8000 训练时长问题 · Issue #16 · BingfengYan/CO-MOT · GitHub
[go: up one dir, main page]

Skip to content

训练时长问题 #16

Open
Open

Description

@WesLee88524

大佬您好,
我最近在复现您的这篇文章,我使用您提供的配置文件在8张T4上训练dancetrack,每个epoch时长竟然需要2天左右,这是正常的吗?
配置文件如下:
--meta_arch motr_unincost
--dataset_file e2e_dance
--epoch 20
--with_box_refine
--lr_drop 8
--lr 2e-4
--lr_backbone 2e-5
--pretrained /pretrained/r50_deformable_detr_plus_iterative_bbox_refinement-checkpoint.pth
--batch_size 1
--sample_mode random_interval
--sample_interval 10
--sampler_lengths 5
--merger_dropout 0
--dropout 0
--random_drop 0.1
--fp_ratio 0.3
--query_interaction_layer GQIM
--num_queries 60
--append_crowd
--use_checkpoint
--mot_path /code/CO-MOT/data
--match_type gmatch
--g_size 3

部分训练log如下:
截屏2024-04-07 21 01 11

Activity

  1. fengxiuyaun commented on Apr 8, 2024

    @fengxiuyaun
    Contributor

    我用v100是训一个epoch大概11个小时,第一个epoch会慢一些

  2. WesLee88524 commented on Apr 8, 2024

    @WesLee88524
    Author

    我用v100是训一个epoch大概11个小时,第一个epoch会慢一些

    感谢!
    请问使用detrex版本的CO-MOT速度会快点吗? 我看那个batch_size可以自己设定,是否可以公布训练代码?我使用那个版本的代码重新训练后测试无结果.

  3. fengxiuyaun commented on Apr 8, 2024

    @fengxiuyaun
    Contributor

    那个我只是简单写了batch_size>1的代码,简单测试了一下没啥大问题,就递交了。不过你可以参考看看有啥我没发现的问题。

  4. WesLee88524 commented on Apr 8, 2024

    @WesLee88524
    Author

    那个我只是简单写了batch_size>1的代码,简单测试了一下没啥大问题,就递交了。不过你可以参考看看有啥我没发现的问题。

    好嘞! 感谢!我再研究下!

  5. WesLee88524 commented on Apr 8, 2024

    @WesLee88524
    Author

    那个我只是简单写了batch_size>1的代码,简单测试了一下没啥大问题,就递交了。不过你可以参考看看有啥我没发现的问题。

    您好,我全部对照了这个版本的CO-MOT和detrex的CO-MOT代码,感觉是什么问题的,训练的过程好像也对,loss都是正常的,但是训练出来的模型评估结果一直为空,请问可能是什么原因呢?我的配置文件如下:
    截屏2024-04-08 17 57 19

    训练的命令为:
    截屏2024-04-08 17 58 47

    评估结果如下:
    截屏2024-04-08 17 52 25

  6. fengxiuyaun commented on Apr 9, 2024

    @fengxiuyaun
    Contributor

    为空啥意思,就是没有识别到?

  7. fengxiuyaun commented on Apr 9, 2024

    @fengxiuyaun
    Contributor

    你可以先跑一下我release的权重,看看代码有啥问题

  8. WesLee88524 commented on Apr 9, 2024

    @WesLee88524
    Author

    对的, 我按照这个配置文件重新训练然后评估,输出文件每帧都是空的,如下图:
    截屏2024-04-09 14 39 43
    , 我用您release的权重直接测试,HOTA在dancetrack val上比论文低4.2%,但是可以跑通. 之后我又对比了各个主要文件,好像都没啥问题. 但训练就是不行😂

    那个我只是简单写了batch_size>1的代码,简单测试了一下没啥大问题,就递交了。不过你可以参考看看有啥我没发现的问题。

    您好,我全部对照了这个版本的CO-MOT和detrex的CO-MOT代码,感觉是什么问题的,训练的过程好像也对,loss都是正常的,但是训练出来的模型评估结果一直为空,请问可能是什么原因呢?我的配置文件如下: 截屏2024-04-08 17 57 19

    训练的命令为: 截屏2024-04-08 17 58 47

    评估结果如下: 截屏2024-04-08 17 52 25

  9. WesLee88524 commented on Apr 9, 2024

    @WesLee88524
    Author

    在使用detrex训练的时候,直接用那个代码会报错,错误及我的修改如下:

    1. TypeError: repr() got an unexpected keyword argument '_repr_indent'
      截屏2024-04-09 20 11 01
      我将之修改为:
      截屏2024-04-09 20 12 14

    2. MS- deformable-attension报错
      截屏2024-04-09 20 18 13
      我将之改为:
      截屏2024-04-09 20 50 07

    3. 形参问题:
      截屏2024-04-09 20 32 16
      我将之改为:
      截屏2024-04-09 20 34 49

    4. 函数缺失
      截屏2024-04-09 20 43 49
      我将本仓库的merge函数加入进去:
      截屏2024-04-09 20 44 46

    至此,可以成功训练.
    我使用的预训练参数是detrex中Deformable-DETR-R50 + Box-Refinement版本, 在加载的时候有很多不匹配的层:
    WARNING [04/09 18:46:39 fvcore.common.checkpoint]: Some model parameters or buffers are not found in the checkpoint: class_embed.0.{bias, weight} class_embed.1.{bias, weight} class_embed.2.{bias, weight} class_embed.3.{bias, weight} class_embed.4.{bias, weight} class_embed.5.{bias, weight} position.weight position_offset.weight query_embed.weight query_embed_offset.weight track_embed.linear1.{bias, weight} track_embed.linear2.{bias, weight} track_embed.linear_feat1.{bias, weight} track_embed.linear_feat2.{bias, weight} track_embed.norm1.{bias, weight} track_embed.norm2.{bias, weight} track_embed.norm_feat.{bias, weight} track_embed.self_attn.out_proj.{bias, weight} track_embed.self_attn.{in_proj_bias, in_proj_weight} transformer.decoder.class_embed.0.{bias, weight} transformer.decoder.class_embed.1.{bias, weight} transformer.decoder.class_embed.2.{bias, weight} transformer.decoder.class_embed.3.{bias, weight} transformer.decoder.class_embed.4.{bias, weight} transformer.decoder.class_embed.5.{bias, weight} transformer.decoder.layers.0.cross_attn.attention_weights.{bias, weight} transformer.decoder.layers.0.cross_attn.output_proj.{bias, weight} transformer.decoder.layers.0.cross_attn.sampling_offsets.{bias, weight} transformer.decoder.layers.0.cross_attn.value_proj.{bias, weight} transformer.decoder.layers.0.linear1.{bias, weight} transformer.decoder.layers.0.linear2.{bias, weight} transformer.decoder.layers.0.norm1.{bias, weight} transformer.decoder.layers.0.norm2.{bias, weight} transformer.decoder.layers.0.norm3.{bias, weight} transformer.decoder.layers.0.self_attn.out_proj.{bias, weight} transformer.decoder.layers.0.self_attn.{in_proj_bias, in_proj_weight} transformer.decoder.layers.1.cross_attn.attention_weights.{bias, weight} transformer.decoder.layers.1.cross_attn.output_proj.{bias, weight} transformer.decoder.layers.1.cross_attn.sampling_offsets.{bias, weight} transformer.decoder.layers.1.cross_attn.value_proj.{bias, weight} transformer.decoder.layers.1.linear1.{bias, weight} transformer.decoder.layers.1.linear2.{bias, weight} transformer.decoder.layers.1.norm1.{bias, weight} transformer.decoder.layers.1.norm2.{bias, weight} transformer.decoder.layers.1.norm3.{bias, weight} transformer.decoder.layers.1.self_attn.out_proj.{bias, weight} transformer.decoder.layers.1.self_attn.{in_proj_bias, in_proj_weight} transformer.decoder.layers.2.cross_attn.attention_weights.{bias, weight} transformer.decoder.layers.2.cross_attn.output_proj.{bias, weight} transformer.decoder.layers.2.cross_attn.sampling_offsets.{bias, weight} transformer.decoder.layers.2.cross_attn.value_proj.{bias, weight} transformer.decoder.layers.2.linear1.{bias, weight} transformer.decoder.layers.2.linear2.{bias, weight} transformer.decoder.layers.2.norm1.{bias, weight} transformer.decoder.layers.2.norm2.{bias, weight} transformer.decoder.layers.2.norm3.{bias, weight} transformer.decoder.layers.2.self_attn.out_proj.{bias, weight} transformer.decoder.layers.2.self_attn.{in_proj_bias, in_proj_weight} transformer.decoder.layers.3.cross_attn.attention_weights.{bias, weight} transformer.decoder.layers.3.cross_attn.output_proj.{bias, weight} transformer.decoder.layers.3.cross_attn.sampling_offsets.{bias, weight} transformer.decoder.layers.3.cross_attn.value_proj.{bias, weight} transformer.decoder.layers.3.linear1.{bias, weight} transformer.decoder.layers.3.linear2.{bias, weight} transformer.decoder.layers.3.norm1.{bias, weight} transformer.decoder.layers.3.norm2.{bias, weight} transformer.decoder.layers.3.norm3.{bias, weight} transformer.decoder.layers.3.self_attn.out_proj.{bias, weight} transformer.decoder.layers.3.self_attn.{in_proj_bias, in_proj_weight} transformer.decoder.layers.4.cross_attn.attention_weights.{bias, weight} transformer.decoder.layers.4.cross_attn.output_proj.{bias, weight} transformer.decoder.layers.4.cross_attn.sampling_offsets.{bias, weight} transformer.decoder.layers.4.cross_attn.value_proj.{bias, weight} transformer.decoder.layers.4.linear1.{bias, weight} transformer.decoder.layers.4.linear2.{bias, weight} transformer.decoder.layers.4.norm1.{bias, weight} transformer.decoder.layers.4.norm2.{bias, weight} transformer.decoder.layers.4.norm3.{bias, weight} transformer.decoder.layers.4.self_attn.out_proj.{bias, weight} transformer.decoder.layers.4.self_attn.{in_proj_bias, in_proj_weight} transformer.decoder.layers.5.cross_attn.attention_weights.{bias, weight} transformer.decoder.layers.5.cross_attn.output_proj.{bias, weight} transformer.decoder.layers.5.cross_attn.sampling_offsets.{bias, weight} transformer.decoder.layers.5.cross_attn.value_proj.{bias, weight} transformer.decoder.layers.5.linear1.{bias, weight} transformer.decoder.layers.5.linear2.{bias, weight} transformer.decoder.layers.5.norm1.{bias, weight} transformer.decoder.layers.5.norm2.{bias, weight} transformer.decoder.layers.5.norm3.{bias, weight} transformer.decoder.layers.5.self_attn.out_proj.{bias, weight} transformer.decoder.layers.5.self_attn.{in_proj_bias, in_proj_weight} transformer.encoder.layers.0.linear1.{bias, weight} transformer.encoder.layers.0.linear2.{bias, weight} transformer.encoder.layers.0.norm1.{bias, weight} transformer.encoder.layers.0.norm2.{bias, weight} transformer.encoder.layers.0.self_attn.attention_weights.{bias, weight} transformer.encoder.layers.0.self_attn.output_proj.{bias, weight} transformer.encoder.layers.0.self_attn.sampling_offsets.{bias, weight} transformer.encoder.layers.0.self_attn.value_proj.{bias, weight} transformer.encoder.layers.1.linear1.{bias, weight} transformer.encoder.layers.1.linear2.{bias, weight} transformer.encoder.layers.1.norm1.{bias, weight} transformer.encoder.layers.1.norm2.{bias, weight} transformer.encoder.layers.1.self_attn.attention_weights.{bias, weight} transformer.encoder.layers.1.self_attn.output_proj.{bias, weight} transformer.encoder.layers.1.self_attn.sampling_offsets.{bias, weight} transformer.encoder.layers.1.self_attn.value_proj.{bias, weight} transformer.encoder.layers.2.linear1.{bias, weight} transformer.encoder.layers.2.linear2.{bias, weight} transformer.encoder.layers.2.norm1.{bias, weight} transformer.encoder.layers.2.norm2.{bias, weight} transformer.encoder.layers.2.self_attn.attention_weights.{bias, weight} transformer.encoder.layers.2.self_attn.output_proj.{bias, weight} transformer.encoder.layers.2.self_attn.sampling_offsets.{bias, weight} transformer.encoder.layers.2.self_attn.value_proj.{bias, weight} transformer.encoder.layers.3.linear1.{bias, weight} transformer.encoder.layers.3.linear2.{bias, weight} transformer.encoder.layers.3.norm1.{bias, weight} transformer.encoder.layers.3.norm2.{bias, weight} transformer.encoder.layers.3.self_attn.attention_weights.{bias, weight} transformer.encoder.layers.3.self_attn.output_proj.{bias, weight} transformer.encoder.layers.3.self_attn.sampling_offsets.{bias, weight} transformer.encoder.layers.3.self_attn.value_proj.{bias, weight} transformer.encoder.layers.4.linear1.{bias, weight} transformer.encoder.layers.4.linear2.{bias, weight} transformer.encoder.layers.4.norm1.{bias, weight} transformer.encoder.layers.4.norm2.{bias, weight} transformer.encoder.layers.4.self_attn.attention_weights.{bias, weight} transformer.encoder.layers.4.self_attn.output_proj.{bias, weight} transformer.encoder.layers.4.self_attn.sampling_offsets.{bias, weight} transformer.encoder.layers.4.self_attn.value_proj.{bias, weight} transformer.encoder.layers.5.linear1.{bias, weight} transformer.encoder.layers.5.linear2.{bias, weight} transformer.encoder.layers.5.norm1.{bias, weight} transformer.encoder.layers.5.norm2.{bias, weight} transformer.encoder.layers.5.self_attn.attention_weights.{bias, weight} transformer.encoder.layers.5.self_attn.output_proj.{bias, weight} transformer.encoder.layers.5.self_attn.sampling_offsets.{bias, weight} transformer.encoder.layers.5.self_attn.value_proj.{bias, weight} transformer.level_embed.
    是否是因为这样才导致训练的模型不好,在推理的时候没有结果?

  10. fengxiuyaun commented on Apr 10, 2024

    @fengxiuyaun
    Contributor
    1. 测试好像没有对齐,不同环境测试会有一些差别,但差别不会很大。这个是我之前跑出来的结果:HOTA DetA AssA DetRe DetPr AssRe AssPr LocA RHOTA HOTA(0) LocA(0) HOTALocA(0) MOTA MOTP MODA CLR_Re CLR_Pr MTR PTR MLR CLR_TP CLR_FN CLR_FP IDSW MT PT ML Frag sMOTA IDF1 IDR IDP IDTP IDFN IDFP Dets GT_Dets IDs GT_IDs
      65.925 77.589 56.232 83.03 89.167 60.291 81.243 90.792 68.317 74.529 88.032 65.609 87.115 89.706 87.696 90.406 97.089 75.092 23.077 1.8315 203548 21600 6103 1308 205 63 5 3712 77.808 68.824 66.456 71.368 149624 75524 60027 209651 225148 671 273
    2. detrex感觉训练改动了很多呢,我记得我在push的时候已经验证过是可以正常训练的,不知道为啥会有这么多错误。会不是是环境不一样。
    3. 由于时间太长,训练log已经删掉了,总感觉哪里有问题。
      我建议还是先对齐测试,通过测试来到环境或者代码有啥问题。假如方便可以加我WX:bphengyan2,这样方便沟通
  11. fufuforu commented on Mar 18, 2025

    @fufuforu

    您好!怎么在八张卡上面跑呢?通过命令行吗?我使用“./tools/simple_inference.sh configs/motrv2ch_uni5cost3ggoon.args ./pretrained_weight/checkpoint0019.pth”可以成功在DanceTrack test上测试,但是只用了一张卡

  12. fengxiuyaun commented on Mar 18, 2025

    @fengxiuyaun
    Contributor

    是的,代码测试只支持单卡。不过可以把代码改为分布式

  13. fufuforu commented on Mar 18, 2025

    @fufuforu

    谢谢您!我在train.sh里看到了支持四张卡训练的代码“python -m torch.distributed.launch --nproc_per_node=4 --master_port 29504 --use_env main.py ${args} --output_dir $OUTPUT_DIR |& tee -a output.log”,我去模仿一下,试试应用到代码测试部分!感谢您的工作以及回复

  14. fufuforu commented on Apr 8, 2025

    @fufuforu

    在使用detrex训练的时候,直接用那个代码会报错,错误及我的修改如下:

    1. TypeError: repr() got an unexpected keyword argument '_repr_indent'
      截屏2024-04-09 20 11 01
      我将之修改为:
      截屏2024-04-09 20 12 14
    2. MS- deformable-attension报错
      截屏2024-04-09 20 18 13
      我将之改为:
      截屏2024-04-09 20 50 07
    3. 形参问题:
      截屏2024-04-09 20 32 16
      我将之改为:
      截屏2024-04-09 20 34 49
    4. 函数缺失
      截屏2024-04-09 20 43 49
      我将本仓库的merge函数加入进去:
      截屏2024-04-09 20 44 46

    至此,可以成功训练. 我使用的预训练参数是detrex中Deformable-DETR-R50 + Box-Refinement版本, 在加载的时候有很多不匹配的层: WARNING [04/09 18:46:39 fvcore.common.checkpoint]: Some model parameters or buffers are not found in the checkpoint: class_embed.0.{bias, weight} class_embed.1.{bias, weight} class_embed.2.{bias, weight} class_embed.3.{bias, weight} class_embed.4.{bias, weight} class_embed.5.{bias, weight} position.weight position_offset.weight query_embed.weight query_embed_offset.weight track_embed.linear1.{bias, weight} track_embed.linear2.{bias, weight} track_embed.linear_feat1.{bias, weight} track_embed.linear_feat2.{bias, weight} track_embed.norm1.{bias, weight} track_embed.norm2.{bias, weight} track_embed.norm_feat.{bias, weight} track_embed.self_attn.out_proj.{bias, weight} track_embed.self_attn.{in_proj_bias, in_proj_weight} transformer.decoder.class_embed.0.{bias, weight} transformer.decoder.class_embed.1.{bias, weight} transformer.decoder.class_embed.2.{bias, weight} transformer.decoder.class_embed.3.{bias, weight} transformer.decoder.class_embed.4.{bias, weight} transformer.decoder.class_embed.5.{bias, weight} transformer.decoder.layers.0.cross_attn.attention_weights.{bias, weight} transformer.decoder.layers.0.cross_attn.output_proj.{bias, weight} transformer.decoder.layers.0.cross_attn.sampling_offsets.{bias, weight} transformer.decoder.layers.0.cross_attn.value_proj.{bias, weight} transformer.decoder.layers.0.linear1.{bias, weight} transformer.decoder.layers.0.linear2.{bias, weight} transformer.decoder.layers.0.norm1.{bias, weight} transformer.decoder.layers.0.norm2.{bias, weight} transformer.decoder.layers.0.norm3.{bias, weight} transformer.decoder.layers.0.self_attn.out_proj.{bias, weight} transformer.decoder.layers.0.self_attn.{in_proj_bias, in_proj_weight} transformer.decoder.layers.1.cross_attn.attention_weights.{bias, weight} transformer.decoder.layers.1.cross_attn.output_proj.{bias, weight} transformer.decoder.layers.1.cross_attn.sampling_offsets.{bias, weight} transformer.decoder.layers.1.cross_attn.value_proj.{bias, weight} transformer.decoder.layers.1.linear1.{bias, weight} transformer.decoder.layers.1.linear2.{bias, weight} transformer.decoder.layers.1.norm1.{bias, weight} transformer.decoder.layers.1.norm2.{bias, weight} transformer.decoder.layers.1.norm3.{bias, weight} transformer.decoder.layers.1.self_attn.out_proj.{bias, weight} transformer.decoder.layers.1.self_attn.{in_proj_bias, in_proj_weight} transformer.decoder.layers.2.cross_attn.attention_weights.{bias, weight} transformer.decoder.layers.2.cross_attn.output_proj.{bias, weight} transformer.decoder.layers.2.cross_attn.sampling_offsets.{bias, weight} transformer.decoder.layers.2.cross_attn.value_proj.{bias, weight} transformer.decoder.layers.2.linear1.{bias, weight} transformer.decoder.layers.2.linear2.{bias, weight} transformer.decoder.layers.2.norm1.{bias, weight} transformer.decoder.layers.2.norm2.{bias, weight} transformer.decoder.layers.2.norm3.{bias, weight} transformer.decoder.layers.2.self_attn.out_proj.{bias, weight} transformer.decoder.layers.2.self_attn.{in_proj_bias, in_proj_weight} transformer.decoder.layers.3.cross_attn.attention_weights.{bias, weight} transformer.decoder.layers.3.cross_attn.output_proj.{bias, weight} transformer.decoder.layers.3.cross_attn.sampling_offsets.{bias, weight} transformer.decoder.layers.3.cross_attn.value_proj.{bias, weight} transformer.decoder.layers.3.linear1.{bias, weight} transformer.decoder.layers.3.linear2.{bias, weight} transformer.decoder.layers.3.norm1.{bias, weight} transformer.decoder.layers.3.norm2.{bias, weight} transformer.decoder.layers.3.norm3.{bias, weight} transformer.decoder.layers.3.self_attn.out_proj.{bias, weight} transformer.decoder.layers.3.self_attn.{in_proj_bias, in_proj_weight} transformer.decoder.layers.4.cross_attn.attention_weights.{bias, weight} transformer.decoder.layers.4.cross_attn.output_proj.{bias, weight} transformer.decoder.layers.4.cross_attn.sampling_offsets.{bias, weight} transformer.decoder.layers.4.cross_attn.value_proj.{bias, weight} transformer.decoder.layers.4.linear1.{bias, weight} transformer.decoder.layers.4.linear2.{bias, weight} transformer.decoder.layers.4.norm1.{bias, weight} transformer.decoder.layers.4.norm2.{bias, weight} transformer.decoder.layers.4.norm3.{bias, weight} transformer.decoder.layers.4.self_attn.out_proj.{bias, weight} transformer.decoder.layers.4.self_attn.{in_proj_bias, in_proj_weight} transformer.decoder.layers.5.cross_attn.attention_weights.{bias, weight} transformer.decoder.layers.5.cross_attn.output_proj.{bias, weight} transformer.decoder.layers.5.cross_attn.sampling_offsets.{bias, weight} transformer.decoder.layers.5.cross_attn.value_proj.{bias, weight} transformer.decoder.layers.5.linear1.{bias, weight} transformer.decoder.layers.5.linear2.{bias, weight} transformer.decoder.layers.5.norm1.{bias, weight} transformer.decoder.layers.5.norm2.{bias, weight} transformer.decoder.layers.5.norm3.{bias, weight} transformer.decoder.layers.5.self_attn.out_proj.{bias, weight} transformer.decoder.layers.5.self_attn.{in_proj_bias, in_proj_weight} transformer.encoder.layers.0.linear1.{bias, weight} transformer.encoder.layers.0.linear2.{bias, weight} transformer.encoder.layers.0.norm1.{bias, weight} transformer.encoder.layers.0.norm2.{bias, weight} transformer.encoder.layers.0.self_attn.attention_weights.{bias, weight} transformer.encoder.layers.0.self_attn.output_proj.{bias, weight} transformer.encoder.layers.0.self_attn.sampling_offsets.{bias, weight} transformer.encoder.layers.0.self_attn.value_proj.{bias, weight} transformer.encoder.layers.1.linear1.{bias, weight} transformer.encoder.layers.1.linear2.{bias, weight} transformer.encoder.layers.1.norm1.{bias, weight} transformer.encoder.layers.1.norm2.{bias, weight} transformer.encoder.layers.1.self_attn.attention_weights.{bias, weight} transformer.encoder.layers.1.self_attn.output_proj.{bias, weight} transformer.encoder.layers.1.self_attn.sampling_offsets.{bias, weight} transformer.encoder.layers.1.self_attn.value_proj.{bias, weight} transformer.encoder.layers.2.linear1.{bias, weight} transformer.encoder.layers.2.linear2.{bias, weight} transformer.encoder.layers.2.norm1.{bias, weight} transformer.encoder.layers.2.norm2.{bias, weight} transformer.encoder.layers.2.self_attn.attention_weights.{bias, weight} transformer.encoder.layers.2.self_attn.output_proj.{bias, weight} transformer.encoder.layers.2.self_attn.sampling_offsets.{bias, weight} transformer.encoder.layers.2.self_attn.value_proj.{bias, weight} transformer.encoder.layers.3.linear1.{bias, weight} transformer.encoder.layers.3.linear2.{bias, weight} transformer.encoder.layers.3.norm1.{bias, weight} transformer.encoder.layers.3.norm2.{bias, weight} transformer.encoder.layers.3.self_attn.attention_weights.{bias, weight} transformer.encoder.layers.3.self_attn.output_proj.{bias, weight} transformer.encoder.layers.3.self_attn.sampling_offsets.{bias, weight} transformer.encoder.layers.3.self_attn.value_proj.{bias, weight} transformer.encoder.layers.4.linear1.{bias, weight} transformer.encoder.layers.4.linear2.{bias, weight} transformer.encoder.layers.4.norm1.{bias, weight} transformer.encoder.layers.4.norm2.{bias, weight} transformer.encoder.layers.4.self_attn.attention_weights.{bias, weight} transformer.encoder.layers.4.self_attn.output_proj.{bias, weight} transformer.encoder.layers.4.self_attn.sampling_offsets.{bias, weight} transformer.encoder.layers.4.self_attn.value_proj.{bias, weight} transformer.encoder.layers.5.linear1.{bias, weight} transformer.encoder.layers.5.linear2.{bias, weight} transformer.encoder.layers.5.norm1.{bias, weight} transformer.encoder.layers.5.norm2.{bias, weight} transformer.encoder.layers.5.self_attn.attention_weights.{bias, weight} transformer.encoder.layers.5.self_attn.output_proj.{bias, weight} transformer.encoder.layers.5.self_attn.sampling_offsets.{bias, weight} transformer.encoder.layers.5.self_attn.value_proj.{bias, weight} transformer.level_embed. 是否是因为这样才导致训练的模型不好,在推理的时候没有结果?

    您好!第四步缺失的merge函数,在哪里能找到呢?

  15. fengxiuyaun commented on Apr 8, 2025

    @fengxiuyaun
    Contributor

    感觉是[Deformable-DETR-R50 + Box-Refinement]与comot实用的backbone不是同一个。

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions

      0