torch-rechub icon indicating copy to clipboard operation
torch-rechub copied to clipboard

A Lighting Pytorch Framework for Recommendation Models, Easy-to-use and Easy-to-extend.

Results 17 torch-rechub issues
Sort by recently updated
recently updated
newest added

您好,请问一下ESMM模型构架的主任务cvr和辅助任务ctr的输入是一样的吗(所有的曝光样本)?因为我看那个模型架构图,它是两个网络分别输入进去的,如果两个网络的输入是一样的,为何架构图不画成MMOE那种一个输入进去,共享底层再分成不同的任务,我这里有点疑惑。

这里的指标全部为0,应该是哪里有bug ![image](https://github.com/datawhalechina/torch-rechub/assets/7457523/079c1ca6-b87d-4f08-9065-39a47e21f7c4)

请问在Ali-CCP数据处理脚本里,对sparse_columns,dense_columns的划分依据是什么?

update ranking model BST version.0.1 (In order to maintain consistency with the DIN model of examples to call data, the `create_seq_features` method has been temporarily restored in `./utils/data.py`) > There...

#50 Implemented a custom CVR loss function to accurately reflect the conditional nature of conversion rate calculations. The loss is computed only for instances where click=1, aligning with the typical...

def generate_dataloader(self, x_test_user, x_all_item, batch_size, num_workers=1): 已将num_workers由8改为1,但运行PLE还是存在以上问题。 应该不是此处num_workers的问题

下载了全量数据集,参数的话,只改了两个tower的dims为[256,128,64],其他没动,不管设置什么seed,最后出来的ctr auc总是0.5啊,cvr的看着还算正常,接近0.6.

请问是否可提供一个onnx部署功能?谢谢

模型输入的是特征纬度的dict,线上java加载推理实在是头疼,ai.djl.pytorch好像也不支持这样的输入格式

如题,在用全量数据进行训练的时候,我发现训练结论总是在baseline的10%左右。效率是不是有点低了?