sampler=sampler,
shuffle=False,
num_workers=4,
drop_last=True)
val_loader = torch.utils.data.DataLoader(val_set,
batch_size=batch_size,
shuffle=False,
num_workers=4,
drop_last=True)
resnet101 = torchvision.models.resnet101(pretrained=True)
resnet101.fc = nn.Linear(2048, 18)
if continue_train == True:
print('Loading network')
load_model(resnet101, checkpoint_dir, 5, was_parallel=was_parallel)
resnet101.cuda(cuda_id)
Loss_function = nn.CrossEntropyLoss()
if parallel:
resnet101 = torch.nn.DataParallel(resnet101, device_ids=cuda_id_list)
total_val_img = np.zeros(18)
total_correct_val_img = np.zeros(18)
#--------------------------training-----------------------
print('evaluating now')
resnet101.eval()
"""
#if using zhang's code
data_transforms = transforms.Compose([
transforms.Resize([224, 224]),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
dataloader = torch.utils.data.DataLoader(testData(transforms=data_transforms),
batch_size=4,
shuffle=False,
num_workers=4)
net = torchvision.models.resnet152(pretrained=True)
net.fc = nn.Linear(2048, 18)
load_model(net, checkpoint_dir, 'latest', was_parallel=was_parallel)
if use_gpu:
net = net.cuda(gpu_id)
net.eval()
id_list = []
id_no = 1
pred_list = []
for inputs in dataloader:
if use_gpu:
inputs = Variable(inputs.cuda(gpu_id))
else:
inputs = Variable(inputs)
def main():
p_dict = dict() # 所有参数
p_dict['args'] = args
### 加载数据
word_index_dict = json.load(open(args.word_index_json))
args.words = { v:k for k,v in word_index_dict.items() }
p_dict['word_index_dict'] = word_index_dict
num_classes = args.num_classes
p_dict['num_classes'] = num_classes
image_label_dict = json.load(open(args.image_label_json))
p_dict['image_label_dict'] = image_label_dict
# 划分数据集
test_filelist = sorted(glob(os.path.join(args.data_dir,'test/plots','*.png')))
trainval_filelist = sorted(glob(os.path.join(args.data_dir,'train/plots','*.png')))
# 两种输入size训练
# train_filelist1: 长宽比小于8:1的图片，经过padding后变成 64*512 的输入
# train_filelist2: 长宽比大于8:1的图片，经过padding,crop后变成 64*1024的输入
'''
train_filelist1, train_filelist2 = [],[]
# 黑名单，这些图片的label是有问题的
black_list = set(json.load(open(args.black_json))['black_list'])
image_hw_ratio_dict = json.load(open(args.image_hw_ratio_json))
for f in trainval_filelist:
image = f.split('/')[-1]
if image in black_list:
continue
r = image_hw_ratio_dict[image]
if r == 0:
train_filelist1.append(f)
else:
train_filelist2.append(f)
val_filelist = train_filelist1[-2048:]
train_filelist = train_filelist1[:-2048]
'''
# train_filelist1 = sorted(trainval_filelist)
# val_filelist = train_filelist1[-2048:]
# train_filelist = train_filelist1[:-2048]
val_filelist = test_filelist
train_filelist = trainval_filelist
# generated_list = glob(os.path.join(args.data_dir.replace('dataset', 'generated_images'), '*_image.png'))
# n_test = 4096
# pretrain_filelist = generated_list[:-n_test]
# preval_filelist = generated_list[-n_test:]
# train_filelist2 = train_filelist2
image_size = args.input_shape
test_dataset = dataloader.DataSet(
test_filelist,
image_label_dict,
num_classes,
# transform=train_transform,
args=args,
image_size=image_size,
phase='test')
test_loader = DataLoader(
dataset=test_dataset,
batch_size=1,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
train_dataset = dataloader.DataSet(
train_filelist,
image_label_dict,
num_classes,
image_size=image_size,
args=args,
phase='train')
train_loader = DataLoader(
dataset=train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
val_dataset = dataloader.DataSet(
val_filelist,
image_label_dict,
num_classes,
image_size=image_size,
args=args,
phase='val')
val_loader = DataLoader(
dataset=val_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
'''
pretrain_dataset = dataloader.DataSet(
pretrain_filelist,
image_label_dict,
num_classes,
image_size=image_size,
word_index_dict = word_index_dict,
args=args,
font_range=[8,32],
margin=10,
rotate_range=[-10., 10. ],
phase='pretrain')
pretrain_loader = DataLoader(
dataset=pretrain_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
preval_dataset = dataloader.DataSet(
preval_filelist,
image_label_dict,
num_classes,
image_size=image_size,
word_index_dict = word_index_dict,
args=args,
font_range=[8,32],
margin=10,
rotate_range=[-10., 10. ],
phase='pretrain')
preval_loader = DataLoader(
dataset=preval_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
'''
p_dict['train_loader'] = train_loader
p_dict['val_loader'] = val_loader
p_dict['test_loader'] = test_loader
# p_dict['pretrain_loader'] = pretrain_loader
# p_dict['train_loader'] = pretrain_loader
# p_dict['val_loader'] = preval_loader
# p_dict['test_loader'] = preval_loader
### 定义模型
cudnn.benchmark = True
framework.build_model(p_dict)
parameters = []
model_dict = p_dict['model_dict']
for model in model_dict.values():
for p in model.parameters():
parameters.append(p)
optimizer = torch.optim.Adam(parameters, lr=args.lr)
p_dict['optimizer'] = optimizer
# model = torch.nn.DataParallel(model).cuda()
# loss = Loss().cuda()
start_epoch = 0
# args.epoch = start_epoch
# print ('best_f1score' + str(best_f1score))
p_dict['epoch'] = 0
p_dict['best_metric'] = 0
### 加载预训练模型与参数
if os.path.exists(args.resume):
function.load_model(p_dict, args.resume)
### 训练及测试模型
if args.phase == 'test':
# 测试输出文字检测结果
test(p_dict)
elif args.phase == 'train':
best_f1score = 0
eval_mode = 'eval'
best_macc = 0
p_dict['best_metric'] = 0
for epoch in range(p_dict['epoch'] + 1, args.epochs):
p_dict['epoch'] = epoch
if best_f1score > 0.9:
args.lr = 0.0001
if best_f1score > 0.9:
args.hard_mining = 1
for param_group in optimizer.param_groups:
param_group['lr'] = args.lr
train_eval(p_dict, 'train')
train_eval(p_dict, 'val')
def main():
p_dict = dict() # All the parameters
p_dict['args'] = args
args.split_nn = args.split_num + args.split_nor * 3
args.vocab_size = args.split_nn * 145 + 1
print 'vocab_size', args.vocab_size
### load data
print 'read data ...'
patient_time_record_dict = py_op.myreadjson(
os.path.join(args.result_dir, 'patient_time_record_dict.json'))
patient_master_dict = py_op.myreadjson(
os.path.join(args.result_dir, 'patient_master_dict.json'))
patient_label_dict = py_op.myreadjson(
os.path.join(args.result_dir, 'patient_label_dict.json'))
patient_train = list(
json.load(open(os.path.join(args.file_dir, args.task, 'train.json'))))
patient_valid = list(
json.load(open(os.path.join(args.file_dir, args.task, 'val.json'))))
if len(patient_train) > len(patient_label_dict):
patients = patient_time_record_dict.keys()
patients = patient_label_dict.keys()
n = int(0.8 * len(patients))
patient_train = patients[:n]
patient_valid = patients[n:]
print 'data loading ...'
train_dataset = dataloader.DataSet(patient_train,
patient_time_record_dict,
patient_label_dict,
patient_master_dict,
args=args,
phase='train')
train_loader = DataLoader(dataset=train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=8,
pin_memory=True)
val_dataset = dataloader.DataSet(patient_valid,
patient_time_record_dict,
patient_label_dict,
patient_master_dict,
args=args,
phase='val')
val_loader = DataLoader(dataset=val_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=8,
pin_memory=True)
p_dict['train_loader'] = train_loader
p_dict['val_loader'] = val_loader
cudnn.benchmark = True
net = lstm.LSTM(args)
if args.gpu:
net = net.cuda()
p_dict['loss'] = loss.Loss().cuda()
else:
p_dict['loss'] = loss.Loss()
parameters = []
for p in net.parameters():
parameters.append(p)
optimizer = torch.optim.Adam(parameters, lr=args.lr)
p_dict['optimizer'] = optimizer
p_dict['model'] = net
start_epoch = 0
# args.epoch = start_epoch
# print ('best_f1score' + str(best_f1score))
p_dict['epoch'] = 0
p_dict['best_metric'] = [0, 0]
### resume pretrained model
if os.path.exists(args.resume):
print 'resume from model ' + args.resume
function.load_model(p_dict, args.resume)
print 'best_metric', p_dict['best_metric']
# return
if args.phase == 'train':
best_f1score = 0
for epoch in range(p_dict['epoch'] + 1, args.epochs):
p_dict['epoch'] = epoch
for param_group in optimizer.param_groups:
param_group['lr'] = args.lr
train_eval(p_dict, 'train')
train_eval(p_dict, 'val')
def main():
args.n_ehr = len(
json.load(
open(os.path.join(args.files_dir, 'demo_index_dict.json'),
'r'))) + 10
args.name_list = json.load(
open(os.path.join(args.files_dir, 'feature_list.json'), 'r'))[1:]
args.input_size = len(args.name_list)
files = sorted(glob(os.path.join(args.data_dir, 'resample_data/*.csv')))
data_splits = json.load(
open(os.path.join(args.files_dir, 'splits.json'), 'r'))
train_files = [
f for idx in [0, 1, 2, 3, 4, 5, 6] for f in data_splits[idx]
]
valid_files = [f for idx in [7] for f in data_splits[idx]]
test_files = [f for idx in [8, 9] for f in data_splits[idx]]
if args.phase == 'test':
train_phase, valid_phase, test_phase, train_shuffle = 'test', 'test', 'test', False
else:
train_phase, valid_phase, test_phase, train_shuffle = 'train', 'valid', 'test', True
train_dataset = data_loader.DataBowl(args, train_files, phase=train_phase)
valid_dataset = data_loader.DataBowl(args, valid_files, phase=valid_phase)
test_dataset = data_loader.DataBowl(args, test_files, phase=test_phase)
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=train_shuffle,
num_workers=args.workers,
pin_memory=True)
valid_loader = DataLoader(valid_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
test_loader = DataLoader(test_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
args.vocab_size = args.input_size + 2
if args.use_unstructure:
args.unstructure_size = len(
py_op.myreadjson(os.path.join(args.files_dir,
'vocab_list.json'))) + 10
# net = icnn.CNN(args)
# net = cnn.CNN(args)
net = lstm.LSTM(args)
# net = torch.nn.DataParallel(net)
# loss = myloss.Loss(0)
loss = myloss.MultiClassLoss(0)
net = _cuda(net, 0)
loss = _cuda(loss, 0)
best_metric = [0, 0]
start_epoch = 0
if args.resume:
p_dict = {'model': net}
function.load_model(p_dict, args.resume)
best_metric = p_dict['best_metric']
start_epoch = p_dict['epoch'] + 1
parameters_all = []
for p in net.parameters():
parameters_all.append(p)
optimizer = torch.optim.Adam(parameters_all, args.lr)
if args.phase == 'train':
for epoch in range(start_epoch, args.epochs):
print('start epoch :', epoch)
t0 = time.time()
train_eval(train_loader, net, loss, epoch, optimizer, best_metric)
t1 = time.time()
print('Running time:', t1 - t0)
best_metric = train_eval(valid_loader,
net,
loss,
epoch,
optimizer,
best_metric,
phase='valid')
print('best metric', best_metric)
elif args.phase == 'test':
train_eval(test_loader, net, loss, 0, optimizer, best_metric, 'test')
def main():
p_dict = dict() # All the parameters
p_dict['args'] = args
args.split_nn = 3 * 5
args.vocab_size = args.split_nn * 145 + 2
print('vocab_size', args.vocab_size)
### load data
print('read data ...')
if args.task == 'mortality':
patient_time_record_dict = py_op.myreadjson(
os.path.join(args.result_dir, 'patient_time_record_dict.json'))
patient_master_dict = py_op.myreadjson(
os.path.join(args.result_dir, 'patient_master_dict.json'))
patient_label_dict = py_op.myreadjson(
os.path.join(args.result_dir, 'patient_label_dict.json'))
if os.path.exists(os.path.join(args.result_dir, 'train.json')):
patient_train = list(
json.load(open(os.path.join(args.result_dir, 'train.json'))))
patient_valid = list(
json.load(open(os.path.join(args.result_dir, 'valid.json'))))
patient_test = list(
json.load(open(os.path.join(args.result_dir, 'test.json'))))
else:
patients = sorted(
set(patient_label_dict.keys()) & set(patient_time_record_dict)
& set(patient_master_dict))
print(len(patient_master_dict), len(patient_label_dict),
len(patient_time_record_dict))
print('There are {:d} patients.'.format(len(patients)))
n_train = int(0.7 * len(patients))
n_valid = int(0.2 * len(patients))
patient_train = patients[:n_train]
patient_valid = patients[n_train:n_train + n_valid]
patient_test = patients[n_train + n_valid:]
args.master_size = len(patient_master_dict[patients[0]])
elif args.task == 'sepsis':
patient_time_record_dict = py_op.myreadjson(
os.path.join(args.result_dir, 'sepsis_time_record_dict.json'))
patient_master_dict = py_op.myreadjson(
os.path.join(args.result_dir, 'patient_master_dict.json'))
patient_label_dict = py_op.myreadjson(
os.path.join(args.result_dir, 'sepsis_label_dict.json'))
sepsis_split = py_op.myreadjson(
os.path.join(args.result_dir, 'sepsis_split.json'))
print(sepsis_split.keys())
sepsis_split = sepsis_split[str(-args.last_time)]
patient_train = sepsis_split['train']
patient_valid = sepsis_split['valid']
print('train: {:d}'.format(len(patient_train)))
print('valid: {:d}'.format(len(patient_valid)))
print('data loading ...')
train_dataset = dataloader.DataSet(patient_train,
patient_time_record_dict,
patient_label_dict,
patient_master_dict,
args=args,
phase='train')
train_loader = DataLoader(dataset=train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=8,
pin_memory=True)
val_dataset = dataloader.DataSet(patient_valid,
patient_time_record_dict,
patient_label_dict,
patient_master_dict,
args=args,
phase='val')
val_loader = DataLoader(dataset=val_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=8,
pin_memory=True)
test_dataset = dataloader.DataSet(patient_test,
patient_time_record_dict,
patient_label_dict,
patient_master_dict,
args=args,
phase='val')
test_loader = DataLoader(dataset=test_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=8,
pin_memory=True)
p_dict['train_loader'] = train_loader
if args.phase == 'train':
p_dict['val_loader'] = val_loader
else:
p_dict['val_loader'] = test_loader
cudnn.benchmark = True
net = attention.Attention(args)
if args.gpu:
net = net.cuda()
p_dict['loss'] = loss.Loss().cuda()
else:
p_dict['loss'] = loss.Loss()
parameters = []
for p in net.parameters():
parameters.append(p)
optimizer = torch.optim.Adam(parameters, lr=args.lr)
p_dict['optimizer'] = optimizer
p_dict['model'] = net
start_epoch = 0
# args.epoch = start_epoch
# print ('best_f1score' + str(best_f1score))
p_dict['epoch'] = 0
p_dict['best_metric'] = [0, 0]
### resume pretrained model
if os.path.exists(args.resume):
print('resume from model ' + args.resume)
function.load_model(p_dict, args.resume)
print('best_metric', p_dict['best_metric'])
if args.phase == 'train':
best_f1score = 0
for epoch in range(p_dict['epoch'] + 1, args.epochs):
p_dict['epoch'] = epoch
for param_group in optimizer.param_groups:
param_group['lr'] = args.lr
train_eval(p_dict, 'train')
train_eval(p_dict, 'val')
log_info = '# task : {:s}; model: {:s} ; last_time: {:d} ; auc: {:3.4f} \n'.format(
args.task, args.model, args.last_time, p_dict['best_metric'][0])
with open('../result/log.txt', 'a') as f:
f.write(log_info)
else:
train_eval(p_dict, 'test')
weight=torch.DoubleTensor(weight)
sampler = torch.utils.data.sampler.WeightedRandomSampler(weight, len(weight))
train_loader = torch.utils.data.DataLoader(train_set, batch_size=batch_size,sampler=sampler,shuffle=False, num_workers=4, drop_last=True)
val_loader = torch.utils.data.DataLoader(val_set, batch_size=batch_size, shuffle=False, num_workers=4, drop_last=True)
resnet101 = torchvision.models.resnet152(pretrained=True)
resnet101.fc=nn.Linear(2048,18)
if continue_train==True:
print('Loading network')
load_model(resnet101,checkpoint_dir,'latest',was_parallel=was_parallel)
loss_list=list(np.loadtxt(os.path.join(checkpoint_dir,'loss_history_train.txt')))
acc_list = list(np.loadtxt(os.path.join(checkpoint_dir, 'acc_history_train.txt')))
loss_list_val = list(np.loadtxt(os.path.join(checkpoint_dir,'loss_history_val.txt')))
acc_list_val = list(np.loadtxt(os.path.join(checkpoint_dir, 'acc_history_val.txt')))
lowest_loss=min(loss_list_val)
start_epoch=len(loss_list)
else:
start_epoch=0
loss_list = []
loss_list_val = []
acc_list=[]
acc_list_val=[]
lowest_loss=9999
resnet101.cuda(cuda_id)
def main():
assert args.dataset in ['DACMI', 'MIMIC']
if args.dataset == 'MIMIC':
args.n_ehr = len(py_op.myreadjson(os.path.join(args.data_dir, args.dataset, 'ehr_list.json')))
args.name_list = py_op.myreadjson(os.path.join(args.file_dir, args.dataset+'_feature_list.json'))[1:]
args.output_size = len(args.name_list)
files = sorted(glob(os.path.join(args.data_dir, args.dataset, 'train_with_missing/*.csv')))
data_splits = py_op.myreadjson(os.path.join(args.file_dir, args.dataset + '_splits.json'))
train_files = [f for idx in [0, 1, 2, 3, 4, 5, 6] for f in data_splits[idx]]
valid_files = [f for idx in [7] for f in data_splits[idx]]
test_files = [f for idx in [8, 9] for f in data_splits[idx]]
if args.phase == 'test':
train_phase, valid_phase, test_phase, train_shuffle = 'test', 'test', 'test', False
else:
train_phase, valid_phase, test_phase, train_shuffle = 'train', 'valid', 'test', True
train_dataset = data_loader.DataBowl(args, train_files, phase=train_phase)
valid_dataset = data_loader.DataBowl(args, valid_files, phase=valid_phase)
test_dataset = data_loader.DataBowl(args, test_files, phase=test_phase)
train_loader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=train_shuffle, num_workers=args.workers, pin_memory=True)
valid_loader = DataLoader(valid_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers, pin_memory=True)
test_loader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.workers, pin_memory=True)
args.vocab_size = (args.output_size + 2) * (1 + args.split_num) + 5
if args.model == 'tame':
net = tame.AutoEncoder(args)
loss = myloss.MSELoss(args)
net = _cuda(net, 0)
loss = _cuda(loss, 0)
best_metric= [0,0]
start_epoch = 0
if args.resume:
p_dict = {'model': net}
function.load_model(p_dict, args.resume)
best_metric = p_dict['best_metric']
start_epoch = p_dict['epoch'] + 1
parameters_all = []
for p in net.parameters():
parameters_all.append(p)
optimizer = torch.optim.Adam(parameters_all, args.lr)
if args.phase == 'train':
for epoch in range(start_epoch, args.epochs):
print('start epoch :', epoch)
train_eval(train_loader, net, loss, epoch, optimizer, best_metric)
best_metric = train_eval(valid_loader, net, loss, epoch, optimizer, best_metric, phase='valid')
print 'best metric', best_metric
elif args.phase == 'test':
folder = os.path.join(args.result_dir, args.dataset, 'imputation_result')
os.system('rm -r ' + folder)
os.system('mkdir ' + folder)
train_eval(train_loader, net, loss, 0, optimizer, best_metric, 'test')
train_eval(valid_loader, net, loss, 0, optimizer, best_metric, 'test')
train_eval(test_loader, net, loss, 0, optimizer, best_metric, 'test')
