From e0a24405e72758dfe1a6413f2a19a8cd85d6aea3 Mon Sep 17 00:00:00 2001
From: Evgeniy Shabalin <36159472+treacker@users.noreply.github.com>
Date: Thu, 22 Dec 2022 01:05:09 +0400
Subject: [PATCH] Fix preprocessing bug

---
 hw4_nv/README.md | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/hw4_nv/README.md b/hw4_nv/README.md
index fe9c8b7..0263f66 100644
--- a/hw4_nv/README.md
+++ b/hw4_nv/README.md
@@ -42,7 +42,6 @@ class MelSpectrogram(nn.Module):
         super(MelSpectrogram, self).__init__()
 
         self.config = config
-
         self.mel_spectrogram = torchaudio.transforms.MelSpectrogram(
             sample_rate=config.sr,
             win_length=config.win_length,
@@ -50,7 +49,8 @@ class MelSpectrogram(nn.Module):
             n_fft=config.n_fft,
             f_min=config.f_min,
             f_max=config.f_max,
-            n_mels=config.n_mels
+            n_mels=config.n_mels,
+            center=False,
         )
 
         # The is no way to set power in constructor in 0.5.0 version.
@@ -72,7 +72,8 @@ class MelSpectrogram(nn.Module):
         :param audio: Expected shape is [B, T]
         :return: Shape is [B, n_mels, T']
         """
-
+        audio = torch.nn.functional.pad(audio.unsqueeze(1), (int((self.config.n_fft-self.config.hop_length)/2), int((self.config.n_fft-self.config.hop_length)/2)), mode='reflect')
+        audio = audio.squeeze(1)
         mel = self.mel_spectrogram(audio) \
             .clamp_(min=1e-5) \
             .log_()