def test_count_by_organ(loaders, models, cls_limit, output_path):
fconv = open(os.path.join(output_path, '{}-count.csv'.format(now)),
'w',
newline="")
w = csv.writer(fconv, delimiter=',')
w.writerow(['id', 'count'])
print('Start testing ...')
colonset.setmode("image")
breastset.setmode("image")
prostateset.setmode("image")
ids = np.array([])
counts = np.array([])
for i in range(3):
organ_ids, _, organ_counts = inference_image(loaders[i],
models[i],
device,
mode='test',
cls_limit=cls_limit,
return_id=True)
for j in range(len(organ_ids)):
print("id: {}, organ: {}, pred: {}".format(int(organ_ids[j]),
organs[i],
int(organ_counts[j])))
ids = np.concatenate((ids, organ_ids))
counts = np.concatenate((counts, organ_counts))
counts = counts[np.argsort(ids.astype(int))]
for i, y in enumerate(counts, start=1):
w.writerow([i, y])
fconv.close()
def test_ensemble(loader, models, epoch, cls_limit, output_path):
fconv = open(os.path.join(output_path,
'{}-count-e{}.csv'.format(now, epoch)),
'w',
newline="")
w = csv.writer(fconv, delimiter=',')
w.writerow(['id', 'count', 'organ'])
outputs = []
for i, m in enumerate(models):
print('Testing {}/{}...'.format(i + 1, len(models)))
testset.setmode("image")
outputs.append(
inference_image(loader,
m,
device,
mode='test',
cls_limit=cls_limit)[1])
# take average as the final result
output = np.asarray(outputs).mean(axis=0).round().astype(int)
for i, y in enumerate(output, start=1):
w.writerow([i, y, testset.organs[i - 1]])
fconv.close()
def test_count(loader, model, epoch, cls_limit, output_path):
fconv = open(os.path.join(output_path,
'{}-count-e{}.csv'.format(now, epoch)),
'w',
newline="")
w = csv.writer(fconv, delimiter=',')
w.writerow(['id', 'count', 'organ'])
print('Start testing ...')
testset.setmode("image")
counts = inference_image(loader,
model,
device,
mode='test',
cls_limit=cls_limit)[1]
for i, y in enumerate(counts, start=1):
w.writerow([i, y, testset.organs[i - 1]])
fconv.close()
def test_tile(loader, model, epoch, reg_limit, reg_loader, output_path):
"""
:param testset: 测试数据集
:param batch_size: Dataloader 打包的小 batch 大小
:param workers: Dataloader 使用的进程数
:param model: 网络模型
:param output_path: 保存模型文件的目录
"""
# 热图中各个 tile 的信息保存在 output_path/<timestamp>-pred-e<epoch>-p<tilesize>-i<interval>-c<threshold>.csv
fconv = open(os.path.join(output_path, '{}-pred-e{}-p{}-i{}-c{}.csv'.format(
now, epoch, args.tile_size, args.interval, args.threshold)), 'w', newline="")
w = csv.writer(fconv)
w.writerow(['tile_size', '{}'.format(testset.tile_size)])
w.writerow(['interval', '{}'.format(testset.interval)])
w.writerow(['idx', 'grid', 'prob'])
fconv.close()
def rank(testset, probs):
"""按概率对 tile 排序，便于与置信度进行比较。"""
groups = np.array(testset.tileIDX)
tiles = np.array(testset.tiles_grid)
order = np.lexsort((probs, groups))
groups = groups[order]
probs = probs[order]
tiles = tiles[order]
# index = np.empty(len(groups), 'bool')
# index[-topk:] = True
# index[:-topk] = groups[topk:] != groups[:-topk]
index = [prob > args.threshold for prob in probs]
return tiles[index], probs[index], groups[index]
print('Start testing ...')
testset.setmode("tile")
probs = inference_tiles(loader, model, device, mode='test')
tiles, probs, groups = rank(testset, probs)
# clear artifact images
if reg_limit:
reg_testset.setmode("image")
model.setmode("image")
with open(os.path.join(output_path, '{}-count-e{}.csv'.format(now, epoch)),
'w', newline="") as f:
w = csv.writer(f, delimiter=',')
w.writerow(['id', 'count', 'organ'])
counts = inference_image(reg_loader, model, device, mode='test')[1]
for i, y in enumerate(counts, start=1):
w.writerow([i, y, reg_testset.organs[i - 1]])
img_indices = np.select([counts != 0], [counts]).nonzero()[0]
indices = [i for i, g in enumerate(groups) if g in img_indices]
tiles = tiles[indices]
probs = probs[indices]
groups = groups[indices]
# 生成热图
fconv = open(os.path.join(output_path, '{}-pred-e{}-p{}-i{}-c{}.csv'.format(
now, epoch, args.tile_size, args.interval, args.threshold)), 'a', newline="")
heatmap(testset, tiles, probs, groups, fconv, output_path)
fconv.close()
from inference import inference_image
# clear artifact images
limit_set = LystoTestset(os.path.join(training_data_path,
"training.h5"),
num_of_imgs=100 if args.debug else 0)
limit_loader = DataLoader(limit_set,
batch_size=args.image_batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
limit_set.setmode("image")
model.setmode("image")
counts = inference_image(limit_loader, model, device, mode='test')[1]
img_indices = np.select([counts != 0], [counts]).nonzero()[0]
indices = [i for i, g in enumerate(groups) if g in img_indices]
tiles = tiles[indices]
groups = groups[indices]
pseudo_masks = generate_masks(dataset,
tiles,
groups,
preprocess=args.preprocess,
output_path=os.path.join(
training_data_path,
args.pseudomask_dir))
trainset = Maskset(os.path.join(training_data_path, "training.h5"),
pseudo_masks,