def main(_):
parser = argparse.ArgumentParser()
parser.add_argument('--action', dest='action', type=str, default='train',
help='actions: train, test, predict or store')
args = parser.parse_args()
if args.action not in ['train', 'test', 'predict', 'store']:
print('invalid action: ', args.action)
print("Please input a action: train, test, predict or store")
else:
if args.action == 'test':
model = Unet(tf.InteractiveSession(), configure())
else:
model = Unet(tf.Session(), configure())
getattr(model, args.action)()
def __init__(self,
nets_RGBM,
nets_RGB=None,
nets_VAE=None,
total_samples=50000,
seed=0):
self.nets_RGBM = nets_RGBM
self.nets_RGB = nets_RGB
self.nets_VAE = nets_VAE
self.zs = nets_RGBM.sample_zs(total_samples, seed)
self.total_samples = total_samples
self.outsize = nets_RGBM.setting['outdim']
sys.path.append('resources/PiCANet-Implementation')
from network import Unet
from dataset import CustomDataset
ckpt = 'resources/PiCANet-Implementation/36epo_383000step.ckpt'
state_dict = torch.load(ckpt)
model = Unet().cuda()
model.load_state_dict(state_dict)
self.model = model
def __init__(self, hparams):
super(Unet3D, self).__init__()
self.hparams = hparams
self.learning_rate = hparams.learning_rate
self.data_set_dir = hparams.data_set_dir
self.loader_kwargs = {
'batch_size': hparams.batch_size,
'num_workers': hparams.num_workers,
'pin_memory': True
}
self.valid_split = hparams.valid_split
num_pool = hparams.num_pool
num_features = hparams.num_features
patch_size = (hparams.patch_x, hparams.patch_y, hparams.patch_z)
def encode_kwargs_fn(level):
num_stacks = max(level, 1)
return {'num_stacks': num_stacks}
paired_features = generate_paired_features(num_pool, num_features)
self.net = Unet(in_channels=1,
out_channels=1,
paired_features=paired_features,
pool_block=ResBlock,
pool_kwargs={'stride': 2},
up_kwargs={'attention': True},
encode_block=ResBlockStack,
encode_kwargs_fn=encode_kwargs_fn,
decode_block=ResBlock)
self.loss_fn = FocalDiceCoefLoss()
self.metrics = {'kd_dsc': DiceCoef()}
self.tr_transform = Compose([
RandomRescaleCrop(0.1,
patch_size,
crop_mode='random',
enforce_label_indices=[1]),
RandomMirror((0.2, 0, 0)),
RandomContrast(0.1),
RandomBrightness(0.1),
RandomGamma(0.1),
CombineLabels([1, 2], 3),
ToTensor()
])
self.vd_transform = Compose(
[RandomCrop(patch_size),
CombineLabels([1, 2], 3),
ToTensor()])
help="save result images as .jpg file. If None -> Not save",
default=None)
args = parser.parse_args()
if args.logdir is None and args.save_dir is None:
print(
"You should specify either --logdir or --save_dir to save results!"
)
assert 0
print(args)
print(os.getcwd())
device = torch.device(args.cuda)
state_dict = torch.load(args.model_dir, map_location=args.cuda)
model = Unet().to(device)
model.load_state_dict(state_dict)
custom_dataset = CustomDataset(root_dir=args.dataset)
dataloader = DataLoader(custom_dataset, args.batch_size, shuffle=False)
os.makedirs(os.path.join(args.save_dir, 'img'), exist_ok=True)
os.makedirs(os.path.join(args.save_dir, 'mask'), exist_ok=True)
if args.logdir is not None:
writer = SummaryWriter(args.logdir)
model.eval()
for i, batch in enumerate(tqdm(dataloader)):
img = batch.to(device)
with torch.no_grad():
pred, loss = model(img)
pred = pred[5].data
pred.requires_grad_(False)
if args.logdir is not None:
default=0.1, type=float)
parser.add_argument('--decay_step', help='Learning rate decrease by lr_decay time per decay_step, default = 7000',
default=7000, type=int)
parser.add_argument('--display_freq', help='display_freq to display result image on Tensorboard',
default=1000, type=int)
args = parser.parse_args()
# TODO : Add multiGPU Model
device = torch.device(args.cuda)
batch_size = args.batch_size
epoch = args.epoch
duts_dataset = PairDataset(args.dataset)
load = args.load
start_iter = 0
model = Unet(cfg).cuda()
#vgg = torchvision.models.vgg16(pretrained=True)
#model.encoder.seq.load_state_dict(vgg.features.state_dict())
now = datetime.datetime.now()
start_epo = 0
#del vgg
if load is not None:
state_dict = torch.load(load, map_location=args.cuda)
#start_iter = int(load.split('epo_')[1].strip('step.ckpt')) + 1
#start_epo = int(load.split('/')[3].split('epo')[0])
#now = datetime.datetime.strptime(load.split('/')[2], '%m%d%H%M')
#print("Loading Model from {}".format(load))
#print("Start_iter : {}".format(start_iter))
'lr_decay': 0.1, # Learning rate decrease by lr_decay time per decay_step, default = 0.1
'decay_step': 7000, # Learning rate decrease by lr_decay time per decay_step, default = 7000
'batch_size': 20, # batchsize, default = 1
'epoch': 20, # epochs, default = 20
'dataset': '', # Directory of your Dataset
'load': None, # Directory of pre-trained model
'display_freq': 1000, # display_freq to display result image on Tensorboard
}
torch.cuda.manual_seed_all(1234)
torch.manual_seed(1234)
start_iter = 0
now = datetime.datetime.now()
start_epo = 0
duts_dataset = CustomDataset(config['dataset'])
model = Unet(cfg,config['mode']).cuda()
# vgg = torchvision.models.vgg16(pretrained=True)
# model.encoder.seq.load_state_dict(vgg.features.state_dict())
# del vgg
if config['load'] is not None:
state_dict = torch.load(config['load'], map_location='cuda')
start_iter = 1#int(config['load'].split('epo_')[1].strip('step.ckpt')) + 1
start_epo = 0#int(config['load'].split('/')[3].split('epo')[0])
now = time.strftime('%Y%m%d-%H%M%S', time.localtime())
print("Loading Model from {}".format(config['load']))
print("Start_iter : {}".format(start_iter))
model.load_state_dict(state_dict)
for cell in model.decoder:
# Make directory where we save model and data
all_dir = os.path.join(CUR_DIR, 'models', f'{DATASET}', 'all')
maybe_mkdir_p(all_dir)
val_dataset = ImageDataset_test(root_dir, ISBI2012=ISBI2012)
# Suffle and load the training set
train_loader = DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True)
val_loader = DataLoader(val_dataset,
batch_size=batch_size,
shuffle=True)
torch.cuda.empty_cache()
unet = Unet().to(device)
if START_FROM == -1: # load latest model
epoch_id = max([
int(name.replace('unet_weight_save_', '').replace('.pth', ''))
for name in os.listdir(os.path.join(all_dir, 'models'))
])
PATH = os.path.join(all_dir, 'models',
'unet_weight_save_{}.pth'.format(epoch_id))
unet.load_state_dict(torch.load(PATH))
print('Number of images used for training:', len(train_dataset))
print(' ')
print('Starting training')
print(' ')
training(unet,
def main():
#Initialize U-net model
print("Initializing Networks...")
haemorrhagesModel = Unet(LEARNING_RATE, IMAGE_SHAPE_3D).get_model()
hardExudatesModel = Unet(LEARNING_RATE, IMAGE_SHAPE_3D).get_model()
microaneurysmsModel = Unet(LEARNING_RATE, IMAGE_SHAPE_3D).get_model()
softExudatesModel = Unet(LEARNING_RATE, IMAGE_SHAPE_3D).get_model()
#Original images
print("Reading original images...")
originalTrain = read_images("Train/original_retinal_images/")
originalTest = read_images("Test/original_retinal_images/")
originalImages = np.concatenate((originalTrain, originalTest), axis=0)
#Train masks
print("Reading train masks...")
originalHaemorrhagesTrain, haemorrhagesTrain = read_masks("Train/masks_Haemorrhages/", originalImages)
originalHardExudatesTrain, hardExudatesTrain = read_masks("Train/masks_Hard_Exudates/", originalImages)
originalMicroaneurysmsTrain, microaneurysmsTrain = read_masks("Train/masks_Microaneurysms/", originalImages)
originalSoftExudatesTrain, softExudatesTrain = read_masks("Train/masks_Soft_Exudates/", originalImages)
#Test masks
print("Reading test masks...")
originalHaemorrhagesTest, haemorrhagesTest = read_masks("Test/masks_Haemorrhages/", originalImages)
originalHardExudatesTest, hardExudatesTest = read_masks("Test/masks_Hard_Exudates/", originalImages)
originalMicroaneurysmsTest, microaneurysmsTest = read_masks("Test/masks_Microaneurysms/", originalImages)
originalSoftExudatesTest, softExudatesTest = read_masks("Test/masks_Soft_Exudates/", originalImages)
#Image Generators
haemorrhagesGen = generate(originalHaemorrhagesTrain, haemorrhagesTrain)
hardExudatesGen = generate(originalHardExudatesTrain, hardExudatesTrain)
microaneurysmsGen = generate(originalMicroaneurysmsTrain, microaneurysmsTrain)
softExudatesGen = generate(originalSoftExudatesTrain, softExudatesTrain)
#Train on generated data
print("Start Training!")
haemorrhagesHistory = haemorrhagesModel.fit_generator(haemorrhagesGen, steps_per_epoch=STEPS_PER_EPOCH, epochs=EPOCHS, validation_data=(originalHaemorrhagesTest, haemorrhagesTest))
hardExudatesHistory = hardExudatesModel.fit_generator(hardExudatesGen, steps_per_epoch=STEPS_PER_EPOCH, epochs=EPOCHS, validation_data=(originalHardExudatesTest, hardExudatesTest))
microaneurysmsHistory = microaneurysmsModel.fit_generator(microaneurysmsGen, steps_per_epoch=STEPS_PER_EPOCH, epochs=EPOCHS, validation_data=(originalMicroaneurysmsTest, microaneurysmsTest))
softExudatesHistory = softExudatesModel.fit_generator(softExudatesGen, steps_per_epoch=STEPS_PER_EPOCH, epochs=EPOCHS, validation_data=(originalSoftExudatesTest, softExudatesTest))
#Save models
print("Saving Models...")
haemorrhagesModel.save("Models/haemorrhages.h5")
hardExudatesModel.save("Models/hardExudates.h5")
microaneurysmsModel.save("Models/microaneurysms.h5")
softExudatesModel.save("Models/softExudates.h5")
#Create loss and metric plots
print("Creating Plots...")
create_plots(haemorrhagesHistory, "Haemorrhages")
create_plots(hardExudatesHistory, "Hard Exudates")
create_plots(microaneurysmsHistory, "Microaneurysms")
create_plots(softExudatesHistory, "Soft Exudates")
#Produce results
print("Producing results...")
haemorrhagesResults = process_masks(haemorrhagesModel.predict(originalHaemorrhagesTest))
write_images(haemorrhagesResults, "Result/masks_Haemorrhages/")
hardExudatesResults = process_masks(hardExudatesModel.predict(originalHardExudatesTest))
write_images(haemorrhagesResults, "Result/masks_Hard_Exudates/")
microaneurysmsResults = process_masks(microaneurysmsModel.predict(originalMicroaneurysmsTest))
write_images(haemorrhagesResults, "Result/masks_Microaneurysms/")
softExudatesResults = process_masks(softExudatesModel.predict(originalSoftExudatesTest))
write_images(haemorrhagesResults, "Result/masks_Soft_Exudates/")
'model_dir': "./models/state_dict/07031644/",
'dataset': '', # Directory of your Dataset
'batch_size': 20, # batchsize, default = 1
'logdir': "./log/", # logdir, log on tensorboard
'which_iter': 1, # "Specific Iter to measure", default=-1
'cont': 0, # "Measure scores from this iter"
'step': 10000, # "Measure scores per this iter step"
}
models = sorted(os.listdir(config['model_dir']),
key=lambda x: int(x.split('epo_')[1].split('step')[0]))
pairdataset = CustomDataset(root_dir=config['dataset'])
dataloader = DataLoader(pairdataset, 8, shuffle=True)
beta_square = 0.3
if config['logdir'] is not None:
writer = SummaryWriter(config['logdir'])
model = Unet(cfg, config['mode']).cuda()
for model_name in models:
model_iter = int(model_name.split('epo_')[1].split('step')[0])
if model_iter % config['step'] != 0:
continue
if model_iter < config['cont']:
continue
if config['which_iter'] > 0 and config['which_iter'] != model_iter:
continue
state_dict = torch.load(os.path.join(config['model_dir'], model_name))
model.load_state_dict(state_dict)
model.eval()
mae = 0
preds = []
masks = []
precs = []
def encode_kwargs_fn(level):
num_stacks = max(level, 1)
return {'num_stacks': num_stacks}
paired_features = generate_paired_features(num_pool, num_features)
model = Unet(in_channels=1,
out_channels=3,
paired_features=paired_features,
pool_block=ResBlock,
pool_kwargs={
'stride': 2
},
up_kwargs={
'attention': True
},
encode_block=ResBlockStack,
encode_kwargs_fn=encode_kwargs_fn,
decode_block=ResBlock).cuda()
patch_size = (160, 160, 80)
optimizer = optim.Adam(model.parameters(), lr=1e-4)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
factor=0.2,
patience=30)
tr_transform = Compose([
GammaTransform((0.9, 1.1)),
result.append(SOUTH)
if -11 <= dy < 0 or dy > 11:
result.append(NORTH)
return result
print("Loading weights (weights.pth). {:.0f} ms - Stage G".format(
(perf_counter() - start_time) * 1e3))
#-------------- CREATE MODEL SECTION ------------------#
model = Unet(encoder_name="efficientnet-b0",
classes=5,
encoder_depth=2,
decoder_channels=(64, 32),
in_channels=64,
encoder_weights=None)
print("Loading weights (weights.pth). {:.0f} ms - Stage H".format(
(perf_counter() - start_time) * 1e3))
model.load_state_dict(
torch.load("halite-imitation-learning-bot/weights.pth",
torch.device(device))['model_state_dict'])
model.to(device)
model.eval()
torch.no_grad()
torch.set_num_threads(os.cpu_count())
# random initialization method
def weights_init_xavier(m):
classname = m.__class__.__name__
# print(classname)
if classname.find('Conv') != -1:
init.xavier_normal(m.weight.data, gain=1)
elif classname.find('Linear') != -1:
init.xavier_normal(m.weight.data, gain=1)
elif classname.find('BatchNorm2d') != -1:
init.uniform(m.weight.data, 1.0, 0.02)
init.constant(m.bias.data, 0.0)
# batch_size = 24
net = Unet().cuda(0)
net.apply(weights_init_xavier)
net.train(True)
optimizerD = optim.Adam(net.parameters(), lr=0.005, betas=(0.5, 0.999))
#Binary Cross Entropy
criterion = nn.BCELoss(size_average=True).cuda(0)
#The path of the data
data_path = './data/images/'
dst = ImageNet_Dataloader(data_path, is_transform=True)
print('length of the dataset', len(dst))
trainloader = data.DataLoader(dst, batch_size=24,shuffle=True)
step_index = 0
beta=(1 - factor),
gamma=0)
return blend
torch.set_printoptions(profile='full')
if __name__ == '__main__':
model_dir = "../pretrained_models/picanet"
models = sorted(os.listdir(model_dir),
key=lambda x: int(x.split('epo_')[1].split('step')[0]))
device = torch.device("cuda")
bdda_model = Unet().to(device)
sage_model = Unet().to(device)
print("Model loaded! Loading Checkpoint...")
bdda_model_name = models[0]
sage_model_name = models[1]
bdda_state_dict = torch.load(os.path.join(model_dir, bdda_model_name))
sage_state_dict = torch.load(os.path.join(model_dir, sage_model_name))
bdda_model.load_state_dict(bdda_state_dict)
sage_model.load_state_dict(sage_state_dict)
print("Checkpoint loaded! Now predicting...")
if args.logdir is None and args.save_dir is None:
print(
"You should specify either --logdir or --save_dir to save results!"
)
assert 0
print(args)
print(os.getcwd())
device = torch.device(args.cuda)
#torch.save(
# model.state_dict(),
# os.path.join(weight_save_dir, '{}epo_{}step.ckpt'.format(epo, iterate))
#)
state_dict = torch.load(args.model_dir, map_location=args.cuda)
model = Unet().to(device)
#model = Unet(cfg).to(device)
#for cell in model.decoder:
# if cell.mode == 'G':
# cell.picanet.renet.vertical.flatten_parameters()
# cell.picanet.renet.horizontal.flatten_parameters()
model.load_state_dict(state_dict) #, strict=False)
custom_dataset = CustomDataset(root_dir=args.dataset)
dataloader = DataLoader(custom_dataset, args.batch_size, shuffle=False)
os.makedirs(os.path.join(args.save_dir, 'img'), exist_ok=True)
os.makedirs(os.path.join(args.save_dir, 'mask'), exist_ok=True)
if args.logdir is not None:
writer = SummaryWriter(args.logdir)
model.eval()
for i, batch in enumerate(tqdm(dataloader)):
img = batch.to(device)
