miemie2013 / miemiedetection

Pytorch and ncnn implementation of PPYOLOE、YOLOX、PPYOLO、PPYOLOv2、PicoDet and so on.

简体中文 | English

miemie2013

个人介绍

好康的照骗比枯燥乏味的文字更具有说服力↓

友情链接

• miemieGAN miemieGAN是咩酱个人开发与维护的图像生成库，以咩酱的名字命名，实现了stylegan2ada等算法，目前文档完善中，欢迎大家试玩。

传送门

算法1群：645796480（人已满）

算法2群：894642886

粉丝群：704991252

关于仓库的疑问尽量在Issues上提，避免重复解答。

B站不定时女装: _糖蜜

知乎不定时谢邀、写文章: 咩咩2013

西瓜视频: 咩咩2013

微信：wer186259

本人微信公众号：miemie_2013

技术博客：https://blog.csdn.net/qq_27311165

AIStudio主页：asasasaaawws

欢迎在GitHub或上面的平台关注我（求粉）~

打赏

如果你觉得我的仓库对你很有帮助，可以给我打钱↓

咩酱爱你哟！

miemiedetection's Issues

关于训练的问题

请问只能训练coco吗，可以训练voc格式的吗

yolox_head_fast.py的get_in_boxes_info函数似乎有一点小错误

这里是不是写错了

用collate_fn实现batch_transforms太耗时了，这句话如何理解

我比较好奇的是：用 collate_fn 实现batch_transforms太耗时了，原因是啥？感觉不合理啊

Train on VOC 2007 Dataset

Can I train the model using VOC 2007 Dataset or do I need to convert the data to coco format?

数据集的问题

只能使用coco数据集吗？

训练自定义数据集时出了个错

An error has been caught in function 'launch', process 'MainProcess' (5932), t
hread 'MainThread' (20804):
Traceback (most recent call last):

File "tools\train.py", line 137, in
launch(
└ <function launch at 0x0000022C3AC6AAF0>

File "D:\YoLo\miemiedetection-main\mmdet\core\launch.py", line 98, in launch
main_func(*args)
│ └ (╒═══════════════════════╤═══════════════════════════════════════════════════════════════════════════════════════════
════════...
└ <function main at 0x0000022C3AC6AE50>

File "tools\train.py", line 115, in main
trainer.train()
│ └ <function Trainer.train at 0x0000022C3D390430>
└ <mmdet.core.trainer.Trainer object at 0x0000022C3D421880>

File "D:\YoLo\miemiedetection-main\mmdet\core\trainer.py", line 101, in train
self.before_train()
│ └ <function Trainer.before_train at 0x0000022C3D3909D0>
└ <mmdet.core.trainer.Trainer object at 0x0000022C3D421880>

File "D:\YoLo\miemiedetection-main\mmdet\core\trainer.py", line 504, in before_train
self.train_loader = self.exp.get_data_loader(
│ │ │ └ <function PPYOLOE_Method_Exp.get_data_loader at 0x0000022C3D3F8B80>
│ │ └ ╒═══════════════════════╤══════════════════════════════════════════════════════════════════════════════
══════════════════════...
│ └ <mmdet.core.trainer.Trainer object at 0x0000022C3D421880>
└ <mmdet.core.trainer.Trainer object at 0x0000022C3D421880>

File "D:\YoLo\miemiedetection-main\mmdet\exp\ppyoloe\ppyoloe_method_base.py", line 260, in get_data_loader
train_dataset = PPYOLOE_COCOTrainDataset(
└ <class 'mmdet.data.datasets.coco.PPYOLOE_COCOTrainDataset'>

File "D:\YoLo\miemiedetection-main\mmdet\data\datasets\coco.py", line 780, in init
train_records = data_clean(train_dataset, train_img_ids, _catid2clsid, train_pre_path, 'train', xy_plus_1=True)
│ │ │ │ └ 'datasets\Go_Black_White_COCO/train2017'
│ │ │ └ {0: 0, 1: 1, 2: 2, 3: 3}
│ │ └ ['000001', '000002', '000004', '000005', '000006', '000007', '000008', '000009', '0000
10', '000011', '000012', '000013', '000...
│ └ <pycocotools.coco.COCO object at 0x0000022C3D5233D0>
└ <function data_clean at 0x0000022C3C6B28B0>

File "D:\YoLo\miemiedetection-main\mmdet\data\datasets\coco.py", line 229, in data_clean
img_anno = coco.loadImgs(img_id)[0]
│ │ └ '000001'
│ └ <function COCO.loadImgs at 0x0000022C3D39A3A0>
└ <pycocotools.coco.COCO object at 0x0000022C3D5233D0>

File "D:\Anaconda3\lib\site-packages\pycocotools\coco.py", line 229, in loadImgs
return [self.imgs[id] for id in ids]
│ │ └ '000001'
│ └ {'000001': {'file_name': '000001.jpg', 'height': 640, 'width': 640, 'id': '000001'}, '000002': {'file_name': '00000
2.jpg', 'h...
└ <pycocotools.coco.COCO object at 0x0000022C3D5233D0>

File "D:\Anaconda3\lib\site-packages\pycocotools\coco.py", line 229, in
return [self.imgs[id] for id in ids]
│ │ │ └ '0'
│ │ └ '0'
│ └ {'000001': {'file_name': '000001.jpg', 'height': 640, 'width': 640, 'id': '000001'}, '000002': {'file_name': '00000
2.jpg', 'h...
└ <pycocotools.coco.COCO object at 0x0000022C3D5233D0>

KeyError: '0'

请教下在训练yolox和ppyoloe的时候，有没有必要放入一些没有目标框的图片做训练集

关于yolox_head_fast.py里面dynamic_k_matching函数的一个小bug

在参考作者代码对官方pipeline进行修改之后，终于统一了数据格式的问题，但是后来又遇到了一个问题，与合理的gt数目有关，源代码是这样的：
# 如果有预测框（花心大萝卜）匹配到了1个以上的gt时，做特殊处理。 if (anchor_matching_gt > 1).float().sum() > 0: # 首先，找到与花心大萝卜具有最小cost的gt。 # 找到 花心大萝卜 的下标（这是在anchor_matching_gt.shape[N, A]中的下标）。假设有R个花心大萝卜。 indexes = torch.where(anchor_matching_gt > 1) index = torch.stack((indexes[0], indexes[1]), 1) # [R, 2] 每个花心大萝卜2个坐标。第0个坐标表示第几张图片，第1个坐标表示第几个格子。 cost_t = cost.permute(0, 2, 1) # [N, G, A] -> [N, A, G] 转置好提取其cost cost2 = self.gather_nd(cost_t, index) # [R, G] 抽出 R个花心大萝卜 与 gt 两两之间的cost。 cost2 = cost2.permute(1, 0) # [G, R] gt 与 R个花心大萝卜 两两之间的cost。 cost_argmin = cost2.argmin(axis=0) # [R, ] 为 每个花心大萝卜 找到 与其cost最小的gt 的下标
我的代码会在进入这个判断后稳定跑飞，报错位置在求取下标index这一步
indexes = torch.where(anchor_matching_gt > 1)
报错信息为：
RuntimeError: numel: integer multiplication overflow
不太清楚这个where计算量在哪里，感觉非常的迷惑，希望能指点一下为什么会出现这个问题

yolox_head_fast.py中当G == 0的返回loss有问题

if G == 0: # 所有图片都没有gt时 obj_targets = torch.zeros((N, A, 1), dtype=torch.float32, device=obj_preds.device) num_fg = 1 # 所有图片都没有gt时，设为1 loss_obj = self.bcewithlog_loss(obj_preds, obj_targets) loss_obj = loss_obj.sum() / num_fg losses = { "loss_obj": loss_obj, } return losses
yolox_head_fast.py中的这一段是不是错了
这个losses和下面的对不上
return ( loss, reg_weight * loss_iou, loss_obj, loss_cls, loss_l1, num_fg / max(num_gts, 1), )

支持ppyoloe plus 的转换么？

我自己在aistudio（白嫖v100）上训练了个ppyoloe plus模型，用ppyoloe 的权重转换脚本转换报错：
脚本命令：

python tools/convert_weights.py -f exps/ppyoloe/ppyoloe_crn_s_300e_coco.py -c model_final.pdparams -oc ppyoloe_crn_s_300e_coco.pth -nc 1

训练rtdert报错

python tools/train.py -f exps/rtdetr/rtdetr_r18vd_6x_coco.py -d 2 -b 20 -eb 24 -w 4 -ew 4 -lrs 0.1
报错如下：

2023-08-22 17:46:14 | INFO | mmdet.core.trainer:493 - ---> start train epoch1
2023-08-22 17:46:16 | ERROR | mmdet.core.trainer:98 - one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor []] is at version 1; expected version 0 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).
2023-08-22 17:46:16 | INFO | mmdet.core.trainer:343 - Training of experiment is done and the best AP is 0.00
2023-08-22 17:46:16 | ERROR | mmdet.core.launch:147 - An error has been caught in function '_distributed_worker', process 'SpawnProcess-1' (478), thread 'MainThread' (139673154561728):
Traceback (most recent call last):
File "", line 1, in
File "/data/anaconda3/envs/miemie_det/lib/python3.8/multiprocessing/spawn.py", line 116, in spawn_main
exitcode = _main(fd, parent_sentinel)
│ │ └ 5
│ └ 8
└ <function _main at 0x7f083058cc10>
File "/data/anaconda3/envs/miemie_det/lib/python3.8/multiprocessing/spawn.py", line 129, in _main
return self._bootstrap(parent_sentinel)
│ │ └ 5
│ └ <function BaseProcess._bootstrap at 0x7f083073dee0>
└
File "/data/anaconda3/envs/miemie_det/lib/python3.8/multiprocessing/process.py", line 315, in _bootstrap
self.run()
│ └ <function BaseProcess.run at 0x7f083073d550>
└
File "/data/anaconda3/envs/miemie_det/lib/python3.8/multiprocessing/process.py", line 108, in run
self._target(*self._args, **self._kwargs)
│ │ │ │ │ └ {}
│ │ │ │ └
│ │ │ └ (<function _distributed_worker at 0x7f07b1e38160>, 0, (<function main at 0x7f077b30d940>, 2, 2, 0, 'nccl', 'tcp://127.0.0.1:5...
│ │ └
│ └ <function _wrap at 0x7f07b1956310>
└
File "/data/anaconda3/envs/miemie_det/lib/python3.8/site-packages/torch/multiprocessing/spawn.py", line 69, in _wrap
fn(i, *args)
│ │ └ (<function main at 0x7f077b30d940>, 2, 2, 0, 'nccl', 'tcp://127.0.0.1:56017', (╒═══════════════════════╤═════════════════════...
│ └ 0
└ <function _distributed_worker at 0x7f07b1e38160>
File "/home/a-bamboo/repositories/miemiedetection/mmdet/core/launch.py", line 147, in _distributed_worker
main_func(*args)
│ └ (╒═══════════════════════╤═══════════════════════════════════════════════════════════════════════════════════════════════════...
└ <function main at 0x7f077b30d940>
File "/home/a-bamboo/repositories/miemiedetection/tools/train.py", line 126, in main
trainer.train()
│ └ <function Trainer.train at 0x7f077a68bc10>
└ <mmdet.core.trainer.Trainer object at 0x7f077a65ce20>
File "/home/a-bamboo/repositories/miemiedetection/mmdet/core/trainer.py", line 96, in train
self.train_in_epoch()
│ └ <function Trainer.train_in_epoch at 0x7f077a68bd30>
└ <mmdet.core.trainer.Trainer object at 0x7f077a65ce20>
File "/home/a-bamboo/repositories/miemiedetection/mmdet/core/trainer.py", line 336, in train_in_epoch
self.train_in_iter()
│ └ <function Trainer.train_in_iter at 0x7f077a68be50>
└ <mmdet.core.trainer.Trainer object at 0x7f077a65ce20>
File "/home/a-bamboo/repositories/miemiedetection/mmdet/core/trainer.py", line 350, in train_in_iter
self.train_one_iter()
│ └ <function Trainer.train_one_iter at 0x7f077a68bee0>
└ <mmdet.core.trainer.Trainer object at 0x7f077a65ce20>
File "/home/a-bamboo/repositories/miemiedetection/mmdet/core/trainer.py", line 462, in train_one_iter
self.scaler.scale(loss).backward()
│ │ │ └ tensor(13467.0713, device='cuda:0', grad_fn=)
│ │ └ <function GradScaler.scale at 0x7f07b2133790>
│ └ <torch.cuda.amp.grad_scaler.GradScaler object at 0x7f077a65ce50>
└ <mmdet.core.trainer.Trainer object at 0x7f077a65ce20>
File "/data/anaconda3/envs/miemie_det/lib/python3.8/site-packages/torch/_tensor.py", line 363, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
│ │ │ │ │ │ │ └ None
│ │ │ │ │ │ └ False
│ │ │ │ │ └ None
│ │ │ │ └ None
│ │ │ └ tensor(13467.0713, device='cuda:0', grad_fn=)
│ │ └ <function backward at 0x7f07b1d6cee0>
│ └ <module 'torch.autograd' from '/data/anaconda3/envs/miemie_det/lib/python3.8/site-packages/torch/autograd/init.py'>
└ <module 'torch' from '/data/anaconda3/envs/miemie_det/lib/python3.8/site-packages/torch/init.py'>
File "/data/anaconda3/envs/miemie_det/lib/python3.8/site-packages/torch/autograd/init.py", line 173, in backward
Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
│ │ └ <method 'run_backward' of 'torch._C._EngineBase' objects>
│ └ <torch._C._EngineBase object at 0x7f07be7d8d80>
└ <class 'torch.autograd.variable.Variable'>

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor []] is at version 1; expected version 0 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).

请问 YOLO 系列算法 支持 TTA吗？

请问 YOLO 系列算法 支持 TTA吗？ 别的mmdetection里都还不支持。

训练中途出现fast_cocoeval问题

你好，很高兴这里有你的辛苦整理，对我有很大的帮助，但是当我浮现代码跑了2个epoch时候，出现如下错误，请问怎么解决啊

File "C:\D_installation_packet\Anaconda\installion_package\envs\yolox\lib\site-packages\torch\utils\cpp_extension.py", line 1079, in load
return _jit_compile(
└ <function _jit_compile at 0x00000213A2B06550>

File "C:\D_installation_packet\Anaconda\installion_package\envs\yolox\lib\site-packages\torch\utils\cpp_extension.py", line 1292, in _jit_compile
_write_ninja_file_and_build_library(
└ <function _write_ninja_file_and_build_library at 0x00000213A2B06670>

File "C:\D_installation_packet\Anaconda\installion_package\envs\yolox\lib\site-packages\torch\utils\cpp_extension.py", line 1404, in _write_ninja_file_and_build_library
_run_ninja_build(
└ <function _run_ninja_build at 0x00000213A2B06AF0>

File "C:\D_installation_packet\Anaconda\installion_package\envs\yolox\lib\site-packages\torch\utils\cpp_extension.py", line 1683, in _run_ninja_build
raise RuntimeError(message) from e
└ "Error building extension 'fast_cocoeval'"

RuntimeError: Error building extension 'fast_cocoeval'

Process finished with exit code 0

请问可以分享ppyoloe m在coco上的训练log吗

您好，我这边在coco上复现ppyoloe m模型，训练过程loss不收敛，评估mAP为0，请问可能是什么原因呢？方便分享一下您在coco上训练的log吗？
训练命令为CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python tools/train.py -f exps/ppyoloe/ppyoloe_crn_m_300e_coco.py -d 8 -b 224 -eb 16 -c CSPResNetb_m_pretrained.pth --fp16

PPYOLO训练时问题

您好，我用自己的数据训练PP-YOLO在300个epoch的时候验证的map突然降到0.0几，本来已经达到0.8多了，请问这是怎么回事？

ImportError: exps/ppyolo/ppyolo_r50vd_2x.py doesn't contains class named 'Exp'

这是怎么回事

critical difference in dfl loss compared to GFL

found some difference of dfl loss compared to the original Gfocal loss
your dfl loss:

original dfl loss:

weight_right should be label minus dis_left rather than 1 minus dis_left.

any idea on the consequence of the change?

转换RT-DETR 为pth文件失败

大佬，参考readme_rtdetr.txt文件，下载转换权重wget https://bj.bcebos.com/v1/paddledet/models/rtdetr_r18vd_dec3_6x_coco.pdparams
运行
python tools/convert_weights.py -f exps/rtdetr/rtdetr_r18vd_6x_coco.py -c rtdetr_r18vd_dec3_6x_coco.pdparams -oc rtdetr_r18vd_dec3_6x_coco.pth -nc 80 --device gpu 报错。

======================== copy weights ======================== transpose param: name='transformer.decoder.layers.0.self_attn.in_proj_weight' transpose param: name='transformer.decoder.layers.0.self_attn.out_proj.weight' transpose param: name='transformer.decoder.layers.0.cross_attn.sampling_offsets.weight' transpose param: name='transformer.decoder.layers.0.cross_attn.attention_weights.weight' transpose param: name='transformer.decoder.layers.0.cross_attn.value_proj.weight' transpose param: name='transformer.decoder.layers.0.cross_attn.output_proj.weight' transpose param: name='transformer.decoder.layers.0.linear1.weight' transpose param: name='transformer.decoder.layers.0.linear2.weight' transpose param: name='transformer.decoder.layers.1.self_attn.in_proj_weight' transpose param: name='transformer.decoder.layers.1.self_attn.out_proj.weight' transpose param: name='transformer.decoder.layers.1.cross_attn.sampling_offsets.weight' transpose param: name='transformer.decoder.layers.1.cross_attn.attention_weights.weight' transpose param: name='transformer.decoder.layers.1.cross_attn.value_proj.weight' transpose param: name='transformer.decoder.layers.1.cross_attn.output_proj.weight' transpose param: name='transformer.decoder.layers.1.linear1.weight' transpose param: name='transformer.decoder.layers.1.linear2.weight' transpose param: name='transformer.decoder.layers.2.self_attn.in_proj_weight' transpose param: name='transformer.decoder.layers.2.self_attn.out_proj.weight' transpose param: name='transformer.decoder.layers.2.cross_attn.sampling_offsets.weight' transpose param: name='transformer.decoder.layers.2.cross_attn.attention_weights.weight' transpose param: name='transformer.decoder.layers.2.cross_attn.value_proj.weight' transpose param: name='transformer.decoder.layers.2.cross_attn.output_proj.weight' transpose param: name='transformer.decoder.layers.2.linear1.weight' transpose param: name='transformer.decoder.layers.2.linear2.weight' Traceback (most recent call last): File "tools/convert_weights.py", line 599, in <module> main(exp, args) File "tools/convert_weights.py", line 571, in main copy(name2, w, model_std) File "tools/convert_weights.py", line 65, in copy assert mul1 == mul2 AssertionError

ModuleNotFoundError: No module named 'yolox'

跑yolox的时候，这里有个小小小bug

关于“regularizer=L2Decay”的torch实现方法

尊敬的老师，您好！
我想请问一下，PP-YOLOE中的WeightDecay是用“regularizer=L2Decay”实现的，PyTorch中好像没有对应的API，这里您是怎么实现的呢
期待您的回复！

大佬，能分享下s模型在coco上的训练日志吗？

如题，我在我们自己框架上复现ppyoloe模型，一直不成功。
把paddl训练的模型转到我们自己的框架上，mAP正常，但复现一直有问题。有哪些需要规避的坑么？
栓Q！

GPU 利用率低

你好，很感谢你的分享。是从老版本的project过来的，训练的时候发现GPU利用率比较低，应该怎么处理呢

ppyoloe_crn_s_300e_coco.pth速度比yolov5s.pt慢很多

您好：
ppyoloe_crn_s_300e_coco.pth在640x640分辨率下测试比yolov5s.pt慢很多，各自的速度分别为5ms和9ms左右。
我的环境是3060ti, pytorch10。
请问您对比过吗

关于yolox_head_fast.py里面dynamic_k_matching函数的一个小bug

在参考作者代码对官方pipeline进行修改之后，终于统一了数据格式的问题，但是后来又遇到了一个问题，与合理的gt数目有关，源代码是这样的：

# 如果有预测框（花心大萝卜）匹配到了1个以上的gt时，做特殊处理。
       if (anchor_matching_gt > 1).float().sum() > 0:

           # 首先，找到与花心大萝卜具有最小cost的gt。

            # 找到 花心大萝卜 的下标（这是在anchor_matching_gt.shape[N, A]中的下标）。假设有R个花心大萝卜。

           indexes = torch.where(anchor_matching_gt > 1)

           index = torch.stack((indexes[0], indexes[1]), 1)  # [R, 2]  每个花心大萝卜2个坐标。第0个坐标表示第几张图片，第1个坐标表示第几个格子。

我的代码会在进入这个判断后稳定跑飞，报错位置在求取下标index这一步

indexes = torch.where(anchor_matching_gt > 1)

报错信息为：

RuntimeError: numel: integer multiplication overflow

不太清楚这个where计算量在哪里，感觉非常的迷惑，希望能指点一下为什么会出现这个问题