Reparameterization trick

[p0p4k]

In the following line,

TTS/TTS/tts/layers/vits/networks.py

Line 287 in d309f50

z = (mean + torch.randn_like(mean) * torch.exp(log_scale)) * x_mask

from VITS architecture, we use VAE. To see reparameterization trick, in action, we predict a mean value and log_scale (log of variance). It is smartly predicted using 2xoutput_channels (since we want output_channel sized values , each value has one mean and corresponding variance). The mean could be directly used to sample but then we cannot backpropogate the errors on the computation graph.
Therefore, we can equivalently do (mean + noise * variance).
Another place this is useful is in sampling for reverse flows.
In this code in VITS :

TTS/TTS/tts/models/vits.py

Line 1154 in c2d15cd

z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * self.inference_noise_scale

It initialized a z shaped mel spec with random noise scaled using m_p and logs_p. This is called z_p then sent to flows to get a predicted z which is further sent to waveform_decoder.
This is just for anyone interested in the math and code relation.