def mnist(root_dir, train):
# Normalize the training set with augmentation
transform_train = transforms.Compose(
[
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Resize((32, 32)),
transforms.Normalize(0.406, 0.225, inplace=True)
]
)
# Normalize the test set same as training set without augmentation
transform_test = transforms.Compose(
[
transforms.ToTensor(),
transforms.Resize((32, 32)),
transforms.Normalize(0.406, 0.225, inplace=True)
]
)
opener = urllib.request.URLopener()
opener.addheader('User-Agent',
'Mozilla/5.0 (Macintosh; Intel Mac OS X 10_9_3) AppleWebKit/537.36 '
'(KHTML, like Gecko) Chrome/35.0.1916.47 Safari/537.36')
dataset = MNIST(root=root_dir, train=train, download=True)
return get_dataset(dataset, train, transform_test, transform_train)
def main():
config_path = './config/cfg_retinanet_r50_fpn_trafficsign_extra_aug.py'
cfg = Config.fromfile(config_path)
dataset = get_dataset(cfg.data.train, TrafficSign)
ana = AnalyzeDataset('traffic_sign', dataset, checkonly=True)
ana.imgcheck(1521)
def __init__(self, model, img_path, use_cuda=True):
self.model = model
self.img = cv2.imread(img_path, 1)
self.use_cuda = use_cuda
self.cfg_path = '../config/cfg_ssd300_vgg16_voc.py'
self.cfg = Config.fromfile(self.cfg_path)
self.cfg.data_root = '../data/VOCdevkit/'
# train
self.cfg.data.train.dataset.ann_file = [
self.cfg.data_root + 'VOC2007/ImageSets/Main/trainval.txt',
self.cfg.data_root + 'VOC2012/ImageSets/Main/trainval.txt'
]
self.cfg.data.train.dataset.img_prefix = [
self.cfg.data_root + 'VOC2007/', self.cfg.data_root + 'VOC2012/'
]
# val
self.cfg.data.val.ann_file = self.cfg.data_root + 'VOC2007/ImageSets/Main/test.txt'
self.cfg.data.val.img_prefix = self.cfg.data_root + 'VOC2007/'
# test
self.cfg.data.test.ann_file = self.cfg.data_root + 'VOC2007/ImageSets/Main/test.txt'
self.cfg.data.test.img_prefix = self.cfg.data_root + 'VOC2007/'
self.dataset = get_dataset(self.cfg.data.train, VOCDataset)
# img preprocess
self.img = np.float32(cv2.resize(self.img,
(224, 224))) / 255 # 变更尺寸，归一化
self.img = self.preprocess_image() # (1, c, h, w)
def train(cfg_path, dataset_class):
"""借用mmcv的Runner框架进行训练，包括里边的hooks作为lr更新，loss计算的工具
1. dataset的数据集输出打包了img/gt_bbox/label/，采用DataContainer封装
2. Dataloader的default_collate用定制collate替换，从而支持dataset的多类型数据
3. DataParallel外壳用定制MMDataparallel替换，从而支持DataContainer
"""
# 初始化2个默认选项
distributed = False
parallel = True
# get cfg
cfg = Config.fromfile(cfg_path)
# set backends
if cfg.get('cudnn_benchmark', False):
torch.backends.cudnn.benchmark = True
# get logger
distributed = False
parallel = True
logger = get_root_logger(cfg.log_level)
logger.info('Distributed training: {}'.format(distributed))
logger.info('DataParallel training: {}'.format(parallel))
# build model & detector
model = OneStageDetector(cfg)
# model = OneStageDetector(cfg)
if not parallel:
model = model.cuda()
else:
model = NNDataParallel(model, device_ids=range(cfg.gpus)).cuda()
# prepare data & dataloader
# Runner要求dataloader放在list里: 使workflow里每个flow对应一个dataloader
dataset = get_dataset(cfg.data.train, dataset_class)
batch_size = cfg.gpus * cfg.data.imgs_per_gpu
num_workers = cfg.gpus * cfg.data.workers_per_gpu
dataloader = [
DataLoader(dataset,
batch_size=batch_size,
sampler=GroupSampler(dataset, cfg.data.imgs_per_gpu),
num_workers=num_workers,
collate_fn=partial(collate,
samples_per_gpu=cfg.data.imgs_per_gpu),
pin_memory=False)
]
# define runner and running type(1.resume, 2.load, 3.train/test)
runner = Runner(model, batch_processor, cfg.optimizer, cfg.work_dir,
cfg.log_level)
runner.register_training_hooks(cfg.lr_config, cfg.optimizer_config,
cfg.checkpoint_config, cfg.log_config)
if cfg.resume_from: # 恢复训练: './work_dirs/ssd300_voc/latest.pth'
runner.resume(cfg.resume_from,
map_location=lambda storage, loc: storage)
elif cfg.load_from: # 加载参数进行测试
runner.load_checkpoint(cfg.load_from)
# 开始训练: 采用workflow来区分train还是test
runner.run(dataloader, cfg.workflow, cfg.total_epochs)
def preprocess_data(self, dataset_class):
self.dataset = get_dataset(self.cfg.data.test, dataset_class)
self.dataloader = build_dataloader(
self.dataset,
imgs_per_gpu=1,
workers_per_gpu=self.cfg.data.workers_per_gpu,
num_gpus=1,
dist=False,
shuffle=False)
def test_voc_dataset():
data_root = '../data/VOCdevkit/' # 指代ssd目录下的data目录
img_norm_cfg = dict(mean=[123.675, 116.28, 103.53],
std=[1, 1, 1],
to_rgb=True)
cfg_train = dict(
type='RepeatDataset',
times=10,
dataset=dict(
type='VOCDataset',
ann_file=[
data_root + 'VOC2007/ImageSets/Main/trainval.txt',
data_root + 'VOC2012/ImageSets/Main/trainval.txt'
],
img_prefix=[data_root + 'VOC2007/', data_root + 'VOC2012/'],
img_scale=(300, 300),
img_norm_cfg=img_norm_cfg,
size_divisor=None,
flip_ratio=0.5,
with_mask=False,
with_crowd=False,
with_label=True,
test_mode=False,
extra_aug=dict(photo_metric_distortion=dict(brightness_delta=32,
contrast_range=(0.5,
1.5),
saturation_range=(0.5,
1.5),
hue_delta=18),
expand=dict(mean=img_norm_cfg['mean'],
to_rgb=img_norm_cfg['to_rgb'],
ratio_range=(1, 4)),
random_crop=dict(min_ious=(0.1, 0.3, 0.5, 0.7, 0.9),
min_crop_size=0.3)),
resize_keep_ratio=False))
trainset = get_dataset(cfg_train, VOCDataset)
classes = trainset.CLASSES
data = trainset[1120] # dict('img', 'img_meta', )
"""已做的数据处理：rgb化，chw化，归一化，tensor化"""
bbox = data['gt_bboxes'].data.numpy()
label = data['gt_labels'].data.numpy()
img = data['img'].data.numpy() # 逆tensor
img1 = img.transpose(1, 2, 0) # 逆chw
img2 = np.clip(
(img1 * img_norm_cfg['std'] + img_norm_cfg['mean']).astype(np.int32),
0, 255) # 逆归一
# plt.imshow(img2)
vis_bbox(img2[..., [2, 0, 1]], bbox, label - 1,
label_names=classes) # vis_bbox内部会bgr转rgb，所以这里要用bgr输入
def cifar10(root_dir, train):
# Normalize the training set with augmentation
transform_train = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
# Normalize the test set same as training set without augmentation
transform_test = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
dataset = CIFAR10(root=root_dir, train=train, download=True)
return get_dataset(dataset, train, transform_test, transform_train)
def test_dataloader():
data_root = '../data/VOCdevkit/' # 指代ssd目录下的data目录
img_norm_cfg = dict(mean=[123.675, 116.28, 103.53],
std=[1, 1, 1],
to_rgb=True)
cfg_train = dict(
type='RepeatDataset',
times=10,
dataset=dict(
type='VOCDataset',
ann_file=[
data_root + 'VOC2007/ImageSets/Main/trainval.txt',
data_root + 'VOC2012/ImageSets/Main/trainval.txt'
],
img_prefix=[data_root + 'VOC2007/', data_root + 'VOC2012/'],
img_scale=(300, 300),
img_norm_cfg=img_norm_cfg,
size_divisor=None,
flip_ratio=0.5,
with_mask=False,
with_crowd=False,
with_label=True,
test_mode=False,
extra_aug=dict(photo_metric_distortion=dict(brightness_delta=32,
contrast_range=(0.5,
1.5),
saturation_range=(0.5,
1.5),
hue_delta=18),
expand=dict(mean=img_norm_cfg['mean'],
to_rgb=img_norm_cfg['to_rgb'],
ratio_range=(1, 4)),
random_crop=dict(min_ious=(0.1, 0.3, 0.5, 0.7, 0.9),
min_crop_size=0.3)),
resize_keep_ratio=False))
trainset = get_dataset(cfg_train, VOCDataset)
dataloader = Dataloader
dataset_type = 'TrafficSign' # 改成trafficsign
data_root = './data/traffic_sign/' # 改成trafficsign
img_norm_cfg = dict(mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True)
trainset_cfg = dict(
type=dataset_type,
ann_file=data_root + 'train_label_fix.csv',
img_prefix=data_root + 'Train_fix/',
# img_scale=(1333, 800), # 把图片缩小
img_scale=(3200, 1800), # 把图片保持尺寸
img_norm_cfg=img_norm_cfg,
size_divisor=32,
with_label=True,
extra_aug=None)
dataset = get_dataset(trainset_cfg, TrafficSign)
ana = AnalyzeDataset('traffic_sign', dataset, checkonly=False)
# ana.cluster_bbox(show=True)
ana.types_bin(show=True)
ana.bbox_size(show=True)
# ana.imgcheck(2)
if dset == 'voc':
dataset_type = 'VOCDataset'
data_root = './data/VOCdevkit/'
img_norm_cfg = dict(mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True)
trainset_cfg = dict(
type=dataset_type,
ann_file=[
def main():
"""针对faster rcnn在voc的评估做微调
1. args parse用直接输入替代
2.
"""
config_path = './config/cfg_ssd300_vgg16_voc.py' # 注意：cfg和模型需要匹配，因为不同数据集类别数不一样，
checkpoint_path = './weights/myssd/epoch_24.pth'
cfg = mmcv.Config.fromfile(config_path)
out_file = 'dataset_eval_result/results.pkl' # 注意这里要选择pkl而不能选择json，因为outputs里边包含array，用json无法保存
eval_type = ['bbox'] # proposal_fast是mmdetection自己的实现
# eval_type = ['proposal','bbox'] # 这几种是coco api的实现包括['proposal','bbox','segm','keypoints']，已跑通
show_result = False # 这里可以设置show=True从而按顺序显示每张图的测试结果(对于少量图片的数据集可以这么玩)
if not out_file.endswith(('.pkl', '.pickle')):
raise ValueError('The output file must be a pkl file.')
# set cudnn_benchmark
if cfg.get('cudnn_benchmark', False):
torch.backends.cudnn.benchmark = True
# cfg.model.pretrained = None
# dataset = obj_from_dict(cfg.data.test, datasets, dict(test_mode=True))
dataset = get_dataset(cfg.data.test, CocoDataset)
cfg.gpus = 1
if cfg.gpus == 1:
# model = build_detector(
# cfg.model, train_cfg=None, test_cfg=cfg.test_cfg)
model = OneStageDetector(cfg)
load_checkpoint(model, checkpoint_path)
model = MMDataParallel(model, device_ids=[0])
data_loader = build_dataloader(
dataset,
imgs_per_gpu=1,
workers_per_gpu=cfg.data.workers_per_gpu,
num_gpus=1,
dist=False,
shuffle=False)
outputs = single_test(model, data_loader, show=show_result)
# outputs结构: [img1,...imgn], len=5000,此为coco val的所有图片
# 每个img结构: [cls1,...clsn], len=80, 此为每个预测的所有类的bbox预测输出
# 每个cls结构: ndarray(n,5), 此为这个cls对应n个bbox，如果该类有预测则n>0，如果该类没有预测则n=0，第一列为置信度？
# 注意：最内层数据结构是ndarray，是不能直接存入json文件，需要转换成data.tolist()
else:
model_args = cfg.model.copy()
model_args.update(train_cfg=None, test_cfg=cfg.test_cfg)
model_type = getattr(detectors, model_args.pop('type'))
outputs = parallel_test(
model_type,
model_args,
checkpoint_path,
dataset,
_data_func,
range(cfg.gpus),
workers_per_gpu=cfg.proc_per_gpu)
# debug
if out_file:
print('writing results to {}'.format(out_file))
mmcv.dump(outputs, out_file) # 先把模型的测试结果输出到文件中: 如果文件不存在会创建
eval_types = eval_type
if eval_types:
print('Starting evaluate {}'.format(' and '.join(eval_types)))
if eval_types == ['proposal_fast']:
result_file = out_file
# coco_eval(result_file, eval_types, dataset.coco) # result_file传入coco_eval()
"""用自己写的evaluation()"""
evaluation(result_file, dataset.coco, eval_types=eval_types)
else:
if not isinstance(outputs[0], dict):
result_file = out_file + '.json'
results2json(dataset, outputs, result_file)
# coco_eval(result_file, eval_types, dataset.coco)
"""用自己写的evaluation()"""
evaluation(result_file, dataset.coco, eval_types=eval_types)
else:
for name in outputs[0]:
print('\nEvaluating {}'.format(name))
outputs_ = [out[name] for out in outputs]
result_file = out_file + '.{}.json'.format(name)
results2json(dataset, outputs_, result_file)
coco_eval(result_file, eval_types, dataset.coco)