Tramac

Hello, are there relevant error logs?

This situation occurred because of the lack of visualization. Firstly, the prediction results were saved in `fusion_Img_L2` and `vote_label` (both nparray format) in `train.py.` You can save them in any...

All data in the `data` folder is used for training. Because the data came from a competition website, where had no public test data, data used in test stage was...

Hi, you can try to change the parameter `batch_size=8`, `learning_rate=0.001` in `train.py`.

They are defined [here](https://github.com/Tramac/pytorch-cam/blob/39c35e39d33c45255dbfdd025eea5476a6a5ff55/torchcam/helper.py#L9).

你好，截取或补充至300帧是有可能造成动作丢失或者引入冗余信息的。300帧图像大概覆盖了300 / 25 = 12 秒的视频时长，我觉得可以统计一下K400数据集的平均时长，如果差别不是很大，这么做我认为是可行的。 该repo中所引用的一篇论文是这么处理的，可以作为参考。

在test阶段，如果将视频切割成csv中所提供的只包含动作的时段，是人为的去除了冗余信息，使测试变得更简单了；取three crop计算平均分理论上是更科学的评估方式，更能体现出模型的鲁棒性。个人建议使用第二种方式更好。

> @swhoosh @Gary2018X We would keep using `` as `eos_token` for Bunny-Llama3. @Isaachhh Why should keep using `` as `eos_token`, if so, doesn't the persistent error affect performance? A very...

> > > @swhoosh @Gary2018X We would keep using `` as `eos_token` for Bunny-Llama3. > > > > > > @Isaachhh Why should keep using `` as `eos_token`, if so,...

> It shouldn't be any like `WARNING: tokenization mismatch: 107 vs. 110. (ignored)` because of [here](https://github.com/BAAI-DCAI/Bunny/blob/main/bunny/train/train.py#L233). Which base LLM did you use? Meta-Llama-3-8B-Insturct, here is my `finetune_lora.sh`: ```bash #!/bin/bash MODEL_TYPE=llama3-8b...