It doesn't reproduce to your accuracy

[ZhangGYGitHub]

1. First I downloaded your Pascal VOC Split 1 After Base Training: Weights
2. and then fine-tuned it in voc_fewshot_split1 seed1
3. But the results are down five points from the paper。

4. Fine-tuning other shots did not achieve the accuracy in the paper. On average, they're down about five points I'd appreciate it if you could tell me why 。I hope you can reply to me as soon as possible --dataset_file voc_base1 \ fewshot_seed=01 num_shot=01 gpu 3090 torch 1.7.1 gcc 7.5.0 Averaged stats: class_error: 16.67 loss: 13.4678 (13.6217) loss_ce: 1.0068 (1.0683) loss_bbox: 0.7345 (0.6811) loss_giou: 0.5437 (0.6277) loss_category_codes_cls: 0.0205 (0.0205) loss_ce_0: 1.0740 (1.0708) loss_bbox_0: 0.7470 (0.6255) loss_giou_0: 0.5046 (0.5959) loss_ce_1: 1.1113 (1.0220) loss_bbox_1: 0.6579 (0.6045) loss_giou_1: 0.4740 (0.5838) loss_ce_2: 1.0879 (1.0172) loss_bbox_2: 0.6594 (0.6091) loss_giou_2: 0.5281 (0.5844) loss_ce_3: 1.0211 (1.0271) loss_bbox_3: 0.6782 (0.6267) loss_giou_3: 0.4912 (0.5870) loss_ce_4: 1.0229 (1.0482) loss_bbox_4: 0.7843 (0.6298) loss_giou_4: 0.5092 (0.5920) loss_ce_unscaled: 0.5034 (0.5341) class_error_unscaled: 0.0000 (9.1140) loss_bbox_unscaled: 0.1469 (0.1362) loss_giou_unscaled: 0.2718 (0.3138) cardinality_error_unscaled: 279.0000 (271.8738) loss_category_codes_cls_unscaled: 0.0041 (0.0041) loss_ce_0_unscaled: 0.5370 (0.5354) loss_bbox_0_unscaled: 0.1494 (0.1251) loss_giou_0_unscaled: 0.2523 (0.2979) cardinality_error_0_unscaled: 294.7500 (293.1469) loss_ce_1_unscaled: 0.5556 (0.5110) loss_bbox_1_unscaled: 0.1316 (0.1209) loss_giou_1_unscaled: 0.2370 (0.2919) cardinality_error_1_unscaled: 297.8750 (295.3336) loss_ce_2_unscaled: 0.5439 (0.5086) loss_bbox_2_unscaled: 0.1319 (0.1218) loss_giou_2_unscaled: 0.2641 (0.2922) cardinality_error_2_unscaled: 295.2500 (292.1838) loss_ce_3_unscaled: 0.5105 (0.5136) loss_bbox_3_unscaled: 0.1356 (0.1253) loss_giou_3_unscaled: 0.2456 (0.2935) cardinality_error_3_unscaled: 295.3750 (292.3367) loss_ce_4_unscaled: 0.5114 (0.5241) loss_bbox_4_unscaled: 0.1569 (0.1260) loss_giou_4_unscaled: 0.2546 (0.2960) cardinality_error_4_unscaled: 290.6250 (288.4062)

   • Novel Categories: Accumulating evaluation results... DONE (t=2.62s). IoU metric: bbox Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.177 Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.304 Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.177

[ZhangGongjie]

After fine-tuning, the log should be like this: (this is my experimental results from seed01 1-shot on Pascal VOC split 1)

{"train_lr": 1.5000000000000005e-06, "train_class_error": 0.0, "train_grad_norm": 58.47532653808594, "train_loss": 1.4763914942741394, "train_loss_bbox": 0.07749657705426216, "train_loss_bbox_0": 0.14321566373109818, "train_loss_bbox_1": 0.09320082515478134, "train_loss_bbox_2": 0.0857505314052105, "train_loss_bbox_3": 0.08199029043316841, "train_loss_bbox_4": 0.07725301757454872, "train_loss_category_codes_cls": 0.11686665192246437, "train_loss_ce": 0.008141814265400171, "train_loss_ce_0": 0.10580113902688026, "train_loss_ce_1": 0.047131314873695374, "train_loss_ce_2": 0.017384656239300966, "train_loss_ce_3": 0.01357174920849502, "train_loss_ce_4": 0.009504480753093958, "train_loss_giou": 0.08753593638539314, "train_loss_giou_0": 0.1434987187385559, "train_loss_giou_1": 0.10159203037619591, "train_loss_giou_2": 0.0932667925953865, "train_loss_giou_3": 0.08680127188563347, "train_loss_giou_4": 0.08638809621334076, "train_cardinality_error_unscaled": 287.76251220703125, "train_cardinality_error_0_unscaled": 287.9250030517578, "train_cardinality_error_1_unscaled": 289.0375061035156, "train_cardinality_error_2_unscaled": 289.54376220703125, "train_cardinality_error_3_unscaled": 289.48126220703125, "train_cardinality_error_4_unscaled": 289.1625061035156, "train_class_error_unscaled": 0.0, "train_loss_bbox_unscaled": 0.015499315224587917, "train_loss_bbox_0_unscaled": 0.02864313218742609, "train_loss_bbox_1_unscaled": 0.018640165217220783, "train_loss_bbox_2_unscaled": 0.017150106839835644, "train_loss_bbox_3_unscaled": 0.016398058272898197, "train_loss_bbox_4_unscaled": 0.01545060332864523, "train_loss_category_codes_cls_unscaled": 0.02337333094328642, "train_loss_ce_unscaled": 0.004070907132700086, "train_loss_ce_0_unscaled": 0.05290056951344013, "train_loss_ce_1_unscaled": 0.023565657436847687, "train_loss_ce_2_unscaled": 0.008692328119650483, "train_loss_ce_3_unscaled": 0.00678587460424751, "train_loss_ce_4_unscaled": 0.004752240376546979, "train_loss_giou_unscaled": 0.04376796819269657, "train_loss_giou_0_unscaled": 0.07174935936927795, "train_loss_giou_1_unscaled": 0.050796015188097954, "train_loss_giou_2_unscaled": 0.04663339629769325, "train_loss_giou_3_unscaled": 0.043400635942816734, "train_loss_giou_4_unscaled": 0.04319404810667038, "test_class_error": 5.472572386649347, "test_loss": 11.130939615926435, "test_loss_bbox": 0.5066791164298211, "test_loss_bbox_0": 0.5212143290427423, "test_loss_bbox_1": 0.5106292650584252, "test_loss_bbox_2": 0.5027764009852563, "test_loss_bbox_3": 0.49650043556767126, "test_loss_bbox_4": 0.4928392595821811, "test_loss_category_codes_cls": 0.06833156943321228, "test_loss_ce": 0.799778366617618, "test_loss_ce_0": 0.7964376487078205, "test_loss_ce_1": 0.7633872688297303, "test_loss_ce_2": 0.7538318927249601, "test_loss_ce_3": 0.7609124298537931, "test_loss_ce_4": 0.7619222026678824, "test_loss_giou": 0.5621161045566682, "test_loss_giou_0": 0.5845255267235541, "test_loss_giou_1": 0.5759262894430468, "test_loss_giou_2": 0.5661270769373064, "test_loss_giou_3": 0.5565041990049424, "test_loss_giou_4": 0.5505002444790256, "test_cardinality_error_unscaled": 286.87172379032256, "test_cardinality_error_0_unscaled": 293.8604334677419, "test_cardinality_error_1_unscaled": 293.77363911290325, "test_cardinality_error_2_unscaled": 293.6239415322581, "test_cardinality_error_3_unscaled": 292.1960181451613, "test_cardinality_error_4_unscaled": 290.13714717741937, "test_class_error_unscaled": 5.472572386649347, "test_loss_bbox_unscaled": 0.10133582312733896, "test_loss_bbox_0_unscaled": 0.10424286586142355, "test_loss_bbox_1_unscaled": 0.10212585292756557, "test_loss_bbox_2_unscaled": 0.10055528025954, "test_loss_bbox_3_unscaled": 0.0993000871952503, "test_loss_bbox_4_unscaled": 0.09856785183712359, "test_loss_category_codes_cls_unscaled": 0.013666314072906971, "test_loss_ce_unscaled": 0.399889183308809, "test_loss_ce_0_unscaled": 0.39821882435391026, "test_loss_ce_1_unscaled": 0.38169363441486515, "test_loss_ce_2_unscaled": 0.37691594636248005, "test_loss_ce_3_unscaled": 0.3804562149268966, "test_loss_ce_4_unscaled": 0.3809611013339412, "test_loss_giou_unscaled": 0.2810580522783341, "test_loss_giou_0_unscaled": 0.29226276336177703, "test_loss_giou_1_unscaled": 0.2879631447215234, "test_loss_giou_2_unscaled": 0.2830635384686532, "test_loss_giou_3_unscaled": 0.2782520995024712, "test_loss_giou_4_unscaled": 0.2752501222395128, "test_coco_eval_bbox": [0.2237539923834384, 0.36212298616476496, 0.23457736703757642, 0.02539442618386005, 0.13298507772985532, 0.28761370973175865, 0.264984217870063, 0.4056139382126699, 0.4253965281576116, 0.10106334841628958, 0.23122574383771002, 0.5370109415978369], "epoch": 699, "n_parameters": 51664028, "evaltype": "novel"}

Your loss after fine-tuning does not seem right. Please check your settings. If you still encounter issues, I would like to help.

However, do take note that I am recently looking for jobs, so I might not be able to check and respond very quickly.

[ZhangGYGitHub]

After fine-tuning, the log should be like this: (this is my experimental results from seed01 1-shot on Pascal VOC split 1)

{"train_lr": 1.5000000000000005e-06, "train_class_error": 0.0, "train_grad_norm": 58.47532653808594, "train_loss": 1.4763914942741394, "train_loss_bbox": 0.07749657705426216, "train_loss_bbox_0": 0.14321566373109818, "train_loss_bbox_1": 0.09320082515478134, "train_loss_bbox_2": 0.0857505314052105, "train_loss_bbox_3": 0.08199029043316841, "train_loss_bbox_4": 0.07725301757454872, "train_loss_category_codes_cls": 0.11686665192246437, "train_loss_ce": 0.008141814265400171, "train_loss_ce_0": 0.10580113902688026, "train_loss_ce_1": 0.047131314873695374, "train_loss_ce_2": 0.017384656239300966, "train_loss_ce_3": 0.01357174920849502, "train_loss_ce_4": 0.009504480753093958, "train_loss_giou": 0.08753593638539314, "train_loss_giou_0": 0.1434987187385559, "train_loss_giou_1": 0.10159203037619591, "train_loss_giou_2": 0.0932667925953865, "train_loss_giou_3": 0.08680127188563347, "train_loss_giou_4": 0.08638809621334076, "train_cardinality_error_unscaled": 287.76251220703125, "train_cardinality_error_0_unscaled": 287.9250030517578, "train_cardinality_error_1_unscaled": 289.0375061035156, "train_cardinality_error_2_unscaled": 289.54376220703125, "train_cardinality_error_3_unscaled": 289.48126220703125, "train_cardinality_error_4_unscaled": 289.1625061035156, "train_class_error_unscaled": 0.0, "train_loss_bbox_unscaled": 0.015499315224587917, "train_loss_bbox_0_unscaled": 0.02864313218742609, "train_loss_bbox_1_unscaled": 0.018640165217220783, "train_loss_bbox_2_unscaled": 0.017150106839835644, "train_loss_bbox_3_unscaled": 0.016398058272898197, "train_loss_bbox_4_unscaled": 0.01545060332864523, "train_loss_category_codes_cls_unscaled": 0.02337333094328642, "train_loss_ce_unscaled": 0.004070907132700086, "train_loss_ce_0_unscaled": 0.05290056951344013, "train_loss_ce_1_unscaled": 0.023565657436847687, "train_loss_ce_2_unscaled": 0.008692328119650483, "train_loss_ce_3_unscaled": 0.00678587460424751, "train_loss_ce_4_unscaled": 0.004752240376546979, "train_loss_giou_unscaled": 0.04376796819269657, "train_loss_giou_0_unscaled": 0.07174935936927795, "train_loss_giou_1_unscaled": 0.050796015188097954, "train_loss_giou_2_unscaled": 0.04663339629769325, "train_loss_giou_3_unscaled": 0.043400635942816734, "train_loss_giou_4_unscaled": 0.04319404810667038, "test_class_error": 5.472572386649347, "test_loss": 11.130939615926435, "test_loss_bbox": 0.5066791164298211, "test_loss_bbox_0": 0.5212143290427423, "test_loss_bbox_1": 0.5106292650584252, "test_loss_bbox_2": 0.5027764009852563, "test_loss_bbox_3": 0.49650043556767126, "test_loss_bbox_4": 0.4928392595821811, "test_loss_category_codes_cls": 0.06833156943321228, "test_loss_ce": 0.799778366617618, "test_loss_ce_0": 0.7964376487078205, "test_loss_ce_1": 0.7633872688297303, "test_loss_ce_2": 0.7538318927249601, "test_loss_ce_3": 0.7609124298537931, "test_loss_ce_4": 0.7619222026678824, "test_loss_giou": 0.5621161045566682, "test_loss_giou_0": 0.5845255267235541, "test_loss_giou_1": 0.5759262894430468, "test_loss_giou_2": 0.5661270769373064, "test_loss_giou_3": 0.5565041990049424, "test_loss_giou_4": 0.5505002444790256, "test_cardinality_error_unscaled": 286.87172379032256, "test_cardinality_error_0_unscaled": 293.8604334677419, "test_cardinality_error_1_unscaled": 293.77363911290325, "test_cardinality_error_2_unscaled": 293.6239415322581, "test_cardinality_error_3_unscaled": 292.1960181451613, "test_cardinality_error_4_unscaled": 290.13714717741937, "test_class_error_unscaled": 5.472572386649347, "test_loss_bbox_unscaled": 0.10133582312733896, "test_loss_bbox_0_unscaled": 0.10424286586142355, "test_loss_bbox_1_unscaled": 0.10212585292756557, "test_loss_bbox_2_unscaled": 0.10055528025954, "test_loss_bbox_3_unscaled": 0.0993000871952503, "test_loss_bbox_4_unscaled": 0.09856785183712359, "test_loss_category_codes_cls_unscaled": 0.013666314072906971, "test_loss_ce_unscaled": 0.399889183308809, "test_loss_ce_0_unscaled": 0.39821882435391026, "test_loss_ce_1_unscaled": 0.38169363441486515, "test_loss_ce_2_unscaled": 0.37691594636248005, "test_loss_ce_3_unscaled": 0.3804562149268966, "test_loss_ce_4_unscaled": 0.3809611013339412, "test_loss_giou_unscaled": 0.2810580522783341, "test_loss_giou_0_unscaled": 0.29226276336177703, "test_loss_giou_1_unscaled": 0.2879631447215234, "test_loss_giou_2_unscaled": 0.2830635384686532, "test_loss_giou_3_unscaled": 0.2782520995024712, "test_loss_giou_4_unscaled": 0.2752501222395128, "test_coco_eval_bbox": [0.2237539923834384, 0.36212298616476496, 0.23457736703757642, 0.02539442618386005, 0.13298507772985532, 0.28761370973175865, 0.264984217870063, 0.4056139382126699, 0.4253965281576116, 0.10106334841628958, 0.23122574383771002, 0.5370109415978369], "epoch": 699, "n_parameters": 51664028, "evaltype": "novel"}

Your loss after fine-tuning does not seem right. Please check your settings. If you still encounter issues, I would like to help.

However, do take note that I am recently looking for jobs, so I might not be able to check and respond very quickly.

十分感谢您的及时回复，我是按照您给出的run_experiments_voc1_50epoch.sh 中28行以后的代码进行的设置，其他并未改动，如果您知道问题的原因，感谢您的告知！

EXP_DIR=exps/voc1 BASE_TRAIN_DIR=${EXP_DIR}/base_train mkdir exps mkdir ${EXP_DIR} mkdir ${BASE_TRAIN_DIR}

fewshot_seed=01 num_shot=10 epoch=500 lr_drop1=300 lr_drop2=450 FS_FT_DIR=${EXP_DIR}/seed${fewshot_seed}${num_shot}shot mkdir ${FS_FT_DIR} for fewshot_seed in 01 do for num_shot in 01 do FS_FT_DIR=${EXP_DIR}/seed${fewshot_seed}${num_shot}shot mkdir ${FS_FT_DIR}

if [ $num_shot -eq 1 ] then epoch=700 lr_drop1=350 lr_drop2=600 lr=1.5e-4 lr_backbone=1.5e-5 elif [ $num_shot -eq 2 ] then epoch=600 lr_drop1=300 lr_drop2=550 lr=1e-4 lr_backbone=1e-5 elif [ $num_shot -eq 3 ] then epoch=600 lr_drop1=300 lr_drop2=550 lr=1e-4 lr_backbone=1e-5 elif [ $num_shot -eq 5 ] then epoch=500 lr_drop1=250 lr_drop2=450 lr=5e-5 lr_backbone=5e-6 elif [ $num_shot -eq 10 ] then epoch=500 lr_drop1=250 lr_drop2=450 lr=5e-5 lr_backbone=5e-6 else exit fi

python -u main.py \ --dataset_file voc_base1 \ --backbone resnet101 \ --num_feature_levels 1 \ --enc_layers 6 \ --dec_layers 6 \ --hidden_dim 256 \ --num_queries 300 \ --batch_size 2 \ --lr ${lr} \ --lr_backbone ${lr_backbone} \ --category_codes_cls_loss \ --resume ${BASE_TRAIN_DIR}/checkpoint.pth \ --fewshot_finetune \ --fewshot_seed ${fewshot_seed} \ --num_shots ${num_shot} \ --epoch ${epoch} \ --lr_drop_milestones ${lr_drop1} ${lr_drop2} \ --warmup_epochs 50 \ --save_every_epoch ${epoch} \ --eval_every_epoch ${epoch} \ --output_dir ${FS_FT_DIR} \ 2>&1 | tee ${FS_FT_DIR}/log.txt done done

[ZhangGongjie]

These commands seem correct.

Did you use 8 GPUs for finetuning? DETR is known for its convergence issues so large batchsizes are needed.

You should run the commands like this.

GPUS_PER_NODE=8 ./tools/run_dist_launch.sh 8 ./fsfinetune.sh

（fsfinetune.sh contains the above commands）

[ZhangGYGitHub]

包含

很抱歉我的计算资源只有1张3090gpu ，batch_size使用的是您代码中默认值 2，是不是因为单卡训练导致的我目前平均下降5个点的问题呢？如果我想更接近您论文中的精度，您有什么建议吗？

[ZhangGongjie]

是的，我认为精度下降是这个原因。我的经验是，对于DETR的模型，单卡finetune也能运行，但是要达到论文汇报的精度，最好用大batchsize finetune （和base training保持同样的batchsize）。

为了让其他人也能参考这次讨论，英文翻译如下： In order to let this discussion benefit everyone, I provide English translation as below.

I think single-GPU fine-tuning is the reason for declined accuracy. To reach paper's reported accuracy, we need to set the batchsize the same as the base training stage, which takes 8 GPUs for training.

Single GPU finetuning can also produce good results, but tend to be lower than 8-GPU finetuning.

[ZhangGongjie]

Your loss is significantly larger than mine, which itself indicates that the finetuning hasn't fully converged.

To use one GPU to improve performance, you may try the following methods: (1) Use amp (Automatic Mixed Precision). (2) Perform 8 feedforwards, then accumulate the gradients, and then perform back propagation once. This should be theoretically equivalent to our fine-tuning setups.

Thank you :)

[ZhangGYGitHub]

Your loss is significantly larger than mine, which itself indicates that the finetuning hasn't fully converged.

To use one GPU to improve performance, you may try the following methods: (1) Use amp (Automatic Mixed Precision). (2) Perform 8 feedforwards, then accumulate the gradients, and then perform back propagation once. This should be theoretically equivalent to our fine-tuning setups.

Thank you :)

Sorry to bother you here. I used the gradient accumulation you mentioned and still does not reach the accuracy of your paper （1）I used gradients to accumulate eight times and got an accuracy of 39.1 （2）Using both the gradient eight times and amp, the following error occurs, and the gradient cannot be decreasing UserWarning: Detected call of lr scheduler.step() before optimizer.step() （3）I would appreciate it if I could post the fine-tuned weights to my email 1844405443@qq.com hope you can reply to me as soon as possible. thanks

[Wang86760]

很抱歉在这里打扰您。 我同样也使用的3090，未能达到论文中的精度，如果可以能否将微调后的权重发到我的电子邮件中1404993818@qq.com 希望您能尽快回复我，我将不胜感激。谢谢