作者您好，我运行python setup.py build_ext --inplace 命令后报如下错误，请问 如何解决 cc1plus: warning: command line option ‘-Wstrict-prototypes’ is valid for C/ObjC but not for C++ In file included from /home/flora/miniconda3/envs/DAL/lib/python3.7/site-packages/numpy/core/include/numpy/ndarraytypes.h:1822:0, from /home/flora/miniconda3/envs/DAL/lib/python3.7/site-packages/numpy/core/include/numpy/ndarrayobject.h:12, from /home/flora/miniconda3/envs/DAL/lib/python3.7/site-packages/numpy/core/include/numpy/arrayobject.h:4, from rbbox_overlaps.cpp:449: /home/flora/miniconda3/envs/DAL/lib/python3.7/site-packages/numpy/core/include/numpy/npy_1_7_deprecated_api.h:17:2: warning: #warning "Using deprecated NumPy API, disable it with " "#define NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION" [-Wcpp] #warning "Using deprecated NumPy API, disable it with "
^ rbbox_overlaps.cpp: In function ‘void __Pyx__ExceptionSave(PyThreadState*, PyObject**, PyObject**, PyObject**)’: rbbox_overlaps.cpp:5670:21: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_type’ type = tstate->exc_type; ^ rbbox_overlaps.cpp:5671:22: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_value’ value = tstate->exc_value; ^ rbbox_overlaps.cpp:5672:19: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_traceback’ tb = tstate->exc_traceback; ^ rbbox_overlaps.cpp: In function ‘void __Pyx__ExceptionReset(PyThreadState, PyObject, PyObject, PyObject*)’: rbbox_overlaps.cpp:5679:24: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_type’ tmp_type = tstate->exc_type; ^ rbbox_overlaps.cpp:5680:25: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_value’ tmp_value = tstate->exc_value; ^ rbbox_overlaps.cpp:5681:22: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_traceback’ tmp_tb = tstate->exc_traceback; ^ rbbox_overlaps.cpp:5682:13: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_type’ tstate->exc_type = type; ^ rbbox_overlaps.cpp:5683:13: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_value’ tstate->exc_value = value; ^ rbbox_overlaps.cpp:5684:13: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_traceback’ tstate->exc_traceback = tb; ^ rbbox_overlaps.cpp: In function ‘int __Pyx__GetException(PyThreadState*, PyObject**, PyObject**, PyObject**)’: rbbox_overlaps.cpp:5739:24: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_type’ tmp_type = tstate->exc_type; ^ rbbox_overlaps.cpp:5740:25: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_value’ tmp_value = tstate->exc_value; ^ rbbox_overlaps.cpp:5741:22: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_traceback’ tmp_tb = tstate->exc_traceback; ^ rbbox_overlaps.cpp:5742:13: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_type’ tstate->exc_type = local_type; ^ rbbox_overlaps.cpp:5743:13: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_value’ tstate->exc_value = local_value; ^ rbbox_overlaps.cpp:5744:13: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_traceback’ tstate->exc_traceback = local_tb; ^ error: command 'gcc' failed with exit status 1

how to eval on train? my data format as {x1,y1,x2,y2,x3,y3,x4,y4,text}, and i want use DLA for text detection, but get mistake(list out of range) on train.py(def eval). the train dataset use IC15dataset,could u write more detail for eval on train.py please.

Thanks great work! I have a question about the angle of rbox: In the paper, i find the angle regression term is . Does the angle of GT rbox (for example, HRSC2016 dataset) in this paper ranges from -90 to 90?

作者好，非常感谢您的工作。我尝试在DOTA1.0上进行训练与精度评价，这里是我训练的超参：

lr

lr0: 0.0001 warmup_lr: 0.00001 warm_epoch:5

setting

num_classes: 15

training

epochs: 24 batch_size: 4 save_interval: 3 test_interval: 1000

这里图片的输入尺寸是1024^2，但是目前我得到的精度是（task1）： mAP: 0.6017705895879002 ap of each class: plane:0.7982196456704045, baseball-diamond:0.6323334007581611, bridge:0.40836332233092576, ground-track-field:0.46187258734002823, small-vehicle:0.6304814611874696, large-vehicle:0.6403297981395368, ship:0.7565095425203512, tennis-court:0.8959681065504155, basketball-court:0.7136081543213204, storage-tank:0.7763719723087148, soccer-ball-field:0.29802453302029897, roundabout:0.5099949396128973, harbor:0.5388697646315033, swimming-pool:0.5819268480975204, helicopter:0.3836847673289567

请问如果达到项目里70+ap我需要做哪些修改？

Hi, I tried to replace RetinaNet with YOLOv5 to get faster speed. The traning loss can decrease but mAP is always low. It cannot detect anything. Before replacement, I can sucessfully get 93% mAP using the default setting (RetinaNet). Could you please tell me if I miss anything here? I read the code about anchor part. It looks like only rectangle anchor is used, but i did not notice where is the way to resize the predicted box to original image corrdinate, and what is the purpose of weights in class BoxCoder weights=(10., 10., 10., 5., 15.)

Here is the part I changed. I make anthor copy of yolo.py in YOLOv5/models. Only change the Detect class to replace cls head and reg head in RetinaNet. I also changed in anchor.py to reduce the pyramid level from 5 to 3 as in YOLOv5. self.pyramid_levels = [3, 4, 5]

class Detect(nn.Module): def init(self, nc=80, anchors=(), ch=()): # detection layer super(Detect, self).init() self.nc = nc + 1 # number of classes self.nr = 5 #xywha self.na = 3 #len(anchors[0]) // 2 # number of anchors self.cls_head = nn.ModuleList(nn.Conv2d(x, self.na * self.nc, 3, 1, 1) for x in ch) # output conv self.reg_head = nn.ModuleList(nn.Conv2d(x, self.na * self.nr, 3, 1, 1) for x in ch) # output conv self.init_weights()

def init_weights(self):
    prior = 0.01
    for m in self.cls_head.modules():
        if isinstance(m, nn.Conv2d):
            m.weight.data.fill_(0)
            m.bias.data.fill_(-math.log((1.0 - prior) / prior))
    for m in self.reg_head.modules():
        if isinstance(m, nn.Conv2d):
            m.weight.data.fill_(0)
            m.bias.data.fill_(0)

def get_cls_feature(self, index, x):
    x = torch.sigmoid(self.cls_head[index](x))
    x = x.permute(0, 2, 3, 1)
    n, w, h, c = x.shape
    x = x.reshape(n, w, h, self.na, self.nc)
    return x.reshape(x.shape[0], -1, self.nc)
def get_reg_feature(self, index, x):
    x = self.reg_head[index](x)
    x = x.permute(0, 2, 3, 1)
    return x.reshape(x.shape[0], -1, self.nr)
def forward(self, x):
    cls_score = torch.cat([self.get_cls_feature(idx, feature) for idx, feature in enumerate(x)], dim=1)
    bbox_pred = torch.cat([self.get_reg_feature(idx, feature) for idx, feature in enumerate(x)], dim=1)
    return [cls_score, bbox_pred]

I define another class to replace class RetinaNet, by only change a few lines.

class YOLOV5(nn.Module): def init(self, cfg='yolov5l.yaml', ch=3, nc=None, export=False): super(YOLOV5, self).init() self.yolov5 = Model(cfg) self.num_classes = 1 + 1 # class number + 1, I only have one class in the dataset self.anchor_generator = Anchors( ratios = np.array([0.5,1,2]), ) self.num_anchors = self.anchor_generator.num_anchors self.loss = IntegratedLoss(func='smooth') self.box_coder = BoxCoder()

def forward(self, ims, gt_boxes=None, test_conf=None,process=None):
    anchors_list, offsets_list, cls_list, var_list = [], [], [], []
    original_anchors = self.anchor_generator(ims)   # (bs, num_all_achors, 5)
    anchors_list.append(original_anchors)
    # features = self.fpn(self.ims_2_features(ims))
    # cls_score = torch.cat([self.cls_head(feature) for feature in features], dim=1)
    # bbox_pred = torch.cat([self.reg_head(feature) for feature in features], dim=1)
    [cls_score, bbox_pred] = self.yolov5(ims)

..........the other parts are the same.

Hi ming71: When I trained the model to 2 iterations with batch 1 of the first epoch for DOTA dataset, I encountered the following problems: Epoch gpu_mem cls reg total targets img_size 0/99 4.38G 2.35 4.93 7.28 33 768: 0%| | 2/1411 [00:00<10:31, 2.23it/s] Traceback (most recent call last): File "/home/mccc/Program_Code/Rotated_Object_Detection/DAL/train.py", line 279, in train_model(arg, hyps) File "/home/mccc/Program_Code/Rotated_Object_Detection/DAL/train.py", line 146, in train_model for i, (ni, batch) in enumerate(pbar): File "/home/mccc/anaconda3/envs/DAL/lib/python3.6/site-packages/tqdm/std.py", line 1176, in iter for obj in iterable: File "/home/mccc/anaconda3/envs/piou/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 322, in next return self._process_next_batch(batch) File "/home/mccc/anaconda3/envs/piou/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 357, in _process_next_batch raise batch.exc_type(batch.exc_msg) IndexError: Traceback (most recent call last): File "/home/mccc/anaconda3/envs/piou/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 106, in _worker_loop samples = collate_fn([dataset[i] for i in batch_indices]) File "/home/mccc/anaconda3/envs/piou/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 106, in samples = collate_fn([dataset[i] for i in batch_indices]) File "/home/mccc/Program_Code/Rotated_Object_Detection/DAL/datasets/dota_dataset.py", line 43, in getitem bboxes = roidb['boxes'][gt_inds, :] IndexError: too many indices for array

Do you have such a problem? My GPU is nvidia RTX 2080Ti , its memory size is 11G. Looking forward to your reply! Thanks.

作者您好，请问如何安装torch-warmup-lr呀 谢谢 (DAL) wen@wen-ubuntu:~/Documents/object_detection/DAL$ pip install torch-warmup-lr==1.0.0 Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple/ ERROR: Could not find a version that satisfies the requirement torch-warmup-lr==1.0.0 ERROR: No matching distribution found for torch-warmup-lr==1.0.0 (DAL) wen@wen-ubuntu:~/Documents/object_detection/DAL$

您好，我想在自己的工程里使用您的计算旋转框iou的函数rbbx_overlaps，但我发现无论我输入什么，输出的iou始终为0。下面是我的测试代码，是我的调用方式有问题吗 import numpy as np from DAL_utils.overlaps_cuda.rbbox_overlaps import rbbx_overlaps from DAL_utils.overlaps.rbox_overlaps import rbox_overlaps

a = np.array([[10, 10, 20, 10, 0]], dtype=np.float32) b = np.array([[10, 10, 20, 10, 0]], dtype=np.float32) c = rbbx_overlaps(a, b) d = rbox_overlaps(a, b) print(c) print(d)

Hi, thanks for your great job! But when i run the code on dota_dataset, there has something wrong with the data processing: Traceback (most recent call last): File "train.py", line 279, in train_model(arg, hyps) File "train.py", line 147, in train_model for i, (ni, batch) in enumerate(pbar): File "/home/cyli/dat01/anaconda3/envs/dal/lib/python3.7/site-packages/tqdm/std.py", line 1166, in iter for obj in iterable: File "/home/cyli/dat01/anaconda3/envs/dal/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 582, in next return self._process_next_batch(batch) File "/home/cyli/dat01/anaconda3/envs/dal/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 608, in _process_next_batch raise batch.exc_type(batch.exc_msg) IndexError: Traceback (most recent call last): File "/home/cyli/dat01/anaconda3/envs/dal/lib/python3.7/site-packages/torch/utils/data/_utils/worker.py", line 99, in _worker_loop samples = collate_fn([dataset[i] for i in batch_indices]) File "/home/cyli/dat01/anaconda3/envs/dal/lib/python3.7/site-packages/torch/utils/data/_utils/worker.py", line 99, in samples = collate_fn([dataset[i] for i in batch_indices]) File "/scratch/cyli/lcy/DAL/datasets/dota_dataset.py", line 40, in getitem bboxes = roidb['boxes'][gt_inds, :] IndexError: too many indices for array: array is 1-dimensional, but 2 were indexed

我是先用了DOTA_devkit进行DOTA数据集的train和val的切分，生成相应的images,labelTxt文件夹，用的标签是DOTA1.5 obb的。之后我也用generage_dataset生成相应的包含trainval所有图像的位置的txt文件，之后修改train.py的default参数如下，然后出现了一下错误。自己不是很明白，所以来这里问下。 if name == 'main':

parser = argparse.ArgumentParser(description='Train a detector')
# config
parser.add_argument('--hyp', type=str, default='hyp.py', help='hyper-parameter path')
# network
parser.add_argument('--backbone', type=str, default='res50')
parser.add_argument('--freeze_bn', type=bool, default=False)
parser.add_argument('--weight', type=str, default='')   # 
parser.add_argument('--multi-scale', action='store_true', help='adjust (67% - 150%) img_size every 10 batches')

# NWPU-VHR10
parser.add_argument('--dataset', type=str, default='DOTA')
parser.add_argument('--train_path', type=str, default='/home/ff/WangZF/datasets/dota/train.txt')
parser.add_argument('--test_path', type=str, default='/home/ff/WangZF/datasets/dota/val.txt')

parser.add_argument('--training_size', type=int, default=800)
parser.add_argument('--resume', action='store_true', help='resume training from last.pth')
parser.add_argument('--load', action='store_true', help='load training from last.pth')
parser.add_argument('--augment', action='store_true', help='data augment')
parser.add_argument('--target_size', type=int, default=[800])   
#

arg = parser.parse_args()
hyps = hyp_parse(arg.hyp)
print(arg)
print(hyps)

train_model(arg, hyps)

/home/ff/anaconda3/envs/pytorch1.6/bin/python /home/ff/WangZF/remoteSense/DAL/train.py fail to speed up training via apex

Namespace(augment=False, backbone='res50', dataset='DOTA', freeze_bn=False, hyp='hyp.py', load=False, multi_scale=False, resume=False, target_size=[800], test_path='/home/ff/WangZF/datasets/dota/val.txt', train_path='/home/ff/WangZF/datasets/dota/train.txt', training_size=800, weight='') {'lr0': 0.0001, 'warmup_lr': 1e-05, 'warm_epoch': 5.0, 'num_classes': 10.0, 'epochs': 100.0, 'batch_size': 2.0, 'save_interval': 5.0, 'test_interval': 5.0} Model Summary: 195 layers, 3.63368e+07 parameters, 3.63368e+07 gradients

 Epoch   gpu_mem       cls       reg     total   targets  img_size

0%| | 0/4043 [00:00<?, ?it/s] Traceback (most recent call last): File "/home/ff/WangZF/remoteSense/DAL/train.py", line 278, in train_model(arg, hyps) File "/home/ff/WangZF/remoteSense/DAL/train.py", line 146, in train_model for i, (ni, batch) in enumerate(pbar): File "/home/ff/anaconda3/envs/pytorch1.6/lib/python3.6/site-packages/tqdm/std.py", line 1193, in iter for obj in iterable: File "/home/ff/anaconda3/envs/pytorch1.6/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 435, in next data = self._next_data() File "/home/ff/anaconda3/envs/pytorch1.6/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 1085, in _next_data return self._process_data(data) File "/home/ff/anaconda3/envs/pytorch1.6/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 1111, in _process_data data.reraise() File "/home/ff/anaconda3/envs/pytorch1.6/lib/python3.6/site-packages/torch/_utils.py", line 428, in reraise raise self.exc_type(msg) IndexError: Caught IndexError in DataLoader worker process 0. Original Traceback (most recent call last): File "/home/ff/anaconda3/envs/pytorch1.6/lib/python3.6/site-packages/torch/utils/data/_utils/worker.py", line 198, in _worker_loop data = fetcher.fetch(index) File "/home/ff/anaconda3/envs/pytorch1.6/lib/python3.6/site-packages/torch/utils/data/_utils/fetch.py", line 44, in fetch data = [self.dataset[idx] for idx in possibly_batched_index] File "/home/ff/anaconda3/envs/pytorch1.6/lib/python3.6/site-packages/torch/utils/data/_utils/fetch.py", line 44, in data = [self.dataset[idx] for idx in possibly_batched_index] File "/home/ff/WangZF/remoteSense/DAL/datasets/dota_dataset.py", line 40, in getitem bboxes = roidb['boxes'][gt_inds, :] IndexError: too many indices for array: array is 1-dimensional, but 2 were indexed

Dear author: thank you for your repo! In the AAAI paper, the experiments on DOTA shows 76.95 mAP with the s2anet-DAL method, I want to know more about the details, such as, whether the DAL applied in the FAM or the ODM module?