From bfbd19c14f81338f2fc7049c4f1fc0e9be77c9e7 Mon Sep 17 00:00:00 2001 From: mrq Date: Thu, 7 Nov 2024 19:58:47 -0600 Subject: [PATCH] ugh --- vall_e/models/base.py | 4 +++- vall_e/samplers.py | 2 +- 2 files changed, 4 insertions(+), 2 deletions(-) diff --git a/vall_e/models/base.py b/vall_e/models/base.py index 41f288d..74a939c 100755 --- a/vall_e/models/base.py +++ b/vall_e/models/base.py @@ -1004,7 +1004,9 @@ class Base(nn.Module): # store dropout mask if "len" in self.capabilities and quant_level == 0: - dropout_mask = _dropout_mask( resps_list[i], p=0.8 ) + t = random.random() + p = math.cos(t * math.pi * 0.5) + dropout_mask = _dropout_mask( resps_list[i], p=p ) inputs[i].append( ("dropout_mask", dropout_mask ) ) # Audio length prediction task diff --git a/vall_e/samplers.py b/vall_e/samplers.py index 2b9b3c2..2ae9cd5 100644 --- a/vall_e/samplers.py +++ b/vall_e/samplers.py @@ -527,7 +527,7 @@ def add_gumbel_noise(t, temperature, device): return (t + torch.Tensor(temperature * np.random.gumbel(size=t.shape)).to(device)) # derived from https://github.com/LeapLabTHU/ImprovedNAT/blob/main/libs/nat_misc.py#L39 -# this +# this provides mostly poor output, but it might just be a matter of how I'm naively training the model for """diffusion""" class SampleScheduler: def __init__( self,