def get_prediction(image_bytes):
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
num_classes = 2
images = get_tensor(image_bytes)
images = images.to(device)
model = DenseNet121(num_classes).to(device)
model.load_state_dict(torch.load("/content/pneumonia_model.pth"))
prediction = model(something.to(device))
activated_features = SaveFeatures(
model._modules.get('densenet121').features.denseblock4.denselayer16)
prediction = model(images)
pred_probabilities = F.softmax(prediction).data.squeeze()
activated_features.remove()
topk(pred_probabilities, 1)
weight_softmax_params = list(
model._modules.get('densenet121').classifier[0].parameters())
weight_softmax = np.squeeze(weight_softmax_params[0].cpu().data.numpy())
#weight_softmax*=0.01
class_idx = topk(pred_probabilities, 1)[1].int()
overlay = getCAM(activated_features.features, weight_softmax, class_idx)
img = images[0].cpu().numpy()[0]
#imshow(img,cmap='gray')
#imshow(skimage.transform.resize(overlay[0], images.shape[2:4]), alpha=0.4, cmap='jet')
return class_idx, skimage.transform.resize(overlay[0], images.shape[2:4])
def get_config(lr=None, bs=None, opt=None, mm=None, wd=None):
if lr == None:
lr = np.random.normal(0.15, 1e-3)
while (lr < 0):
lr = np.random.normal(0.15, 1e-3)
if bs == None:
bs = random.choice(BATCHES)
if opt == None:
opt = random.choice(OPTIMS)
if mm == None:
mm = np.random.normal(0.9, 1e-2)
if wd == None:
wd = np.random.normal(5e-6, 1e-7)
model = DenseNet121()
optimiser = None
# Make optimiser
if opt == 'Adam':
optimiser = torch.optim.Adam(model.parameters(),
lr=lr,
weight_decay=wd)
elif opt == 'SGD':
optimiser = torch.optim.SGD(model.parameters(),
lr=lr,
weight_decay=wd,
momentum=mm)
elif opt == 'RMSprop':
optimiser = torch.optim.RMSprop(model.parameters(),
lr=lr,
weight_decay=wd,
momentum=mm)
return HP_Config(bs, lr, optimiser, opt, mm, wd, model)
def get_result(threshold=[0.2]):
submit = pd.read_csv(config['sample_submission'])
if not os.path.exists("./predict_result2/"):
os.mkdir("./predict_result2")
state_dict = torch.load(config['load_model_path'])
epoch = ntpath.basename(config['load_model_path'])
net = DenseNet121()
print(config['load_model_path'])
net.load_state_dict(state_dict['state_dict']) #load the model
net.cuda()
net.eval()
for thresh in threshold:
predicted = []
draw_predict = []
for name in tqdm(submit['Id']):
path = os.path.join(config['path_to_test'], name)
image1 = load_image(path,
(config['SIZE'], config['SIZE'], 3)) / 255.
image2 = np.fliplr(image1)
image3 = np.flipud(image1)
image4 = np.rot90(image1)
image5 = np.rot90(image4)
image6 = np.rot90(image5)
images = np.stack([image1, image2, image3, image4, image5,
image6]).transpose(0, 3, 1, 2).copy()
images = torch.from_numpy(images)
score_predict = net(images.float().cuda())
score_predict = F.sigmoid(score_predict) #sigmoid激活
score_predict = torch.mean(score_predict,
dim=0).detach().cpu().numpy() #转成numpy
draw_predict.append(score_predict)
label_predict = np.arange(28)[score_predict >= thresh]
if int(label_predict.shape[0]) == 0:
print("the label is None")
print(score_predict)
label_predict = np.arange(28)[np.argsort(score_predict)[-2:]]
print(label_predict)
str_predict_label = ' '.join(str(l) for l in label_predict)
predicted.append(str_predict_label)
submit['Predicted'] = predicted
np.save(
'./predict_result2/draw_predict_DenseNet121_512_' + str(thresh) +
'_' + str(epoch) + '.npy', score_predict)
submit.to_csv("./predict_result2/" + str(epoch) + ".csv", index=False)
def main():
torch.manual_seed(0)
train_dataset_info=read_train_dataset_info()
indexes=np.arange(train_dataset_info.shape[0])
np.random.shuffle(indexes)
train_indexes, valid_indexes = train_test_split(indexes, test_size=0.1, random_state=8)
train_dataset=DataLoad_protein(train_dataset_info[train_indexes],config['batch_size'],(config['SIZE'],config['SIZE'],4),
augument=True)
train_loader=DataLoader(
dataset=train_dataset,
batch_size=config['batch_size'],
pin_memory=True,
num_workers=config['num_workers'],
)
net=DenseNet121(pretrained=True,start_first="start_training")
start_epoch=config['start_epoch']
if config['resume'] is not None:
print("load the model",config['resume'])
model=torch.load(config['resume'])
start_epoch=model['epoch']
net.load_state_dict(model['state_dict'])
net.cuda()
first_conv=net.state_dict()['base_model.features.denseblock3.denselayer22.conv2.weight'].clone()
second_classifier=net.state_dict()['S.conv1.weight'].clone()
second_separ=net.state_dict()['classifier.weight'].clone()
loss=BCEWithLogitsLoss()
opt=torch.optim.Adam(
filter(lambda p:p.requires_grad,net.parameters()),
lr=1e-3,
)
for epoch in range(start_epoch+1,start_epoch+6):
print("epoch is ",epoch)
train(train_loader,net,loss,epoch,opt,config['save_freq'],config['save_dir'])
first_conv_after=net.state_dict()['base_model.features.denseblock3.denselayer22.conv2.weight'].clone()
second_classifier_after=net.state_dict()['S.conv1.weight'].clone()
second_separ_after=net.state_dict()['classifier.weight'].clone()
print(net.state_dict()['base_model.features.denseblock3.denselayer22.conv2.weight'])
print("the training parameters",net.state_dict()['S.conv1.weight'],net.state_dict()['classifier.weight'])
print(torch.equal(first_conv_after,first_conv))
print(torch.equal(second_classifier,second_classifier_after))
print(torch.equal(second_separ,second_separ_after))
def build_network(train_dataset, val_dataset):
train_dataset = DataLoad_protein(train_dataset,
config['batch_size'],
(config['SIZE'], config['SIZE'], 4),
augument=True)
val_dataset = DataLoad_protein(val_dataset,
config['batch_size'],
(config['SIZE'], config['SIZE'], 4),
augument=False)
train_loader = DataLoader(
dataset=train_dataset,
shuffle=True,
batch_size=config['batch_size'],
pin_memory=True,
num_workers=config['num_workers'],
)
validation_loader = DataLoader(dataset=val_dataset,
shuffle=True,
batch_size=3,
pin_memory=True,
num_workers=config['num_workers'])
net = DenseNet121(pretrained=True)
start_epoch = config['start_epoch']
print("load the fine model", config['finetune_model'])
fine_model = torch.load(config['finetune_model'])
net.load_state_dict(fine_model['state_dict'])
if config['resume'] is not None:
model = torch.load(config['resume'])
start_epoch = model['epoch']
net.load_state_dict(model['state_dict'])
opt = torch.optim.Adam(net.parameters(), lr=config['lr']) #without
loss = BCEWithLogitsLoss()
return train_loader, validation_loader, net, loss, opt, start_epoch
def main():
"""
Script entrypoint
"""
t_start = datetime.now()
header = ["Start Time", "End Time", "Duration (s)"]
row = [t_start.strftime(DEFAULT_DATE_TIME_FORMAT)]
dnn = DenseNet121()
# show class indices
print('****************')
for cls, idx in dnn.train_batches.class_indices.items():
print('Class #{} = {}'.format(idx, cls))
print('****************')
print(dnn.model.summary())
dnn.train(t_start,
epochs=dnn.num_epochs,
batch_size=dnn.batch_size,
training=dnn.train_batches,
validation=dnn.valid_batches)
# save trained weights
dnn.model.save(dnn.file_weights + 'old')
dnn.model.save_weights(dnn.file_weights)
with open(dnn.file_architecture, 'w') as f:
f.write(dnn.model.to_json())
t_end = datetime.now()
difference_in_seconds = get_difference_in_seconds(t_start, t_end)
row.append(t_end.strftime(DEFAULT_DATE_TIME_FORMAT))
row.append(str(difference_in_seconds))
append_row_to_csv(complete_run_timing_file, header)
append_row_to_csv(complete_run_timing_file, row)
def main(args):
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# initialize and load the model
model = DenseNet121(N_CLASSES).to(device)
if torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model).to(device)
if os.path.isfile(args.model_path):
model.load_state_dict(torch.load(args.model_path, map_location=device))
print('model state has loaded')
else:
print('=> model state file not found')
model.train(False)
dummy_input = torch.randn(args.batch_size, 3, 224, 224)
torch_out = model(dummy_input)
torch.onnx.export(model,
dummy_input,
'model/densenet121.onnx',
verbose=False)
print('ONNX model exported.')
def main(args):
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print('Using %s device.' % device)
# initialize and load the model
net = DenseNet121(N_CLASSES).to(device)
if torch.cuda.device_count() > 1:
net = torch.nn.DataParallel(net).to(device)
net.load_state_dict(torch.load(args.model_path, map_location=device))
print('model state has loaded')
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)
# initialize the ground truth and output tensor
gt = torch.FloatTensor().to(device)
pred = torch.FloatTensor().to(device)
# switch to evaluate mode
net.eval()
for index, (data, target) in enumerate(test_loader):
start_time = timeit.default_timer()
target = target.to(device)
bs, n_crops, c, h, w = data.size()
data = data.view(-1, c, h, w).to(device)
with torch.no_grad():
output = net(data)
output_mean = output.view(bs, n_crops, -1).mean(1)
gt = torch.cat((gt, target))
pred = torch.cat((pred, output_mean))
print('\rbatch %03d/%03d %6.3fsec' %
(index, len(test_loader), (timeit.default_timer() - start_time)),
end='')
AUCs = []
for i in range(N_CLASSES):
AUCs.append(roc_auc_score(gt.cpu()[:, i], pred.cpu()[:, i]))
print('The average AUC is %6.3f' % np.mean(AUCs))
for i in range(N_CLASSES):
print('The AUC of %s is %6.3f' % (CLASS_NAMES[i], AUCs[i]))
def main():
cfg = Config()
# Redirect logs to both console and file.
if cfg.log_to_file:
ReDirectSTD(cfg.stdout_file, 'stdout', False)
ReDirectSTD(cfg.stderr_file, 'stderr', False)
# Lazily create SummaryWriter
writer = None
TVT, TMO = set_devices(cfg.sys_device_ids)
if cfg.seed is not None:
set_seed(cfg.seed)
# Dump the configurations to log.
import pprint
print('-' * 60)
print('cfg.__dict__')
pprint.pprint(cfg.__dict__)
print('-' * 60)
###########
# Dataset #
###########
if not cfg.only_test:
train_set = create_dataset(**cfg.train_set_kwargs)
# The combined dataset does not provide val set currently.
val_set = None if (cfg.dataset == 'combined' or
cfg.model_type != 'resnet50') else create_dataset(
**cfg.val_set_kwargs)
###########
# Models #
###########
if cfg.add_softmax_loss:
model = DenseNet121_classifier(751)
else:
if cfg.model_type == 'resnet50':
model = Model(last_conv_stride=cfg.last_conv_stride)
elif cfg.model_type == 'densenet121':
model = DenseNet121()
elif cfg.model_type == 'preActResnet50':
model = PreActResNet50()
elif cfg.model_type == 'resnet50mid':
model = resnet50mid()
#Output the embedding size
#input = Variable(torch.FloatTensor(32, 3, 256, 128))
#out = model(input)
#print('Model is ', str(cfg.model_type), 'embedding size is ', out.shape)
# Model wrapper
model_w = DataParallel(model)
#############################
# Criteria and Optimizers #
#############################
tri_loss = TripletLoss(margin=cfg.margin)
optimizer = optim.Adam(model.parameters(),
lr=cfg.base_lr,
weight_decay=cfg.weight_decay)
# Bind them together just to save some codes in the following usage.
modules_optims = [model, optimizer]
# May Transfer Models and Optims to Specified Device. Transferring optimizer
# is to cope with the case when you load the checkpoint to a new device.
TMO(modules_optims)
#Softmax loss
criterian_softmax = CrossEntropyLabelSmooth(751)
########
# Test #
########
def validate():
if val_set.extract_feat_func is None:
val_set.set_feat_func(ExtractFeature(model_w, TVT))
print('\n=========> Test on validation set <=========\n')
mAP, cmc_scores, _, _ = val_set.eval(
normalize_feat=cfg.normalize_feature,
to_re_rank=False,
verbose=False)
print()
return mAP, cmc_scores[0]
############
# Training #
############
start_ep = 0
for ep in range(start_ep, cfg.total_epochs):
# Adjust Learning Rate
if cfg.lr_decay_type == 'exp':
adjust_lr_exp(optimizer, cfg.base_lr, ep + 1, cfg.total_epochs,
cfg.exp_decay_at_epoch)
else:
adjust_lr_staircase(optimizer, cfg.base_lr, ep + 1,
cfg.staircase_decay_at_epochs,
cfg.staircase_decay_multiply_factor)
may_set_mode(modules_optims, 'train')
# For recording precision, satisfying margin, etc
prec_meter = AverageMeter()
sm_meter = AverageMeter()
dist_ap_meter = AverageMeter()
dist_an_meter = AverageMeter()
loss_meter = AverageMeter()
ep_st = time.time()
step = 0
epoch_done = False
while not epoch_done:
step += 1
step_st = time.time()
ims, im_names, labels, mirrored, epoch_done = train_set.next_batch(
)
ims_var = Variable(TVT(torch.from_numpy(ims).float()))
labels_t = TVT(torch.from_numpy(labels).long())
if cfg.add_softmax_loss:
feat, v = model_w(ims_var)
else:
feat = model_w(ims_var)
loss, p_inds, n_inds, dist_ap, dist_an, dist_mat = global_loss(
tri_loss,
feat,
labels_t,
normalize_feature=cfg.normalize_feature)
if cfg.add_softmax_loss:
softmax_loss = criterian_softmax(v, labels_t)
loss = (1 - cfg.softmax_loss_weight
) * loss + cfg.softmax_loss_weight * softmax_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
############
# Step Log #
############
# precision
prec = (dist_an > dist_ap).data.float().mean()
# the proportion of triplets that satisfy margin
sm = (dist_an > dist_ap + cfg.margin).data.float().mean()
# average (anchor, positive) distance
d_ap = dist_ap.data.mean()
# average (anchor, negative) distance
d_an = dist_an.data.mean()
prec_meter.update(prec)
sm_meter.update(sm)
dist_ap_meter.update(d_ap)
dist_an_meter.update(d_an)
loss_meter.update(to_scalar(loss))
if step % cfg.steps_per_log == 0:
time_log = '\tStep {}/Ep {}, {:.2f}s'.format(
step,
ep + 1,
time.time() - step_st,
)
tri_log = (', prec {:.2%}, sm {:.2%}, '
'd_ap {:.4f}, d_an {:.4f}, '
'loss {:.4f}'.format(
prec_meter.val,
sm_meter.val,
dist_ap_meter.val,
dist_an_meter.val,
loss_meter.val,
))
log = time_log + tri_log
print(log)
#############
# Epoch Log #
#############
time_log = 'Ep {}, {:.2f}s'.format(ep + 1, time.time() - ep_st)
tri_log = (', prec {:.2%}, sm {:.2%}, '
'd_ap {:.4f}, d_an {:.4f}, '
'loss {:.4f}'.format(
prec_meter.avg,
sm_meter.avg,
dist_ap_meter.avg,
dist_an_meter.avg,
loss_meter.avg,
))
log = time_log + tri_log
print(log)
##########################
# Test on Validation Set #
##########################
mAP, Rank1 = 0, 0
if ((ep + 1) % cfg.epochs_per_val == 0) and (val_set is not None):
mAP, Rank1 = validate()
# save ckpt
if cfg.log_to_file:
save_weights(modules_optims[0], cfg.ckpt_file)
from model import DenseNet121
from extras import SaveFeatures, getCAM
#device supporting the training
device=torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
# the two classes here are pneumonia and non pneumonia
# Hyper parameters.
num_classes=2
num_epochs=1000
batch_size=64
learning_rate=0.001
data_dir="/content/drive/My Drive/Projects/Pneumonia detection/chest_xray/chest_xray/"
train_loader, val_loader,test_loader=loadTrainData(data_dir)
model=DenseNet121(num_classes).to(device) # pretrained model can be loaded too. Just add pretrained =True
criterion=nn.CrossEntropyLoss()
optimizer=torch.optim.Adam(model.parameters(),lr=learning_rate)
"""##Model"""
#Training process
total_step=len(train_loader)
for epoch in range(num_epochs):
for i,(images,labels) in enumerate(train_loader):
# batch pair of image and label using data loader
images=images.to(device)
labels=labels.to(device)
im = im[:, :, None]
padding = 0
if im.shape[0] > im.shape[1]:
padding = (int((im.shape[0] - im.shape[1]) / 2), 0)
else:
padding = (0, int((im.shape[1] - im.shape[0]) / 2))
data_transforms = transforms.Compose([
transforms.ToPILImage(),
transforms.Pad(padding, fill=0),
])
pim = data_transforms(im)
return (pim)
densenet = DenseNet121(14)
# densenet = addDropout(densenet, p=0)
saved_model_path = "../Models/model.pth.tar"
densenet.load_state_dict(load_dictionary(saved_model_path, map_location='cpu'))
if torch.cuda.is_available():
densenet = densenet.cuda()
name = "Exp_64_512_0.00001_RandomLabel_4.0"
#Path to the experiment (it would be the github)
ExpDir = "/media/vince/MILA/ChestXrays/ALI/model/" + name
ExpDir = "/network/home/frappivi/ChestXrays/ALI/model/" + name
isize = 64
LS = 512 #Latent Space Size
ColorsNumber = 1 #Number of color (always 1 for x-ray)
isize = 64
def main_worker(local_rank, ngpus, args):
best_prec1 = .0
dist.init_process_group(backend=args.dist_backend,
init_method=args.dist_url)
print(f'local_rank: {local_rank}\n')
torch.cuda.set_device(local_rank)
# IMPORTANT: we need to set the random seed in each process so that the models are initialized with the same weights
# Reference: https://yangkky.github.io/2019/07/08/distributed-pytorch-tutorial.html
# torch.cuda.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# build model
densenet = DenseNet121(in_channels=3, growth_rate=args.growth_rate, \
compression_rate=args.compression_rate, \
num_classes=args.num_classes)
densenet.cuda(local_rank)
densenet = nn.parallel.DistributedDataParallel(densenet, \
device_ids=[local_rank], \
output_device=local_rank)
# Reference: https://github.com/pytorch/examples/blob/master/imagenet/main.py
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_transform = transforms.Compose([
transforms.RandomResizedCrop(args.image_width),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(), normalize
])
trainset = torchvision.datasets.ImageFolder(root=os.path.join(
args.dataset_root, 'train'),
transform=train_transform)
val_transform = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(args.image_width),
transforms.ToTensor(), normalize
])
valset = torchvision.datasets.ImageFolder(root=os.path.join(
args.dataset_root, 'val'),
transform=val_transform)
# num_replicas: int, Number of processes participating in distributed training. By default, world_size is retrieved from the current distributed group.
train_sampler = torch.utils.data.distributed.DistributedSampler(
trainset, num_replicas=ngpus)
batch_size = args.batch_size // ngpus
num_workers = args.num_workers // ngpus
train_data = DataLoader(
trainset,
batch_size=batch_size,
shuffle=False, # when sampler is specified, shuffle should be False
num_workers=num_workers,
pin_memory=True,
sampler=train_sampler)
val_data = DataLoader(valset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers,
pin_memory=True)
criterion = nn.CrossEntropyLoss().cuda(local_rank)
optimizer = optim.SGD(densenet.parameters(), lr=args.lr, momentum=args.momentum, \
weight_decay=args.weight_decay)
# Reference: https://discuss.pytorch.org/t/what-does-torch-backends-cudnn-benchmark-do/5936/2
# this is useful for cudnn finding optimal set of algorithms for particular configurations
# and accelerate training when the input sizes do not change over iteration.
cudnn.backend = True
for epoch in range(args.epochs):
train_sampler.set_epoch(epoch)
adjust_lr(args, optimizer, epoch)
losses, top1, top5 = train(densenet, train_data, criterion, optimizer,
epoch, local_rank, args)
if args.tensorboard:
log_value('train_loss', losses.avg, epoch)
log_value('top1_acc', top1.avg, epoch)
log_value('top5_acc', top5.avg, epoch)
# validate the model every epoch
prec1 = validate(args, val_data, densenet, criterion, epoch)
is_best = prec1.avg > best_prec1
best_prec1 = max(prec1.avg, best_prec1)
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': densenet.state_dict(),
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict()
}, is_best)
def main():
N_CLASSES = 14
CLASS_NAMES = ['Atelectasis',
'Cardiomegaly',
'Effusion',
'Infiltration',
'Mass',
'Nodule',
'Pneumonia',
'Pneumothorax',
'Consolidation',
'Edema',
'Emphysema',
'Fibrosis',
'Pleural_Thickening',
'Hernia']
# initialize model
device = utils.get_device()
model = DenseNet121(N_CLASSES).to(device)
checkpoint = torch.load(args.checkpoint)
model.load_state_dict(checkpoint['model_state_dict'])
# initialize test loader
test_dataset = ChestXrayDataSet(data_dir=args.path_to_images,
image_list_file=args.test_list,
transform=transforms_test)
test_loader = DataLoader(dataset=test_dataset, batch_size=args.batch_size,
shuffle=False, num_workers=args.num_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()
with torch.no_grad():
for i, (inp, target) in enumerate(test_loader):
target = target.cuda()
gt = torch.cat((gt, target), 0)
bs, c, h, w = inp.size()
input_var = torch.autograd.Variable(inp.view(-1, c, h, w).cuda())
output = model(input_var)
output_mean = output.view(bs, -1)
pred = torch.cat((pred, output_mean.data), 0)
gt_np = gt.cpu().numpy()
pred_np = sigmoid(pred.cpu().numpy())
Y_t = [] #labels for each anomaly
for i in range(N_CLASSES):
Y_t.append([])
for x in gt_np:
Y_t[i].append(x[i])
Y_pred = [] #preds for each anomaly
for j in range(N_CLASSES):
Y_pred.append([])
for y in pred_np:
Y_pred[j].append(y[j])
AUCs = [] # AUCs for each
for i in range(N_CLASSES):
auc = roc_auc_score(Y_t[i], Y_pred[i])
AUCs.append(auc)
matrices=[] #for each
for i in range(14):
matrix = confusion_matrix(Y_t[i], np.asarray(Y_pred[i])>0.6)
matrices.append(matrix)
class_names = ['no disease', 'disease']
fig = plt.figure(figsize = (20,20))
for i in range(14):
plt.subplot(4,4,i+1)
df_cm = pd.DataFrame(
matrices[i], index=class_names, columns=class_names)
heatmap = sns.heatmap(df_cm, annot=True, fmt="d").set_title(CLASS_NAMES[i])
plt.ylabel('True label')
plt.xlabel('Predicted label')
plt.show()
fig.savefig(os.path.join(args.test_outdir,'confusion_matrix.pdf'))
fig, axes2d = plt.subplots(nrows=2, ncols=7,
sharex=True, sharey=True,figsize = (12, 4))
for i, row in enumerate(axes2d):
for j, cell in enumerate(row):
if i==0:
x=i+j
else:
x=13-i*j
fpr, tpr, threshold = roc_curve(Y_t[x], Y_pred[x])
roc_auc = auc(fpr, tpr)
cell.plot(fpr, tpr, 'b', label = 'AUC = %0.2f' % roc_auc)
cell.legend(loc = 'lower right', handlelength=0,handletextpad=0,frameon=False, prop={'size': 8})
cell.plot([0, 1], [0, 1],'r--')
plt.xlim([0, 1])
plt.ylim([0, 1])
cell.set_title(CLASS_NAMES[x],fontsize=10)
if i == len(axes2d) - 1:
cell.set_xlabel('False positive rate')
if j == 0:
cell.set_ylabel('True negative rate')
fig.tight_layout(pad=1.0)
plt.show()
fig.savefig(os.path.join(args.test_outdir,'roc_auc.pdf'))
def main():
args = get_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpus
SEED = 42
utils.set_global_seed(SEED)
utils.prepare_cudnn(deterministic=True)
num_classes = 14
#define datasets
train_dataset = ChestXrayDataSet(
data_dir=args.path_to_images,
image_list_file=args.train_list,
transform=transforms_train,
)
val_dataset = ChestXrayDataSet(
data_dir=args.path_to_images,
image_list_file=args.val_list,
transform=transforms_val,
)
loaders = {
'train':
DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers),
'valid':
DataLoader(val_dataset,
batch_size=2,
shuffle=False,
num_workers=args.num_workers)
}
logdir = args.log_dir #where model weights and logs are stored
#define model
model = DenseNet121(num_classes)
if len(args.gpus) > 1:
model = nn.DataParallel(model)
device = utils.get_device()
runner = SupervisedRunner(device=device)
optimizer = RAdam(model.parameters(), lr=args.lr, weight_decay=0.0003)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
factor=0.25,
patience=2)
weights = torch.Tensor(
[10, 100, 30, 8, 40, 40, 330, 140, 35, 155, 110, 250, 155,
200]).to(device)
criterion = BCEWithLogitsLoss(pos_weight=weights)
class_names = [
'Atelectasis', 'Cardiomegaly', 'Effusion', 'Infiltration', 'Mass',
'Nodule', 'Pneumonia', 'Pneumothorax', 'Consolidation', 'Edema',
'Emphysema', 'Fibrosis', 'Pleural_Thickening', 'Hernia'
]
runner.train(
model=model,
logdir=logdir,
criterion=criterion,
optimizer=optimizer,
scheduler=scheduler,
loaders=loaders,
num_epochs=args.epochs,
# We can specify the callbacks list for the experiment;
# For this task, we will check AUC and accuracy
callbacks=[
AUCCallback(
input_key="targets",
output_key='logits',
prefix='auc',
class_names=class_names,
num_classes=num_classes,
activation='Sigmoid',
),
AccuracyCallback(
input_key="targets",
output_key="logits",
prefix="accuracy",
accuracy_args=[1],
num_classes=14,
threshold=0.5,
activation='Sigmoid',
),
],
main_metric='auc/_mean',
minimize_metric=False,
verbose=True,
)
from sklearn.metrics import roc_auc_score
small_set = False
num_classes = 14
class_names = labels = [
'Atelectasis', 'Consolidation', 'Infiltration', 'Pneumothorax', 'Edema',
'Emphysema', 'Fibrosis', 'Effusion', 'Pneumonia', 'Pleural_Thickening',
'Cardiomegaly', 'Mass', 'Nodule', 'Hernia'
]
data_dir = 'data_entry_labels.csv'
batch_size = 32
max_epoch = 50
print('Loading Model')
model = DenseNet121(num_classes).cuda()
normalize = transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
ts = transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
])
if small_set == True:
train_list = 'train_list_small.txt'
val_list = 'val_list_small.txt'
else:
train_list = 'train_list.txt'
query_cam, query_label = get_id(query_path)
######################################################################
# Load Collected data Trained model
print('-------test-----------')
###########
# Models #
###########
if opt.add_softmax_loss:
model_structure = DenseNet121_classifier(751)
else:
if opt.model_type == 'resnet50':
model_structure = Model(last_conv_stride=opt.last_conv_stride)
elif opt.model_type == 'densenet121':
model_structure = DenseNet121()
elif opt.model_type == 'preActResnet50':
model_structure = PreActResNet50()
elif opt.model_type == 'resnet50mid':
model_structure = resnet50mid()
print('Model selected: ', opt.model_type)
model = load_network(model_structure)
# Remove the final fc layer and classifier layer
if not opt.add_softmax_loss:
model.model.fc = nn.Sequential()
model.classifier = nn.Sequential()
# Change to test mode
def main_worker(gpu, args):
"""
@param: gpu - index of the gpu on a single node, here its range is [0, args.gpus-1]
"""
# IMPORTANT: we need to set the random seed in each process so that the models are initialized with the same weights
# Reference: https://yangkky.github.io/2019/07/08/distributed-pytorch-tutorial.html
# torch.cuda.manual_seed(args.seed)
# for distributed training, rank needs to be global rank among all processes
rank = args.node_rank * args.gpus + gpu
dist.init_process_group(backend=args.dist_backend, \
init_method=args.dist_url, \
world_size=args.world_size, \
rank=rank)
# build model
densenet = DenseNet121(in_channels=3, growth_rate=args.growth_rate, \
compression_rate=args.compression_rate, \
num_classes=args.num_classes)
# torch.cuda.device(gpu)
# densenet.cuda(gpu)
densenet.cuda()
# densenet = nn.parallel.DistributedDataParallel(densenet, device_ids=[gpu])
densenet = nn.parallel.DistributedDataParallel(densenet)
# Reference: https://github.com/pytorch/examples/blob/master/imagenet/main.py
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_transform = transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(), normalize
])
trainset = torchvision.datasets.ImageFolder(root=os.path.join(
args.dataset_root, 'train'),
transform=train_transform)
val_transform = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(), normalize
])
valset = torchvision.datasets.ImageFolder(root=os.path.join(
args.dataset_root, 'val'),
transform=val_transform)
train_sampler = torch.utils.data.distributed.DistributedSampler(
trainset, num_replicas=args.world_size, rank=rank)
args.batch_size = int(args.batch_size / args.gpus)
args.num_workers = int(args.num_workers / args.gpus)
train_data = torch.utils.data.DataLoader(
trainset,
batch_size=args.batch_size,
shuffle=False, # when sampler is specified, shuffle should be False
num_workers=args.num_workers,
pin_memory=True,
sampler=train_sampler)
val_data = torch.utils.data.DataLoader(valset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
pin_memory=True)
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(densenet.parameters(), lr=args.lr, momentum=args.momentum, \
weight_decay=args.weight_decay)
global best_prec1
# Reference: https://discuss.pytorch.org/t/what-does-torch-backends-cudnn-benchmark-do/5936/2
# this is useful for cudnn finding optimal set of algorithms for particular configurations
# and accelerate training when the input sizes do not change over iteration.
cudnn.backend = True
for epoch in range(args.epochs):
train_sampler.set_epoch(epoch)
adjust_lr(args, optimizer, epoch)
train(densenet, train_data, criterion, optimizer, epoch, args)
if args.tensorboard:
log_value('train_loss', losses.avg, epoch)
log_value('train_acc', top1.avg, epoch)
# validate the model every epoch
prec1 = validate(args, val_data, densenet, criterion, epoch)
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': densenet.state_dict(),
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict()
}, is_best)