)
train_loss += loss.data[0]
pred = output.data.max(1, keepdim=True)[1]
train_correct += pred.eq(target.data.view_as(pred)).sum()
loss.backward()
self.optimizer.step()
# Print information about current step
current_progress = int(100 * (batch_idx + 1) * self.batch_size /
len(self.train_loader.dataset))
if current_progress is not progress and current_progress % 5 == 0:
progress = current_progress
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, (batch_idx + 1) * len(data),
len(self.train_loader.dataset), current_progress,
loss.data[0]))
train_loss /= (len(self.train_loader.dataset) / self.batch_size)
train_correct /= len(self.train_loader.dataset)
train_correct *= 100
# Print information about current epoch
print(
'Train Epoch: {} \tCorrect: {:3.2f}%\tAverage loss: {:.6f}'.format(
epoch, train_correct, train_loss))
# Save snapshot
torch.save(
{
'model': self.model.state_dict(),
'optimizer': self.optimizer.state_dict()
}, './trained_models/{}_{}.pth'.format(self.dataset, epoch))
def start(self):
"""
Start training the network
:return: None
"""
for epoch in range(1, self.epochs + 1):
self.reduce_learning_rate(epoch)
self.train(epoch)
assert os.path.isdir('checkpoint'), 'Error: no checkpoint directory found!'
checkpoint = torch.load('./checkpoint/ckpt.pth')
net.load_state_dict(checkpoint['model'])
best_acc = checkpoint['accuracy']
start_epoch = checkpoint['epoch']
else:
best_acc = 0
start_epoch = 0
# hyperparameters
batch_size= args.batch_size
epoch_num = args.epochs
optimizer = optim.Adam(model.parameters(), lr= args.lr, weight_decay=args.weight_decay)
train_loader = DataLoader(data_train, shuffle=True,batch_size= batch_size, num_workers=8)
test_loader = DataLoader(data_test, shuffle=True, batch_size= batch_size, num_workers=8)
def train(epoch):
model.train() # set model in training mode (need this because of dropout)
correct = 0
loss = 0
for batch_id, (data, label) in enumerate(train_loader):
data = data.to(device)
target = label.to(device)
optimizer.zero_grad()
import torch.nn as nn
from torch.autograd import Variable
from data_process import MyDataset
from torchvision import transforms, utils
from torch.utils.data import DataLoader
import torch
train_data = MyDataset(txt='./data/train.txt', transform=transforms.ToTensor())
train_loader = DataLoader(train_data, batch_size=50, shuffle=True) #返回的是迭代器
test_data = MyDataset(txt='./data/val.txt', transform=transforms.ToTensor())
test_loader = DataLoader(test_data, batch_size=50)
model = AlexNet().cuda() #使用gpu，将模型加载到显存
print(model)
#print(list(model.parameters()))
optimizer = optim.Adam(model.parameters(), lr=0.001)
loss_func = nn.CrossEntropyLoss()
#开始训练
for epoch in range(30):
print('epoch {}'.format(epoch + 1))
#training------------------------
train_loss = 0
train_accu = 0
for batch_x, batch_y in train_loader:
batch_x, batch_y = batch_x.cuda(), batch_y.cuda() #数据加载到显存
out = model(batch_x)
loss = loss_func(out, batch_y)
train_loss += loss.item()
def train(train_loader, eval_loader, opt):
print('==> Start training...')
summary_writer = SummaryWriter('./runs/' + str(int(time.time())))
is_cuda = torch.cuda.is_available()
model = AlexNet()
if is_cuda:
model = model.cuda()
optimizer = optim.SGD(
params=model.parameters(),
lr=opt.base_lr,
momentum=0.9,
)
criterion = nn.CrossEntropyLoss()
best_eval_acc = -0.1
losses = AverageMeter()
accuracies = AverageMeter()
global_step = 0
for epoch in range(1, opt.epochs + 1):
# train
model.train()
for batch_idx, (inputs, targets) in enumerate(train_loader):
global_step += 1
if is_cuda:
inputs = inputs.cuda()
targets = targets.cuda()
outputs = model(inputs)
loss = criterion(outputs, targets)
losses.update(loss.item(), outputs.shape[0])
summary_writer.add_scalar('train/loss', loss, global_step)
_, preds = torch.max(outputs, dim=1)
acc = preds.eq(targets).sum().item() / len(targets)
accuracies.update(acc)
summary_writer.add_scalar('train/acc', acc, global_step)
optimizer.zero_grad()
loss.backward()
optimizer.step()
summary_writer.add_scalar('lr', optimizer.param_groups[0]['lr'],
global_step)
print(
'==> Epoch: %d; Average Train Loss: %.4f; Average Train Acc: %.4f'
% (epoch, losses.avg, accuracies.avg))
# eval
model.eval()
losses.reset()
accuracies.reset()
for batch_idx, (inputs, targets) in enumerate(eval_loader):
if is_cuda:
inputs = inputs.cuda()
targets = targets.cuda()
outputs = model(inputs)
loss = criterion(outputs, targets)
losses.update(loss.item(), outputs.shape[0])
_, preds = torch.max(outputs, dim=1)
acc = preds.eq(targets).sum().item() / len(targets)
accuracies.update(acc)
summary_writer.add_scalar('eval/loss', losses.avg, global_step)
summary_writer.add_scalar('eval/acc', accuracies.avg, global_step)
if accuracies.avg > best_eval_acc:
best_eval_acc = accuracies.avg
torch.save(model, './weights/best.pt')
print(
'==> Epoch: %d; Average Eval Loss: %.4f; Average/Best Eval Acc: %.4f / %.4f'
% (epoch, losses.avg, accuracies.avg, best_eval_acc))
def main():
#gpus=[4,5,6,7]
gpus = [0]
print("GPUs :", gpus)
print("prepare data")
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_tfs = transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(), normalize
])
val_tfs = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(), normalize
])
train_ds = datasets.ImageFolder('/home/gw/data/imagenet_10/train',
train_tfs)
val_ds = datasets.ImageFolder('/home/gw/data/imagenet_10/val', val_tfs)
train_ld = torch.utils.data.DataLoader(train_ds,
batch_size=256,
shuffle=True,
num_workers=4,
pin_memory=True)
val_ld = torch.utils.data.DataLoader(val_ds,
batch_size=64,
shuffle=False,
num_workers=4,
pin_memory=True)
print("construct model")
#model = ResNet50()
#model=torchvision.models.AlexNet()
model = AlexNet()
#model = torch.nn.DataParallel(model, device_ids=gpus).cuda(gpus[0])
model.cuda()
criterion = nn.CrossEntropyLoss().cuda(gpus[0])
optimizer = torch.optim.SGD(model.parameters(),
0.01,
momentum=0.875,
weight_decay=3.0517578125e-05)
model.train()
print("begin trainning")
for epoch in range(0, 50):
batch_time = AverageMeter('Time', ':6.3f')
data_time = AverageMeter('Data', ':6.3f')
losses = AverageMeter('Loss', ':.4e')
top1 = AverageMeter('Acc@1', ':6.2f')
top5 = AverageMeter('Acc@5', ':6.2f')
progress = ProgressMeter(len(train_ld),
[batch_time, data_time, losses, top1, top5],
prefix="Epoch: [{}]".format(epoch))
end = time.time()
for i, (images, labels) in enumerate(train_ld):
data_time.update(time.time() - end)
print('image shape: ', images.shape)
print('labels shape: ', labels.shape)
images = images.cuda(gpus[0], non_blocking=True)
labels = labels.cuda(gpus[0], non_blocking=True)
outputs = model(images)
loss = criterion(outputs, labels)
# measure accuracy
acc1, acc5 = accuracy(outputs, labels, topk=(1, 5))
losses.update(loss.item(), images.size(0))
top1.update(acc1[0], images.size(0))
top5.update(acc5[0], images.size(0))
optimizer.zero_grad()
loss.backward()
optimizer.step()
batch_time.update(time.time() - end)
end = time.time()
if i % 10 == 0:
progress.display(i)
dataset = torchvision.datasets.ImageFolder(path, transform=transformation)
loader = torch.utils.data.DataLoader(dataset,
batch_size=64,
num_workers=0,
shuffle=shuffle)
return loader
# https://github.com/fastai/imagenette
# https://s3.amazonaws.com/fast-ai-imageclas/imagenette2.tgz
train_loader = train_dataset(r'data/imagenette2-320/train', shuffle=True)
val_loader = val_dataset(r'data/imagenette2-320/val', shuffle=False)
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(net.parameters(), lr=1e-4, momentum=0.9)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
mode='max',
factor=0.1,
patience=2,
verbose=True,
threshold=0.1,
threshold_mode='rel',
cooldown=0,
min_lr=0,
eps=1e-08)
def train(epoch):
net.train()
num_workers=2 ) #################################################################### # # Load the model # #################################################################### model = AlexNet().to(device) criterion = nn.CrossEntropyLoss() # Stochastic gradient descent optimizer = optim.SGD( model.parameters(), lr=0.01, weight_decay=0.0005, momentum=0.9, ) def test_model(model, epoch): correct = 0 total = 0 with torch.no_grad(): for i, data in enumerate(test_loader, 0): images, labels, _ = data images, labels = images.to(device), labels.to(device)
name = k[7:] # remove `module.`
# name = k[9:] # remove `module.1.`
new_check_point[name] = v
net.load_state_dict(new_check_point)
net = net.to(device)
if device == 'cuda':
net = torch.nn.DataParallel(net)
cudnn.benchmark = True
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=args.lr, momentum=0.9, weight_decay=5e-4)
#
# # Training
# def main():
# AV = [[] for _ in range(5)]
# MAV = [[] for _ in range(5)]
# with torch.no_grad():
# for batch_idx, (inputs, targets) in enumerate(trainloader):
# print("Processing {} batches in {}".format(batch_idx, 25000//args.bs))
# inputs, targets = inputs.to(device), targets.to(device)
# outputs = net(inputs)
# for score, t in zip(outputs, targets):
# if np.argmax(score) == t:
# AV[t].append(score.unsqueeze(dim=0))
# for i in range(5):
classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse',
'ship', 'truck')
# Model
print('==> Building model..')
net = AlexNet(5)
net = net.to(device)
if device == 'cuda':
net = torch.nn.DataParallel(net)
cudnn.benchmark = True
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(),
lr=args.lr,
momentum=0.9,
weight_decay=5e-4)
# Training
def train(epoch):
adjust_learning_rate(optimizer, epoch, args.lr)
print('\nEpoch: %d Learning rate: %f' %
(epoch, optimizer.param_groups[0]['lr']))
net.train()
train_loss = 0
correct = 0
total = 0
image_datasets = {
'train': ImageFolder(train_dir, transform=data_transforms['train']),
'val': ImageFolder(val_dir, transform=data_transforms['val']),
}
data_loader = {
x: torch.utils.data.DataLoader(image_datasets[x],
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
for x in ['train', 'val']
}
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(),
lr=learning_rate,
weight_decay=weight_decay)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
'min',
verbose=True)
model = model.to(device)
model = train(model,
data_loader,
criterion,
optimizer,
scheduler,
num_epochs=num_epoches)
n_datapoints = len(model.train_epoch_loss)
def main():
progress = default_progress()
experiment_dir = 'experiment/miniplaces'
# Here's our data
train_loader = torch.utils.data.DataLoader(CachedImageFolder(
'dataset/miniplaces/simple/train',
transform=transforms.Compose([
transforms.Resize(128),
transforms.RandomCrop(119),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(IMAGE_MEAN, IMAGE_STDEV)
])),
batch_size=64,
shuffle=True,
num_workers=6,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(CachedImageFolder(
'dataset/miniplaces/simple/val',
transform=transforms.Compose([
transforms.Resize(128),
transforms.CenterCrop(119),
transforms.ToTensor(),
transforms.Normalize(IMAGE_MEAN, IMAGE_STDEV)
])),
batch_size=512,
shuffle=False,
num_workers=6,
pin_memory=True)
# Create a simplified AlexNet with half resolution.
model = AlexNet(first_layer='conv1',
last_layer='fc8',
layer_sizes=dict(fc6=2048, fc7=2048),
output_channels=100,
half_resolution=True,
include_lrn=False,
split_groups=False).cuda()
# Use Kaiming initialization for the weights
for name, val in model.named_parameters():
if 'weight' in name:
init.kaiming_uniform_(val)
else:
# Init positive bias in many layers to avoid dead neurons.
assert 'bias' in name
init.constant_(
val, 0 if any(
name.startswith(layer)
for layer in ['conv1', 'conv3', 'fc8']) else 1)
# An abbreviated training schedule: 40000 batches.
# TODO: tune these hyperparameters.
# init_lr = 0.002
init_lr = 0.002
# max_iter = 40000 - 34.5% @1
# max_iter = 50000 - 37% @1
# max_iter = 80000 - 39.7% @1
# max_iter = 100000 - 40.1% @1
max_iter = 100000
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(
model.parameters(),
lr=init_lr,
momentum=0.9, # 0.9,
# weight_decay=0.001)
weight_decay=0.001)
iter_num = 0
best = dict(val_accuracy=0.0)
model.train()
# Oh, hold on. Let's actually resume training if we already have a model.
checkpoint_filename = 'miniplaces.pth.tar'
best_filename = 'best_%s' % checkpoint_filename
best_checkpoint = os.path.join(experiment_dir, best_filename)
try_to_resume_training = False
if try_to_resume_training and os.path.exists(best_checkpoint):
checkpoint = torch.load(os.path.join(experiment_dir, best_filename))
iter_num = checkpoint['iter']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
best['val_accuracy'] = checkpoint['accuracy']
def save_checkpoint(state, is_best):
filename = os.path.join(experiment_dir, checkpoint_filename)
ensure_dir_for(filename)
torch.save(state, filename)
if is_best:
shutil.copyfile(filename,
os.path.join(experiment_dir, best_filename))
def validate_and_checkpoint():
model.eval()
val_loss, val_acc = AverageMeter(), AverageMeter()
for input, target in progress(val_loader):
# Load data
input_var, target_var = [
Variable(d.cuda(non_blocking=True)) for d in [input, target]
]
# Evaluate model
with torch.no_grad():
output = model(input_var)
loss = criterion(output, target_var)
_, pred = output.max(1)
accuracy = (target_var.eq(pred)
).data.float().sum().item() / input.size(0)
val_loss.update(loss.data.item(), input.size(0))
val_acc.update(accuracy, input.size(0))
# Check accuracy
post_progress(l=val_loss.avg, a=val_acc.avg)
# Save checkpoint
save_checkpoint(
{
'iter': iter_num,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
'accuracy': val_acc.avg,
'loss': val_loss.avg,
}, val_acc.avg > best['val_accuracy'])
best['val_accuracy'] = max(val_acc.avg, best['val_accuracy'])
post_progress(v=val_acc.avg)
# Here is our training loop.
while iter_num < max_iter:
for input, target in progress(train_loader):
# Track the average training loss/accuracy for each epoch.
train_loss, train_acc = AverageMeter(), AverageMeter()
# Load data
input_var, target_var = [
Variable(d.cuda(non_blocking=True)) for d in [input, target]
]
# Evaluate model
output = model(input_var)
loss = criterion(output, target_var)
train_loss.update(loss.data.item(), input.size(0))
# Perform one step of SGD
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Also check training set accuracy
_, pred = output.max(1)
accuracy = (target_var.eq(pred)).data.float().sum().item() / (
input.size(0))
train_acc.update(accuracy)
remaining = 1 - iter_num / float(max_iter)
post_progress(l=train_loss.avg,
a=train_acc.avg,
v=best['val_accuracy'])
# Advance
iter_num += 1
if iter_num >= max_iter:
break
# Linear learning rate decay
lr = init_lr * remaining
for param_group in optimizer.param_groups:
param_group['lr'] = lr
# Ocassionally check validation set accuracy and checkpoint
if iter_num % 1000 == 0:
validate_and_checkpoint()
model.train()
class Solver(object):
def __init__(self, config):
self.model = None
self.name = config.name
self.lr = config.lr
self.momentum = config.momentum
self.beta = config.beta
self.max_alpha = config.max_alpha
self.epochs = config.epochs
self.patience = config.patience
self.N = config.N
self.batch_size = config.batch_size
self.random_labels = config.random_labels
self.use_bn = config.batchnorm
self.criterion = None
self.optimizer = None
self.scheduler = None
self.device = None
self.cuda = config.cuda
self.train_loader = None
self.test_loader = None
def load_data(self):
# ToTensor scales pixel values from [0,255] to [0,1]
mean_var = (125.3 / 255, 123.0 / 255,
113.9 / 255), (63.0 / 255, 62.1 / 255, 66.7 / 255)
transform = transforms.Compose([
transforms.CenterCrop(28),
transforms.ToTensor(),
transforms.Normalize(*mean_var, inplace=True)
])
train_set = torchvision.datasets.CIFAR10(root='./data',
train=True,
download=DOWNLOAD,
transform=transform)
test_set = torchvision.datasets.CIFAR10(root='./data',
train=False,
download=DOWNLOAD,
transform=transform)
if self.random_labels:
np.random.shuffle(train_set.targets)
np.random.shuffle(test_set.targets)
assert self.N <= 50000
if self.N < 50000:
train_set.data = train_set.data[:self.N]
# downsize the test set to improve speed for small N
test_set.data = test_set.data[:self.N]
self.train_loader = torch.utils.data.DataLoader(
dataset=train_set,
batch_size=self.batch_size,
shuffle=True,
drop_last=True)
self.test_loader = torch.utils.data.DataLoader(
dataset=test_set,
batch_size=self.batch_size,
shuffle=False,
drop_last=True)
def load_model(self):
if self.cuda:
self.device = torch.device('cuda')
cudnn.benchmark = True
else:
self.device = torch.device('cpu')
self.model = AlexNet(device=self.device,
B=self.batch_size,
max_alpha=self.max_alpha,
use_bn=self.use_bn).to(self.device)
self.optimizer = optim.SGD(self.model.parameters(),
lr=self.lr,
momentum=self.momentum)
self.scheduler = optim.lr_scheduler.StepLR(self.optimizer,
step_size=140)
self.criterion = nn.NLLLoss().to(self.device)
def getIw(self):
# Iw should be normalized with respect to N
# via reparameterization, we optimize alpha with only 1920 dimensions
# but Iw should scale with the dimension of the weights
return 7 * 7 * 64 * 384 / 1920 * self.model.getIw() / self.batch_size
def do_batch(self, train, epoch):
loader = self.train_loader if train else self.test_loader
total_ce, total_Iw, total_loss = 0, 0, 0
total_correct = 0
total = 0
pbar = tqdm(loader)
num_batches = len(loader)
for batch_num, (data, target) in enumerate(pbar):
data, target = data.to(self.device), target.to(self.device)
if train:
self.optimizer.zero_grad()
output = self.model(data)
# NLLLoss is averaged across observations for each minibatch
ce = self.criterion(torch.log(output + EPS), target)
Iw = self.getIw()
loss = ce + 0.5 * self.beta * Iw
if train:
loss.backward()
self.optimizer.step()
total_ce += ce.item()
total_Iw += Iw.item()
total_loss += loss.item()
prediction = torch.max(
output,
1) # second param "1" represents the dimension to be reduced
total_correct += np.sum(
prediction[1].cpu().numpy() == target.cpu().numpy())
total += target.size(0)
a = self.model.get_a()
pbar.set_description('Train' if train else 'Test')
pbar.set_postfix(N=self.N,
b=self.beta,
ep=epoch,
acc=100. * total_correct / total,
loss=total_loss / num_batches,
ce=total_ce / num_batches,
Iw=total_Iw / num_batches,
a=a)
return total_correct / total, total_loss / num_batches, total_ce / num_batches, total_Iw / num_batches, a
def train(self, epoch):
self.model.train()
return self.do_batch(train=True, epoch=epoch)
def test(self, epoch):
self.model.eval()
with torch.no_grad():
return self.do_batch(train=False, epoch=epoch)
def save(self, name=None):
model_out_path = (name or self.name) + ".pth"
# torch.save(self.model, model_out_path)
# print("Checkpoint saved to {}".format(model_out_path))
def run(self):
self.load_data()
self.load_model()
results = []
best_acc, best_ep = -1, -1
for epoch in range(1, self.epochs + 1):
# print("\n===> epoch: %d/200" % epoch)
train_acc, train_loss, train_ce, train_Iw, train_a = self.train(
epoch)
self.scheduler.step(epoch)
test_acc, test_loss, test_ce, test_Iw, test_a = self.test(epoch)
results.append([
self.N, self.beta, train_acc, test_acc, train_loss, test_loss,
train_ce, test_ce, train_Iw, test_Iw, train_a, test_a
])
if test_acc > best_acc:
best_acc, best_ep = test_acc, epoch
if self.patience >= 0: # early stopping
if best_ep < epoch - self.patience:
break
with open(self.name + '.csv', 'a') as f:
w = csv.writer(f)
w.writerows(results)
self.save()
return train_acc, test_acc
def run_experiment(args):
torch.manual_seed(args.seed)
if not args.no_cuda:
torch.cuda.manual_seed(args.seed)
# Dataset
if args.dataset == 'mnist':
train_loader, test_loader, _, val_data = prepare_mnist(args)
else:
create_val_img_folder(args)
train_loader, test_loader, _, val_data = prepare_imagenet(args)
idx_to_class = {i: c for c, i in val_data.class_to_idx.items()}
# Model & Criterion
if args.model == 'AlexNet':
if args.pretrained:
model = models.__dict__['alexnet'](pretrained=True)
# Change the last layer
in_f = model.classifier[-1].in_features
model.classifier[-1] = nn.Linear(in_f, args.classes)
else:
model = AlexNet(args.classes)
criterion = nn.CrossEntropyLoss(size_average=False)
else:
model = SVM(args.features, args.classes)
criterion = MultiClassHingeLoss(margin=args.margin, size_average=False)
if not args.no_cuda:
model.cuda()
# Load saved model and test on it
if args.load:
model.load_state_dict(torch.load(args.model_path))
val_acc = test(model, criterion, test_loader, 0, [], [], idx_to_class,
args)
# Optimizer
if args.optimizer == 'adam':
optimizer = optim.Adam(model.parameters())
else:
optimizer = optim.SGD(model.parameters(),
lr=args.lr,
momentum=args.momentum)
total_minibatch_count = 0
val_acc = 0
train_losses, train_accs = [], []
val_losses, val_accs = [], []
# Train and test
for epoch in range(1, args.epochs + 1):
total_minibatch_count = train(model, criterion, optimizer,
train_loader, epoch,
total_minibatch_count, train_losses,
train_accs, args)
val_acc = test(model, criterion, test_loader, epoch, val_losses,
val_accs, idx_to_class, args)
# Save model
if args.save:
if not os.path.exists(args.models_dir):
os.makedirs(args.models_dir)
filename = '_'.join(
[args.prefix, args.dataset, args.model, 'model.pt'])
torch.save(model.state_dict(), os.path.join(args.models_dir, filename))
# Plot graphs
fig, axes = plt.subplots(1, 4, figsize=(13, 4))
axes[0].plot(train_losses)
axes[0].set_title('Loss')
axes[1].plot(train_accs)
axes[1].set_title('Acc')
axes[1].set_ylim([0, 1])
axes[2].plot(val_losses)
axes[2].set_title('Val loss')
axes[3].plot(val_accs)
axes[3].set_title('Val Acc')
axes[3].set_ylim([0, 1])
# Images don't show on Ubuntu
# plt.tight_layout()
# Save results
if not os.path.exists(args.results_dir):
os.makedirs(args.results_dir)
filename = '_'.join([args.prefix, args.dataset, args.model, 'plot.png'])
fig.suptitle(filename)
fig.savefig(os.path.join(args.results_dir, filename))
logger = Logger('./largecnnlogs')
lossfunction = nn.MSELoss()
dataset = Rand_num()
sampler = RandomSampler(dataset)
loader = DataLoader(dataset,
batch_size=20,
sampler=sampler,
shuffle=False,
num_workers=1,
drop_last=True)
net = AlexNet(3)
#net.load_state_dict(torch.load(SAVE_PATH))
net.cuda()
optimizer = optim.Adam(net.parameters(), lr=0.001)
for epoch in range(10000):
for i, data in enumerate(loader, 0):
net.zero_grad()
video, labels = data
video = video.view(-1, 3, 227, 227)
labels = labels.view(-1, 3)
labels = torch.squeeze(Variable(labels.float().cuda()))
video = torch.squeeze(Variable((video.float() / 256).cuda()))
net.train()
outputs = net.forward(video)
loss = lossfunction(outputs, labels)
loss.backward()
optimizer.step()
if i == 0:
torch.save(net.state_dict(), SAVE_PATH)
def run():
global args
args = parser.parse_args()
if args.resume:
print 'Resuming from checkpoint!'
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
# load datasets, split train and test
'''
train_dataset = SketchData(root=path,
train=True,
transform=transforms.ToTensor(),
target_transform=transforms.ToTensor(),
)
test_dataset = SketchData(root=path,
train=False,
transform=transforms.ToTensor(),
target_transform=transforms.ToTensor(),
)
'''
train_dataset = SketchData(
root=path,
train=True,
transform=None,
target_transform=None,
)
val_dataset = SketchData(
root=path,
train=False,
transform=None,
target_transform=None,
)
train_loader = torch.utils.data.DataLoader(dataset=train_dataset,
batch_size=args.b,
shuffle=True)
val_loader = torch.utils.data.DataLoader(dataset=val_dataset,
batch_size=args.b,
shuffle=False)
# create model, set parameters, optimiser, loss
model = AlexNet()
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.wd)
criterion = nn.CrossEntropyLoss()
best_prec = 0
for epoch in range(args.epochs):
adjust_learning_rate(optimizer, epoch)
train(train_loader, model, criterion, optimizer, epoch)
precision = validate(val_loader, model, criterion)
best_prec = max(precision.data[0], best_prec)
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'best_prec1': best_prec,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
}, (precision.data[0] > best_prec))
pretrained=False)
elif args.model == 'resnext':
model = resnext50(img_channel=3, num_classes=19)
elif args.model == 'alexnet':
model = AlexNet(num_classes=19)
else:
raise ValueError("This utility can't train that kind of model.")
logging.info("Setting dataset")
torch.cuda.empty_cache()
train_dataset = TrainImageDataset(
args.data,
224,
224,
)
optimizer = torch.optim.Adam(model.parameters(), args.lr)
# scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=10, gamma=0.1)
criterion = nn.CrossEntropyLoss()
config = wandb.config
# wandb config specification
config.learning_rate = args.lr
config.batch_size = args.batch_size
config.model = args.model
logging.info("Training...")
train_loss, train_acc, test_acc, test_loss = train_for_classification(
net=model,
dataset=train_dataset,
batch_size=args.batch_size,
optimizer=optimizer,
train=False,
download=True,
transform=transform)
# Define test batch data
testloader = torch.utils.data.DataLoader(
testset,
batch_size=BATCH_SIZE,
shuffle=True,
)
# Define loss function loss function and optimization method (using SGD)
net = AlexNet().to(device)
criterion = nn.CrossEntropyLoss(
) # Cross entropy loss function, usually used for multi-classification problems
optimizer = optim.SGD(net.parameters(), lr=LR, momentum=0.9)
# Train and save model parameters
def train():
for epoch in range(EPOCH):
sum_loss = 0.0
# Data read
for i, data in enumerate(trainloader):
inputs, labels = data
inputs, labels = inputs.to(device), labels.to(device)
# Gradient clear
optimizer.zero_grad()
def train_generic_model(model_name="alexnet",
dataset="custom",
num_classes=-1,
batch_size=8,
is_transform=1,
num_workers=2,
lr_decay=1,
l2_reg=0,
hdf5_path="dataset-bosch-224x224.hdf5",
trainset_dir="./TRAIN_data_224_v8",
testset_dir="./TEST_data_224_v8",
convert_grey=False):
CHKPT_PATH = "./checkpoint_{}.PTH".format(model_name)
print("CUDA:")
print(torch.cuda.is_available())
if is_transform:
trans_ls = []
if convert_grey:
trans_ls.append(transforms.Grayscale(num_output_channels=1))
trans_ls.extend([
transforms.Resize((224, 224)),
# transforms.RandomCrop((224, 224)),
# transforms.Grayscale(num_output_channels=1),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
transform = transforms.Compose(trans_ls)
else:
transform = None
print("DATASET FORMAT: {}".format(dataset))
print("TRAINSET PATH: {}".format(trainset_dir))
print("TESTSET PATH: {}".format(testset_dir))
print("HDF5 PATH: {}".format(hdf5_path))
if dataset == "custom":
trainset = torchvision.datasets.ImageFolder(root=trainset_dir,
transform=transform)
train_size = len(trainset)
testset = torchvision.datasets.ImageFolder(root=testset_dir,
transform=transform)
test_size = len(testset)
elif dataset == "cifar":
trainset = torchvision.datasets.CIFAR10(root="CIFAR_TRAIN_data",
train=True,
download=True,
transform=transform)
train_size = len(trainset)
testset = torchvision.datasets.CIFAR10(root="CIFAR_TEST_data",
train=False,
download=True,
transform=transform)
test_size = len(testset)
elif dataset == "hdf5":
if num_workers == 1:
trainset = Hdf5Dataset(hdf5_path,
transform=transform,
is_test=False)
else:
trainset = Hdf5DatasetMPI(hdf5_path,
transform=transform,
is_test=False)
train_size = len(trainset)
if num_workers == 1:
testset = Hdf5Dataset(hdf5_path, transform=transform, is_test=True)
else:
testset = Hdf5DatasetMPI(hdf5_path,
transform=transform,
is_test=True)
test_size = len(testset)
train_loader = torch.utils.data.DataLoader(trainset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
test_loader = torch.utils.data.DataLoader(testset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
if model_name == "alexnet":
net = AlexNet(num_classes=num_classes)
elif model_name == "lenet5":
net = LeNet5(num_classes=num_classes)
elif model_name == "stn-alexnet":
net = STNAlexNet(num_classes=num_classes)
elif model_name == "stn-lenet5":
net = LeNet5STN(num_classes=num_classes)
elif model_name == "capsnet":
net = CapsuleNet(num_classes=num_classes)
elif model_name == "convneta":
net = ConvNetA(num_classes=num_classes)
elif model_name == "convnetb":
net = ConvNetB(num_classes=num_classes)
elif model_name == "convnetc":
net = ConvNetC(num_classes=num_classes)
elif model_name == "convnetd":
net = ConvNetD(num_classes=num_classes)
elif model_name == "convnete":
net = ConvNetE(num_classes=num_classes)
elif model_name == "convnetf":
net = ConvNetF(num_classes=num_classes)
elif model_name == "convnetg":
net = ConvNetG(num_classes=num_classes)
elif model_name == "convneth":
net = ConvNetH(num_classes=num_classes)
elif model_name == "convneti":
net = ConvNetI(num_classes=num_classes)
elif model_name == "convnetj":
net = ConvNetJ(num_classes=num_classes)
elif model_name == "convnetk":
net = ConvNetK(num_classes=num_classes)
elif model_name == "convnetl":
net = ConvNetL(num_classes=num_classes)
elif model_name == "convnetm":
net = ConvNetM(num_classes=num_classes)
elif model_name == "convnetn":
net = ConvNetN(num_classes=num_classes)
elif model_name == "resnet18":
net = models.resnet18(pretrained=False, num_classes=num_classes)
print(net)
if torch.cuda.is_available():
net = net.cuda()
if model_name == "capsnet":
criterion = CapsuleLoss()
else:
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(),
lr=LEARNING_RATE,
momentum=0.9,
weight_decay=l2_reg)
if lr_decay:
scheduler = ReduceLROnPlateau(optimizer, 'min')
best_acc = 0
from_epoch = 0
if os.path.exists(CHKPT_PATH):
print("Checkpoint Found: {}".format(CHKPT_PATH))
state = torch.load(CHKPT_PATH)
net.load_state_dict(state['state_dict'])
optimizer.load_state_dict(state['optimizer'])
best_acc = state['best_accuracy']
from_epoch = state['epoch']
for epoch in range(from_epoch, NUM_EPOCHS):
#print("Epoch: {}/{}".format(epoch + 1, NUM_EPOCHS))
epoch_loss = 0
correct = 0
for i, data in enumerate(train_loader, 0):
#print("Train \t Epoch: {}/{} \t Batch: {}/{}".format(epoch + 1,
# NUM_EPOCHS,
# i + 1,
# ceil(train_size / BATCH_SIZE)))
inputs, labels = data
inputs, labels = Variable(inputs).type(torch.FloatTensor),\
Variable(labels).type(torch.LongTensor)
if model_name == "capsnet":
inputs = augmentation(inputs)
ground_truth = torch.eye(num_classes).index_select(
dim=0, index=labels)
if torch.cuda.is_available():
inputs = inputs.cuda()
labels = labels.cuda()
optimizer.zero_grad()
if model_name == "capsnet":
classes, reconstructions = net(inputs, ground_truth)
loss = criterion(inputs, ground_truth, classes,
reconstructions)
else:
outputs = net(inputs)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
epoch_loss += loss.data[0]
if model_name != "capsnet":
log_outputs = F.softmax(outputs, dim=1)
else:
log_outputs = classes
pred = log_outputs.data.max(1, keepdim=True)[1]
correct += pred.eq(labels.data.view_as(pred)).sum()
print(
"Epoch: {} \t Training Loss: {:.4f} \t Training Accuracy: {:.2f} \t {}/{}"
.format(epoch + 1, epoch_loss / train_size,
100 * correct / train_size, correct, train_size))
correct = 0
test_loss = 0
for i, data in enumerate(test_loader, 0):
# print("Test \t Epoch: {}/{} \t Batch: {}/{}".format(epoch + 1,
# NUM_EPOCHS,
# i + 1,
# ceil(test_size / BATCH_SIZE)))
inputs, labels = data
inputs, labels = Variable(inputs).type(
torch.FloatTensor), Variable(labels).type(torch.LongTensor)
if model_name == "capsnet":
inputs = augmentation(inputs)
ground_truth = torch.eye(num_classes).index_select(
dim=0, index=labels)
if torch.cuda.is_available():
inputs = inputs.cuda()
labels = labels.cuda()
if model_name == "capsnet":
classes, reconstructions = net(inputs)
loss = criterion(inputs, ground_truth, classes,
reconstructions)
else:
outputs = net(inputs)
loss = criterion(outputs, labels)
test_loss += loss.data[0]
if model_name != "capsnet":
log_outputs = F.softmax(outputs, dim=1)
else:
log_outputs = classes
pred = log_outputs.data.max(1, keepdim=True)[1]
correct += pred.eq(labels.data.view_as(pred)).sum()
print(
"Epoch: {} \t Testing Loss: {:.4f} \t Testing Accuracy: {:.2f} \t {}/{}"
.format(epoch + 1, test_loss / test_size,
100 * correct / test_size, correct, test_size))
if correct >= best_acc:
if not os.path.exists("./models"):
os.mkdir("./models")
torch.save(
net.state_dict(),
"./models/model-{}-{}-{}-{}-val-acc-{:.2f}-train-{}-test-{}-epoch-{}.pb"
.format(model_name, dataset, hdf5_path, str(datetime.now()),
100 * correct / test_size,
trainset_dir.replace(" ", "_").replace("/", "_"),
testset_dir.replace(" ", "_").replace("/",
"_"), epoch + 1))
best_acc = max(best_acc, correct)
# save checkpoint path
state = {
'epoch': epoch,
'state_dict': net.state_dict(),
'optimizer': optimizer.state_dict(),
'best_accuracy': best_acc
}
torch.save(state, CHKPT_PATH)
if lr_decay:
# Note that step should be called after validate()
scheduler.step(test_loss)
print('Finished Training')
print("")
print("")
