The pytorch-enformer has much lower performance than the tensorflow-enformer

[biginfor]

Dear developers, thank you very much for developing the pytorch version of enformer！
However, when I use the example data provided by basenji and my own ATAC-Seq dataset, the predictive power of pytorch-enformer compared to the predictive power of tensorflow-enformer (referring to the Pearson correlation coefficient) drops by around 0.2, (0.3 vs 0.5 for the validation set) . I tried to initialize my model weights following the weight initialization method of the tensorflow-enformer, but things didn't get radically better, can anyone tell me why?

Best,
Eli