calculate hop_length dynamically

core/audio.py

@@ -23,15 +23,15 @@ def __init__(self, device='cuda'):
         self.linear_transform = ta.transforms.Spectrogram(
             n_fft=2*cfg.audio.win_length,
             win_length=cfg.audio.win_length,
-            hop_length=cfg.audio.hop_length,
+            hop_length=int(cfg.audio.segment_len * cfg.audio.sampling_rate / cfg.audio.spec_width),
             power=1
         ).to(self.device)
 
         self.mel_transform = ta.transforms.MelSpectrogram(
             sample_rate=cfg.audio.sampling_rate,
             n_fft=2*cfg.audio.win_length,
             win_length=cfg.audio.win_length,
-            hop_length=cfg.audio.hop_length,
+            hop_length=int(cfg.audio.segment_len * cfg.audio.sampling_rate / cfg.audio.spec_width),
             f_min=cfg.audio.min_audio_freq,
             f_max=cfg.audio.max_audio_freq,
             n_mels=cfg.audio.spec_height,

core/base_config.py

@@ -11,7 +11,6 @@ class Audio:
     spec_height = 192       # spectrogram height
     spec_width = 384        # spectrogram width (3 * 128)
     sampling_rate = 37120
-    hop_length = int(segment_len * sampling_rate /spec_width)
     win_length = 2048
     min_audio_freq = 200    # need this low for American Bittern
     max_audio_freq = 13000  # need this high for Chestnut-backed Chickadee "seet series"
@@ -102,12 +101,6 @@ class Inference:
     confirmed_if_seconds = 8     # need at least this many confirmed seconds >= raised threshold
     lower_min_factor = .6        # if so, include all labels with score >= this * min_score
 
-    # map our species names to the names used by eBird for location/date processing
-    ebird_names = {
-        "American Goshawk": "Northern Goshawk",
-        "Black-crowned Night Heron": "Black-crowned Night-Heron",
-    }
-
     # Low/high/band-pass filters can be used during inference and have to be enabled and configured here.
     # Inference will then use the max prediction per species, with and without the filter(s).
     # Using a single filter adds ~50% to elapsed time for large datasets, but less for small ones where