From bf678a8399d1c7c4d9b47290e62d0720c0c7fd36 Mon Sep 17 00:00:00 2001
From: "paul.best" <paul.best@lis-lab.fr>
Date: Wed, 12 Jun 2024 12:22:22 +0200
Subject: [PATCH] small fixes
---
pesto/data.py | 8 ++++++++
pesto/loader.py | 2 +-
pesto/model.py | 2 +-
pesto/utils/hcqt.py | 2 +-
4 files changed, 11 insertions(+), 3 deletions(-)
diff --git a/pesto/data.py b/pesto/data.py
index 3044bbd..ebdd151 100644
--- a/pesto/data.py
+++ b/pesto/data.py
@@ -2,6 +2,7 @@ from typing import Optional
import torch
import torch.nn as nn
+from torchaudio import functional as taudio
from .utils import HarmonicCQT
@@ -45,6 +46,9 @@ class Preprocessor(nn.Module):
# if the sampling rate is provided, instantiate the CQT kernels
if sampling_rate is not None:
+ top_bin = hcqt_kwargs['fmin'] * 2 ** (hcqt_kwargs['n_bins'] / hcqt_kwargs['bins_per_semitone'] / 12)
+ if sampling_rate / 2 < top_bin:
+ sampling_rate = int(round(top_bin*1e-3) * 2e3)
self.hcqt_sr = sampling_rate
self._reset_hcqt_kernels()
@@ -60,6 +64,10 @@ class Preprocessor(nn.Module):
torch.Tensor: log-magnitude CQT of batch of CQTs,
shape (batch_size?, num_timesteps, num_harmonics, num_freqs)
"""
+
+ if sr < self.hcqt_sr:
+ x, sr = taudio.resample(x, sr, self.hcqt_sr), self.hcqt_sr
+
# compute CQT from input waveform, and invert dims for (time_steps, num_harmonics, freq_bins)
# in other words, time becomes the batch dimension, enabling efficient processing for long audios.
complex_cqt = torch.view_as_complex(self.hcqt(x, sr=sr)).permute(0, 3, 1, 2)
diff --git a/pesto/loader.py b/pesto/loader.py
index a270615..1f0c77d 100644
--- a/pesto/loader.py
+++ b/pesto/loader.py
@@ -44,7 +44,7 @@ def load_model(checkpoint: str,
model = PESTO(encoder,
preprocessor=preprocessor,
crop_kwargs=hparams["pitch_shift"],
- reduction=hparams["reduction"])
+ reduction="awa")
model.load_state_dict(state_dict, strict=False)
model.eval()
diff --git a/pesto/model.py b/pesto/model.py
index b9cf5ed..748b6ee 100644
--- a/pesto/model.py
+++ b/pesto/model.py
@@ -138,7 +138,7 @@ class Resnet1d(nn.Module):
Args:
x (torch.Tensor): shape (batch, channels, freq_bins)
"""
- x = self.layernorm(x)
+ x = self.layernorm(x.unsqueeze(1))
x = self.conv1(x)
for p in range(0, self.n_prefilt_layers - 1):
diff --git a/pesto/utils/hcqt.py b/pesto/utils/hcqt.py
index f7ff561..2701936 100644
--- a/pesto/utils/hcqt.py
+++ b/pesto/utils/hcqt.py
@@ -104,7 +104,7 @@ def create_cqt_kernels(
freqs = fmin * 2.0 ** (np.r_[0:n_bins] / float(bins_per_octave))
else:
- warnings.warn("If fmax is given, n_bins will be ignored", SyntaxWarning)
+ #warnings.warn("If fmax is given, n_bins will be ignored", SyntaxWarning)
n_bins = np.ceil(
bins_per_octave * np.log2(fmax / fmin)
) # Calculate the number of bins
--
GitLab