validation_dataset = ChestXrayDataSetMultiLabel(DATA_DIR,
VALIDATION_IMAGE_LIST,
transform=transform_val)
validation_dataloader = DataLoader(dataset=validation_dataset,
batch_size=BATCH_SIZE,
shuffle=False,
num_workers=4,
pin_memory=True,
drop_last=True)
test_dataset = ChestXrayDataSetMultiLabel(
data_dir=DATA_DIR,
image_list_file=TEST_IMAGE_LIST,
transform=transforms.Compose([
transforms.Resize(400),
transforms.TenCrop(299),
transforms.Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops])),
transforms.Lambda(lambda crops: torch.stack([
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
(crop) for crop in crops
]))
]))
test_dataloader = DataLoader(dataset=test_dataset,
batch_size=BATCH_SIZE,
shuffle=False,
num_workers=4,
pin_memory=True,
drop_last=True)
preprocess_hflip = transforms.Compose([
transforms.Resize(image_size),
transforms.CenterCrop(input_size),
HorizontalFlip(),
transforms.ToTensor(), normalize
])
def temp_trans(torch_image):
torch_image = transforms.ToTensor()(torch_image)
return normalize(torch_image)
preprocess_tencrop = transforms.Compose([
transforms.Resize(image_size),
transforms.TenCrop(input_size),
transforms.Lambda(
lambda crops: torch.stack([temp_trans(crop) for crop in crops])),
# transforms.ToTensor(),
# normalize
])
def main():
TTA2_preprocess = [preprocess, preprocess_hflip]
TTA10_preprocess = [preprocess_tencrop]
TTA12_preprocess = [preprocess, preprocess_hflip, preprocess_tencrop]
id = 0
print("testing {}.....".format(ckp_path))
transforms.ToTensor(),
normalize,
]),
'val_em': transforms.Compose([
# transforms.Resize(290),
# transforms.RandomAffine(degrees=0, shear=10),
# transforms.CenterCrop(250),
transforms.RandomResizedCrop(224, scale=(0.7, 1)),
transforms.ColorJitter(brightness=0.4, contrast=0.4),
# transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]),
'test_v1': transforms.Compose([
transforms.Resize(180),
transforms.TenCrop(160),
transforms.ToTensor(),
normalize,
]),
'test_v2': transforms.Compose([
transforms.Resize(250),
transforms.TenCrop(224),
transforms.ToTensor(),
normalize,
]),
'rtest_v1': transforms.Compose([
transforms.Resize(180),
transforms.TenCrop(160),
transforms.ToTensor(),
normalize,
]),
def main():
best_prec1 = 0
if not os.path.isdir(args.save_dir):
os.makedirs(args.save_dir)
# log = open(os.path.join(args.save_dir, '{}.{}.log'.format(args.arch,args.prefix)), 'a')
# create model
print("=> creating model '{}'\n".format(args.arch))
model = imagenet_models.__dict__[args.arch](1000)
print("=> Model : {}\n".format(args.arch))
# print("=> Model : {}\n".format(model))
# print("=> parameter : {}\n".format(args))
if args.arch.startswith('alexnet') or args.arch.startswith('vgg'):
model.features = torch.nn.DataParallel(model.features)
model.cuda()
else:
model = torch.nn.DataParallel(model).cuda()
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay,
nesterov=True)
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume,
map_location=lambda storage, loc: storage)
args.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
del checkpoint['state_dict'][
'module.conv1.weights'] # fix this bug
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
for state in optimizer.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.cuda()
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
cudnn.benchmark = True
# Data loading code
traindir = os.path.join(args.data, 'train')
valdir = os.path.join(args.data, 'val-pytorch')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_dataset = datasets.ImageFolder(
traindir,
transforms.Compose([
transforms.RandomSizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]))
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True,
sampler=None)
val_loader = torch.utils.data.DataLoader(datasets.ImageFolder(
valdir,
transforms.Compose([
transforms.Scale(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])),
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
val_loader_10 = torch.utils.data.DataLoader(datasets.ImageFolder(
valdir,
transforms.Compose([
transforms.Scale(256),
transforms.TenCrop(224),
transforms.Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops])),
transforms.Lambda(
lambda crops: torch.stack([normalize(crop) for crop in crops]))
])),
batch_size=16,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
if args.evaluate:
validate(val_loader, model, criterion)
return
best_p1, best_p5, best_c10_p1, best_c10_p5 = 0, 0, 0, 0
filename = os.path.join(
args.save_dir,
'checkpoint.{}.{}.pth.tar'.format(args.arch, args.prefix))
bestname = os.path.join(
args.save_dir, 'best.{}.{}.pth.tar'.format(args.arch, args.prefix))
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer,
milestones=milestones,
gamma=0.1)
start_time = time.time()
epoch_time = AverageMeter()
for epoch in range(args.start_epoch, args.epochs):
need_hour, need_mins, need_secs = convert_secs2time(
epoch_time.val * (args.epochs - epoch))
need_time = '[Need: {:02d}:{:02d}:{:02d}]'.format(
need_hour, need_mins, need_secs)
print(' [{:s}] :: {:3d}/{:3d} ----- [{:s}] {:s}'.format(
args.arch, epoch, args.epochs, time_string(), need_time))
scheduler.step()
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch)
# evaluate on validation set
prec1, prec5 = validate(val_loader, model, criterion, epoch)
c10_p1, c10_p5 = validate_10crop(val_loader_10, model, criterion,
epoch)
# store the best
best_p1 = prec1 if prec1 > best_p1 else best_p1
best_p5 = prec5 if prec5 > best_p5 else best_p5
best_c10_p1 = c10_p1 if c10_p1 > best_c10_p1 else best_c10_p1
best_c10_p5 = c10_p5 if c10_p5 > best_c10_p5 else best_c10_p5
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
}, is_best, filename, bestname)
# measure elapsed time
epoch_time.update(time.time() - start_time)
start_time = time.time()
# loss_history, top1_history, top5_history
print('training loss', loss_history)
print('training top1 accuracy', top1_history)
print('training top5 accuracy', top5_history)
print('testing top1 accuracy', test_top1_history)
print('testing top5 accuracy', test_top5_history)
print('10-croping top1', crop10_top1_history)
print('10-croping top5', crop10_top5_history)
print('Best single-crop top 1: {0:.2f}\n'
'Best single-crop top 5: {1:.2f}\n'
'Best 10-crop top 1: {2:.2f}\n'
'Best 10-crop top 5: {3:.2f}\n'.format(best_p1, best_p5, best_c10_p1,
best_c10_p5))
cut_size = 44
total_epoch = 250
path = os.path.join(opt.dataset + '_' + opt.model)
# Data
print('==> Preparing data..')
transform_train = transforms.Compose([
transforms.RandomCrop(44),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
])
transform_test = transforms.Compose([
transforms.TenCrop(cut_size),
transforms.Lambda(lambda crops: torch.stack([transforms.ToTensor()(crop) for crop in crops])),
])
trainset = FER2013(split = 'Training', transform=transform_train)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=opt.bs, shuffle=True, num_workers=1)
PublicTestset = FER2013(split = 'PublicTest', transform=transform_test)
PublicTestloader = torch.utils.data.DataLoader(PublicTestset, batch_size=opt.bs, shuffle=False, num_workers=1)
PrivateTestset = FER2013(split = 'PrivateTest', transform=transform_test)
PrivateTestloader = torch.utils.data.DataLoader(PrivateTestset, batch_size=opt.bs, shuffle=False, num_workers=1)
# Model
if opt.model == 'VGG19':
net = VGG('VGG19')
elif opt.model == 'Resnet18':
net = ResNet18()
def main():
cudnn.benchmark = True
# initialize and load the model
model = DenseNet121(N_CLASSES).cuda()
model = torch.nn.DataParallel(model).cuda()
if os.path.isfile(CKPT_PATH):
print("=> loading checkpoint")
checkpoint = torch.load(CKPT_PATH)
#*************************debug*************************
state_dict=checkpoint['state_dict']
#print("before")
#print(next(iter(state_dict)))
#print("=> transforming parameter key names")
pattern=re.compile(
r'^(.*denselayer\d+\.(?:norm|relu|conv))\.((?:[12])\.(?:weight|bias|running_mean|running_var))$'
)
for key in list(state_dict.keys()):
res=pattern.match(key)
if res:
new_key=res.group(1)+res.group(2)
state_dict[new_key]=state_dict[key]
del state_dict[key]
#*************************debug*************************
#print("after")
#print(next(iter(state_dict)))
model.load_state_dict(state_dict)
print("=> loaded checkpoint")
else:
print("=> no checkpoint found")
normalize = transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])
test_dataset = ChestXrayDataSet(data_dir=DATA_DIR,
image_list_file=TEST_IMAGE_LIST,
transform=transforms.Compose([
transforms.Resize(256),
transforms.TenCrop(224),
transforms.Lambda
(lambda crops: torch.stack([transforms.ToTensor()(crop) for crop in crops])),
transforms.Lambda
(lambda crops: torch.stack([normalize(crop) for crop in crops]))
]))
test_loader = DataLoader(dataset=test_dataset, batch_size=BATCH_SIZE,
shuffle=False, num_workers=8, pin_memory=True)
# initialize the ground truth and output tensor
gt = torch.FloatTensor()
gt = gt.cuda()
pred = torch.FloatTensor()
pred = pred.cuda()
# switch to evaluate mode
model.eval()
with torch.no_grad():
for i, (inp, target) in enumerate(test_loader):
target = target.cuda()
gt = torch.cat((gt, target), 0)
bs, n_crops, c, h, w = inp.size()
input_var = torch.autograd.Variable(inp.view(-1, c, h, w).cuda())
#input_var = torch.autograd.Variable(inp.view(-1, c, h, w).cuda(), volatile=True)
output = model(input_var)
output_mean = output.view(bs, n_crops, -1).mean(1)
pred = torch.cat((pred, output_mean.data), 0)
AUROCs = compute_AUCs(gt, pred)
AUROC_avg = np.array(AUROCs).mean()
print('The average AUROC is {AUROC_avg:.3f}'.format(AUROC_avg=AUROC_avg))
for i in range(N_CLASSES):
print('The AUROC of {} is {}'.format(CLASS_NAMES[i], AUROCs[i]))
def test(test_pth):
CLASS_NAMES = [
'Opacity', 'Cardiomegaly', 'Pulmonary Atelectasis', 'Calcinosis',
'Lung/hypoinflation', 'Calcified Granuloma',
'Thoracic Vertebrae/degenerative', 'Lung/hyperdistention',
'Spine/degenerative', 'Aorta/tortuous', 'Pleural Effusion',
'Atherosclerosis', 'Airspace Disease', 'Granulomatous Disease',
'Nodule', 'Scoliosis'
]
# CLASS_NAMES=['abnormal']
cudnn.benchmark = True
net = Net().cuda()
net = torch.nn.DataParallel(net).cuda()
modelCheckpoint = torch.load(test_pth)
net.load_state_dict(modelCheckpoint['state_dict'])
normalize = transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])
transformList = []
transformList.append(transforms.Resize([256, 256]))
transformList.append(transforms.TenCrop([224, 224]))
transformList.append(
transforms.Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops])))
transformList.append(
transforms.Lambda(
lambda crops: torch.stack([normalize(crop) for crop in crops])))
transformSequence = transforms.Compose(transformList)
datasetTest = DatasetGenerator(pathImageDirectory=pathDirData,
pathDatasetFile=pathFileTest,
transform=transformSequence)
dataLoaderTest = DataLoader(dataset=datasetTest,
batch_size=32,
num_workers=8,
shuffle=False,
pin_memory=True)
outGT = torch.FloatTensor().cuda()
outPRED = torch.FloatTensor().cuda()
for i, (input1, input2, target) in enumerate(dataLoaderTest):
target = target.cuda()
outGT = torch.cat((outGT, target), 0)
bs, n_crops, c, h, w = input1.size()
varInput1 = torch.autograd.Variable(input1.view(-1, c, h, w).cuda(),
volatile=True)
varInput2 = torch.autograd.Variable(input2.view(-1, c, h, w).cuda(),
volatile=True)
out = net(varInput1, varInput2)
outMean = out.view(bs, n_crops, -1).mean(1)
outPRED = torch.cat((outPRED, outMean.data), 0)
plt_roc(outGT, outPRED)
aurocIndividual = computeAUROC(outGT, outPRED, 16)
aurocMean = np.array(aurocIndividual).mean()
print('AUROC mean ', aurocMean)
for i in range(0, len(aurocIndividual)):
print(CLASS_NAMES[i], ' ', aurocIndividual[i])
return
def __init__(self,
train,
means=None,
stds=None,
size=None,
resize=None,
scale=None,
ratio=None,
colorjitter=None,
random_grayscale=None,
random_hflip=None,
tencrops=False):
means = means if means is not None else cfg.tfm_means
stds = stds if stds is not None else cfg.tfm_stds
size = size if size is not None else cfg.tfm_size
resize = resize if resize is not None else cfg.tfm_resize
scale = scale if scale is not None else cfg.tfm_scale
ratio = ratio if ratio is not None else cfg.tfm_ratio
colorjitter = colorjitter if colorjitter is not None else cfg.tfm_colorjitter
random_grayscale = random_grayscale if random_grayscale is not None else cfg.tfm_random_grayscale
random_hflip = random_hflip if random_hflip is not None else cfg.tfm_random_hflip
self.transforms = []
if train:
# size transform
if scale is not None:
logger.debug('Training samples will be '
'random resized with scale [%s] and ratio [%s] '
'then cropped to size [%s]' %
(scale, ratio, size))
self.transforms.append(
transforms.RandomResizedCrop(size=size,
scale=scale,
ratio=ratio))
else:
logger.debug(
'Training samples will be resized to [%s] and then '
'random cropped into [%s]' % (resize, size))
self.transforms.append(transforms.Resize(size=resize))
self.transforms.append(transforms.RandomCrop(size))
# colorjitter
if colorjitter is not None:
logger.debug(
'Training samples will be enhanced with colorjitter: '
'[%s]' % str(colorjitter))
self.transforms.append(transforms.ColorJitter(*colorjitter))
# grayscale
if random_grayscale > 0:
logger.debug('Training samples will be randomly converted to '
'grayscale with probability [%f]' %
random_grayscale)
self.transforms.append(
transforms.RandomGrayscale(p=random_grayscale))
# random hflip
if random_hflip > 0:
logger.debug(
'Training samples will be random horizontally flip '
'with probability [%f]' % random_hflip)
self.transforms.append(
transforms.RandomHorizontalFlip(p=random_hflip))
else:
self.transforms.append(transforms.Resize(resize))
if not tencrops:
logger.debug(
'Testing samples will be resized to [%s] and then '
'center crop to [%s]' % (resize, size))
self.transforms.append(transforms.CenterCrop(size))
else:
self.transforms.append(transforms.TenCrop(size))
logger.debug(
'Testing sampels will be resized to [%s] and then '
'ten cropped to [%s]' % (resize, size))
to_tensor = transforms.ToTensor()
# to tensor and normalize
if means is not None and stds is not None:
logger.debug('Samples will be normalised with means: [%s] '
'and stds: [%s]' % (means, stds))
normalise = transforms.Normalize(means, stds)
if train or not tencrops:
self.transforms.append(to_tensor)
self.transforms.append(normalise)
else:
self.transforms.append(
transforms.Lambda(lambda crops: torch.stack(
[normalise(to_tensor(crop)) for crop in crops])))
else:
logger.debug('Samples will not be normalised')
if train or not tencrops:
self.transforms.append(to_tensor)
else:
self.transforms.append(
transforms.Lambda(lambda crops: torch.stack(
[to_tensor(crop) for crop in crops])))
import pandas as pd pathTestData='../../processed_data/test' IRID_normalize = transforms.Normalize([0.511742964836, 0.243537961753, 0.0797484182405], [0.223165616204, 0.118469339976, 0.0464971614141]) IMAGENET_normalize = transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) #-------------------- SETTINGS: DATASET BUILDERS datasetTest = DatasetGenerator(pathImageDirectory=pathTestData, transform=None) dataLoaderTest = DataLoader(dataset=datasetTest, batch_size=1, num_workers=8, shuffle=False, pin_memory=False) #---------------------custom transforms transformListIRID = [] transformListIMAGENET = [] transformListIMAGENET.append(transforms.ToPILImage()) transformListIMAGENET.append(transforms.Resize(256)) transformListIMAGENET.append(transforms.TenCrop(224)) transformListIMAGENET.append(transforms.Lambda(lambda crops: torch.stack([transforms.ToTensor()(crop) for crop in crops]))) transformListIMAGENET.append(transforms.Lambda(lambda crops: torch.stack([IMAGENET_normalize(crop) for crop in crops]))) transformIMAGENET = transforms.Compose(transformListIMAGENET) imgs, labs, paths = next(iter(dataLoaderTest)) print (imgs.size()) # simgs = Image.fromarray(simgs) print (transformIMAGENET(imgs.squeeze()).unsqueeze(0).size())
def __init__(self, dataset, valsize=640, phase='train'):
self.valsize = valsize
if dataset == 'NYU':
data_dir = '/home/gzx/RGBD/NYU-DV2/nyudv2_data_label_extractor/'
input_list = sio.loadmat(
data_dir +
'splits.mat')['trainNdxs'][:, 0].tolist() + sio.loadmat(
data_dir + 'splits.mat')['testNdxs'][::5][:, 0].tolist(
) if phase == 'train' else sio.loadmat(
data_dir + 'splits.mat')['testNdxs']
self.image_list = [
data_dir + 'data/images/' + 'img_%d.png' % (input + 5000)
for input in input_list
]
self.label_list = [
data_dir + 'data/label40/' + 'img_%d.png' % (input + 5000)
for input in input_list
]
self.depth_list = [
data_dir + 'data/depth/' + 'img_%d.png' % (input + 5000)
for input in input_list
]
elif dataset == 'SUN':
sun_phase = phase if phase == 'train' else 'test'
data_dir = '/home/gzx/RGBD/'
input_list = range(1, 5286) if phase == 'train' else range(1, 5051)
self.image_list = [
data_dir + 'SUN/%s_image/' % sun_phase + 'img-%06d.jpg' % input
for input in input_list
]
self.label_list = [
data_dir + 'SUN/sunrgbd-meta-data/labels/%s/' % sun_phase +
'img-%06d.png' %
(input if sun_phase == 'test' else input + 5050)
for input in input_list
]
self.depth_list = [
data_dir + 'SUN/sunrgbd_%s_depth/' % sun_phase +
'%d.png' % input for input in input_list
]
else:
assert (1 < 0)
self.dataset = dataset
def base_transform(t='img'):
if t == 'label':
interpolation = Image.NEAREST
else:
interpolation = Image.BILINEAR
return {
'train':
transforms.Compose([
transforms.Resize(720, interpolation=interpolation),
transforms.RandomRotation(15, resample=interpolation),
transforms.RandomResizedCrop(640,
scale=(0.8, 1.5),
ratio=(1.2, 0.85),
interpolation=interpolation),
transforms.RandomHorizontalFlip(),
]),
'val':
transforms.Compose([
transforms.Resize((int(0.75 * self.valsize), self.valsize),
interpolation=interpolation),
]),
'test':
transforms.Compose([
transforms.Resize(720, interpolation=interpolation),
transforms.Resize((640, 640), interpolation=interpolation),
]),
}[phase]
img_transform = {
'train':
transforms.Compose([
base_transform(),
#transforms.ColorJitter(brightness=0.2, contrast=0.3, saturation=0.3, hue=0.05),
transforms.ToTensor(),
#transforms.Lambda(lambda x: x + torch.randn_like(x) * 0.02),
]),
'val':
transforms.Compose([
base_transform(),
transforms.ToTensor(),
]),
'test':
transforms.Compose([
transforms.TenCrop(640),
transforms.Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops])),
]),
}[phase]
def image_transform(image, rand_seed=None):
if rand_seed:
random.seed(rand_seed)
image = img_transform(image).permute(1, 2, 0)
image = (image - IMG_MEAN) / IMG_STD
image = image.permute(2, 0, 1)
if image.size(0) == 1:
image = image.repeat(3, 1, 1)
return image
self.image_transform = image_transform
def depth_transform(depth, rand_seed=None):
if rand_seed:
random.seed(rand_seed)
depth = np.array(base_transform('label')(depth))[np.newaxis, :, :]
return torch.from_numpy(depth.astype(np.float32))
self.depth_transform = depth_transform
def label_transform(label, rand_seed=None):
if rand_seed:
random.seed(rand_seed)
label = np.array(
base_transform('label')(label)).astype('uint8') - 1
#print (label)
if label.ndim == 3:
label = label[:, :, 0]
return torch.from_numpy(label.astype(np.int64))
self.label_transform = label_transform
def main(modelfile):
model_xml = os.path.join('model', modelfile)
model_bin = model_xml.replace('.xml', '.bin')
log.info('Creating Inference Engine')
ie = IECore()
net = ie.read_network(model=model_xml, weights=model_bin)
log.info('Preparing input blobs')
input_blob = next(iter(net.input_info))
out_blob = next(iter(net.outputs))
net.batch_size = (args.batch_size * N_CROPS)
n, c, h, w = net.input_info[input_blob].input_data.shape
# for image load
normalize = transforms.Normalize(
[0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])
test_dataset = ChestXrayDataSet(data_dir=DATA_DIR,
image_list_file=TEST_IMAGE_LIST,
transform=transforms.Compose([
transforms.Resize(256),
transforms.TenCrop(224),
transforms.Lambda
(lambda crops: torch.stack([transforms.ToTensor()(crop) for crop in crops])),
transforms.Lambda
(lambda crops: torch.stack([normalize(crop) for crop in crops]))
]))
test_loader = DataLoader(dataset=test_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=False)
gt = torch.FloatTensor()
pred = torch.FloatTensor()
# loading model to the plugin
log.info('Loading model to the plugin')
#config = {'CPU_THREADS_NUM': '48', 'CPU_THROUGHPUT_STREAMS': 'CPU_THROUGHPUT_AUTO'}
config = {'CPU_THROUGHPUT_STREAMS': '%d' % args.cpu_throughput_streams}
exec_net = ie.load_network(network=net, device_name='CPU', config=config, num_requests=args.num_requests)
# Number of requests
infer_requests = exec_net.requests
print('reqeuest len', len(infer_requests))
request_queue = InferRequestsQueue(infer_requests, out_blob)
start_time = timeit.default_timer()
for i, (inp, target) in enumerate(test_loader):
bs, n_crops, c, h, w = inp.size()
images = inp.view(-1, c, h, w).numpy()
if bs != args.batch_size:
images2 = np.zeros(shape=(args.batch_size * n_crops, c, h, w))
images2[:bs*n_crops, :c, :h, :w] = images
images = images2
infer_request = request_queue.get_idle_request()
infer_request.start_async({input_blob: images}, bs, target)
if i == 20:
break
# wait the latest inference executions
request_queue.wait_all()
for i, queue in enumerate(request_queue.requests):
# print(i, queue)
gt = torch.cat((gt, queue.get_ground_truth()), 0)
pred = torch.cat((pred, queue.get_prediction()), 0)
print('Elapsed time: %0.2f sec.' % (timeit.default_timer() - start_time))
AUCs = [roc_auc_score(gt.cpu()[:, i], pred.cpu()[:, i]) for i in range(N_CLASSES)]
AUC_avg = np.array(AUCs).mean()
print('The average AUC is {AUC_avg:.3f}'.format(AUC_avg=AUC_avg))
for i in range(N_CLASSES):
print('The AUC of {} is {:.3f}'.format(CLASS_NAMES[i], AUCs[i]))
time_elapsed // 60, time_elapsed % 60))
print('Best val Acc: {:4f}'.format(best_acc))
model.load_state_dict(best_model_wts)
return model, val_acc_history
crop_size = 44
transform_train = transforms.Compose([
transforms.RandomCrop(crop_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor()
])
transform_test = transforms.Compose([
transforms.TenCrop(crop_size),
transforms.Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops]))
])
if args.checkpoint is None:
start_epoch = 0
model = VGG()
optimizer = optim.SGD(model.parameters(),
lr=args.lr,
momentum=0.9,
weight_decay=5e-4)
else:
checkpoint = torch.load(args.checkpoint)
start_epoch = checkpoint['epoch'] + 1
print('\nLoaded checkpoint from epoch %d.\n' % start_epoch)
def Mytrainsform():
transform1 = transforms.Compose([
transforms.TenCrop(44),
transforms.Lambda(lambda crops: torch.stack([transforms.ToTensor()(crop) for crop in crops])),
])
return transform1
validation_dataset = ChestXrayDataSetMultiLabel(DATA_DIR,
VALIDATION_IMAGE_LIST,
transform=transform_val)
validation_dataloader = DataLoader(dataset=validation_dataset,
batch_size=BATCH_SIZE,
shuffle=False,
num_workers=4,
pin_memory=True,
drop_last=True)
test_dataset = ChestXrayDataSetMultiLabel(
data_dir=DATA_DIR,
image_list_file=TEST_IMAGE_LIST,
transform=transforms.Compose([
transforms.Resize(256),
transforms.TenCrop(224),
transforms.Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops])),
transforms.Lambda(lambda crops: torch.stack([
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
(crop) for crop in crops
]))
]))
test_dataloader = DataLoader(dataset=test_dataset,
batch_size=BATCH_SIZE,
shuffle=False,
num_workers=4,
pin_memory=True,
drop_last=True)
def main():
# initialize and load the model
cudnn.benchmark = True
if USE_DENSENET:
model = DenseNet121(N_CLASSES).cuda()
else:
model = ResNet18(N_CLASSES).cuda()
print(model)
model = torch.nn.DataParallel(model).cuda()
if (USE_DENSENET):
if os.path.isfile(CKPT_PATH):
print("=> loading checkpoint")
checkpoint = torch.load(CKPT_PATH)
model.load_state_dict(checkpoint['state_dict'])
print("=> loaded checkpoint")
else:
print("=> no checkpoint found")
normalize = transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])
#read training data and train
train_dataset = ChestXrayDataSet(data_dir=DATA_DIR,
image_list_file=TRAIN_IMAGE_LIST,
transform=transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(), normalize
]))
train_loader = DataLoader(dataset=train_dataset,
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=8,
pin_memory=True)
criterion = nn.BCELoss().cuda()
optimizer = optim.Adam(model.parameters())
for epoch in range(0, RUNS):
print("Epoch " + str(epoch + 1))
train_run(model, train_loader, optimizer, criterion, epoch)
test_dataset = ChestXrayDataSet(
data_dir=DATA_DIR,
image_list_file=TEST_IMAGE_LIST,
transform=transforms.Compose([
transforms.Resize(256),
transforms.TenCrop(224),
transforms.Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops])),
transforms.Lambda(
lambda crops: torch.stack([normalize(crop) for crop in crops]))
]))
test_loader = DataLoader(dataset=test_dataset,
batch_size=BATCH_SIZE,
shuffle=False,
num_workers=8,
pin_memory=True)
# initialize the ground truth and output tensor
gt = torch.FloatTensor()
gt = gt.cuda()
pred = torch.FloatTensor()
pred = pred.cuda()
# switch to evaluate mode
model.eval()
for i, (inp, target) in enumerate(test_loader):
target = target.cuda()
gt = torch.cat((gt, target), 0)
bs, n_crops, c, h, w = inp.size()
input_var = torch.autograd.Variable(inp.view(-1, c, h, w).cuda(),
volatile=True)
output = model(input_var)
output_mean = output.view(bs, n_crops, -1).mean(1)
pred = torch.cat((pred, output_mean.data), 0)
AUROCs = compute_AUCs(gt, pred)
AUROC_avg = np.array(AUROCs).mean()
print('The average AUROC is {AUROC_avg:.3f}'.format(AUROC_avg=AUROC_avg))
for i in range(N_CLASSES):
print('The AUROC of {} is {}'.format(CLASS_NAMES[i], AUROCs[i]))
def main():
cudnn.benchmark = True
# initialize and load the model
model = DenseNet121(N_CLASSES).cuda()
model = torch.nn.DataParallel(model).cuda()
#if os.path.isfile(CKPT_PATH):
if 0:
print("=> loading checkpoint")
checkpoint = torch.load(CKPT_PATH)
model.load_state_dict(checkpoint['state_dict'])
print("=> loaded checkpoint")
else:
print("=> no checkpoint found")
normalize = transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])
test_dataset = ChestXrayDataSet(
data_dir=DATA_DIR,
image_list_file=TEST_IMAGE_LIST,
transform=transforms.Compose([
transforms.Resize(256),
transforms.TenCrop(224),
transforms.Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops])),
transforms.Lambda(
lambda crops: torch.stack([normalize(crop) for crop in crops]))
]))
test_loader = DataLoader(dataset=test_dataset,
batch_size=BATCH_SIZE,
shuffle=False,
num_workers=LOADER_WORKERS,
pin_memory=True)
# initialize the ground truth and output tensor
gt = torch.FloatTensor()
gt = gt.cuda()
pred = torch.FloatTensor()
pred = pred.cuda()
# switch to evaluate mode
model.eval()
print("starting loop")
try:
for i, (inp, target) in enumerate(test_loader):
target = target.cuda()
gt = torch.cat((gt, target), 0)
bs, n_crops, c, h, w = inp.size()
input_var = torch.autograd.Variable(inp.view(-1, c, h, w).cuda(),
volatile=True)
output = model(input_var)
output_mean = output.view(bs, n_crops, -1).mean(1)
pred = torch.cat((pred, output_mean.data), 0)
if not i % 100:
print("iteration " + str(i))
except:
print('error in iteration: ' + str(i))
AUROCs = compute_AUCs(gt, pred)
AUROC_avg = np.array(AUROCs).mean()
print('The average AUROC is {AUROC_avg:.3f}'.format(AUROC_avg=AUROC_avg))
for i in range(N_CLASSES):
print('The AUROC of {} is {}'.format(CLASS_NAMES[i], AUROCs[i]))
cut_size = 44
total_epoch = 250
path = os.path.join(opt.dataset + '_2_' + opt.model)
# Data
print('==> Preparing data..')
transform_train = tfs.Compose([
tfs.RandomCrop(cut_size),
tfs.RandomHorizontalFlip(),
tfs.ToTensor(),
])
transform_test = tfs.Compose([
tfs.TenCrop(cut_size),
tfs.Lambda(
lambda crops: torch.stack([tfs.ToTensor()(crop) for crop in crops])),
])
trainset = FER2013(split='Training', transform=transform_train)
trainloader = torch.utils.data.DataLoader(trainset,
batch_size=opt.bs,
shuffle=True,
num_workers=1)
PublicTestset = FER2013(split='PublicTest', transform=transform_test)
PublicTestloader = torch.utils.data.DataLoader(PublicTestset,
batch_size=opt.bs,
shuffle=False,
num_workers=1)
PrivateTestset = FER2013(split='PrivateTest', transform=transform_test)
def get_data(data_path):
with open(data_path, 'rb') as f:
train_test_paths_labels = pickle.load(f)
train_paths = train_test_paths_labels[0]
val_paths = train_test_paths_labels[1]
train_labels = train_test_paths_labels[3]
val_labels = train_test_paths_labels[4]
train_num_each = train_test_paths_labels[6]
val_num_each = train_test_paths_labels[7]
print('train_paths : {:6d}'.format(len(train_paths)))
print('train_labels : {:6d}'.format(len(train_labels)))
print('valid_paths : {:6d}'.format(len(val_paths)))
print('valid_labels : {:6d}'.format(len(val_labels)))
train_labels = np.asarray(train_labels, dtype=np.int64)
val_labels = np.asarray(val_labels, dtype=np.int64)
train_transforms = None
test_transforms = None
if use_flip == 0:
train_transforms = transforms.Compose([
transforms.Resize((250, 250)),
transforms.RandomCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.44893518,0.3226702,0.34424525],[0.22357443,0.18503027,0.1900281])
])
elif use_flip == 1:
train_transforms = transforms.Compose([
transforms.Resize((250, 250)),
RandomCrop(224),
ColorJitter(brightness=0.1, contrast=0.1, saturation=0.1, hue=0.05),
RandomHorizontalFlip(),
RandomRotation(5),
transforms.ToTensor(),
transforms.Normalize([0.44893518,0.3226702,0.34424525],[0.22357443,0.18503027,0.1900281])
])
if crop_type == 0:
test_transforms = transforms.Compose([
transforms.Resize((250, 250)),
transforms.RandomCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.44893518,0.3226702,0.34424525],[0.22357443,0.18503027,0.1900281])
])
elif crop_type == 1:
test_transforms = transforms.Compose([
transforms.Resize((250, 250)),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.44893518,0.3226702,0.34424525],[0.22357443,0.18503027,0.1900281])
])
elif crop_type == 2:
test_transforms = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize([0.44893518,0.3226702,0.34424525],[0.22357443,0.18503027,0.1900281])
])
elif crop_type == 5:
test_transforms = transforms.Compose([
transforms.Resize((250, 250)),
transforms.FiveCrop(224),
Lambda(lambda crops: torch.stack([transforms.ToTensor()(crop) for crop in crops])),
Lambda(
lambda crops: torch.stack(
[transforms.Normalize([0.44893518,0.3226702,0.34424525],[0.22357443,0.18503027,0.1900281])(crop) for crop in crops]))
])
elif crop_type == 10:
test_transforms = transforms.Compose([
transforms.Resize((250, 250)),
transforms.TenCrop(224),
Lambda(lambda crops: torch.stack([transforms.ToTensor()(crop) for crop in crops])),
Lambda(
lambda crops: torch.stack(
[transforms.Normalize([0.44893518,0.3226702,0.34424525],[0.22357443,0.18503027,0.1900281])(crop) for crop in crops]))
])
train_dataset = CholecDataset(train_paths, train_labels, train_transforms)
val_dataset = CholecDataset(val_paths, val_labels, test_transforms)
# test_dataset = CholecDataset(test_paths, test_labels, test_transforms)
return train_dataset, train_num_each, val_dataset, val_num_each
norm_mean = [0.485, 0.456, 0.406]
norm_std = [0.229, 0.224, 0.225]
train_transform = transforms.Compose([
transforms.Resize((256)), # (256, 256) 区别,(256)表示短边到256，长宽比不变
transforms.CenterCrop(256),
transforms.RandomCrop(224),
transforms.RandomHorizontalFlip(p=0.5), # p表示水平翻转概率
transforms.ToTensor(),
transforms.Normalize(norm_mean, norm_std),
])
normalizes = transforms.Normalize(norm_mean, norm_std)
valid_transform = transforms.Compose([
transforms.Resize((256, 256)),
transforms.TenCrop(224, vertical_flip=False),
transforms.Lambda(lambda crops: torch.stack(
[normalizes(transforms.ToTensor()(crop)) for crop in crops])),
]) # 5维张量：[B,10,C,H,W]
# 构建MyDataset实例
train_data = CatDogDataset(data_dir=data_dir,
mode="train",
transform=train_transform)
valid_data = CatDogDataset(data_dir=data_dir,
mode="valid",
transform=valid_transform)
# 构建DataLoder
train_loader = DataLoader(dataset=train_data,
batch_size=BATCH_SIZE,
def test(pathDirData, pathFileTest, pathModel, nnArchitecture,
nnClassCount, nnIsTrained, trBatchSize, transResize, transCrop,
launchTimeStamp):
CLASS_NAMES = [
'Atelectasis', 'Cardiomegaly', 'Effusion', 'Infiltration', 'Mass',
'Nodule', 'Pneumonia', 'Pneumothorax', 'Consolidation', 'Edema',
'Emphysema', 'Fibrosis', 'Pleural_Thickening', 'Hernia'
]
cudnn.benchmark = True
#-------------------- SETTINGS: NETWORK ARCHITECTURE, MODEL LOAD
if nnArchitecture == 'DENSE-NET-121':
model = DenseNet121(nnClassCount, nnIsTrained).cuda()
elif nnArchitecture == 'DENSE-NET-169':
model = DenseNet169(nnClassCount, nnIsTrained).cuda()
elif nnArchitecture == 'DENSE-NET-201':
model = DenseNet201(nnClassCount, nnIsTrained).cuda()
model = torch.nn.DataParallel(model).cuda()
modelCheckpoint = torch.load(pathModel)
model.load_state_dict(modelCheckpoint['state_dict'])
#-------------------- SETTINGS: DATA TRANSFORMS, TEN CROPS
normalize = transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])
#-------------------- SETTINGS: DATASET BUILDERS
transformList = []
transformList.append(transforms.Resize(transResize))
transformList.append(transforms.TenCrop(transCrop))
transformList.append(
transforms.Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops])))
transformList.append(
transforms.Lambda(lambda crops: torch.stack(
[normalize(crop) for crop in crops])))
transformSequence = transforms.Compose(transformList)
datasetTest = DatasetGenerator(pathImageDirectory=pathDirData,
pathDatasetFile=pathFileTest,
transform=transformSequence)
dataLoaderTest = DataLoader(dataset=datasetTest,
batch_size=trBatchSize,
num_workers=0,
shuffle=False,
pin_memory=False)
outGT = torch.FloatTensor().cuda()
outPRED = torch.FloatTensor().cuda()
model.eval()
for i, (input, target) in enumerate(dataLoaderTest):
target = target.cuda()
outGT = torch.cat((outGT, target), 0)
bs, n_crops, c, h, w = input.size()
varInput = torch.autograd.Variable(input.view(-1, c, h, w).cuda())
out = model(varInput)
outMean = out.view(bs, n_crops, -1).mean(1)
outPRED = torch.cat((outPRED, outMean.data), 0)
aurocIndividual = ChexnetTrainer.computeAUROC(outGT, outPRED,
nnClassCount)
aurocMean = np.array(aurocIndividual).mean()
print('AUROC mean ', aurocMean)
for i in range(0, len(aurocIndividual)):
print(CLASS_NAMES[i], ' ', aurocIndividual[i])
return
def predict(pathDirData, pathFileTest, pathModel, nnArchitecture,
nnClassCount, nnIsTrained, trBatchSize, transResize, transCrop,
launchTimeStamp):
cudnn.benchmark = True
#-------------------- SETTINGS: NETWORK ARCHITECTURE, MODEL LOAD
if nnArchitecture == 'DENSE-NET-121':
model = DenseNet121(nnClassCount, nnIsTrained).cuda()
elif nnArchitecture == 'DENSE-NET-169':
model = DenseNet169(nnClassCount, nnIsTrained).cuda()
elif nnArchitecture == 'DENSE-NET-201':
model = DenseNet201(nnClassCount, nnIsTrained).cuda()
model = torch.nn.DataParallel(model).cuda()
modelCheckpoint = torch.load(pathModel)
# https://github.com/KaiyangZhou/deep-person-reid/issues/23
pattern = re.compile(
r'^(.*denselayer\d+\.(?:norm|relu|conv))\.((?:[12])\.(?:weight|bias|running_mean|running_var))$'
)
state_dict = modelCheckpoint['state_dict']
for key in list(state_dict.keys()):
res = pattern.match(key)
if res:
new_key = res.group(1) + res.group(2)
state_dict[new_key] = state_dict[key]
del state_dict[key]
model.load_state_dict(state_dict)
#-------------------- SETTINGS: DATA TRANSFORMS, TEN CROPS
normalize = transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])
#-------------------- SETTINGS: DATASET BUILDERS
transformList = []
transformList.append(transforms.Resize(transResize))
transformList.append(transforms.TenCrop(transCrop))
transformList.append(
transforms.Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops])))
transformList.append(
transforms.Lambda(lambda crops: torch.stack(
[normalize(crop) for crop in crops])))
transformSequence = transforms.Compose(transformList)
datasetTest = DatasetGenerator(pathImageDirectory=pathDirData,
pathDatasetFile=pathFileTest,
transform=transformSequence)
dataLoaderTest = DataLoader(dataset=datasetTest,
batch_size=trBatchSize,
shuffle=False,
pin_memory=True) #num_workers=N
outGT = torch.FloatTensor().cuda()
outPRED = torch.FloatTensor().cuda()
model.eval()
for i, (input, target) in enumerate(dataLoaderTest):
target = target.cuda()
outGT = torch.cat((outGT, target), 0)
bs, n_crops, c, h, w = input.size()
varInput = torch.autograd.Variable(input.view(-1, c, h, w).cuda(),
volatile=True)
out = model(varInput)
outMean = out.view(bs, n_crops, -1).mean(1)
outPRED = torch.cat((outPRED, outMean.data), 0)
pneumonia_probas = []
for p in outPRED.cpu().data.numpy()[:, CLASS_NAMES.index('Pneumonia')]:
pneumonia_probas.append(p)
return pneumonia_probas
def main():
global args
args = parser.parse_args()
#fix random seeds
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
np.random.seed(args.seed)
best_prec1 = 0
# load model
model = load_model(args.model)
cudnn.benchmark = True
# freeze the features layers
for block in model.module:
try:
for param in block.parameters():
param.requires_grad = False
except:
for layer in block:
for param in layer.parameters():
param.requires_grad = False
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()
# data loading code
traindir = os.path.join(args.data, 'train')
valdir = os.path.join(args.data, 'val_in_folders')
valdir_double = os.path.join(args.data, 'val_in_double_folders')
valdir_list = []
# Load in AoA table if needed
if args.aoaval:
aoalist = pd.read_csv('matchingAoA_ImageNet_excel.csv')
for index, row in aoalist.iterrows():
node = row['node']
aoa = float(row['aoa'])
if not math.isnan(aoa):
valdir_list.append({
'node': node,
'pth': os.path.join(valdir_double, node),
'aoa': aoa
})
else:
print('Not found %s' % node)
#valdir_list=valdir_list[:5] trim for testing
print('Using %d validation categories for aoa' % len(valdir_list))
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
# Can't do validation if the tencrops option is chosne a
if args.tencrops:
transformations_val = [
transforms.Resize(256),
transforms.TenCrop(224),
transforms.Lambda(lambda crops: torch.stack(
[normalize(transforms.ToTensor()(crop)) for crop in crops])),
]
else:
transformations_val = [
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(), normalize
]
transformations_train = [
transforms.Resize(256),
transforms.CenterCrop(256),
transforms.RandomCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(), normalize
]
train_dataset = datasets.ImageFolder(
traindir, transform=transforms.Compose(transformations_train))
val_dataset = datasets.ImageFolder(
valdir, transform=transforms.Compose(transformations_val))
# Load up individual categories for AoA validation
if args.aoaval:
val_list_dataset = []
val_list_loader = []
for entry in valdir_list:
val_list_dataset.append(
datasets.ImageFolder(
entry['pth'],
transform=transforms.Compose(transformations_val)))
val_list_loader.append(
torch.utils.data.DataLoader(val_list_dataset[-1],
batch_size=50,
shuffle=False,
num_workers=args.workers))
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=int(args.batch_size /
2),
shuffle=False,
num_workers=args.workers)
# logistic regression
reglog = RegLog(args.conv, len(train_dataset.classes)).cuda()
optimizer = torch.optim.SGD(filter(lambda x: x.requires_grad,
reglog.parameters()),
args.lr,
momentum=args.momentum,
weight_decay=10**args.weight_decay)
# create logs
exp_log = os.path.join(args.exp, 'log')
if not os.path.isdir(exp_log):
os.makedirs(exp_log)
loss_log = Logger(os.path.join(exp_log, 'loss_log'))
prec1_log = Logger(os.path.join(exp_log, 'prec1'))
prec5_log = Logger(os.path.join(exp_log, 'prec5'))
for epoch in range(args.epochs):
end = time.time()
# If savedmodel already exists, load this
savedmodelpth = os.path.join(args.exp, 'model_best.pth.tar')
if os.path.exists(savedmodelpth):
print('Loading saved decoder %s' % savedmodelpth)
model_with_decoder = torch.load(savedmodelpth)
reglog.load_state_dict(model_with_decoder['reglog_state_dict'])
else:
# train for one epoch
train(train_loader, model, reglog, criterion, optimizer, epoch)
# evaluate on validation set
prec1, prec5, loss = validate(val_loader, model, reglog, criterion)
loss_log.log(loss)
prec1_log.log(prec1)
prec5_log.log(prec5)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
if is_best:
filename = 'model_best.pth.tar'
else:
filename = 'checkpoint.pth.tar'
torch.save(
{
'epoch': epoch + 1,
'arch': 'alexnet',
'state_dict': model.state_dict(),
'reglog_state_dict':
reglog.state_dict(), # Also save decoding layers
'prec5': prec5,
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
},
savedmodelpth)
if args.aoaval:
# Validate individual categories, so loss can be compared to AoA
# # To check weights loaded OK
# # evaluate on validation set
# prec1, prec5, loss = validate(val_loader, model, reglog, criterion)
# loss_log.log(loss)
# prec1_log.log(prec1)
# prec5_log.log(prec5)
aoares = {}
for idx, row in enumerate(zip(valdir_list, val_list_loader)):
# evaluate on validation set
print("AOA validation %d/%d" % (idx, len(valdir_list)))
prec1_tmp, prec5_tmp, loss_tmp = validate(
row[1], model, reglog, criterion)
aoares[row[0]['node']] = {
'prec1': float(prec1_tmp),
'prec5': float(prec5_tmp),
'loss': float(loss_tmp),
'aoa': row[0]['aoa']
}
# Save to JSON
aoapth = os.path.join(args.exp, 'aoaresults.json')
with open(aoapth, 'w') as f:
json.dump(aoares, f)
# Data
print('==> Preparing data..')
transform_train = transforms.Compose([
transforms.RandomCrop(44),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
utils.Cutout(n_holes=1, length=13),
#transforms.Normalize((0.589499, 0.45687565, 0.40699387),
#(0.25357702, 0.23312956, 0.23275192)),
transforms.Normalize((0.56010324, 0.43693307, 0.39122295),
(0.23726934, 0.21260591, 0.20737909)), #Augmentation
])
transform_test = transforms.Compose([
transforms.TenCrop(44),
#transforms.Lambda(lambda crops: torch.stack([transforms.ToTensor()(crop) for crop in crops])),
transforms.Lambda(lambda crops: torch.stack([transforms.Normalize(
mean=[0.589499, 0.45687565, 0.40699387], std=[0.25357702, 0.23312956, 0.23275192])
(transforms.ToTensor()(crop)) for crop in crops])),
])
trainset = RAF(split = 'Training', transform=transform_train)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=opt.train_bs, shuffle=True, num_workers=1)
PrivateTestset = RAF(split = 'PrivateTest', transform=transform_test)
PrivateTestloader = torch.utils.data.DataLoader(PrivateTestset, batch_size=opt.test_bs, shuffle=False, num_workers=1)
# Model
if opt.model == 'VGG19':
net = VGG('VGG19')
def load_imagenet(data_root,
batch_size_train,
batch_size_val,
augement,
num_workers=12,
ret='loader',
**kwargs):
"""
augment: {'train': ['rresize-1crop', '1resize-1crop', '2resize-1crop', '1resize-0crop'][0],
'val': ['center-crop'][0]}
'rresize-1crop' random resized and crop, used in torch.transforms.RandomResizedCrop().
'2resize-1crop' used in resnet paper, resize to (256 or 480), then crop to 224, then flip.
MinMaxResizedCrop: 先缩放再裁切 Size'=Size*scale*1
minmax: (256, 480) (300, 500)
下界太小，导致物体有效像素占比太小，物体圆满性被破坏！ 切太稀！
上界太大，会导致图片被切的太碎，物体完整性被破坏！ 切太碎！
RandomResizedCrop: 先裁切再缩放 Size'=Size*scale*β_auto_crop
scale： (1/5, 4/5)=(0.2~0.8)
下界太小，会导致图片被切的太碎，物体完整性被破坏！ 切太碎！
上界太大，导致物体有效像素占比太小，物体圆满性被破坏！ 切太稀！
"""
# titanxp_root /data/dataset/ImageNetDownSample/ 64x64 32x32
# 1080ti_root /data0/ImageNet_ILSVRC2012/
# K40_root no
traindir = os.path.join(data_root, 'train')
valdir = os.path.join(data_root, 'val')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
insize = augement.get('insize', None)
xscale = augement.get('xscale', []) # for train
vscale = augement.get('vscale', []) # for validate
assert insize is not None, '输入图片的尺寸 <insize @cfg.data_augment> 未设置.'
assert isinstance(
xscale,
(list, tuple)), '<xscale @cfg.data_augment> must be list or tuple'
assert isinstance(
vscale,
(list, tuple)), '<vscale @cfg.data_augment> must be list or tuple'
if augement['train'] == 'rresize-1crop':
# +随机尺寸变换 + 随机形状变换 +随机平移变换
scale = augement.get('scale', (0.08, 1))
ratio = augement.get('ratio', (3. / 4, 4. / 3))
resize_crop = [transforms.RandomResizedCrop(insize, scale, ratio)]
elif augement['train'] == 'rotate-rresize-1crop':
degree = augement.get('degree', (-5, 5))
scale = augement.get('scale', (0.08, 1))
ratio = augement.get('ratio', (3. / 4, 4. / 3))
if tuple(degree) == (0, 0):
resize_crop = [transforms.RandomResizedCrop(insize, scale, ratio)]
else:
resize_crop = [
transforms.RandomRotation(degree),
transforms.RandomResizedCrop(insize, scale, ratio)
]
elif augement['train'] == 'rotate-rresize-1crop-color':
degree = augement.get('degree', (-5, 5))
scale = augement.get('scale', (0.08, 1))
ratio = augement.get('ratio', (3. / 4, 4. / 3))
brightness = augement.get('brightness', 0.2)
contrast = augement.get('contrast', 0.2)
saturation = augement.get('saturation', 0.2)
hue = augement.get('hue', 0.2)
if tuple(degree) == (0, 0):
resize_crop = [transforms.RandomResizedCrop(insize, scale, ratio)]
else:
resize_crop = [
transforms.RandomRotation(degree),
transforms.RandomResizedCrop(insize, scale, ratio)
]
color_jitter = transforms.ColorJitter(brightness, contrast, saturation,
hue)
resize_crop.append(color_jitter)
elif augement['train'] == 'xresize-1crop':
# +固定X尺寸变换 -形状变换 +随机平移变换
assert len(xscale) >= 1 and min(xscale) >= insize
resize_list = [transforms.Resize(x) for x in xscale]
resize_crop = [
transforms.RandomChoice(resize_list),
transforms.RandomCrop(insize)
]
elif augement['train'] == 'minmax-resize-1crop':
assert len(xscale) == 2 and min(xscale) > insize
resize_crop = [
MinMaxResize(min(xscale), max(xscale)),
transforms.RandomCrop(insize)
]
elif augement['train'] == 'rotate-minmax-resize-1crop':
assert len(xscale) == 2 and min(xscale) > insize
degree = augement.get('degree', (-5, 5))
if tuple(degree) == (0, 0):
resize_crop = [
MinMaxResize(min(xscale), max(xscale)),
transforms.RandomCrop(insize)
]
else:
resize_crop = [
transforms.RandomRotation(degree),
MinMaxResize(min(xscale), max(xscale)),
transforms.RandomCrop(insize)
]
elif augement['train'] == '1resize-0crop':
resize_crop = [transforms.Resize(insize)]
elif augement['train'] == '0resize-1crop':
resize_crop = [transforms.RandomCrop(insize)]
elif augement['train'] == 'xresize-1affine-1crop':
# +固定X尺寸变换 + 仿射变换 -形状变换 +随机平移变换
assert len(xscale) >= 1 and min(xscale) >= insize
resize_list = [transforms.Resize(x) for x in xscale]
resize_crop = [
transforms.RandomChoice(resize_list),
transforms.RandomAffine(degrees=(-15, 15),
translate=None,
scale=None,
shear=(-5, 5)),
transforms.RandomCrop(insize)
]
elif augement['train'] == 'xresize-raffine-1crop':
# +固定X尺寸变换 + 随机仿射变换 -形状变换 +随机平移变换
assert len(xscale) >= 1 and min(xscale) >= insize
resize_list = [transforms.Resize(x) for x in xscale]
resize_crop = [
transforms.RandomChoice(resize_list),
transforms.RandomApply(
[transforms.RandomAffine(degrees=(-15, 15), shear=(-5, 5))]),
transforms.RandomCrop(insize)
]
elif augement['train'] == 'xresize-caffine-1crop':
# +固定X尺寸变换 + 仿射变换 -形状变换 +随机平移变换
assert len(xscale) >= 1 and min(xscale) >= insize
resize_list = [transforms.Resize(x) for x in xscale]
resize_crop = [
transforms.RandomChoice(resize_list),
transforms.RandomChoice([
transforms.RandomAffine(degrees=15, shear=None),
transforms.RandomAffine(degrees=0, shear=10)
]),
transforms.RandomCrop(insize)
]
else:
raise NotImplementedError('Unknown Resize & Crop Method %s ...' %
(augement, ))
if augement.get('color', None):
other_process = [
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
ColorAugmentation(),
normalize,
]
else:
other_process = [
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]
train_transform = transforms.Compose(resize_crop + other_process)
if augement['val'] == '1resize-1crop':
assert len(vscale) == 1 and vscale[0] >= insize
val_transform = transforms.Compose([
transforms.Resize(vscale[0]),
transforms.CenterCrop(insize),
transforms.ToTensor(),
normalize,
])
elif augement['val'] == '1resize-0crop':
assert insize > 0
val_transform = transforms.Compose([
transforms.Resize(insize),
transforms.ToTensor(),
normalize,
])
elif augement['val'] == '1resize-1crop-color':
assert len(vscale) == 1 and vscale[0] >= insize
val_transform = transforms.Compose([
transforms.Resize(vscale[0]),
transforms.CenterCrop(insize),
transforms.ToTensor(),
ColorAugmentation(),
normalize,
])
elif augement['val'] == '1resize-5crop':
assert len(vscale) == 1 and vscale[0] >= insize
val_transform = transforms.Compose([
transforms.Resize(vscale[0]),
transforms.FiveCrop(insize),
transforms.Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops])),
transforms.Lambda(lambda crops: torch.stack(
[normalize(crop) for crop in crops])),
])
elif augement['val'] == '1resize-10crop':
assert len(vscale) == 1 and vscale[0] >= insize
val_transform = transforms.Compose([
transforms.Resize(vscale[0]),
transforms.TenCrop(insize),
transforms.Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops])),
transforms.Lambda(lambda crops: torch.stack(
[normalize(crop) for crop in crops])),
])
else:
raise NotImplementedError('Unknown Resize & Crop Method %s ...' %
(augement, ))
print('\nWarning: Please Assure Train-Transform is -->\n' +
repr(train_transform))
print('\nWarning: Please Assure Val-Transform is -->\n' +
repr(val_transform))
train_dataset = datasets.ImageFolder(traindir, train_transform)
val_dataset = datasets.ImageFolder(valdir, val_transform)
test_dataset = val_dataset
if ret == 'dataset':
return train_dataset, val_dataset, test_dataset
elif ret == 'loader':
train_loader = torch.utils.data.DataLoader(dataset=train_dataset,
batch_size=batch_size_train,
shuffle=True,
num_workers=num_workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(dataset=val_dataset,
batch_size=batch_size_val,
shuffle=False,
num_workers=16,
pin_memory=False)
test_loader = val_loader
return train_loader, val_loader, test_loader
else:
raise NotImplementedError('<ret> must be loader or dataset!')
def load_imagenet(data_root,
bsize_train,
bsize_val,
augement,
num_workers=12,
result='loader',
**kwargs):
"""
augment: {'train': ['rresize-1crop', '1resize-1crop', '2resize-1crop', '1resize-0crop'][0],
'val': ['center-crop'][0]}
imsize / insize = 0.875, eg. 256/0.875=224
"""
traindir = os.path.join(data_root, 'train')
valdir = os.path.join(data_root, 'val')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
imsize = augement.get('imsize', None)
insize = augement.get('insize', None)
interp = augement.get('interp', 'bilinear')
assert imsize is not None, '数据扩增的图片尺寸 <imsize @cfg.data_augment> 未设置.'
assert insize is not None, '网络输入的图片尺寸 <insize @cfg.data_augment> 未设置.'
assert imsize >= insize, 'imsize must > insize.'
assert interp in ['linear', 'bilinear', 'bicubic']
interp = getattr(Image, interp.upper())
if augement['train'] == 'rotate-rresize-1crop':
degree = augement.get('degree', (0, 0))
scale = augement.get('scale', (0.08, 1))
ratio = augement.get('ratio', (3. / 4, 4. / 3))
resize_crop = []
if tuple(degree) != (0, 0):
resize_crop.append(transforms.RandomRotation(degree))
resize_crop.append(
transforms.RandomResizedCrop(insize, scale, ratio, interp))
elif augement['train'] == 'rotate-rresize-1crop-colorj':
degree = augement.get('degree', (0, 0))
scale = augement.get('scale', (0.08, 1))
ratio = augement.get('ratio', (3. / 4, 4. / 3))
brightness = augement.get('brightness', 0.2)
contrast = augement.get('contrast', 0.2)
saturation = augement.get('saturation', 0.2)
hue = augement.get('hue', 0.2)
resize_crop = []
if tuple(degree) != (0, 0):
resize_crop.append(transforms.RandomRotation(degree))
resize_crop.append(
transforms.RandomResizedCrop(insize, scale, ratio, interp))
color_jitter = transforms.ColorJitter(brightness, contrast, saturation,
hue)
resize_crop.append(color_jitter)
elif augement['train'] == 'rotate-mmresize-1crop':
mmsize = augement.get('mmsize', [])
assert len(mmsize) == 2 and min(mmsize) > insize
degree = augement.get('degree', (0, 0))
resize_crop = []
if tuple(degree) != (0, 0):
resize_crop.append(transforms.RandomRotation(degree))
resize_crop.append(MinMaxResize(min(mmsize), max(mmsize), interp))
resize_crop.append(transforms.RandomCrop(insize))
elif augement['train'] == 'mmresize-raffine-1crop':
mmsize = augement.get('mmsize', [])
assert len(mmsize) == 2 and min(mmsize) > insize
resize_crop = [
MinMaxResize(min(mmsize), max(mmsize), interp),
transforms.RandomApply(
[transforms.RandomAffine(degrees=(-15, 15), shear=(-5, 5))]),
transforms.RandomCrop(insize)
]
else:
raise NotImplementedError('Unknown Resize & Crop Method %s ...' %
(augement, ))
other_process = [
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]
if augement.get('color', None):
other_process.insert(-1, ColorAugmentation())
train_transform = transforms.Compose(resize_crop + other_process)
if augement['val'] == '1resize-0crop':
assert insize > 0
val_transform = transforms.Compose([
transforms.Resize(insize, interp),
transforms.ToTensor(),
normalize,
])
elif augement['val'] == '1resize-1crop':
val_transform = transforms.Compose([
transforms.Resize(imsize, interp),
transforms.CenterCrop(insize),
transforms.ToTensor(),
normalize,
])
elif augement['val'] == '1resize-10crop':
val_transform = transforms.Compose([
transforms.Resize(imsize, interp),
transforms.TenCrop(insize),
transforms.Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops])),
transforms.Lambda(lambda crops: torch.stack(
[normalize(crop) for crop in crops])),
])
else:
raise NotImplementedError('Unknown Resize & Crop Method %s ...' %
(augement, ))
# print('\nWarning: Please Assure Train-Transform is -->\n' + repr(train_transform))
# print('\nWarning: Please Assure Val-Transform is -->\n' + repr(val_transform))
train_dataset = datasets.ImageFolder(traindir, train_transform)
val_dataset = datasets.ImageFolder(valdir, val_transform)
test_dataset = val_dataset
if result == 'dataset':
return train_dataset, val_dataset, test_dataset
train_loader = torch.utils.data.DataLoader(dataset=train_dataset,
batch_size=bsize_train,
shuffle=True,
num_workers=num_workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(dataset=val_dataset,
batch_size=bsize_val,
shuffle=False,
num_workers=num_workers,
pin_memory=True)
test_loader = torch.utils.data.DataLoader(dataset=test_dataset,
batch_size=bsize_val,
shuffle=False,
num_workers=num_workers,
pin_memory=True)
return train_loader, val_loader, test_loader
def main():
args = parser.parse_args()
print(args)
# fix random seeds
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
np.random.seed(args.seed)
# create model and move it to gpu
model = load_model(args.model)
model.top_layer = nn.Linear(model.top_layer.weight.size(1), 20)
model.cuda()
cudnn.benchmark = True
# what partition of the data to use
if args.split == 'train':
args.test = 'val'
elif args.split == 'trainval':
args.test = 'test'
# data loader
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
dataset = VOC2007_dataset(args.vocdir,
split=args.split,
transform=transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.RandomResizedCrop(
224,
scale=(args.min_scale, args.max_scale),
ratio=(1, 1)),
transforms.ToTensor(),
normalize,
]))
loader = torch.utils.data.DataLoader(dataset,
batch_size=16,
shuffle=False,
num_workers=24,
pin_memory=True)
print('PASCAL VOC 2007 ' + args.split + ' dataset loaded')
# re initialize classifier
for y, m in enumerate(model.classifier.modules()):
if isinstance(m, nn.Linear):
m.weight.data.normal_(0, 0.01)
m.bias.data.fill_(0.1)
model.top_layer.bias.data.fill_(0.1)
if args.fc6_8:
# freeze some layers
for param in model.features.parameters():
param.requires_grad = False
# unfreeze batchnorm scaling
if args.train_batchnorm:
for layer in model.modules():
if isinstance(layer, torch.nn.BatchNorm2d):
for param in layer.parameters():
param.requires_grad = True
# set optimizer
optimizer = torch.optim.SGD(
filter(lambda x: x.requires_grad, model.parameters()),
lr=args.lr,
momentum=0.9,
weight_decay=args.wd,
)
criterion = nn.BCEWithLogitsLoss(reduction='none')
print('Start training')
it = 0
losses = AverageMeter()
while it < args.nit:
it = train(
loader,
model,
optimizer,
criterion,
args.fc6_8,
losses,
it=it,
total_iterations=args.nit,
stepsize=args.stepsize,
)
print('Evaluation')
if args.eval_random_crops:
transform_eval = [
transforms.RandomHorizontalFlip(),
transforms.RandomResizedCrop(224,
scale=(args.min_scale,
args.max_scale),
ratio=(1, 1)),
transforms.ToTensor(),
normalize,
]
else:
transform_eval = [
transforms.Resize(256),
transforms.TenCrop(224),
transforms.Lambda(lambda crops: torch.stack(
[normalize(transforms.ToTensor()(crop)) for crop in crops]))
]
print('Train set')
train_dataset = VOC2007_dataset(
args.vocdir,
split=args.split,
transform=transforms.Compose(transform_eval))
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=1,
shuffle=False,
num_workers=24,
pin_memory=True,
)
evaluate(train_loader, model, args.eval_random_crops)
print('Test set')
test_dataset = VOC2007_dataset(
args.vocdir,
split=args.test,
transform=transforms.Compose(transform_eval))
test_loader = torch.utils.data.DataLoader(
test_dataset,
batch_size=1,
shuffle=False,
num_workers=24,
pin_memory=True,
)
evaluate(test_loader, model, args.eval_random_crops)
def test_cnn(MODEL_NAME, MODEL_NAME_TARGET, BATCH_SIZE, N_LABELS,
PATH_TO_IMAGES, DEBUG_MODE, CHECKPOINT_PATH,
CHECKPOINT_PATH_TARGET):
"""
Train torchvision model to NIH data given high level hyperparameters.
Args:
MODEL_NAME: model name
MODEL_NAME_TARGET: the other model name
BATCH_SIZE: number of batch data per training
N_LABELS: number of class labels
PATH_TO_IMAGES: path to NIH images
DEBUG_MODE: if true then no log will be created
CHECKPOINT_PATH: load checkpoint path
CHECKPOINT_PATH_TARGET: load the other checkpoint path
Returns:
# preds: torchvision model predictions on test fold with ground truth for comparison
# aucs: AUCs for each train,test tuple
"""
# use imagenet mean,std for normalization
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
normalize = transforms.Normalize(mean=mean, std=std)
# define torchvision transforms
data_transforms = {
'train':
transforms.Compose([
transforms.Resize(256),
transforms.TenCrop(224),
transforms.Lambda(lambda crops: torch.stack(
[normalize(transforms.ToTensor()(crop)) for crop in crops]))
]),
'val':
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean, std)
]),
'test':
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean, std)
])
}
# create train/val dataloaders
transformed_datasets = {
x: datasets.ImageFolder(os.path.join(PATH_TO_IMAGES, x),
data_transforms[x])
for x in ['test']
}
dataloaders = {
x: torch.utils.data.DataLoader(transformed_datasets[x],
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=0)
for x in ['test']
}
# please do not attempt to train without GPU as will take excessively long
if not use_gpu:
raise ValueError("Error, requires GPU")
model = init_model(MODEL_NAME, N_LABELS)
model = load_checkpoint(model, CHECKPOINT_PATH)
model = model.cuda()
if (CHECKPOINT_PATH_TARGET):
model_target = init_model(MODEL_NAME_TARGET, N_LABELS)
model_target = load_checkpoint(model_target, CHECKPOINT_PATH_TARGET)
model_target = model_target.cuda()
loading_bar = ''
dataloaders_length = len(dataloaders['test'])
for i in range(dataloaders_length):
loading_bar += '-'
model_labels = []
model_pred = []
model_target_pred = []
model_pred_bin = []
model_target_pred_bin = []
for phase in ['test']:
for data in dataloaders[phase]:
loading_bar = f'={loading_bar}'
loading_bar = loading_bar[:dataloaders_length]
print(f'Testing: {loading_bar}', end='\r')
inputs, labels = data
if phase == 'test':
inputs = inputs.cuda()
labels = labels.cpu().data.numpy()
model_labels.extend(labels)
outputs = model(inputs)
outputs_pred = torch.max(outputs, dim=1)[1].cpu().data.numpy()
model_pred.extend(outputs_pred)
if (CHECKPOINT_PATH_TARGET):
outputs_target = model_target(inputs)
outputs_target_pred = torch.max(
outputs_target, dim=1)[1].cpu().data.numpy()
model_target_pred.extend(outputs_target_pred)
print('')
for i, _ in enumerate(model_labels):
model_pred_bin.append(1 if model_labels[i] == model_pred[i] else 0)
if (CHECKPOINT_PATH_TARGET):
model_target_pred_bin.append(1 if model_labels[i] ==
model_target_pred[i] else 0)
print(accuracy_score(model_labels, model_pred))
print(f1_score(model_labels, model_pred, average='micro'))
if (CHECKPOINT_PATH_TARGET):
print(accuracy_score(model_labels, model_target_pred))
print(f1_score(model_labels, model_target_pred, average='micro'))
tp = 0
fp = 0
tn = 0
fn = 0
for i, _ in enumerate(model_pred_bin):
if model_pred_bin[i] == model_target_pred_bin[i]:
if model_pred_bin[i] == 1:
tp += 1
else:
tn += 1
else:
if model_pred_bin[i] == 0:
fp += 1
else:
fn += 1
print(f"True positive = {tp}")
print(f"False positive = {fp}")
print(f"False negative = {fn}")
print(f"True negative = {tn}")
print("Finish testing")
def get_data(data_path):
with open(data_path, 'rb') as f:
train_test_paths_labels = pickle.load(f)
train_paths = train_test_paths_labels[0]
val_paths = train_test_paths_labels[1]
test_paths = train_test_paths_labels[2]
train_labels = train_test_paths_labels[3]
val_labels = train_test_paths_labels[4]
test_labels = train_test_paths_labels[5]
train_num_each = train_test_paths_labels[6]
val_num_each = train_test_paths_labels[7]
test_num_each = train_test_paths_labels[8]
print('train_paths : {:6d}'.format(len(train_paths)))
print('train_labels : {:6d}'.format(len(train_labels)))
print('valid_paths : {:6d}'.format(len(val_paths)))
print('valid_labels : {:6d}'.format(len(val_labels)))
print('test_paths : {:6d}'.format(len(test_paths)))
print('test_labels : {:6d}'.format(len(test_labels)))
train_labels = np.asarray(train_labels, dtype=np.int64)
val_labels = np.asarray(val_labels, dtype=np.int64)
test_labels = np.asarray(test_labels, dtype=np.int64)
if use_flip == 0:
train_transforms = transforms.Compose([
RandomCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.3456, 0.2281, 0.2233],
[0.2528, 0.2135, 0.2104])
])
elif use_flip == 1:
train_transforms = transforms.Compose([
RandomCrop(224),
RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.3456, 0.2281, 0.2233],
[0.2528, 0.2135, 0.2104])
])
if crop_type == 0:
test_transforms = transforms.Compose([
RandomCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.3456, 0.2281, 0.2233],
[0.2528, 0.2135, 0.2104])
])
elif crop_type == 1:
test_transforms = transforms.Compose([
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.3456, 0.2281, 0.2233],
[0.2528, 0.2135, 0.2104])
])
elif crop_type == 5:
test_transforms = transforms.Compose([
transforms.FiveCrop(224),
Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops])),
Lambda(lambda crops: torch.stack([
transforms.Normalize([0.3456, 0.2281, 0.2233],
[0.2528, 0.2135, 0.2104])(crop)
for crop in crops
]))
])
elif crop_type == 10:
test_transforms = transforms.Compose([
transforms.TenCrop(224),
Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops])),
Lambda(lambda crops: torch.stack([
transforms.Normalize([0.3456, 0.2281, 0.2233],
[0.2528, 0.2135, 0.2104])(crop)
for crop in crops
]))
])
train_dataset = CholecDataset(train_paths, train_labels, train_transforms)
val_dataset = CholecDataset(val_paths, val_labels, test_transforms)
test_dataset = CholecDataset(test_paths, test_labels, test_transforms)
return train_dataset, train_num_each, val_dataset, val_num_each, test_dataset, test_num_each
def main(args):
#
# Image preprocessing
if args.feature_type == 'plain':
if args.extractor == 'resnet152caffe-original':
# Use custom transform:
transform = transforms.Compose([
transforms.Resize((args.crop_size, args.crop_size)),
# Swap color space from RGB to BGR and subtract caffe-specific
# channel values from each pixel
transforms.Lambda(lambda img: np.array(img, dtype=np.float32)[
..., [2, 1, 0]] - [103.939, 116.779, 123.68]),
# Create a torch tensor and put channels first:
transforms.Lambda(
lambda img: torch.from_numpy(img).permute(2, 0, 1)),
# Cast tensor to correct type:
transforms.Lambda(lambda img: img.type('torch.FloatTensor'))
])
else:
# Default transform
transform = transforms.Compose([
transforms.Resize((args.crop_size, args.crop_size)),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406),
(0.229, 0.224, 0.225))
])
elif args.feature_type == 'avg' or args.feature_type == 'max':
# Try with no normalization
# Try with subtracting 0.5 from all values
# See example here: https://pytorch.org/docs/stable/torchvision/transforms.html
if args.normalize == 'default':
transform = transforms.Compose([
transforms.Resize((args.image_size, args.image_size)),
# 10-crop implementation as described in PyTorch documentation:
transforms.TenCrop((args.crop_size, args.crop_size)),
# Apply next two transforms to each crop in turn and then stack them
# to a single tensor:
transforms.Lambda(lambda crops: torch.stack([
transforms.Normalize((0.485, 0.456, 0.406),
(0.229, 0.224, 0.225))
(transforms.ToTensor()(crop)) for crop in crops
]))
])
elif args.normalize == 'skip':
transform = transforms.Compose([
transforms.Resize((args.image_size, args.image_size)),
transforms.TenCrop((args.crop_size, args.crop_size)),
transforms.Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops]))
])
elif args.normalize == 'subtract_half':
transform = transforms.Compose([
transforms.Resize((args.image_size, args.image_size)),
transforms.TenCrop((args.crop_size, args.crop_size)),
transforms.Lambda(lambda crops: torch.stack(
[transforms.ToTensor()(crop) for crop in crops]) - 0.5)
])
else:
print("Invalid normalization parameter")
sys.exit(1)
else:
print("Invalid feature type specified {}".args.feature_type)
sys.exit(1)
print("Creating features of type: {}".format(args.feature_type))
# Get dataset parameters and vocabulary wrapper:
dataset_configs = DatasetParams(args.dataset_config_file)
dataset_params = dataset_configs.get_params(args.dataset)
# We want to only get the image file name, not the full path:
for i in dataset_params:
i.config_dict['return_image_file_name'] = True
# We ask it to iterate over images instead of all (image, caption) pairs
data_loader, _ = get_loader(dataset_params,
vocab=None,
transform=transform,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
ext_feature_sets=None,
skip_images=False,
iter_over_images=True)
extractor = FeatureExtractor(args.extractor, True).to(device).eval()
# To open an lmdb handle and prepare it for the right size
# it needs to fit the total number of elements in the dataset
# so we set a map_size to a largish value here:
map_size = 1e12
lmdb_path = None
file_name = None
if args.output_file:
file_name = args.output_file
else:
file_name = '{}-{}-{}-normalize-{}.lmdb'.format(
args.dataset, args.extractor, args.feature_type, args.normalize)
os.makedirs(args.output_dir, exist_ok=True)
lmdb_path = os.path.join(args.output_dir, file_name)
# Check that we are not overwriting anything
if os.path.exists(lmdb_path):
print(
'ERROR: {} exists, please remove it first if you really want to replace it.'
.format(lmdb_path))
sys.exit(1)
print("Preparing to store extracted features to {}...".format(lmdb_path))
print("Starting to extract features from dataset {} using {}...".format(
args.dataset, args.extractor))
show_progress = sys.stderr.isatty()
# If feature shape is not 1-dimensional, store feature shape metadata:
if isinstance(extractor.output_dim, np.ndarray):
with lmdb.open(lmdb_path, map_size=map_size) as env:
with env.begin(write=True) as txn:
txn.put(str('@vdim').encode('ascii'), extractor.output_dim)
for i, (images, _, _, image_ids,
_) in enumerate(tqdm(data_loader, disable=not show_progress)):
images = images.to(device)
# If we are dealing with cropped images, image dimensions are: bs, ncrops, c, h, w
if images.dim() == 5:
bs, ncrops, c, h, w = images.size()
# fuse batch size and ncrops:
raw_features = extractor(images.view(-1, c, h, w))
if args.feature_type == 'avg':
# Average over crops:
features = raw_features.view(bs, ncrops,
-1).mean(1).data.cpu().numpy()
elif args.feature_type == 'max':
# Max over crops:
features = raw_features.view(bs, ncrops,
-1).max(1)[0].data.cpu().numpy()
# Otherwise our image dimensions are bs, c, h, w
else:
features = extractor(images).data.cpu().numpy()
# Write to LMDB object:
with lmdb.open(lmdb_path, map_size=map_size) as env:
with env.begin(write=True) as txn:
for j, image_id in enumerate(image_ids):
# If output dimension is not a scalar, flatten the array.
# When retrieving this feature from the LMDB, developer must take
# care to reshape the feature back to the correct dimensions!
if isinstance(extractor.output_dim, np.ndarray):
_feature = features[j].flatten()
# Otherwise treat it as is:
else:
_feature = features[j]
txn.put(str(image_id).encode('ascii'), _feature)
# Print log info
if not show_progress and ((i + 1) % args.log_step == 0):
print('Batch [{}/{}]'.format(i + 1, len(data_loader)))
sys.stdout.flush()
def data_loader(args):
mean_vals = [0.485, 0.456, 0.406]
std_vals = [0.229, 0.224, 0.225]
tsfm_train = transforms.Compose([
transforms.Resize((args.resize_size, args.resize_size)),
transforms.RandomCrop(args.crop_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean_vals, std_vals)
])
if args.tencrop == 'True':
func_transforms = [
transforms.Resize(args.resize_size),
transforms.TenCrop(args.crop_size),
transforms.Lambda(lambda crops: torch.stack([
transforms.Normalize(mean_vals, std_vals)
(transforms.ToTensor()(crop)) for crop in crops
])),
]
else:
func_transforms = []
# print input_size, crop_size
if args.resize_size == 0 or args.crop_size == 0:
pass
else:
func_transforms.append(
transforms.Resize((args.resize_size, args.resize_size)))
func_transforms.append(transforms.CenterCrop(args.crop_size))
func_transforms.append(transforms.ToTensor())
func_transforms.append(transforms.Normalize(mean_vals, std_vals))
tsfm_test = transforms.Compose(func_transforms)
if args.dataset == 'ILSVRC':
with_image = False
else:
with_image = True
img_train = CUBClsDataset(root=args.data,
datalist=args.train_list,
transform=tsfm_train,
with_image=with_image)
img_test = CUBCamDataset(root=args.data,
datalist=args.test_list,
transform=tsfm_test,
with_image=with_image)
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
img_train)
else:
train_sampler = None
train_loader = DataLoader(img_train,
batch_size=args.batch_size,
shuffle=(train_sampler is None),
sampler=train_sampler,
num_workers=args.workers)
val_loader = DataLoader(img_test,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers)
return train_loader, val_loader, train_sampler