def test_soft_nms_device_and_dtypes_cpu():
"""
CommandLine:
xdoctest -m tests/test_soft_nms.py test_soft_nms_device_and_dtypes_cpu
"""
iou_thr = 0.7
base_dets = np.array([[49.1, 32.4, 51.0, 35.9, 0.9],
[49.3, 32.9, 51.0, 35.3, 0.9],
[35.3, 11.5, 39.9, 14.5, 0.4],
[35.2, 11.7, 39.7, 15.7, 0.3]])
# CPU can handle float32 and float64
dets = base_dets.astype(np.float32)
new_dets, inds = soft_nms(dets, iou_thr)
assert dets.dtype == new_dets.dtype
assert len(inds) == len(new_dets) == 4
dets = torch.FloatTensor(base_dets)
new_dets, inds = soft_nms(dets, iou_thr)
assert dets.dtype == new_dets.dtype
assert len(inds) == len(new_dets) == 4
dets = base_dets.astype(np.float64)
new_dets, inds = soft_nms(dets, iou_thr)
assert dets.dtype == new_dets.dtype
assert len(inds) == len(new_dets) == 4
dets = torch.DoubleTensor(base_dets)
new_dets, inds = soft_nms(dets, iou_thr)
assert dets.dtype == new_dets.dtype
assert len(inds) == len(new_dets) == 4
print("cur test passed!")
def get_det_score(score, json_dets, dataset, model_configs, config_idx,
**kwargs):
results = []
for file_index, file_dets in enumerate(json_dets):
all_dets = []
config_dets = file_dets[config_idx]
config = model_configs[config_idx]
dets = []
for index, det in enumerate(config_dets):
if len(det) == 0:
det = np.empty((0, 5))
else:
det = np.array(det)
det = det[det[:, 4] >= score]
det[:, 4] *= config["{}_weight".format(CLASSES[index])]
dets.append(det)
for index, det in enumerate(dets):
if len(dets[index]) == 0:
continue
dets[index] = soft_nms(dets[index], 0.5)[0]
results.append(dets)
_, stats, aps = dataset.evaluate(results, 'bbox', **kwargs)
return stats[0], aps
def debugging_test():
from mmdet.ops.nms.nms_wrapper import soft_nms
import numpy as np
a = soft_nms(np.array([[0.5, 0.5, 0.75, 0.75, 0.6], [0.5, 0.5, 1, 1, 0.4]],
dtype=np.float32),
iou_thr=0.5,
min_score=0.05)
print(a)
def main():
args = parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
config_file = open(args.config)
configs = json.load(config_file)
config_file.close()
infer_results = []
dataset = MyDataset(args.data_dir)
file_count = len(dataset)
dataloader = DataLoader(dataset=dataset, batch_size=1, shuffle=False, num_workers=1)
models = []
model_configs = []
for config in configs:
if not config['use']:
continue
print('config = {}, model = {}'.format(config['config'], config['model']))
model = init_detector(config['config'], config['model'], device=torch.device('cuda', 0))
models.append(model)
model_configs.append(config)
start = time.time()
for image_index, data in enumerate(dataloader):
im, filename = data[0][0].numpy(), data[1][0]
image_id = filename.split('.')[0]
all_dets = []
for config, model in zip(model_configs, models):
dets = inference_detector(model, im)
for index, det in enumerate(dets):
dets[index] = det[det[:,4]>=config[CLASSES[index]]]
dets[index][:,4] *= config["{}_weight".format(CLASSES[index])]
if len(all_dets) == 0:
all_dets = dets
else:
for det_index, (all_det, det) in enumerate(zip(all_dets, dets)):
all_dets[det_index] = np.concatenate((all_det, det), axis=0)
for index, all_det in enumerate(all_dets):
all_dets[index] = soft_nms(all_dets[index], 0.5, min_score=0.001)[0][0:100]
for index, bboxes in enumerate(all_dets):
category = CLASSES[index]
for bbox in bboxes:
infer_results.append((category, image_id+'.xml', str(bbox[4]), str(bbox[0]), str(bbox[1]), str(bbox[2]), str(bbox[3])))
end = time.time()
print('\r', '{}/{}, use time:{}'.format(image_index, file_count, end -start), end='', flush=True)
start = end
csvfile = open('../shengna_results.csv', 'w')
writer = csv.writer(csvfile)
writer.writerow(['name', 'image_id', 'confidence', 'xmin', 'ymin', 'xmax', 'ymax'])
writer.writerows(infer_results)
csvfile.close()
def main():
args = parse_args()
config_file = open(args.config)
configs = json.load(config_file)
config_file.close()
model_configs = []
for config in configs:
if config['use']:
print('config = {}, model = {}'.format(config['config'],
config['model']))
model_configs.append(config)
g_config = config['config']
cfg = mmcv.Config.fromfile(g_config)
coco_dataset = build_dataset(cfg.data.test)
with open('shengna_config/assemble_results.json') as fp:
results = []
json_dets = json.load(fp)
for file_index, file_dets in enumerate(json_dets):
all_dets = []
for config_index, config_dets in enumerate(file_dets):
config = model_configs[config_index]
if config['use'] == False:
continue
dets = []
for index, det in enumerate(config_dets):
if len(det) == 0:
det = np.empty((0, 5))
else:
det = np.array(det)
det = det[det[:, 4] >= config[CLASSES[index]]]
det[:, 4] *= config["{}_weight".format(CLASSES[index])]
dets.append(det)
if len(all_dets) == 0:
all_dets = dets
else:
for det_index, (all_det,
det) in enumerate(zip(all_dets, dets)):
all_dets[det_index] = np.concatenate((all_det, det),
axis=0)
for index, all_det in enumerate(all_dets):
#all_dets[index] = soft_nms(all_dets[index], 0.5, min_score=0.001)[0]
all_dets[index] = soft_nms(all_dets[index],
0.5,
min_score=0.001)[0][0:100]
results.append(all_dets)
kwargs = {} if args.options is None else args.options
coco_dataset.evaluate(results, 'bbox', **kwargs)
def get_results(config_weights, json_dets, model_configs):
print("\n\n\n\n current weights = {}".format(config_weights))
results = []
for file_index, file_dets in enumerate(json_dets):
all_dets = []
for config_index, config_dets in enumerate(file_dets):
config = model_configs[config_index]
if config['use'] == False:
continue
dets = []
for index, det in enumerate(config_dets):
if len(det) == 0:
det = np.empty((0, 5))
else:
det = np.array(det)
weight = config_weights[config_index]
if weight == 0:
det = np.empty((0, 5))
else:
det = det[det[:, 4] >= config[CLASSES[index]]]
det[:, 4] *= weight
dets.append(det)
if len(all_dets) == 0:
all_dets = dets
else:
for det_index, (all_det, det) in enumerate(zip(all_dets,
dets)):
all_dets[det_index] = np.concatenate((all_det, det),
axis=0)
for index, all_det in enumerate(all_dets):
if len(all_det) == 0:
continue
all_dets[index] = soft_nms(all_dets[index], 0.5)[0][0:100]
results.append(all_dets)
return results
for filename in tqdm.tqdm(deal_nms['image_id'].unique()):
img_id = name2id[filename]
base_dets = deal_nms[deal_nms['image_id'] == filename]
#lis[jl] = filename
for defect_label, value in underwater_classes.items(): # 查找标签，对应标签进行融合
base_dets_1 = base_dets[base_dets['name'] == defect_label]
dets = torch.FloatTensor(
np.array(
base_dets_1[['xmin', 'ymin', 'xmax', 'ymax',
'confidence']])).cuda()
iou_thr = 0.5
# surpressed, inds = nms(dets, iou_thr)
surpressed, inds = soft_nms(dets, iou_thr)
for press in surpressed:
x1, y1, x2, y2, score = press.cpu().numpy()[:]
x1, y1, x2, y2 = round(float(x1), 2), round(float(y1), 2), round(
float(x2), 2), round(float(y2), 2) # save 0.00
result.append({
'image_id': img_id,
'bbox': [x1, y1, x2 - x1, y2 - y1],
'category_id': int(value),
'score': float(score)
})
#for k in lis.keys():
# lis_img = {}
# lis_img['image_id'] = lis[k]
def main():
args = parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
config_file = open(args.config)
configs = json.load(config_file)
config_file.close()
results = []
json_results = []
image_dir = os.path.join(args.data_dir, 'JPEGImages')
anno_dir = os.path.join(args.data_dir, 'Annotations')
dataset = MyDataset(args.file_path, image_dir)
file_count = len(dataset)
dataloader = DataLoader(dataset=dataset,
batch_size=1,
shuffle=False,
num_workers=1)
models = []
model_configs = []
for config in configs:
if not config['use']:
continue
print('config = {}, model = {}'.format(config['config'],
config['model']))
model = init_detector(config['config'],
config['model'],
device=torch.device('cuda', 0))
models.append(model)
model_configs.append(config)
g_config = config['config']
print(g_config)
cfg = mmcv.Config.fromfile(g_config)
coco_dataset = build_dataset(cfg.data.test)
category_counts = {}
start = time.time()
for image_index, data in enumerate(dataloader):
im, filename = data[0][0].numpy(), data[1][0]
all_dets = []
all_config_dets = []
for config, model in zip(model_configs, models):
dets = inference_detector(model, im)
json_dets = []
for det in dets:
json_dets.append(det.tolist())
all_config_dets.append(json_dets)
for index, det in enumerate(dets):
dets[index] = det[det[:, 4] >= config[CLASSES[index]]]
dets[index][:, 4] *= config["{}_weight".format(CLASSES[index])]
if len(all_dets) == 0:
all_dets = dets
else:
for det_index, (all_det, det) in enumerate(zip(all_dets,
dets)):
all_dets[det_index] = np.concatenate((all_det, det),
axis=0)
json_results.append(all_config_dets)
for index, all_det in enumerate(all_dets):
all_dets[index] = soft_nms(all_dets[index], 0.5,
min_score=0.001)[0][0:100]
results.append(all_dets)
end = time.time()
print('\r',
'{}/{}, use time:{}'.format(image_index, file_count,
end - start),
end='',
flush=True)
start = end
#mmcv.dump(results, 'eval/assemble_result.pkl')
kwargs = {} if args.options is None else args.options
coco_dataset.evaluate(results, 'bbox', **kwargs)
with open(args.result, 'w') as fp:
json.dump(json_results, fp, indent=4, separators=(',', ': '))