diff --git a/README.md b/README.md index fc4f24e..54ade1d 100644 --- a/README.md +++ b/README.md @@ -38,7 +38,7 @@ class AudioPipeline(torch.nn.Module): device = "cpu" -hifigan = torch.hub.load("vtuber-plan/hifi-gan:v0.2.0", "hifigan_48k", force_reload=True).to(device) +hifigan = torch.hub.load("vtuber-plan/hifi-gan:v0.2.1", "hifigan_48k", force_reload=True).to(device) # Load audio wav, sr = torchaudio.load("test.wav") diff --git a/hifigan/hub/__init__.py b/hifigan/hub/__init__.py index 5ead691..6d070d5 100644 --- a/hifigan/hub/__init__.py +++ b/hifigan/hub/__init__.py @@ -1,5 +1,5 @@ CKPT_URLS = { - "hifigan-48k": "https://github.com/vtuber-plan/hifi-gan/releases/download/v0.2.0/hifigan-48k-C8FDBD55FE7700384955A6EC41AF1D84.pt", + "hifigan-48k": "https://github.com/vtuber-plan/hifi-gan/releases/download/v0.2.1/hifigan-48k-C8FDBD55FE7700384955A6EC41AF1D84.pt", } import torch from ..model.generators.generator import Generator @@ -17,9 +17,9 @@ def hifigan_48k( [1,3,5], [1,3,5] ], - upsample_rates=[8,8,4,2], + upsample_rates=[8,8,2,2,2], upsample_initial_channel=512, - upsample_kernel_sizes=[16,16,8,4] + upsample_kernel_sizes=[16,16,4,4,4] ) if pretrained: checkpoint = torch.hub.load_state_dict_from_url( diff --git a/test.py b/test.py index d6240d6..59f3b70 100644 --- a/test.py +++ b/test.py @@ -51,7 +51,7 @@ def load_local(): return model.net_g def load_remote(): - return torch.hub.load("vtuber-plan/hifi-gan:v0.2.0", "hifigan_48k", force_reload=True) + return torch.hub.load("vtuber-plan/hifi-gan:v0.2.1", "hifigan_48k", force_reload=True) device = "cpu"