Failure to achieve the given mIoU result in the first stage

[1rua11]

Thank you very much for your work! But there are some queries that I would like to get your help on, in the CAM generation step why is the first instruction（python main.py --model deit_small_WeakTr_patch16_224 \ --data-path data \ --data-set VOC12 \ --img-ms-list voc12/train_id.txt \ --cam-npy-dir WeakTr_results/WeakTr/attn-patchrefine-npy \ --output_dir WeakTr_results/WeakTr \ --reduction 8 \ --pool-type max \ --lr 6e-4 \ --weight-decay 0.03 \） using “train_id.txt” instead of “train_aug_id.txt” , also the result 69.4% of the VOC that you have given is using “deit_small_WeakTr_patch16224” or “deit small_WeakTr_AAF_RandWeight_patch16_224” ? I followed the instructions you provided and used “deit_small_WeakTr_patch16_224” and the VOC result for the first step (End-to-End CAM Generation) is 67.7%, not 69.4% .

[Unrealluver]

Thanks for your interest in our work! But I am sorry that I could not fully understand your questions. Could you please list your questions one by one? Or directly describe them in Chinese?

[1rua11]

作者您好！

      好的，抱歉我没有表述清楚。我将用中文表述一下我的疑问。      （1）在训练阶段的第一步端到端CAM生成中所给的VOC的训练结果miou为69.4%，我按照所给的程序和操作没有复现出这个结果，我不清楚问题出在哪里，所以我猜想是否是使用的模型不对，我使用的是     deit_small_WeakTr_patch16_224模型，结果为67.7%。      （2）同样在训练阶段的第一步CAM生成步骤里，VOC训练时使用的是train_id.txt，包含1464张图片，为什么不是使用10582张图片呢？具体指令如下：

祝好！

------------------ 原始邮件 ------------------ 发件人: "hustvl/WeakTr" @.>; 发送时间: 2023年10月11日(星期三) 中午1:22 @.>; @.**@.>; 主题: Re: [hustvl/WeakTr] Failure to achieve the given mIoU result in the first stage (Issue #21)

Thanks for your interest in our work! But I am sorry that I could not fully understand your questions. Could you please list your questions one by one? Or directly describe them in Chinese?

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

[Unrealluver]

Thank you for reaching out and trying our methods!

For your first question, the performance discrepancy might be attributed to variations in hardware specifications. We conducted our training on a Nvidia A400 (16G) GPU. Should your GPU be of a different model or specification, there could be variances in performance outcomes. To mitigate this, you have a couple of options:

• Utilize our pre-trained checkpoint to generate CAM directly, eliminating the need to retrain the model on your hardware.
• Adjust the hyperparameters to optimize performance based on your specific GPU.

Regarding your second question, we only use train_id.txt to evaluate the model's performance during the training process, and the training data is train_aug_id.txt.

If you have any other questions, feel free to communicate with us.

---

感谢您联系并尝试我们的方法！

对于您的第一个问题，性能差异可能归因于硬件规格的变化。我们在 Nvidia A400 (16G) GPU 上进行了训练。如果您的 GPU 具有不同的型号或规格，则性能结果可能会有所不同。为了缓解这种情况，您有几个选择：

• 利用我们预先训练的检查点直接生成 CAM，无需在硬件上重新训练模型。
• 调整超参数以根据您的特定 GPU 优化性能。

关于您的第二个问题，我们仅使用 train_id.txt 来评估模型在训练过程中的性能，训练数据为 train_aug_id.txt。

如果您还有其他问题，请随时与我们沟通。

[Unrealluver]

I will close this issue. If there are more questions, you are welcome to raise issues :)

[zbb1111]

非常感谢您出色的工作，我有几个问题想请教一下，希望得到您的回复！！！！谢谢！！！ 1.对于training，我看您的回答训练的是train_aug_id,但是这里是train_id,是否应该改成rain_aug_id？ 2. Generate CAM，这一步是生成的coarse CAM，还是fine CAM 3. 这一步对应的结果是66.2%？ 4. CRF post-processing，这一步是什么意思？是得到MASK 76.5%这一步吗？