Size mismatches in released model

TTS (Text-to-Speech)
#1

Must be doing something stupid here. Trying to run one of the released multispeaker models:

Multi-Speaker-Tacotron2 DDC VCTK 2136433

when I checkout 2136433 and try a test with synthesize.py I still get a bunch of layer size mistmatches.

size mismatch for decoder.attention_rnn.weight_ih: copying a param with shape torch.Size([4096, 1024]) from checkpoint, the shape in current model is torch.Size([4096, 896])

(seems to be a delta of 128)…tried a few other commits around the time of the model and no luck. Can anyone point out my mistake?

EDIT: I also noticed that the link to Universal WaveGrad on the wiki doesn’t point to a google drive link like the rest of the vocoder models. Would like to try it if possible.

#2

you need to use the right commit version with those models. Commit versions has been given on the table.

#3

Thanks erogol,
I checked out 2136433 (21364331d22b09a347bb069f98744dea78ca899f) when I tried, do I have that wrong?

#4

Ah, I think the commits on the wiki differ from the notebooks? For the two tacotron2 multispeaker models:

notebook:
!git clone https://github.com/Mozilla/TTS TTS_repo

%cd TTS_repo

!git checkout ea976b0

wiki:
4873601
2136433

Things seem to work fine with ea976b0.

1 Like