Marcin Junczys-Dowmunt
Marcin Junczys-Dowmunt
Maybe also option to save the valid set to a given filename, with path as another option to the same option: ``` --valid-heldout 5000 path/to/saved/valid.prefix. ``` which depending on the...
Should be doable with the new type interface. Can you recommend someone?
Sooo, what do we do about this?
Lots of progress actually. Have something that's close to working. Testing and patching things up.
Just pushed to the `types` branch. Everything that uses transformer should work with `--fp16` or the equivalent `--precision float16 float32 --cost-scaling 7 2000 2 0.05`. I think the RNNs like...
OK, GRU and LSTMs with fp16 should work too, not properly tested.
--fp16 flag should now work in master for inference in marian-decoder and marian-scorer. Still working on it for training. Should come soon-ish.
@msf Are you trying to decode on the GPU? If so then just use the normal model. You have been converting things to special CPU models. @GokulNC Can you open...
@msf Oh, OK. Don't use the fp16 flag in that case, that's GPU fp16 only.
Hi, the conversion script on huggingface is their implementation. I did help a bit, but we do not really provide conversions for other frameworks. Mostly because we don't use them...