def default_classical_scorings(task="predict"):
if task == 'predict':
scorings = (mm.Accuracy(tensor=False),
mm.BalancedAccuracy(tensor=False),
mm.F1Score(average='macro', tensor=False),
mm.Precision(average='macro', tensor=False),
mm.Recall(average='macro', tensor=False),
mm.ROCAUC(average='macro', tensor=False))
else:
scorings = (mm.CIndex(tensor=False, hazard=True), )
return scorings
def main(args):
if args.cuda and not torch.cuda.is_available():
raise ValueError("GPUs are not available, please run at cpu mode")
# init
data = tileData(args.root, args.img_rows, args.img_cols)
evaluators = [
metrics.OAAcc(),
metrics.Precision(),
metrics.Recall(),
metrics.F1Score(),
metrics.Kappa(),
metrics.Jaccard()
]
# prediction
for checkpoint in args.checkpoints:
model, is_multi = load_checkpoint(checkpoint, args.cuda)
performs = [[] for i in range(len(evaluators))]
for idx in range(len(data)):
print("Handling {} by {} \r".format(data.files[idx], checkpoint))
x, y, shapes = data.slice_by_id(idx)
# generate prediction
with torch.set_grad_enabled(False):
for step in range(0, x.shape[0], args.batch_size):
x_batch = x[step:step + args.batch_size]
y_batch = y[step:step + args.batch_size]
if args.cuda:
x_batch = x_batch.cuda()
y_batch = y_batch.cuda()
if is_multi:
y_pred = model(x_batch)[0].detach()
else:
y_pred = model(x_batch).detach()
# get performance
for i, evaluator in enumerate(evaluators):
performs[i].append(
evaluator(y_pred, y_batch)[0].item())
performs = [(sum(p) / len(p)) for p in performs]
performs = pd.DataFrame(
[[time.strftime("%h_%d"), checkpoint] + performs],
columns=['time', 'checkpoint'] + [repr(x) for x in evaluators])
# save performance
log_path = os.path.join(Result_DIR, "patchPerforms.csv")
if os.path.exists(log_path):
perform = pd.read_csv(log_path)
else:
perform = pd.DataFrame([])
perform = perform.append(performs, ignore_index=True)
perform.to_csv(log_path, index=False, float_format="%.3f")
def main():
warnings.filterwarnings('ignore')
# config
parser = argparse.ArgumentParser()
parser.add_argument('-s',
'--save',
default='./save',
help='保存的文件夹路径，如果有重名，会在其后加-来区别')
parser.add_argument('-is',
'--image_size',
default=224,
type=int,
help='patch会被resize到多大，默认时224 x 224')
parser.add_argument('-vts',
'--valid_test_size',
default=(0.1, 0.1),
type=float,
nargs=2,
help='训练集和测试集的大小，默认时0.1, 0.1')
parser.add_argument('-bs',
'--batch_size',
default=32,
type=int,
help='batch size，默认时32')
parser.add_argument('-nw',
'--num_workers',
default=12,
type=int,
help='多进程数目，默认时12')
parser.add_argument('-lr',
'--learning_rate',
default=0.0001,
type=float,
help='学习率大小，默认时0.0001')
parser.add_argument('-e',
'--epoch',
default=10,
type=int,
help='epoch 数量，默认是10')
parser.add_argument('--reduction',
default='mean',
help='聚合同一bag的instances时的聚合方式，默认时mean')
parser.add_argument('--multipler',
default=2.0,
type=float,
help="为了平衡pos和neg，在weight再乘以一个大于1的数，默认是2.0")
args = parser.parse_args()
save = args.save
image_size = (args.image_size, args.image_size)
valid_size, test_size = args.valid_test_size
batch_size = args.batch_size
num_workers = args.num_workers
lr = args.learning_rate
epoch = args.epoch
reduction = args.reduction
multipler = args.multipler
# ----- 读取数据 -----
neg_dir = './DATA/TCT/negative'
pos_dir = './DATA/TCT/positive'
dat = MilData.from2dir(neg_dir, pos_dir)
train_transfer = transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.Resize(image_size),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
test_transfer = transforms.Compose([
transforms.Resize(image_size),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
train_dat, valid_dat, test_dat = dat.split_by_bag(
test_size,
valid_size,
train_transfer=train_transfer,
valid_transfer=test_transfer,
test_transfer=test_transfer)
dataloaders = {
'train':
data.DataLoader(train_dat,
batch_size=batch_size,
num_workers=num_workers,
shuffle=True),
'valid':
data.DataLoader(
valid_dat,
batch_size=batch_size,
num_workers=num_workers,
),
'test':
data.DataLoader(
test_dat,
batch_size=batch_size,
num_workers=num_workers,
)
}
# ----- 构建网络和优化器 -----
net = NormalCnn()
criterion = nn.BCELoss(reduction='none')
optimizer = optim.Adam(net.parameters(), lr=lr)
scorings = [
mm.Loss(),
mm.Recall(reduction=reduction),
mm.ROCAUC(reduction=reduction),
mm.BalancedAccuracy(reduction=reduction),
mm.F1Score(reduction=reduction),
mm.Precision(reduction=reduction),
mm.Accuracy(reduction=reduction)
]
# ----- 训练网络 -----
try:
net, hist, weighter = train(net,
criterion,
optimizer,
dataloaders,
epoch=epoch,
metrics=scorings,
weighter_multipler=multipler)
test_hist = evaluate(net, dataloaders['test'], criterion, scorings)
except Exception as e:
import ipdb
ipdb.set_trace() # XXX BREAKPOINT
# 保存结果
dirname = check_update_dirname(save)
torch.save(net.state_dict(), os.path.join(dirname, 'model.pth'))
torch.save(weighter, os.path.join(dirname, 'weigher.pth'))
pd.DataFrame(hist).to_csv(os.path.join(dirname, 'train.csv'))
with open(os.path.join(dirname, 'config.json'), 'w') as f:
json.dump(args.__dict__, f)
with open(os.path.join(dirname, 'test.json'), 'w') as f:
json.dump(test_hist, f)
def main(args):
if args.cuda and not torch.cuda.is_available():
raise ValueError("GPUs are not available, please run at cpu mode")
# init
data = tileData(args.root, args.img_rows, args.img_cols)
evaluators = [
metrics.OAAcc(),
metrics.Precision(),
metrics.Recall(),
metrics.F1Score(),
metrics.Kappa(),
metrics.Jaccard()
]
# prediction
for checkpoint in args.checkpoints:
model, is_multi = load_checkpoint(checkpoint, args.cuda)
Save_DIR = os.path.join(Result_DIR, "area", checkpoint.split("_")[0])
if not os.path.exists(Save_DIR):
os.makedirs(Save_DIR)
performs = [[] for i in range(len(evaluators))]
for idx in range(len(data)):
print("Handling {} by {} \r".format(data.files[idx], checkpoint))
x, y, shapes = data.slice_by_id(idx)
# get prediction
y_pred = []
with torch.set_grad_enabled(False):
for step in range(0, x.shape[0], args.batch_size):
x_batch = x[step:step + args.batch_size]
if args.cuda:
x_batch = x_batch.cuda()
# generate prediction
if is_multi:
y_pred.append(model(x_batch)[0].detach())
else:
y_pred.append(model(x_batch).detach())
y_pred = torch.cat(y_pred, 0)
if args.cuda:
y_pred = y_pred.cpu()
assert y_pred.shape[0] == x.shape[
0], "All data should be iterated."
del x
pred_img = vision.slices_to_img(
vision.ytensor_to_slices(y_pred, data.cmap), shapes)
# y_img = vision.slices_to_img(
# vision.ytensor_to_slices(y, data.cmap), data.shapes)
# merge slices into image & save result image
imsave(os.path.join(Save_DIR, data.files[idx]),
pred_img,
compress=6)
# N tensor 2 one
# pred_tensor = vision.tensors_to_tensor(y_pred, shapes)
# y_tensor = vision.tensors_to_tensor(y, shapes)
# get performance
for idx, evaluator in enumerate(evaluators):
# a = evaluator(pred_tensor, y_tensor)[0].item()
# b = evaluator(y_pred, y)[0].item()
# print("{} => One : {} ; N : {}".format(repr(evaluator), a, b))
performs[idx].append(evaluator(y_pred, y)[0].item())
performs = [(sum(p) / len(p)) for p in performs]
performs = pd.DataFrame(
[[time.strftime("%h_%d"), checkpoint] + performs],
columns=['time', 'checkpoint'] + [repr(x) for x in evaluators])
# save performance
log_path = os.path.join(Result_DIR, "areaPerforms.csv")
if os.path.exists(log_path):
perform = pd.read_csv(log_path)
else:
perform = pd.DataFrame([])
perform = perform.append(performs, ignore_index=True)
perform.to_csv(log_path, index=False, float_format="%.3f")
def main(args):
if args.cuda and not torch.cuda.is_available():
raise ValueError("GPUs are not available, please run at cpu mode")
evaluators = [
metrics.OAAcc(),
metrics.Precision(),
metrics.Recall(),
metrics.F1Score(),
metrics.Kappa(),
metrics.Jaccard()
]
for checkpoint in args.checkpoints:
print("Handling by {} ...\r".format(checkpoint))
Save_DIR = os.path.join(Result_DIR, 'single', checkpoint.split("_")[0])
if not os.path.exists(Save_DIR):
os.makedirs(Save_DIR)
# initialize datasets
infos = checkpoint.split('_')[0].split('-')
_, valset = load_dataset(infos[2], "IM")
print("Testing with {}-Dataset: {} examples".format(
infos[2], len(valset)))
# Load checkpoint
model, is_multi = load_checkpoint(checkpoint, args.cuda)
# load data
data_loader = DataLoader(
valset,
1,
num_workers=4,
shuffle=False,
pin_memory=True,
)
performs = [[] for i in range(len(evaluators))]
imgsets = []
with torch.set_grad_enabled(False):
for idx, sample in enumerate(data_loader):
# get tensors from sample
x = sample["src"]
y = sample["tar"]
if args.cuda:
x = x.cuda()
y = y.cuda()
if is_multi:
gen_y = model(x)[0]
else:
gen_y = model(x)
# get performance
for i, evaluator in enumerate(evaluators):
performs[i].append(
evaluator(gen_y.detach(), y.detach())[0].item())
if args.cuda:
x = x.detach().cpu()
y = x.detach().cpu()
gen_y = gen_y.detach().cpu()
x = x.numpy()[0].transpose((1, 2, 0))
y = y.numpy()[0].transpose((1, 2, 0))
gen_y = gen_y.numpy()[0].transpose((1, 2, 0))
x_img = valset._src2img(x, whitespace=False)
y_img = valset._tar2img(y, whitespace=False)
gen_img = valset._tar2img(gen_y, whitespace=False)
canny_x = vision.canny_edge(x_img)
canny_y = vision.canny_edge(y_img)
canny_gen = vision.canny_edge(gen_img)
# mask_pair = vision.pair_to_rgb(gen_img, y_img, args.color)
canny_pair = vision.pair_to_rgb(canny_y,
canny_x,
args.color,
use_dilation=True,
disk_value=args.disk)
edge_pair = vision.pair_to_rgb(canny_gen,
canny_y,
args.color,
use_dilation=True,
disk_value=args.disk)
imgsets.append([
vision.add_barrier(x_img, args.spaces),
vision.add_barrier(canny_pair, args.spaces),
# vision.add_barrier(mask_pair, args.spaces),
vision.add_barrier(edge_pair, args.spaces),
])
if len(imgsets) >= args.disp_cols * args.gen_nb:
break
# visualization
for i in range(args.gen_nb):
imgset = []
for j in range(args.disp_cols):
imgset.append(
np.concatenate(imgsets[i * args.disp_cols + j],
axis=0))
vis_img = np.concatenate(imgset, axis=1)
name = "{}_canny_segmap_edge_{}.png".format(
checkpoint.split('_')[0], i)
imsave(os.path.join(Save_DIR, name),
vision.add_barrier(vis_img, args.spaces))
print("Saving {} ...".format(name))
def main():
# ----- 根据data来读取不同的数据和不同的loss、metrics -----
if config.args.data == 'brca':
rna = RnaData.predicted_data(config.brca_cli, config.brca_rna,
{'PAM50Call_RNAseq': 'pam50'})
rna.transform(tf.LabelMapper(config.brca_label_mapper))
out_shape = len(config.brca_label_mapper)
criterion = nn.CrossEntropyLoss()
scorings = (mm.Loss(), mm.Accuracy(), mm.BalancedAccuracy(),
mm.F1Score(average='macro'), mm.Precision(average='macro'),
mm.Recall(average='macro'), mm.ROCAUC(average='macro'))
elif config.args.data == 'survival':
if os.path.exists('./DATA/temp_pan.pth'):
rna = RnaData.load('./DATA/temp_pan.pth')
else:
rna = RnaData.survival_data(config.pan_cli, config.pan_rna,
'_OS_IND', '_OS')
out_shape = 1
if config.args.loss_type == 'cox':
criterion = NegativeLogLikelihood()
elif config.args.loss_type == 'svm':
criterion = SvmLoss(rank_ratio=config.args.svm_rankratio)
scorings = (mm.Loss(), mm.CIndex())
rna.transform(tf.ZeroFilterCol(0.8))
rna.transform(tf.MeanFilterCol(1))
rna.transform(tf.StdFilterCol(0.5))
norm = tf.Normalization()
rna.transform(norm)
# ----- 构建网络和优化器 -----
inpt_shape = rna.X.shape[1]
if config.args.net_type == 'mlp':
net = MLP(inpt_shape, out_shape, config.args.hidden_num,
config.args.block_num).cuda()
elif config.args.net_type == 'atten':
net = SelfAttentionNet(inpt_shape, out_shape, config.args.hidden_num,
config.args.bottle_num, config.args.block_num,
config.args.no_res, config.act,
config.args.no_head, config.args.no_bottle,
config.args.no_atten,
config.args.dropout_rate).cuda()
elif config.args.net_type == 'resnet':
net = ResidualNet(inpt_shape, out_shape, config.args.hidden_num,
config.args.bottle_num,
config.args.block_num).cuda()
# ----- 训练网络，cross validation -----
split_iterator = rna.split_cv(config.args.test_size,
config.args.cross_valid)
train_hists = []
test_hists = []
for split_index, (train_rna, test_rna) in enumerate(split_iterator):
print('##### save: %s, split: %d #####' %
(config.args.save, split_index))
# 从train中再分出一部分用作验证集，决定停止
train_rna, valid_rna = train_rna.split(0.1)
dats = {
'train': train_rna.to_torchdat(),
'valid': valid_rna.to_torchdat(),
}
dataloaders = {
k: data.DataLoader(v, batch_size=config.args.batch_size)
for k, v in dats.items()
}
test_dataloader = data.DataLoader(test_rna.to_torchdat(),
batch_size=config.args.batch_size)
# 网络训练前都进行一次参数重置，避免之前的训练的影响
net.reset_parameters()
# train
optimizer = optim.Adamax(net.parameters(),
lr=config.args.learning_rate)
lrs = config.lrs(optimizer)
net, hist = train(
net,
criterion,
optimizer,
dataloaders,
epoch=config.args.epoch,
metrics=scorings,
l2=config.args.l2,
standard_metric_index=config.args.standard_metric_index,
scheduler=lrs)
# test
test_res = evaluate(net, criterion, test_dataloader, metrics=scorings)
# 将多次训练的结果保存到一个df中
hist = pd.DataFrame(hist)
hist['split_index'] = split_index
train_hists.append(hist)
# 保存多次test的结果
test_res['split_index'] = split_index
test_hists.append(test_res)
# 每个split训练的模型保存为一个文件
torch.save(net.state_dict(),
os.path.join(config.save_dir, 'model%d.pth' % split_index))
# 保存train的结果
train_hists = pd.concat(train_hists)
train_hists.to_csv(os.path.join(config.save_dir, 'train.csv'))
# 保存test的结果
test_hists = pd.DataFrame(test_hists)
test_hists.to_csv(os.path.join(config.save_dir, 'test.csv'))