Cannot repreduce the result of fashion by training from scratch

[Jarvisss]

Hi, anyone has trained the model from scratch?

Followed IMAGE_GENERATION.md, I download the same dataset of fashion, and train the model from scratch,

I first pretrain the flow generator for ~500k iters and then train the whole network for 100K iterations using the default parameters,

This is the result I get using my model.

But by using the pretrained model provided by the author, the result is much better:

Could anyone please tell me how many iterations should I train to get result as good as the pretrained model?

[RenYurui]

Hi Jarvisss,
It seems that the difference is caused by the flow generator.
Does the sampling correctness loss coverage after ~500k iters?
I think training the flow generator for a longer time may help with the results.
Thanks.

[Jarvisss]

Maybe I should train another 500k iters, Thanks for your advice,
I wonder what is the sampling correctness loss when it converges in your experiment ant the value?

[Jarvisss]

Another interesting thing is, if I use bilinear sampling in the sample correctness loss, the loss even cannot converge