britekit 0.0.11__tar.gz → 0.1.0__tar.gz

{britekit-0.0.11 → britekit-0.1.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: britekit
-Version: 0.0.11
+Version: 0.1.0
 Summary: Core functions for bioacoustic recognizers.
 Project-URL: Documentation, https://github.com/jhuus/BriteKit#readme
 Project-URL: Issues, https://github.com/jhuus/BriteKit/issues
@@ -21,14 +21,14 @@ Requires-Python: >=3.8
 Requires-Dist: click<8.3,>=8.1
 Requires-Dist: librosa<2.0,>=0.10
 Requires-Dist: lightning<2.6,>=2.5
-Requires-Dist: matplotlib<3.11,>=3.9
-Requires-Dist: numpy<2.4,>=2.2
+Requires-Dist: matplotlib<3.11,>=3.8
+Requires-Dist: numpy<2.4,>=2.0
 Requires-Dist: omegaconf<2.4,>=2.3
 Requires-Dist: onnx<2.0,>=1.18
-Requires-Dist: pandas<2.3,>=2.2
+Requires-Dist: pandas<2.3,>=2.0
 Requires-Dist: pyinaturalist<1.0,>=0.20
 Requires-Dist: pytorch-lightning<2.6,>=2.5
-Requires-Dist: scipy<2.0,>=1.15
+Requires-Dist: scipy<2.0,>=1.13
 Requires-Dist: soundfile<1.0,>=0.13
 Requires-Dist: tensorboard<3.0,>=2.19
 Requires-Dist: timm<2.0,>=1.0.17

{britekit-0.0.11 → britekit-0.1.0}/britekit/__about__.py

@@ -1,4 +1,4 @@
 # SPDX-FileCopyrightText: 2025-present Jan Huus <jhuus1@gmail.com>
 #
 # SPDX-License-Identifier: MIT
-__version__ = "0.0.11"
+__version__ = "0.1.0"

{britekit-0.0.11 → britekit-0.1.0}/britekit/commands/_xeno.py

@@ -4,6 +4,7 @@ import json
 import logging
 import os
 from typing import Optional
+from urllib.parse import quote
 
 import click
 
@@ -101,6 +102,8 @@ def xeno(
     else:
         name = f'en:"={name.lower()}"'
 
+    name = quote(name)
+
     # get list of recordings
     recordings: list = []
     page = 0
@@ -137,6 +140,9 @@ def xeno(
         if not os.path.exists(outfile):
             logging.info(f"Downloading {outfile}")
             url = recording["file"]
+            if not url:
+                url = f"https:{recording['url']}/download"
+
             response = requests.get(url)
             with open(outfile, "wb") as mp3:
                 mp3.write(response.content)

{britekit-0.0.11 → britekit-0.1.0}/britekit/install/yaml/base_config.yaml

@@ -5,6 +5,7 @@ audio: !!python/object:britekit.core.base_config.Audio
   db_power: 1.0
   decibels: false
   freq_scale: mel
+  log_freq_gain: 0.6
   max_freq: 8000
   min_freq: 100
   power: 1.0

{britekit-0.0.11 → britekit-0.1.0}/install/yaml/base_config.yaml

@@ -5,6 +5,7 @@ audio: !!python/object:britekit.core.base_config.Audio
   db_power: 1.0
   decibels: false
   freq_scale: mel
+  log_freq_gain: 0.6
   max_freq: 8000
   min_freq: 100
   power: 1.0

{britekit-0.0.11 → britekit-0.1.0}/pyproject.toml

@@ -26,19 +26,19 @@ classifiers = [
 ]
 dependencies = [
   "click>=8.1,<8.3",
-  "numpy>=2.2,<2.4",
+  "numpy>=2.0,<2.4",
   "onnx>=1.18,<2.0",
-  "pandas>=2.2,<2.3",
+  "pandas>=2.0,<2.3",
   "pyinaturalist>=0.20,<1.0",
   "soundfile>=0.13,<1.0",
   "torch>=2.5,<2.9",
   "torchaudio>=2.5,<2.9",
   "librosa>=0.10,<2.0",
   "lightning>=2.5,<2.6",
-  "matplotlib>=3.9,<3.11",
+  "matplotlib>=3.8,<3.11",
   "omegaconf>=2.3,<2.4",
   "pytorch_lightning>=2.5,<2.6",
-  "scipy>=1.15,<2.0",
+  "scipy>=1.13,<2.0",
   "tensorboard>=2.19,<3.0",
   "timm>=1.0.17,<2.0",
   "yt-dlp>=2025.6.25",

{britekit-0.0.11 → britekit-0.1.0}/src/britekit/core/audio.py

@@ -264,7 +264,6 @@ class Audio:
         import numpy as np
         import torch
 
-        """Create a filterbank for log spectrograms."""
         f_min = self.cfg.audio.min_freq
         f_max = self.cfg.audio.max_freq
         n_bins = self.cfg.audio.spec_height
@@ -282,13 +281,19 @@ class Audio:
 
         filters = []
         for log_cf in log2_centers:
+            # Gaussian filter in log2(f)
             weight = np.exp(-0.5 * ((fft_log2 - log_cf) / sigma_log2) ** 2)
-            weight_sum = np.sum(weight)
-            if weight_sum > 0:
-                filters.append(weight / weight_sum)
-            else:
-                # Fallback: use uniform weights if sum is zero
-                filters.append(weight)
+
+            # Normalize per filter (contrast-preserving)
+            weight /= np.sum(weight) + 1e-12
+
+            # Boost power in higher frequencies, so more like mel scale;
+            # increase log_freq_gain to increase the boost
+            cf_hz = 2**log_cf
+            gain = (cf_hz / f_min) ** self.cfg.audio.log_freq_gain
+            weight *= gain
+
+            filters.append(weight)
 
         filters = np.array(filters)
         return torch.tensor(filters, dtype=torch.float32).to(self.device)
@@ -323,28 +328,41 @@ class Audio:
         tensor = torch.from_numpy(signal).to(self.device)
 
         if freq_scale == "log":
-            spec = self.linear_transform(tensor)
-            spec = torch.matmul(self.log2_filterbank, spec.squeeze(0)).unsqueeze(0)
+            spec = self.linear_transform(tensor)  # [1, n_freqs, n_frames]
+            spec = torch.matmul(
+                self.log2_filterbank, spec.squeeze(0)
+            )  # [n_mels, n_frames]
+            spec = spec.unsqueeze(0).unsqueeze(1)  # [1, 1, n_mels, n_frames]
+
         elif freq_scale == "mel":
-            spec = self.mel_transform(tensor)
+            spec = self.mel_transform(tensor).unsqueeze(1)  # [1, 1, n_mels, T]
+
         elif freq_scale == "linear":
             spec = self.linear_transform(tensor)
-
             freqs = torch.fft.rfftfreq(
                 2 * self.win_length, d=1 / self.cfg.audio.sampling_rate
-            )  # [freq_bins]
+            )
             mask = (freqs >= self.cfg.audio.min_freq) & (
                 freqs <= self.cfg.audio.max_freq
             )
-            spec = spec[:, mask, :]  # shape: [channel, selected_freq_bins, time_frames]
-            spec = spec.unsqueeze(1)
-            spec = F.interpolate(
-                spec,
-                size=(self.cfg.audio.spec_height, self.cfg.audio.spec_width),
-                mode="bilinear",
-                align_corners=False,
-            )
-            spec = spec.squeeze(1)
+            spec = spec[:, mask, :].unsqueeze(1)  # [1, 1, F_sel, T]
+
+        # downsample frequency to spec_height (energy-preserving)
+        spec = F.interpolate(
+            spec,
+            size=(self.cfg.audio.spec_height, spec.shape[-1]),
+            mode="area",
+        )
+
+        # pad or crop to spec_width
+        T = spec.shape[-1]
+        if T < self.cfg.audio.spec_width:
+            pad_width = self.cfg.audio.spec_width - T
+            spec = F.pad(spec, (0, pad_width))  # pad on the right (time axis)
+        else:
+            spec = spec[..., : self.cfg.audio.spec_width]
+
+        spec = spec.squeeze(1)
 
         if decibels:
             spec = ta.transforms.AmplitudeToDB(stype="power", top_db=top_db)(spec)

{britekit-0.0.11 → britekit-0.1.0}/src/britekit/core/base_config.py

@@ -23,6 +23,7 @@ class Audio:
     decibels: bool = False  # Use decibel amplitude scale?
     top_db: float = 80  # Parameter to decibel conversion
     db_power: float = 1.0  # Raise to this exponent after convert to decibels
+    log_freq_gain: float = 0.6  # Boost loudness of higher frequencies with log scale
 
 
 @dataclass
@@ -61,7 +62,7 @@ class Training:
     # "adam" with decay = 1e-6
     # "adamp" with decay = 0
     optimizer: str = "radam"  # Any timm optimizer
-    opt_weight_decay: float = 1e-6  # Weight decay option (L2 normalization)
+    opt_weight_decay: float = 1e-6  # Weight decay option (L2 regularization)
     opt_beta1: float = 0.9  # Optimizer parameter
     opt_beta2: float = 0.999  # Optimizer parameter
 

{britekit-0.0.11 → britekit-0.1.0}/src/britekit/core/trainer.py

@@ -47,7 +47,9 @@ class Trainer:
         val_rocs = []
         for k in range(self.cfg.train.num_folds):
             logger = TensorBoardLogger(
-                save_dir="logs", name=f"fold-{k}", default_hp_metric=False
+                save_dir="logs",
+                name=None if self.cfg.train.num_folds == 1 else f"fold-{k}",
+                default_hp_metric=False,
             )
             version = (
                 logger.version

{britekit-0.0.11 → britekit-0.1.0}/src/britekit/models/base_model.py

@@ -177,7 +177,7 @@ class BaseModel(pl.LightningModule):
         else:
             preds = torch.softmax(seg_logits, dim=1)
 
-        self.log("val_loss", loss, on_step=False, on_epoch=True, prog_bar=True)
+        self.log("val_loss", loss, on_step=False, on_epoch=True, prog_bar=False)
 
         roc_auc = metrics.roc_auc_score(y.cpu(), preds.cpu(), average="micro")
         self.log("val_roc", roc_auc, on_step=False, on_epoch=True, prog_bar=True)