zgcr / simpleaicv_pytorch_training_examples Goto Github PK

Hi, it is an amazing job! Could you please introduce how to build the celeba-hq dataset?

where to download the pretrained Regnet models ?

thanks

I just wonder to know what is your computer hardware device, I chose the i5-12400f and RTX4080 to train the just simple Conv model, just have 5 layers, but the speed is so slow, and the training time will cost many years about 100 epochs. And I try to ues AutoDL to train this model , the cpu is 100% utilize but the gpu is just 40%, and the training speed is also very slow.

我按照您的编程环境，在运行代码时出现了这个错误，希望你能解答一下

您好，想问一下，为啥我在训练retinanet的时候总是执行一段出现几个warning之后就自动停下来了，而且也不报错；我一开始以为是用来apex的问题，设置为false之后还是自动停下来了；后面我给换成多卡的也是同样的，请问大佬知道是为啥嘛？
`root@container-ab78119f3c-c31dcd5b:~/SimpleAICV-pytorch-ImageNet-COCO-training-master/detection_training/coco/res50_retinanet_retinaresize800# sh train.sh
======================1======================
No pretrained model file!
loading annotations into memory...
Done (t=16.43s)
creating index...
index created!
Dataset Size:117266
Dataset Class Num:80
loading annotations into memory...
Done (t=0.51s)
creating index...
index created!
Dataset Size:5000
Dataset Class Num:80
======================2======================
Selected optimization level O1: Insert automatic casts around Pytorch functions and Tensor methods.

Defaults for this optimization level are:
enabled : True
opt_level : O1
cast_model_type : None
patch_torch_functions : True
keep_batchnorm_fp32 : None
master_weights : None
loss_scale : dynamic
Processing user overrides (additional kwargs that are not None)...
After processing overrides, optimization options are:
enabled : True
opt_level : O1
cast_model_type : None
patch_torch_functions : True
keep_batchnorm_fp32 : None
master_weights : None
loss_scale : dynamic
Warning: multi_tensor_applier fused unscale kernel is unavailable, possibly because apex was installed without --cuda_ext --cpp_ext. Using Python fallback. Original ImportError was: ModuleNotFoundError("No module named 'amp_C'")
Warning: apex was installed without --cpp_ext. Falling back to Python flatten and unflatten.`

想拜读一下代码，但是clone时出错

(base) PS D:\code> git clone https://github.com/zgcr/SimpleAICV_pytorch_training_examples.git
Cloning into 'SimpleAICV_pytorch_training_examples'...
remote: Enumerating objects: 2761, done.
remote: Counting objects: 100% (2181/2181), done.
remote: Compressing objects: 100% (970/970), done.
remote: Total 2761 (delta 1257), reused 2010 (delta 1129), pack-reused 580
Receiving objects: 100% (2761/2761), 35.31 MiB | 2.01 MiB/s, done.
Resolving deltas: 100% (1588/1588), done.
fatal: cannot create directory at 'simpleAICV/detection/compile_multiscale_deformable_attention/build/temp.linux-x86_64-3.8/root/code/SimpleAICV_pytorch_training_examples_on_ImageNet_COCO_ADE20K/z_dino_main/dino_multiscale_deformable_attention_compile': Filename too long
warning: Clone succeeded, but checkout failed.
You can inspect what was checked out with 'git status'
and retry with 'git restore --source=HEAD :/'

(base) PS D:\code>

Hi,
Nice work. I have a question. What is the difference between resnet_imagenet_DataParallel_train_example and resnet_imagenet_DistributedDataParallel_train_example?

Thank you very much for your outstanding work. However, when I use Dataparrel for training, the GPU will take up more and more time, and then CUDA will start stop the program. May I ask why？

作者您好，请问一下您的代码当中是否可以自定义模型呢，我想自己定义一个resnet110网络是否可以通过将参数修改成[3,4,26,3]来实现呢

您好，我只有一张显卡应该怎样设置目标检测里面的train.sh呢？
感谢！

Could you please share me with your pretrained model of Vovnet series model and RegNet series model? thank you very much!!!!

感谢博主贡献这么nice的代码，我想下载个预训练的模型，但是链接都不行，麻烦大神看到了修复下，灰常感谢。

Thanks

First, thanks author for the great work. It is a great tool to conduct ablation studies. (I even think you can write a paper about that, after adding some more training option, e.g., few-shot, zero-shot learning.)

However, there seems to have a bug about distillation training. That is, when i finished downloading ResNet-34 weights, loading them to the teacher model, the accuracy of it seems really low. It claims 0.298% top-1 accuracy on ImageNet-1K.

I did not checked the code yet, but i suspect it is because the weights you published has a different order of output classes. Could you kindly check this out?

Hi, thanks for your great contributions. I have a question about the implementation of RetinaNet. In losses.py, it seems that the reg_head directly output the absolute position of bounding boxes and l1 loss was calculated by the difference between ground truth bbox positions and reg_head output. Is my understanding correct ？

你好：
你说多次训练会有波动，我这边darknet53在imagenet数据集上训练，现在得到最好的结果top1acc：76.5%，不知道是不是算波动范围内？我这边的训练配置和你是一样的，除了我是用分布式训练，四张卡，batchsize=124×4这个有区别吧。
我这边训练的脚本地址：https://github.com/njustczr/darknet53

Hello. I'm using 4 2080Ti and I wonder how to set num_workers properly?

你好，我主要想学习COCO数据集的一个加载方式，看见你写的很好，但是对cocodataset中一些内容有疑问，比如coco数据集中COCODataPrefetcher()这个类是干嘛的呢，还有这个文件中的coco_class_color干什么用的呢

这个可能是导致你精度比它低一点点的原因吧

您好，我想请教一下，为什么咱们自己训练下的权重和pytorch官方给出的预训练权重会相差很大。原本pytorch官方的预训练参数我能到0.616.现在拿咱们这个模型训出来的预训练参数性能只能到0.499. 我是哪一步出错了吗。因为我这个还挺依赖预训练参数。

你好再次来打扰你了 我在训练时候train.info.log中反馈的是训练到8700轮不给反馈信息了
2021-12-03 15:46:39 - train: epoch 0001, iter [08200, 58633], lr: 0.000100, total_loss: 0.4340, cls_loss: 0.2691, reg_loss: 0.1649
2021-12-03 15:47:37 - train: epoch 0001, iter [08300, 58633], lr: 0.000100, total_loss: 0.6410, cls_loss: 0.4634, reg_loss: 0.1775
2021-12-03 15:48:35 - train: epoch 0001, iter [08400, 58633], lr: 0.000100, total_loss: 0.5121, cls_loss: 0.2628, reg_loss: 0.2494
2021-12-03 15:49:28 - train: epoch 0001, iter [08500, 58633], lr: 0.000100, total_loss: 0.4244, cls_loss: 0.2080, reg_loss: 0.2165
2021-12-03 15:50:28 - train: epoch 0001, iter [08600, 58633], lr: 0.000100, total_loss: 0.5233, cls_loss: 0.3370, reg_loss: 0.1864
2021-12-03 15:51:25 - train: epoch 0001, iter [08700, 58633], lr: 0.000100, total_loss: 0.9907, cls_loss: 0.6687, reg_loss: 0.3220
而且也没有生成权重 训练几次都是在这个地方卡主了 不知道是该继续训练还是哪里需要改动
请问这是怎么一回事呢

你好，我看到了你在CSDN上使用IoU loss训练RetinaNet的文章，很详细，但是我有个问题：
改动的地方是直接把smooth L1 loss改成IoU loss就可以了吗？我自己训练的话起始分类损失是1.228，IoU损失到了11.56，感觉差的有点大，请问是什么原因？有什么好办法解决吗？

作者您好能提供国内下载源吗

hello，
我对比了一下centernet源码和你的repo里的centernet，发现用你的repo训练centernet比源码一个epoch耗时长很多，大概一个64batchsize的iter需要20s，centernet源码几乎是秒级。
对比了下代码好像没有大的区别，请问你知道为啥么

你好，我看到你的代码里面是评价的正确率，而你的github上表格写的错误率，它们之和等于1？？

我训练时如果不用apex 如下：
loading annotations into memory...
Done (t=20.29s)
creating index...
index created!
loading annotations into memory...
Done (t=2.85s)
creating index...
index created!
如果用了的话 还会显示
Defaults for this optimization level are:
enabled : True
opt_level : O1
cast_model_type : None
patch_torch_functions : True
keep_batchnorm_fp32 : None
master_weights : None
loss_scale : dynamic
Processing user overrides (additional kwargs that are not None)...
After processing overrides, optimization options are:
enabled : True
opt_level : O1
cast_model_type : None
patch_torch_functions : True
keep_batchnorm_fp32 : None
master_weights : None
loss_scale : dynamic
Warning: multi_tensor_applier fused unscale kernel is unavailable, possibly because apex was installed without --cuda_ext --cpp_ext. Using Python fallback. Original ImportError was: ModuleNotFoundError("No module named 'amp_C'")
Gradient overflow. Skipping step, loss scaler 0 reducing loss scale to 32768.0
Gradient overflow. Skipping step, loss scaler 0 reducing loss scale to 16384.0
Gradient overflow. Skipping step, loss scaler 0 reducing loss scale to 8192.0
Gradient overflow. Skipping step, loss scaler 0 reducing loss scale to 4096.0
Gradient overflow. Skipping step, loss scaler 0 reducing loss scale to 2048.0

然后就不显示别的了 请问这是在训练还是卡住不动了 如果是卡住是什么引起的呢 我的训练环境是3080ti batch设置为2

我把pretrained设置为True之后，出现如下错误，请问我该如何解决？

Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.

Traceback (most recent call last):

File "../../../tools/train_detection_model.py", line 205, in

main()

File "../../../tools/train_detection_model.py", line 46, in main

from train_config import config

File "./train_config.py", line 19, in

class config:

File "./train_config.py", line 28, in config

'num_classes': num_classes,

File "/home/cc631/hailong/code/Dilated-FPN/simpleAICV-pytorch-ImageNet-COCO-training/simpleAICV/detection/models/retinanet.py", line 145, in resnet50_retinanet

return _retinanet('resnet50', pretrained, **kwargs)

File "/home/cc631/hailong/code/Dilated-FPN/simpleAICV-pytorch-ImageNet-COCO-training/simpleAICV/detection/models/retinanet.py", line 131, in _retinanet

map_location=torch.device('cpu')), model)

File "/home/cc631/anaconda3/envs/pytorch1.7/lib/python3.7/site-packages/torch/serialization.py", line 581, in load

with _open_file_like(f, 'rb') as opened_file:

File "/home/cc631/anaconda3/envs/pytorch1.7/lib/python3.7/site-packages/torch/serialization.py", line 230, in _open_file_like

return _open_file(name_or_buffer, mode)

File "/home/cc631/anaconda3/envs/pytorch1.7/lib/python3.7/site-packages/torch/serialization.py", line 211, in init

super(_open_file, self).init(open(name, mode))

FileNotFoundError: [Errno 2] No such file or directory: 'empty'

The yolox backbone in this codebase without focus operation, the shape of stem between https://github.com/Megvii-BaseDetection/YOLOX and this codebase is different.
The stem of yolox_m backbone in https://github.com/Megvii-BaseDetection/YOLOX:

The stem of yolox_m backbone in this codebase:

class Focus(nn.Module):
"""Focus width and height information into channel space."""

def __init__(self, in_channels, out_channels, ksize=1, stride=1, act="silu"):
    super().__init__()
    self.conv = BaseConv(in_channels * 4, out_channels, ksize, stride, act=act)

def forward(self, x):
    # shape of x (b,c,w,h) -> y(b,4c,w/2,h/2)
    patch_top_left = x[..., ::2, ::2]
    patch_top_right = x[..., ::2, 1::2]
    patch_bot_left = x[..., 1::2, ::2]
    patch_bot_right = x[..., 1::2, 1::2]
    x = torch.cat(
        (
            patch_top_left,
            patch_bot_left,
            patch_top_right,
            patch_bot_right,
        ),
        dim=1,
    )
    return self.conv(x)