Daoyuan Chen

这是因为基于fvcore和ctx.data_batch =[x, y]的假设不符合，触发了异常。FS在这种场景下会将该值设置为负数，来提醒用户flops并未计算成功，需要实现合适的flops_per_sample函数。可以参考general_torch_trainer的[实现](https://github.com/alibaba/FederatedScope/blob/480b67de851df2fa02b7cc2189f1803d947998d7/federatedscope/core/trainers/torch_trainer.py#L346)，以及相关的[搜索](https://github.com/search?q=repo%3Aalibaba%2FFederatedScope%20flops_per_sample%20%3D%20-1&type=code)

@Alan-Qin zeyu, you can have a look when you are available.

resolved in #359

Thank you for the pull request! To enhance this PR, kindly refer to the [Data-Juicer Developer's Guide](https://github.com/modelscope/data-juicer/blob/main/docs/DeveloperGuide.md#how-to-guide-for-developers) for instructions. Should you have any questions, feel free to ask them here;...

Thanks for your interest! You can find the details on how to use our provided script to train your own classifier, as well as information about its underlying architecture (based...

Resolved in the this [optimization](https://github.com/modelscope/data-juicer/pull/684).

Thank you for reaching out. You can try `pip install modelscope[framework]` to install the necessary download-related [requirements](https://github.com/modelscope/modelscope/blob/master/requirements/framework.txt).

> > 能更具体描述一下你的场景么？ > > [@cyruszhang](https://github.com/cyruszhang) 比如：kafka->filter1->filter2->mapper1->files or kafka，就是这种，数据会源源不断的送进kafka，需要持续消费kafka，算子间一直不停计算并在最后输出成文件（可能多个）或继续送去kafka @xiedeyantu We are currently developing support for similar streaming/micro-batch scenarios in another open-source library focused on reinforcement learning, [Trinity-RFT](https://github.com/modelscope/Trinity-RFT). The implementation...

> [@yxdyc](https://github.com/yxdyc) Thank you very much for your detailed reply. It is very helpful to me. I would like to ask further. The Trinity-RFT project is also very complicated. Will...

It seems that this requires support of diamond dependency graph? @cyruszhang @pan-x-c Otherwise, maybe we can have some workarounds like extending current Aggregator/Group OPs, or introducing something like UnionOP? @HYLcool