Hi, I found a problem of FFCV, which may not the fault of the library, but this library suffers from it.
I use the same model to evaluate the accuracy on different validation sets. The results show compression reduces the performance significantly, but it's too much for FFCV.
I applied FFCV to accelerate SSL. The performance dropped for both pretraining and finetuning stages.
We set the JPEG compression quality at 90.0% to maintain high visual fidelity, write mode to jpg, the maximal resolution to 1000.
First, I hope that a lossless compression is supported.
Second, I wonder any solution to reduce the compression shidt.
Hi, I found a problem of FFCV, which may not the fault of the library, but this library suffers from it.
I use the same model to evaluate the accuracy on different validation sets. The results show compression reduces the performance significantly, but it's too much for FFCV.![image](https://github.com/libffcv/ffcv/assets/7524356/cf954f4e-813f-4dc4-94d7-168e3236d9ba)
I applied FFCV to accelerate SSL. The performance dropped for both pretraining and finetuning stages.
We set the JPEG compression quality at 90.0% to maintain high visual fidelity, write mode to jpg, the maximal resolution to 1000.
First, I hope that a lossless compression is supported. Second, I wonder any solution to reduce the compression shidt.