def main(config_path, type, n_cluster, n_neighbors):
"""
Ablation study on Contrastive or AE representation.
"""
# Load config file
cfg = Config(settings=None)
cfg.load_config(config_path)
# Update n_cluster and n_neighbors if provided
if n_cluster > 0:
cfg.settings['n_cluster'] = n_cluster
if n_neighbors > 0:
cfg.settings['n_neighbors'] = n_neighbors
# Get path to output
OUTPUT_PATH = cfg.settings['PATH']['OUTPUT'] + cfg.settings[
'Experiment_Name'] + datetime.today().strftime('%Y_%m_%d_%Hh%M') + '/'
# make output dir
if not os.path.isdir(OUTPUT_PATH + 'models/'):
os.makedirs(OUTPUT_PATH + 'model/', exist_ok=True)
if not os.path.isdir(OUTPUT_PATH + 'results/'):
os.makedirs(OUTPUT_PATH + 'results/', exist_ok=True)
if not os.path.isdir(OUTPUT_PATH + 'logs/'):
os.makedirs(OUTPUT_PATH + 'logs/', exist_ok=True)
for seed_i, seed in enumerate(cfg.settings['seeds']):
############################### Set Up #################################
# initialize logger
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger()
try:
logger.handlers[1].stream.close()
logger.removeHandler(logger.handlers[1])
except IndexError:
pass
logger.setLevel(logging.INFO)
formatter = logging.Formatter(
'%(asctime)s | %(levelname)s | %(message)s')
log_file = OUTPUT_PATH + 'logs/' + f'log_{seed_i+1}.txt'
file_handler = logging.FileHandler(log_file)
file_handler.setLevel(logging.INFO)
file_handler.setFormatter(formatter)
logger.addHandler(file_handler)
# print path
logger.info(f"Log file : {log_file}")
logger.info(f"Data path : {cfg.settings['PATH']['DATA']}")
logger.info(f"Outputs path : {OUTPUT_PATH}" + "\n")
# Set seed
if seed != -1:
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
logger.info(
f"Set seed {seed_i+1:02}/{len(cfg.settings['seeds']):02} to {seed}"
)
# set number of thread
if cfg.settings['n_thread'] > 0:
torch.set_num_threads(cfg.settings['n_thread'])
# check if GPU available
cfg.settings['device'] = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
# Print technical info in logger
logger.info(f"Device : {cfg.settings['device']}")
logger.info(f"Number of thread : {cfg.settings['n_thread']}")
############################### Split Data #############################
# Load data informations
df_info = pd.read_csv(cfg.settings['PATH']['DATA_INFO'])
df_info = df_info.drop(df_info.columns[0], axis=1)
# remove low contrast images (all black)
df_info = df_info[df_info.low_contrast == 0]
# Train Validation Test Split
spliter = MURA_TrainValidTestSplitter(
df_info,
train_frac=cfg.settings['Split']['train_frac'],
ratio_known_normal=cfg.settings['Split']['known_normal'],
ratio_known_abnormal=cfg.settings['Split']['known_abnormal'],
random_state=42)
spliter.split_data(verbose=False)
train_df = spliter.get_subset('train')
valid_df = spliter.get_subset('valid')
test_df = spliter.get_subset('test')
# print info to logger
for key, value in cfg.settings['Split'].items():
logger.info(f"Split param {key} : {value}")
logger.info("Split Summary \n" +
str(spliter.print_stat(returnTable=True)))
############################### Load model #############################
if type == 'SimCLR':
net = Encoder(MLP_Neurons_layer=[512, 256, 128])
init_key = 'repr_net_dict'
elif type == 'AE':
net = AE_net(MLP_Neurons_layer_enc=[512, 256, 128],
MLP_Neurons_layer_dec=[128, 256, 512],
output_channels=1)
init_key = 'ae_net_dict'
net = net.to(cfg.settings['device'])
pretrain_state_dict = torch.load(cfg.settings['model_path_name'] +
f'{seed_i+1}.pt',
map_location=cfg.settings['device'])
net.load_state_dict(pretrain_state_dict[init_key])
logger.info('Model weights successfully loaded from ' +
cfg.settings['model_path_name'] + f'{seed_i+1}.pt')
ablation_model = NearestNeighbors(
net,
kmeans_reduction=cfg.settings['kmeans_reduction'],
batch_size=cfg.settings['batch_size'],
n_job_dataloader=cfg.settings['num_worker'],
print_batch_progress=cfg.settings['print_batch_progress'],
device=cfg.settings['device'])
############################### Training ###############################
# make dataset
train_dataset = MURA_Dataset(
train_df,
data_path=cfg.settings['PATH']['DATA'],
load_mask=True,
load_semilabels=True,
output_size=cfg.settings['Split']['img_size'],
data_augmentation=False)
valid_dataset = MURA_Dataset(
valid_df,
data_path=cfg.settings['PATH']['DATA'],
load_mask=True,
load_semilabels=True,
output_size=cfg.settings['Split']['img_size'],
data_augmentation=False)
test_dataset = MURA_Dataset(
test_df,
data_path=cfg.settings['PATH']['DATA'],
load_mask=True,
load_semilabels=True,
output_size=cfg.settings['Split']['img_size'],
data_augmentation=False)
logger.info("Online preprocessing pipeline : \n" +
str(train_dataset.transform) + "\n")
# Train
ablation_model.train(train_dataset,
n_cluster=cfg.settings['n_cluster'])
# Evaluate
logger.info(
f'--- Validation with {cfg.settings["n_neighbors"]} neighbors')
ablation_model.evaluate(valid_dataset,
n_neighbors=cfg.settings['n_neighbors'],
mode='valid')
logger.info(f'--- Test with {cfg.settings["n_neighbors"]} neighbors')
ablation_model.evaluate(test_dataset,
n_neighbors=cfg.settings['n_neighbors'],
mode='test')
############################## Save Results ############################
ablation_model.save_results(OUTPUT_PATH +
f'results/results_{seed_i+1}.json')
logger.info("Results saved at " + OUTPUT_PATH +
f"results/results_{seed_i+1}.json")
ablation_model.save_model(OUTPUT_PATH +
f'model/ablation_{seed_i+1}.pt')
logger.info("model saved at " + OUTPUT_PATH +
f"model/abaltion_{seed_i+1}.pt")
# save config file
cfg.settings['device'] = str(cfg.settings['device'])
cfg.save_config(OUTPUT_PATH + 'config.json')
logger.info("Config saved at " + OUTPUT_PATH + "config.json")
def main(ratio_known_abnormal_i):
"""
Joint DSAD experiment with different fraction of known abnormal sample : 0.0%
0.25% 0.5% 1.0% 5.0% and 10.0% over a single replicate.
"""
# initialize logger
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger()
try:
logger.handlers[1].stream.close()
logger.removeHandler(logger.handlers[1])
except IndexError:
pass
logger.setLevel(logging.INFO)
formatter = logging.Formatter('%(asctime)s | %(levelname)s | %(message)s')
log_file = OUTPUT_PATH + 'logs/' + f'log_{ratio_known_abnormal_i:.2%}.txt'
file_handler = logging.FileHandler(log_file)
file_handler.setLevel(logging.INFO)
file_handler.setFormatter(formatter)
logger.addHandler(file_handler)
# print path and main docstring with experiment summary
logger.info('Brief summary of experiment : \n' + main.__doc__)
logger.info(f'Log file : {log_file}')
logger.info(f'Data path : {DATA_PATH}')
logger.info(f'Outputs path : {OUTPUT_PATH}' + '\n')
############################## Make datasets ###############################
# load data_info
df_info = pd.read_csv(DATA_INFO_PATH)
df_info = df_info.drop(df_info.columns[0], axis=1)
# remove low contrast images (all black)
df_info = df_info[df_info.low_contrast == 0]
# Train Validation Test Split
spliter = MURA_TrainValidTestSplitter(
df_info,
train_frac=train_frac,
ratio_known_normal=ratio_known_normal,
ratio_known_abnormal=ratio_known_abnormal_i,
random_state=42)
spliter.split_data(verbose=False)
train_df = spliter.get_subset('train')
valid_df = spliter.get_subset('valid')
test_df = spliter.get_subset('test')
# make datasets
train_dataset = MURA_Dataset(train_df,
data_path=DATA_PATH,
load_mask=True,
load_semilabels=True,
output_size=img_size,
data_augmentation=True)
valid_dataset = MURA_Dataset(valid_df,
data_path=DATA_PATH,
load_mask=True,
load_semilabels=True,
output_size=img_size,
data_augmentation=False)
test_dataset = MURA_Dataset(test_df,
data_path=DATA_PATH,
load_mask=True,
load_semilabels=True,
output_size=img_size,
data_augmentation=False)
# print info to logger
logger.info(f'Train fraction : {train_frac:.0%}')
logger.info(f'Fraction knonw normal : {ratio_known_normal:.0%}')
logger.info(f'Fraction known abnormal : {ratio_known_abnormal_i:.0%}')
logger.info('Split Summary \n' + str(spliter.print_stat(returnTable=True)))
logger.info('Online preprocessing pipeline : \n' +
str(train_dataset.transform) + '\n')
################################ Set Up ####################################
# Set seed
if seed != -1:
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
logger.info(f'Set seed to {seed}')
# set number of thread
if n_thread > 0:
torch.set_num_threads(n_thread)
# print info in logger
logger.info(f'Device : {device}')
logger.info(f'Number of thread : {n_thread}')
logger.info(
f'Number of dataloader worker for Joint DeepSAD : {n_jobs_dataloader}'
+ '\n')
######################### Networks Initialization ##########################
net = AE_SVDD_Hybrid(pretrain_ResNetEnc=ae_pretrain,
output_channels=ae_out_size[0],
return_svdd_embed=True)
net = net.to(device)
# add info to logger
logger.info(f'Network : {net.__class__.__name__}')
logger.info(f'Autoencoder pretrained on ImageNet : {ae_pretrain}')
logger.info(f'DeepSAD eta : {eta}')
logger.info('Network architecture: \n' +
summary_string(net, (1, img_size, img_size),
device=str(device),
batch_size=batch_size) + '\n')
# initialization of the Model
jointDeepSAD = JointDeepSAD(net, eta=eta)
if model_path_to_load:
jointDeepSAD.load_model(model_path_to_load, map_location=device)
logger.info(f'Model Loaded from {model_path_to_load}' + '\n')
################################ Training ##################################
# add parameter info
logger.info(f'Joint DeepSAD number of epoch : {n_epoch}')
logger.info(
f'Joint DeepSAD number of pretraining epoch: {n_epoch_pretrain}')
logger.info(f'Joint DeepSAD learning rate : {lr}')
logger.info(f'Joint DeepSAD learning rate milestone : {lr_milestone}')
logger.info(f'Joint DeepSAD weight_decay : {weight_decay}')
logger.info(f'Joint DeepSAD optimizer : Adam')
logger.info(f'Joint DeepSAD batch_size {batch_size}')
logger.info(
f'Joint DeepSAD number of dataloader worker : {n_jobs_dataloader}')
logger.info(
f'Joint DeepSAD criterion weighting : {criterion_weight[0]} Reconstruction loss + {criterion_weight[1]} SVDD embdedding loss'
+ '\n')
# train DeepSAD
jointDeepSAD.train(train_dataset,
lr=lr,
n_epoch=n_epoch,
n_epoch_pretrain=n_epoch_pretrain,
lr_milestone=lr_milestone,
batch_size=batch_size,
weight_decay=weight_decay,
device=device,
n_jobs_dataloader=n_jobs_dataloader,
print_batch_progress=print_batch_progress,
criterion_weight=criterion_weight,
valid_dataset=None)
# validate DeepSAD
jointDeepSAD.validate(valid_dataset,
device=device,
n_jobs_dataloader=n_jobs_dataloader,
print_batch_progress=print_batch_progress,
criterion_weight=criterion_weight)
# test DeepSAD
jointDeepSAD.test(test_dataset,
device=device,
n_jobs_dataloader=n_jobs_dataloader,
print_batch_progress=print_batch_progress,
criterion_weight=criterion_weight)
# save results
jointDeepSAD.save_results(
OUTPUT_PATH +
f'results/JointDeepSAD_results_{ratio_known_abnormal_i:.2%}.json')
logger.info(
'Test results saved at ' + OUTPUT_PATH +
f'results/JointDeepSAD_results_{ratio_known_abnormal_i:.2%}.json')
# save model
jointDeepSAD.save_model(
OUTPUT_PATH +
f'model/JointDeepSAD_model_{ratio_known_abnormal_i:.2%}.pt')
logger.info('Model saved at ' + OUTPUT_PATH +
f'model/JointDeepSAD_model_{ratio_known_abnormal_i:.2%}.pt')
def main(seed_i):
"""
Train jointly the AutoEncoder and the DeepSAD model following Lukas Ruff et
al. (2019) work adapted to the MURA dataset (preprocessing inspired from the
work of Davletshina et al. (2020)). The network structure is a ResNet18
AutoEncoder until the Adaptative average pooling layer. The AE embdedding is
thus (512, 16, 16). This embdedding is further processed through 3 convolutional
layers (specific to the SVDD embdedding generation) to provide the SVDD
embdedding of 512. The network is trained with two loss functions: a masked MSE
loss for the reconstruction and the DeepSAD loss on the embedding. The two
losses are scaled to be comparable by perfoming one forward pass prior the
training. The Encoder is not initialized with weights trained on ImageNet.
The AE masked reconstruction loss is not computed for known abnormal sample
so that the AE learn to reconstruct normal samples only. The network input is
masked with the mask : only the body part is kept and the background is set
to zero. The AE is pretrained over 5 epochs in order to improve the initialization
of the hypersphere center (we hypothetize that with a pretrained AE the
hypersphere center estimation will be more meaningful). Note that the 'affine'
parameters of BatchNorm2d layers has been changed to False in this implementation.
"""
# initialize logger
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger()
try:
logger.handlers[1].stream.close()
logger.removeHandler(logger.handlers[1])
except IndexError:
pass
logger.setLevel(logging.INFO)
formatter = logging.Formatter('%(asctime)s | %(levelname)s | %(message)s')
log_file = OUTPUT_PATH + 'logs/' + f'log_{seed_i+1}.txt'
file_handler = logging.FileHandler(log_file)
file_handler.setLevel(logging.INFO)
file_handler.setFormatter(formatter)
logger.addHandler(file_handler)
# print path and main docstring with experiment summary
logger.info('Brief summary of experiment : \n' + main.__doc__)
logger.info(f'Log file : {log_file}')
logger.info(f'Data path : {DATA_PATH}')
logger.info(f'Outputs path : {OUTPUT_PATH}' + '\n')
############################## Make datasets ###############################
# load data_info
df_info = pd.read_csv(DATA_INFO_PATH)
df_info = df_info.drop(df_info.columns[0], axis=1)
# remove low contrast images (all black)
df_info = df_info[df_info.low_contrast == 0]
#df_info = df_info[df_info.body_part == 'HAND']
# Train Validation Test Split
spliter = MURA_TrainValidTestSplitter(
df_info,
train_frac=train_frac,
ratio_known_normal=ratio_known_normal,
ratio_known_abnormal=ratio_known_abnormal,
random_state=42)
spliter.split_data(verbose=False)
train_df = spliter.get_subset('train')
valid_df = spliter.get_subset('valid')
test_df = spliter.get_subset('test')
# make datasets
train_dataset = MURA_Dataset(train_df,
data_path=DATA_PATH,
load_mask=True,
load_semilabels=True,
output_size=img_size,
data_augmentation=True)
valid_dataset = MURA_Dataset(valid_df,
data_path=DATA_PATH,
load_mask=True,
load_semilabels=True,
output_size=img_size,
data_augmentation=False)
test_dataset = MURA_Dataset(test_df,
data_path=DATA_PATH,
load_mask=True,
load_semilabels=True,
output_size=img_size,
data_augmentation=False)
# print info to logger
logger.info(f'Train fraction : {train_frac:.0%}')
logger.info(f'Fraction knonw normal : {ratio_known_normal:.0%}')
logger.info(f'Fraction known abnormal : {ratio_known_abnormal:.0%}')
logger.info('Split Summary \n' + str(spliter.print_stat(returnTable=True)))
logger.info('Online preprocessing pipeline : \n' +
str(train_dataset.transform) + '\n')
################################ Set Up ####################################
# Set seed
seed = seeds[seed_i]
if seed != -1:
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
logger.info(f'Set seed {seed_i+1:02}/{n_seeds:02} to {seed}')
# set number of thread
if n_thread > 0:
torch.set_num_threads(n_thread)
# print info in logger
logger.info(f'Device : {device}')
logger.info(f'Number of thread : {n_thread}')
logger.info(
f'Number of dataloader worker for Joint DeepSAD : {n_jobs_dataloader}'
+ '\n')
######################### Networks Initialization ##########################
net = AE_SVDD_Hybrid(pretrain_ResNetEnc=ae_pretrain,
output_channels=ae_out_size[0],
return_svdd_embed=True)
net = net.to(device)
# add info to logger
logger.info(f'Network : {net.__class__.__name__}')
logger.info(f'Autoencoder pretrained on ImageNet : {ae_pretrain}')
logger.info(f'DeepSAD eta : {eta}')
logger.info('Network architecture: \n' +
summary_string(net, (1, img_size, img_size),
device=str(device),
batch_size=batch_size) + '\n')
# initialization of the Model
jointDeepSAD = JointDeepSAD(net, eta=eta)
if model_path_to_load:
jointDeepSAD.load_model(model_path_to_load, map_location=device)
logger.info(f'Model Loaded from {model_path_to_load}' + '\n')
################################ Training ##################################
# add parameter info
logger.info(f'Joint DeepSAD number of epoch : {n_epoch}')
logger.info(
f'Joint DeepSAD number of pretraining epoch: {n_epoch_pretrain}')
logger.info(f'Joint DeepSAD learning rate : {lr}')
logger.info(f'Joint DeepSAD learning rate milestone : {lr_milestone}')
logger.info(f'Joint DeepSAD weight_decay : {weight_decay}')
logger.info(f'Joint DeepSAD optimizer : Adam')
logger.info(f'Joint DeepSAD batch_size {batch_size}')
logger.info(
f'Joint DeepSAD number of dataloader worker : {n_jobs_dataloader}')
logger.info(
f'Joint DeepSAD criterion weighting : {criterion_weight[0]} Reconstruction loss + {criterion_weight[1]} SVDD embdedding loss'
+ '\n')
# train DeepSAD
jointDeepSAD.train(train_dataset,
lr=lr,
n_epoch=n_epoch,
n_epoch_pretrain=n_epoch_pretrain,
lr_milestone=lr_milestone,
batch_size=batch_size,
weight_decay=weight_decay,
device=device,
n_jobs_dataloader=n_jobs_dataloader,
print_batch_progress=print_batch_progress,
criterion_weight=criterion_weight)
# validate DeepSAD
jointDeepSAD.validate(valid_dataset,
device=device,
n_jobs_dataloader=n_jobs_dataloader,
print_batch_progress=print_batch_progress,
criterion_weight=criterion_weight)
# test DeepSAD
jointDeepSAD.test(test_dataset,
device=device,
n_jobs_dataloader=n_jobs_dataloader,
print_batch_progress=print_batch_progress,
criterion_weight=criterion_weight)
# save results
jointDeepSAD.save_results(OUTPUT_PATH +
f'results/JointDeepSAD_results_{seed_i+1}.json')
logger.info('Test results saved at ' + OUTPUT_PATH +
f'results/JointDeepSAD_results_{seed_i+1}.json' + '\n')
# save model
jointDeepSAD.save_model(OUTPUT_PATH +
f'model/JointDeepSAD_model_{seed_i+1}.pt')
logger.info('Model saved at ' + OUTPUT_PATH +
f'model/JointDeepSAD_model_{seed_i+1}.pt')
def main(config_path):
"""
Train a DSAD on the MURA dataset using a SimCLR pretraining.
"""
# Load config file
cfg = Config(settings=None)
cfg.load_config(config_path)
# Get path to output
OUTPUT_PATH = cfg.settings['PATH']['OUTPUT'] + cfg.settings[
'Experiment_Name'] + datetime.today().strftime('%Y_%m_%d_%Hh%M') + '/'
# make output dir
if not os.path.isdir(OUTPUT_PATH + 'models/'):
os.makedirs(OUTPUT_PATH + 'model/', exist_ok=True)
if not os.path.isdir(OUTPUT_PATH + 'results/'):
os.makedirs(OUTPUT_PATH + 'results/', exist_ok=True)
if not os.path.isdir(OUTPUT_PATH + 'logs/'):
os.makedirs(OUTPUT_PATH + 'logs/', exist_ok=True)
for seed_i, seed in enumerate(cfg.settings['seeds']):
############################### Set Up #################################
# initialize logger
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger()
try:
logger.handlers[1].stream.close()
logger.removeHandler(logger.handlers[1])
except IndexError:
pass
logger.setLevel(logging.INFO)
formatter = logging.Formatter(
'%(asctime)s | %(levelname)s | %(message)s')
log_file = OUTPUT_PATH + 'logs/' + f'log_{seed_i+1}.txt'
file_handler = logging.FileHandler(log_file)
file_handler.setLevel(logging.INFO)
file_handler.setFormatter(formatter)
logger.addHandler(file_handler)
# print path
logger.info(f"Log file : {log_file}")
logger.info(f"Data path : {cfg.settings['PATH']['DATA']}")
logger.info(f"Outputs path : {OUTPUT_PATH}" + "\n")
# Set seed
if seed != -1:
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
logger.info(
f"Set seed {seed_i+1:02}/{len(cfg.settings['seeds']):02} to {seed}"
)
# set number of thread
if cfg.settings['n_thread'] > 0:
torch.set_num_threads(cfg.settings['n_thread'])
# check if GPU available
cfg.settings['device'] = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
# Print technical info in logger
logger.info(f"Device : {cfg.settings['device']}")
logger.info(f"Number of thread : {cfg.settings['n_thread']}")
############################### Split Data #############################
# Load data informations
df_info = pd.read_csv(cfg.settings['PATH']['DATA_INFO'])
df_info = df_info.drop(df_info.columns[0], axis=1)
# remove low contrast images (all black)
df_info = df_info[df_info.low_contrast == 0]
# Train Validation Test Split
spliter = MURA_TrainValidTestSplitter(
df_info,
train_frac=cfg.settings['Split']['train_frac'],
ratio_known_normal=cfg.settings['Split']['known_normal'],
ratio_known_abnormal=cfg.settings['Split']['known_abnormal'],
random_state=42)
spliter.split_data(verbose=False)
train_df = spliter.get_subset('train')
valid_df = spliter.get_subset('valid')
test_df = spliter.get_subset('test')
# print info to logger
for key, value in cfg.settings['Split'].items():
logger.info(f"Split param {key} : {value}")
logger.info("Split Summary \n" +
str(spliter.print_stat(returnTable=True)))
############################# Build Model #############################
# make networks
net_CLR = SimCLR_net(
MLP_Neurons_layer=cfg.settings['SimCLR']['MLP_head'])
net_CLR = net_CLR.to(cfg.settings['device'])
net_DSAD = SimCLR_net(
MLP_Neurons_layer=cfg.settings['DSAD']['MLP_head'])
net_DSAD = net_DSAD.to(cfg.settings['device'])
# print network architecture
net_architecture = summary_string(
net_CLR, (1, cfg.settings['Split']['img_size'],
cfg.settings['Split']['img_size']),
batch_size=cfg.settings['SimCLR']['batch_size'],
device=str(cfg.settings['device']))
logger.info("SimCLR net architecture: \n" + net_architecture + '\n')
net_architecture = summary_string(
net_DSAD, (1, cfg.settings['Split']['img_size'],
cfg.settings['Split']['img_size']),
batch_size=cfg.settings['DSAD']['batch_size'],
device=str(cfg.settings['device']))
logger.info("DSAD net architecture: \n" + net_architecture + '\n')
# make model
clr_DSAD = SimCLR_DSAD(net_CLR,
net_DSAD,
tau=cfg.settings['SimCLR']['tau'],
eta=cfg.settings['DSAD']['eta'])
############################# Train SimCLR #############################
# make datasets
train_dataset_CLR = MURADataset_SimCLR(
train_df,
data_path=cfg.settings['PATH']['DATA'],
output_size=cfg.settings['Split']['img_size'],
mask_img=True)
valid_dataset_CLR = MURADataset_SimCLR(
valid_df,
data_path=cfg.settings['PATH']['DATA'],
output_size=cfg.settings['Split']['img_size'],
mask_img=True)
test_dataset_CLR = MURADataset_SimCLR(
test_df,
data_path=cfg.settings['PATH']['DATA'],
output_size=cfg.settings['Split']['img_size'],
mask_img=True)
logger.info("SimCLR Online preprocessing pipeline : \n" +
str(train_dataset_CLR.transform) + "\n")
# Load model if required
if cfg.settings['SimCLR']['model_path_to_load']:
clr_DSAD.load_repr_net(
cfg.settings['SimCLR']['model_path_to_load'],
map_location=cfg.settings['device'])
logger.info(
f"SimCLR Model Loaded from {cfg.settings['SimCLR']['model_path_to_load']}"
+ "\n")
# print Train parameters
for key, value in cfg.settings['SimCLR'].items():
logger.info(f"SimCLR {key} : {value}")
# Train SimCLR
clr_DSAD.train_SimCLR(
train_dataset_CLR,
valid_dataset=None,
n_epoch=cfg.settings['SimCLR']['n_epoch'],
batch_size=cfg.settings['SimCLR']['batch_size'],
lr=cfg.settings['SimCLR']['lr'],
weight_decay=cfg.settings['SimCLR']['weight_decay'],
lr_milestones=cfg.settings['SimCLR']['lr_milestone'],
n_job_dataloader=cfg.settings['SimCLR']['num_worker'],
device=cfg.settings['device'],
print_batch_progress=cfg.settings['print_batch_progress'])
# Evaluate SimCLR to get embeddings
clr_DSAD.evaluate_SimCLR(
valid_dataset_CLR,
batch_size=cfg.settings['SimCLR']['batch_size'],
n_job_dataloader=cfg.settings['SimCLR']['num_worker'],
device=cfg.settings['device'],
print_batch_progress=cfg.settings['print_batch_progress'],
set='valid')
clr_DSAD.evaluate_SimCLR(
test_dataset_CLR,
batch_size=cfg.settings['SimCLR']['batch_size'],
n_job_dataloader=cfg.settings['SimCLR']['num_worker'],
device=cfg.settings['device'],
print_batch_progress=cfg.settings['print_batch_progress'],
set='test')
# save repr net
clr_DSAD.save_repr_net(OUTPUT_PATH + f'model/SimCLR_net_{seed_i+1}.pt')
logger.info("SimCLR model saved at " + OUTPUT_PATH +
f"model/SimCLR_net_{seed_i+1}.pt")
# save Results
clr_DSAD.save_results(OUTPUT_PATH + f'results/results_{seed_i+1}.json')
logger.info("Results saved at " + OUTPUT_PATH +
f"results/results_{seed_i+1}.json")
######################## Transfer Encoder Weight #######################
clr_DSAD.transfer_encoder()
############################## Train DSAD ##############################
# make dataset
train_dataset_AD = MURA_Dataset(
train_df,
data_path=cfg.settings['PATH']['DATA'],
load_mask=True,
load_semilabels=True,
output_size=cfg.settings['Split']['img_size'])
valid_dataset_AD = MURA_Dataset(
valid_df,
data_path=cfg.settings['PATH']['DATA'],
load_mask=True,
load_semilabels=True,
output_size=cfg.settings['Split']['img_size'])
test_dataset_AD = MURA_Dataset(
test_df,
data_path=cfg.settings['PATH']['DATA'],
load_mask=True,
load_semilabels=True,
output_size=cfg.settings['Split']['img_size'])
logger.info("DSAD Online preprocessing pipeline : \n" +
str(train_dataset_AD.transform) + "\n")
# Load model if required
if cfg.settings['DSAD']['model_path_to_load']:
clr_DSAD.load_AD(cfg.settings['DSAD']['model_path_to_load'],
map_location=cfg.settings['device'])
logger.info(
f"DSAD Model Loaded from {cfg.settings['DSAD']['model_path_to_load']} \n"
)
# print Train parameters
for key, value in cfg.settings['DSAD'].items():
logger.info(f"DSAD {key} : {value}")
# Train DSAD
clr_DSAD.train_AD(
train_dataset_AD,
valid_dataset=valid_dataset_AD,
n_epoch=cfg.settings['DSAD']['n_epoch'],
batch_size=cfg.settings['DSAD']['batch_size'],
lr=cfg.settings['DSAD']['lr'],
weight_decay=cfg.settings['DSAD']['weight_decay'],
lr_milestone=cfg.settings['DSAD']['lr_milestone'],
n_job_dataloader=cfg.settings['DSAD']['num_worker'],
device=cfg.settings['device'],
print_batch_progress=cfg.settings['print_batch_progress'])
logger.info('--- Validation')
clr_DSAD.evaluate_AD(
valid_dataset_AD,
batch_size=cfg.settings['DSAD']['batch_size'],
n_job_dataloader=cfg.settings['DSAD']['num_worker'],
device=cfg.settings['device'],
print_batch_progress=cfg.settings['print_batch_progress'],
set='valid')
logger.info('--- Test')
clr_DSAD.evaluate_AD(
test_dataset_AD,
batch_size=cfg.settings['DSAD']['batch_size'],
n_job_dataloader=cfg.settings['DSAD']['num_worker'],
device=cfg.settings['device'],
print_batch_progress=cfg.settings['print_batch_progress'],
set='test')
# save DSAD
clr_DSAD.save_AD(OUTPUT_PATH + f'model/DSAD_{seed_i+1}.pt')
logger.info("model saved at " + OUTPUT_PATH +
f"model/DSAD_{seed_i+1}.pt")
########################## Save Results ################################
# save Results
clr_DSAD.save_results(OUTPUT_PATH + f'results/results_{seed_i+1}.json')
logger.info("Results saved at " + OUTPUT_PATH +
f"results/results_{seed_i+1}.json")
# save config file
cfg.settings['device'] = str(cfg.settings['device'])
cfg.save_config(OUTPUT_PATH + 'config.json')
logger.info("Config saved at " + OUTPUT_PATH + "config.json")
def main(seed_i):
"""
Implementation of the unsupervised DROCC model proposed by Goyal et al (2020).
The model uses a binary classifier with a ResNet18 backbone. The output is a
logit that serves as anomaly score. The loss is computed as Binary Cross
Entropy loss with logit. The training consist of few epoch trained only with
the normal samples. Then each epoch starts with the generation of adversarial
examples. The adversarial search is performed only on normal samples as we
want the network to learn the manifold of normal samples. It uses a slightly
modifided projection gradient descent algorithm. Then the samples and
adversarial samples are passed through the network similarly to a standard
classification task. Note that the input samples are masked with the mask
generated in the preprocesing steps.
"""
# initialize logger
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger()
try:
logger.handlers[1].stream.close()
logger.removeHandler(logger.handlers[1])
except IndexError:
pass
logger.setLevel(logging.INFO)
formatter = logging.Formatter('%(asctime)s | %(levelname)s | %(message)s')
log_file = OUTPUT_PATH + 'logs/' + f'log_{seed_i+1}.txt'
file_handler = logging.FileHandler(log_file)
file_handler.setLevel(logging.INFO)
file_handler.setFormatter(formatter)
logger.addHandler(file_handler)
# print path and main docstring with experiment summary
logger.info('Brief summary of experiment : \n' + main.__doc__)
logger.info(f'Log file : {log_file}')
logger.info(f'Data path : {DATA_PATH}')
logger.info(f'Outputs path : {OUTPUT_PATH}' + '\n')
############################## Make datasets ###############################
# load data_info
df_info = pd.read_csv(DATA_INFO_PATH)
df_info = df_info.drop(df_info.columns[0], axis=1)
# remove low contrast images (all black)
df_info = df_info[df_info.low_contrast == 0]
# Train Validation Test Split
spliter = MURA_TrainValidTestSplitter(
df_info,
train_frac=train_frac,
ratio_known_normal=ratio_known_normal,
ratio_known_abnormal=ratio_known_abnormal,
random_state=42)
spliter.split_data(verbose=False)
train_df = spliter.get_subset('train')
valid_df = spliter.get_subset('valid')
test_df = spliter.get_subset('test')
# make datasets
train_dataset = MURA_Dataset(train_df,
data_path=DATA_PATH,
load_mask=True,
load_semilabels=True,
output_size=img_size)
valid_dataset = MURA_Dataset(valid_df,
data_path=DATA_PATH,
load_mask=True,
load_semilabels=True,
output_size=img_size)
test_dataset = MURA_Dataset(test_df,
data_path=DATA_PATH,
load_mask=True,
load_semilabels=True,
output_size=img_size)
# print info to logger
logger.info(f'Train fraction : {train_frac:.0%}')
logger.info(f'Fraction knonw normal : {ratio_known_normal:.0%}')
logger.info(f'Fraction known abnormal : {ratio_known_abnormal:.0%}')
logger.info('Split Summary \n' + str(spliter.print_stat(returnTable=True)))
logger.info('Online preprocessing pipeline : \n' +
str(train_dataset.transform) + '\n')
################################ Set Up ####################################
# Set seed
seed = seeds[seed_i]
if seed != -1:
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
logger.info(f'Set seed {seed_i+1:02}/{n_seeds:02} to {seed}')
# set number of thread
if n_thread > 0:
torch.set_num_threads(n_thread)
# print info in logger
logger.info(f'Device : {device}')
logger.info(f'Number of thread : {n_thread}')
logger.info(
f'Number of dataloader worker for {Experiment_Name} : {n_jobs_dataloader}'
+ '\n')
######################### Networks Initialization ##########################
net = ResNet18_binary(pretrained=pretrain)
net = net.to(device)
# add info to logger
logger.info(f'Network : {net.__class__.__name__}')
logger.info(f'ResNet18 pretrained on ImageNet : {pretrain}')
logger.info('Network architecture: \n' +
summary_string(net, (1, img_size, img_size),
device=str(device),
batch_size=batch_size) + '\n')
# initialization of the Model
drocc = DROCC(net, r)
if model_path_to_load:
drocc.load_model(model_path_to_load, map_location=device)
logger.info(f'Model Loaded from {model_path_to_load}' + '\n')
################################ Training ##################################
# add parameter info
logger.info(f'{Experiment_Name} radius r : {r}')
logger.info(f'{Experiment_Name} gamma : {gamma}')
logger.info(f'{Experiment_Name} adversarial importance mu : {mu}')
logger.info(f'{Experiment_Name} number of initial epoch : {n_epoch_init}')
logger.info(f'{Experiment_Name} number of epoch : {n_epoch}')
logger.info(
f'{Experiment_Name} number of adversarial search epoch: {n_epoch_adv}')
logger.info(f'{Experiment_Name} learning rate : {lr}')
logger.info(
f'{Experiment_Name} adversarial search learning rate : {lr_adv}')
logger.info(f'{Experiment_Name} learning rate milestone : {lr_milestone}')
logger.info(f'{Experiment_Name} weight_decay : {weight_decay}')
logger.info(f'{Experiment_Name} optimizer : Adam')
logger.info(f'{Experiment_Name} batch_size {batch_size}')
logger.info(
f'{Experiment_Name} number of dataloader worker : {n_jobs_dataloader}')
# train DROCC
drocc.train(train_dataset,
gamma=gamma,
mu=mu,
lr=lr,
lr_adv=lr_adv,
lr_milestone=lr_milestone,
weight_decay=weight_decay,
n_epoch=n_epoch,
n_epoch_init=n_epoch_init,
n_epoch_adv=n_epoch_adv,
batch_size=batch_size,
device=device,
n_jobs_dataloader=n_jobs_dataloader,
LFOC=use_LFOC,
print_batch_progress=print_batch_progress)
# validate DROCC
drocc.validate(valid_dataset,
device=device,
n_jobs_dataloader=n_jobs_dataloader,
print_batch_progress=print_batch_progress)
# test DROCC
drocc.test(test_dataset,
device=device,
n_jobs_dataloader=n_jobs_dataloader,
print_batch_progress=print_batch_progress)
# save results
drocc.save_results(OUTPUT_PATH +
f'results/{Experiment_Name}_results_{seed_i+1}.json')
logger.info('Test results saved at ' + OUTPUT_PATH +
f'results/{Experiment_Name}_results_{seed_i+1}.json' + '\n')
# save model
drocc.save_model(OUTPUT_PATH +
f'model/{Experiment_Name}_model_{seed_i+1}.pt')
logger.info('Model saved at ' + OUTPUT_PATH +
f'model/{Experiment_Name}_model_{seed_i+1}.pt')
def main(seed_i):
"""
Extension of the deep multi-sphere SVDD to semi-supervised settings inpired
from the DSAD of Ruff et al. (2020).
MSAD loss changed to use the sqrt(dist) for normal samples and 1/(dist^2)
for abnormal samples. The network is pretrained for longer (30 epochs) to get
a better KMeans initialization. Anomaly score is dist - R
"""
# initialize logger
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger()
try:
logger.handlers[1].stream.close()
logger.removeHandler(logger.handlers[1])
except IndexError:
pass
logger.setLevel(logging.INFO)
formatter = logging.Formatter('%(asctime)s | %(levelname)s | %(message)s')
log_file = OUTPUT_PATH + 'logs/' + f'log_{seed_i+1}.txt'
file_handler = logging.FileHandler(log_file)
file_handler.setLevel(logging.INFO)
file_handler.setFormatter(formatter)
logger.addHandler(file_handler)
# print path and main docstring with experiment summary
logger.info('Brief summary of experiment : \n' + main.__doc__)
logger.info(f'Log file : {log_file}')
logger.info(f'Data path : {DATA_PATH}')
logger.info(f'Outputs path : {OUTPUT_PATH}' + '\n')
############################## Make datasets ###############################
# load data_info
df_info = pd.read_csv(DATA_INFO_PATH)
df_info = df_info.drop(df_info.columns[0], axis=1)
# remove low contrast images (all black)
df_info = df_info[df_info.low_contrast == 0]
# Train Validation Test Split
spliter = MURA_TrainValidTestSplitter(
df_info,
train_frac=train_frac,
ratio_known_normal=ratio_known_normal,
ratio_known_abnormal=ratio_known_abnormal,
random_state=42)
spliter.split_data(verbose=False)
train_df = spliter.get_subset('train')
valid_df = spliter.get_subset('valid')
test_df = spliter.get_subset('test')
# make datasets
train_dataset = MURA_Dataset(train_df,
data_path=DATA_PATH,
load_mask=True,
load_semilabels=True,
output_size=img_size)
valid_dataset = MURA_Dataset(valid_df,
data_path=DATA_PATH,
load_mask=True,
load_semilabels=True,
output_size=img_size)
test_dataset = MURA_Dataset(test_df,
data_path=DATA_PATH,
load_mask=True,
load_semilabels=True,
output_size=img_size)
# print info to logger
logger.info(f'Train fraction : {train_frac:.0%}')
logger.info(f'Fraction knonw normal : {ratio_known_normal:.0%}')
logger.info(f'Fraction known abnormal : {ratio_known_abnormal:.0%}')
logger.info('Split Summary \n' + str(spliter.print_stat(returnTable=True)))
logger.info('Online preprocessing pipeline : \n' +
str(train_dataset.transform) + '\n')
################################ Set Up ####################################
# Set seed
seed = seeds[seed_i]
if seed != -1:
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
logger.info(f'Set seed {seed_i+1:02}/{n_seeds:02} to {seed}')
# set number of thread
if n_thread > 0:
torch.set_num_threads(n_thread)
# print info in logger
logger.info(f'Device : {device}')
logger.info(f'Number of thread : {n_thread}')
logger.info(
f'Number of dataloader worker for {Experiment_Name} : {n_jobs_dataloader}'
+ '\n')
######################### Networks Initialization ##########################
net = AE_SVDD_Hybrid(pretrain_ResNetEnc=ae_pretrain,
output_channels=ae_out_size[0],
return_svdd_embed=True)
net = net.to(device)
# add info to logger
logger.info(f'Network : {net.__class__.__name__}')
logger.info(f'Autoencoder pretrained on ImageNet : {ae_pretrain}')
logger.info('Network architecture: \n' +
summary_string(net, (1, img_size, img_size),
device=str(device),
batch_size=batch_size) + '\n')
# initialization of the Model
jointDMSAD = joint_DMSAD(net, eta=eta, gamma=gamma)
if model_path_to_load:
jointDMSAD.load_model(model_path_to_load, map_location=device)
logger.info(f'Model Loaded from {model_path_to_load}' + '\n')
################################ Training ##################################
# add parameter info
logger.info(f'{Experiment_Name} eta : {eta}')
logger.info(f'{Experiment_Name} gamma : {gamma}')
logger.info(f'{Experiment_Name} number of epoch : {n_epoch}')
logger.info(
f'{Experiment_Name} number of pretraining epoch: {n_epoch_pretrain}')
logger.info(
f'{Experiment_Name} number of initial hypersphere: {n_sphere_init}')
logger.info(f'{Experiment_Name} learning rate : {lr}')
logger.info(f'{Experiment_Name} learning rate milestones : {lr_milestone}')
logger.info(f'{Experiment_Name} weight_decay : {weight_decay}')
logger.info(f'{Experiment_Name} optimizer : Adam')
logger.info(f'{Experiment_Name} batch_size {batch_size}')
logger.info(
f'{Experiment_Name} number of dataloader worker : {n_jobs_dataloader}')
logger.info(
f'{Experiment_Name} criterion weighting : {criterion_weight[0]} Reconstruction loss + {criterion_weight[1]} MSAD embdedding loss'
)
logger.info(
f'{Experiment_Name} reset scaling epoch : {reset_scaling_epoch}')
# train DMSAD
jointDMSAD.train(train_dataset,
valid_dataset=valid_dataset,
n_sphere_init=n_sphere_init,
n_epoch=n_epoch,
n_epoch_pretrain=n_epoch_pretrain,
lr=lr,
weight_decay=weight_decay,
lr_milestone=lr_milestone,
criterion_weight=criterion_weight,
reset_scaling_epoch=reset_scaling_epoch,
batch_size=batch_size,
n_jobs_dataloader=n_jobs_dataloader,
device=device,
print_batch_progress=print_batch_progress)
# validate DMSAD
jointDMSAD.validate(valid_dataset,
batch_size=batch_size,
n_jobs_dataloader=n_jobs_dataloader,
criterion_weight=criterion_weight,
device=device,
print_batch_progress=print_batch_progress)
# test DMSAD
jointDMSAD.test(test_dataset,
batch_size=batch_size,
n_jobs_dataloader=n_jobs_dataloader,
criterion_weight=criterion_weight,
device=device,
print_batch_progress=print_batch_progress)
# save results
jointDMSAD.save_results(
OUTPUT_PATH + f'results/{Experiment_Name}_results_{seed_i+1}.json')
logger.info('Test results saved at ' + OUTPUT_PATH +
f'results/{Experiment_Name}_results_{seed_i+1}.json' + '\n')
# save model
jointDMSAD.save_model(OUTPUT_PATH +
f'model/{Experiment_Name}_model_{seed_i+1}.pt')
logger.info('Model saved at ' + OUTPUT_PATH +
f'model/{Experiment_Name}_model_{seed_i+1}.pt')
def main(seed_i):
"""
Train a DeepSAD model following Lukas Ruff et al. (2019) work and code structure
adapted to the MURA dataset (preprocessing inspired from the work of Davletshina
et al. (2020)). The DeepSAD network structure is a ResNet18 Encoder. The Encoder
is pretrained via Autoencoder training. The Autoencoder itself is not initialized
with weights trained on ImageNet. The best threshold on the scores is defined
using the validation set as the one maximizing the F1-score. The ROC AUC is
reported on the test and validation set.
"""
# initialize logger
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger()
try:
logger.handlers[1].stream.close()
logger.removeHandler(logger.handlers[1])
except IndexError:
pass
logger.setLevel(logging.INFO)
formatter = logging.Formatter('%(asctime)s | %(levelname)s | %(message)s')
log_file = OUTPUT_PATH + 'logs/' + f'log_{seed_i+1}.txt'
file_handler = logging.FileHandler(log_file)
file_handler.setLevel(logging.INFO)
file_handler.setFormatter(formatter)
logger.addHandler(file_handler)
# print path
logger.info('Brief summary of experiment : \n' + main.__doc__)
if Note is not None: logger.info(Note + '\n')
logger.info(f'Log file : {log_file}')
logger.info(f'Data path : {DATA_PATH}')
logger.info(f'Outputs path : {OUTPUT_PATH}' + '\n')
############################## Make datasets ###############################
# load data_info
df_info = pd.read_csv(DATA_INFO_PATH)
df_info = df_info.drop(df_info.columns[0], axis=1)
# remove low contrast images (all black)
df_info = df_info[df_info.low_contrast == 0]
# Train Validation Test Split
spliter = MURA_TrainValidTestSplitter(
df_info,
train_frac=train_frac,
ratio_known_normal=ratio_known_normal,
ratio_known_abnormal=ratio_known_abnormal,
random_state=42)
spliter.split_data(verbose=False)
train_df = spliter.get_subset('train')
valid_df = spliter.get_subset('valid')
test_df = spliter.get_subset('test')
# make datasets
train_dataset = MURA_Dataset(train_df,
data_path=DATA_PATH,
load_mask=True,
load_semilabels=True,
output_size=img_size)
valid_dataset = MURA_Dataset(valid_df,
data_path=DATA_PATH,
load_mask=True,
load_semilabels=True,
output_size=img_size)
test_dataset = MURA_Dataset(test_df,
data_path=DATA_PATH,
load_mask=True,
load_semilabels=True,
output_size=img_size)
# print info to logger
logger.info(f'Train fraction : {train_frac:.0%}')
logger.info(f'Fraction knonw normal : {ratio_known_normal:.0%}')
logger.info(f'Fraction known abnormal : {ratio_known_abnormal:.0%}')
logger.info('Split Summary \n' + str(spliter.print_stat(returnTable=True)))
logger.info('Online preprocessing pipeline : \n' +
str(train_dataset.transform) + '\n')
################################ Set Up ####################################
# Set seed
seed = seeds[seed_i]
if seed != -1:
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
logger.info(f'Set seed {seed_i+1:02}/{n_seeds:02} to {seed}')
# set number of thread
if n_thread > 0:
torch.set_num_threads(n_thread)
# print info in logger
logger.info(f'Device : {device}')
logger.info(f'Number of thread : {n_thread}')
logger.info(
f'Number of dataloader worker for DeepSAD : {n_jobs_dataloader}')
logger.info(
f'Autoencoder number of dataloader worker : {ae_n_jobs_dataloader}' +
'\n')
######################### Networks Initialization ##########################
ae_net = AE_ResNet18(embed_dim=embed_dim,
pretrain_ResNetEnc=ae_pretrain,
output_size=ae_out_size)
ae_net = ae_net.to(device)
net = ResNet18_Encoder(embed_dim=embed_dim, pretrained=False)
net = net.to(device)
# initialization of the Model
deepSAD = DeepSAD(net, ae_net=ae_net, eta=eta)
# add info to logger
logger.info(f'Autoencoder : {ae_net.__class__.__name__}')
logger.info(f'Encoder : {net.__class__.__name__}')
logger.info(f'Embedding dimension : {embed_dim}')
logger.info(f'Autoencoder pretrained on ImageNet : {ae_pretrain}')
logger.info(f'DeepSAD eta : {eta}')
logger.info(
'Autoencoder architecture: \n' +
summary_string(ae_net, (1, img_size, img_size), device=str(device)) +
'\n')
if model_path_to_load:
deepSAD.load_model(model_path_to_load,
load_ae=True,
map_location=device)
logger.info(f'Model Loaded from {model_path_to_load}' + '\n')
############################## Pretraining #################################
logger.info(f'Pretraining DeepSAD via Autoencoder : {pretrain}')
if pretrain:
# add parameter info
logger.info(f'Autoencoder number of epoch : {ae_n_epoch}')
logger.info(f'Autoencoder learning rate : {ae_lr}')
logger.info(f'Autoencoder learning rate milestone : {ae_lr_milestone}')
logger.info(f'Autoencoder weight_decay : {ae_weight_decay}')
logger.info(f'Autoencoder optimizer : Adam')
logger.info(f'Autoencoder batch_size {ae_batch_size}' + '\n')
# train AE
deepSAD.pretrain(train_dataset,
valid_dataset,
test_dataset,
lr=ae_lr,
n_epoch=ae_n_epoch,
lr_milestone=ae_lr_milestone,
batch_size=ae_batch_size,
weight_decay=ae_weight_decay,
device=device,
n_jobs_dataloader=ae_n_jobs_dataloader,
print_batch_progress=print_batch_progress)
################################ Training ##################################
# add parameter info
logger.info(f'DeepSAD number of epoch : {n_epoch}')
logger.info(f'DeepSAD learning rate : {lr}')
logger.info(f'DeepSAD learning rate milestone : {lr_milestone}')
logger.info(f'DeepSAD weight_decay : {weight_decay}')
logger.info(f'DeepSAD optimizer : Adam')
logger.info(f'DeepSAD batch_size {batch_size}')
logger.info(f'DeepSAD number of dataloader worker : {n_jobs_dataloader}' +
'\n')
# train DeepSAD
deepSAD.train(train_dataset,
lr=lr,
n_epoch=n_epoch,
lr_milestone=lr_milestone,
batch_size=batch_size,
weight_decay=weight_decay,
device=device,
n_jobs_dataloader=n_jobs_dataloader,
print_batch_progress=print_batch_progress)
# validate DeepSAD
deepSAD.validate(valid_dataset,
device=device,
n_jobs_dataloader=n_jobs_dataloader,
print_batch_progress=print_batch_progress)
# test DeepSAD
deepSAD.test(test_dataset,
device=device,
n_jobs_dataloader=n_jobs_dataloader,
print_batch_progress=print_batch_progress)
# save results
deepSAD.save_results(OUTPUT_PATH +
f'results/DeepSAD_results_{seed_i+1}.json')
logger.info('Test results saved at ' + OUTPUT_PATH +
f'results/DeepSAD_results_{seed_i+1}.json' + '\n')
# save model
deepSAD.save_model(OUTPUT_PATH + f'model/DeepSAD_model_{seed_i+1}.pt')
logger.info('Model saved at ' + OUTPUT_PATH +
f'model/DeepSAD_model_{seed_i+1}.pt')
def main(seed_i):
"""
Implementation of the unsupervised ARAE model proposed by Salehi et al (2020).
This unsupervised method apply a projected gradient descent algorithm to find
a more meaningful lattent space for the autoencoder. The encoder composed of
a ResNet18 encoder. The decoder is composed of a mirrored ResNet18. The latent
space has dimension (16,16,512).
"""
# initialize logger
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger()
try:
logger.handlers[1].stream.close()
logger.removeHandler(logger.handlers[1])
except IndexError:
pass
logger.setLevel(logging.INFO)
formatter = logging.Formatter('%(asctime)s | %(levelname)s | %(message)s')
log_file = OUTPUT_PATH + 'logs/' + f'log_{seed_i+1}.txt'
file_handler = logging.FileHandler(log_file)
file_handler.setLevel(logging.INFO)
file_handler.setFormatter(formatter)
logger.addHandler(file_handler)
# print path and main docstring with experiment summary
logger.info('Brief summary of experiment : \n' + main.__doc__)
logger.info(f'Log file : {log_file}')
logger.info(f'Data path : {DATA_PATH}')
logger.info(f'Outputs path : {OUTPUT_PATH}' + '\n')
############################## Make datasets ###############################
# load data_info
df_info = pd.read_csv(DATA_INFO_PATH)
df_info = df_info.drop(df_info.columns[0], axis=1)
# remove low contrast images (all black)
df_info = df_info[df_info.low_contrast == 0]
# keep only hands
df_info = df_info[df_info.body_part == 'HAND']
# Train Validation Test Split
spliter = MURA_TrainValidTestSplitter(df_info, train_frac=train_frac,
ratio_known_normal=ratio_known_normal,
ratio_known_abnormal=ratio_known_abnormal, random_state=42)
spliter.split_data(verbose=False)
train_df = spliter.get_subset('train')
valid_df = spliter.get_subset('valid')
test_df = spliter.get_subset('test')
# make datasets
train_dataset = MURA_Dataset(train_df, data_path=DATA_PATH, load_mask=True,
load_semilabels=True, output_size=img_size)
valid_dataset = MURA_Dataset(valid_df, data_path=DATA_PATH, load_mask=True,
load_semilabels=True, output_size=img_size)
test_dataset = MURA_Dataset(test_df, data_path=DATA_PATH, load_mask=True,
load_semilabels=True, output_size=img_size)
# print info to logger
logger.info(f'Train fraction : {train_frac:.0%}')
logger.info(f'Fraction knonw normal : {ratio_known_normal:.0%}')
logger.info(f'Fraction known abnormal : {ratio_known_abnormal:.0%}')
logger.info('Split Summary \n' + str(spliter.print_stat(returnTable=True)))
logger.info('Online preprocessing pipeline : \n' + str(train_dataset.transform) + '\n')
################################ Set Up ####################################
# Set seed
seed = seeds[seed_i]
if seed != -1:
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
logger.info(f'Set seed {seed_i+1:02}/{n_seeds:02} to {seed}')
# set number of thread
if n_thread > 0:
torch.set_num_threads(n_thread)
# print info in logger
logger.info(f'Device : {device}')
logger.info(f'Number of thread : {n_thread}')
logger.info(f'Number of dataloader worker for {Experiment_Name} : {n_jobs_dataloader}' + '\n')
######################### Networks Initialization ##########################
net = AE_ResNet18(pretrain_ResNetEnc=pretrain, output_channels=ae_output_size[0])
net = net.to(device)
# add info to logger
logger.info(f'Network : {net.__class__.__name__}')
logger.info(f'ResNet18 pretrained on ImageNet : {pretrain}')
logger.info('Network architecture: \n' + summary_string(net, (1, img_size, img_size), device=str(device), batch_size=batch_size) + '\n')
# initialization of the Model
arae = ARAE(net, gamma, epsilon)
if model_path_to_load:
arae.load_model(model_path_to_load, map_location=device)
logger.info(f'Model Loaded from {model_path_to_load}' + '\n')
################################ Training ##################################
# add parameter info
logger.info(f'{Experiment_Name} epsilon : {epsilon}')
logger.info(f'{Experiment_Name} adversarial importance gamma : {gamma}')
logger.info(f'{Experiment_Name} number of epoch : {n_epoch}')
logger.info(f'{Experiment_Name} number of adversarial search epoch: {n_epoch_adv}')
logger.info(f'{Experiment_Name} learning rate : {lr}')
logger.info(f'{Experiment_Name} adversarial search learning rate : {lr_adv}')
method_adv = 'PGD' if use_PGD else 'FGSM'
logger.info(f'{Experiment_Name} adversarial search method : {method_adv}')
logger.info(f'{Experiment_Name} learning rate milestone : {lr_milestone}')
logger.info(f'{Experiment_Name} weight_decay : {weight_decay}')
logger.info(f'{Experiment_Name} optimizer : Adam')
logger.info(f'{Experiment_Name} batch_size {batch_size}')
logger.info(f'{Experiment_Name} number of dataloader worker : {n_jobs_dataloader}')
# train DROCC
arae.train(train_dataset, lr=lr, lr_adv=lr_adv, lr_milestone=lr_milestone,
weight_decay=weight_decay, n_epoch=n_epoch, n_epoch_adv=n_epoch_adv, use_PGD=use_PGD,
batch_size=batch_size, device=device, n_jobs_dataloader=n_jobs_dataloader,
print_batch_progress=print_batch_progress, valid_dataset=valid_dataset)
# validate DROCC
arae.validate(valid_dataset, device=device,
n_jobs_dataloader=n_jobs_dataloader,
print_batch_progress=print_batch_progress)
# test DROCC
arae.test(test_dataset, device=device,
n_jobs_dataloader=n_jobs_dataloader,
print_batch_progress=print_batch_progress)
# save results
arae.save_results(OUTPUT_PATH + f'results/{Experiment_Name}_results_{seed_i+1}.json')
logger.info('Test results saved at ' + OUTPUT_PATH + f'results/{Experiment_Name}_results_{seed_i+1}.json' + '\n')
# save model
arae.save_model(OUTPUT_PATH + f'model/{Experiment_Name}_model_{seed_i+1}.pt')
logger.info('Model saved at ' + OUTPUT_PATH + f'model/{Experiment_Name}_model_{seed_i+1}.pt')