From 9a9c3cafba2619db931212cddc22373a745c5e0c Mon Sep 17 00:00:00 2001 From: James Betker Date: Tue, 17 May 2022 18:14:52 -0600 Subject: [PATCH] Make feature encoder a bit more descriptive --- codes/models/audio/mel2vec.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/codes/models/audio/mel2vec.py b/codes/models/audio/mel2vec.py index 6e768cd3..827e347b 100644 --- a/codes/models/audio/mel2vec.py +++ b/codes/models/audio/mel2vec.py @@ -366,8 +366,9 @@ class Mel2Vec(nn.Module): self.input_blocks = nn.Sequential(nn.Conv1d(mel_input_channels, inner_dim//2, kernel_size=5, padding=2, stride=2), nn.GroupNorm(num_groups=8, num_channels=inner_dim//2, affine=True), nn.SiLU(), - nn.Conv1d(inner_dim//2, inner_dim, kernel_size=3, padding=1, stride=2), - nn.GroupNorm(num_groups=8, num_channels=inner_dim, affine=True), + nn.Conv1d(inner_dim//2, inner_dim, kernel_size=3, padding=1, stride=2), + nn.SiLU(), + nn.Conv1d(inner_dim, inner_dim, kernel_size=3, padding=1), nn.SiLU(), ) self.projector = Mel2Vec2FeatureProjection(inner_dim, dropout)