Lots of stuff. Syncing to home.

python/save_env_sd_conversion.py

@@ -0,0 +1,82 @@
+import glob
+import numpy as np
+import matplotlib.pyplot as plt
+from thunderhopper.modeltools import load_data
+from thunderhopper.filters import sosfilter
+from IPython import embed
+
+# GENERAL SETTINGS:
+target = 'Omocestus_rufipes'
+data_path = glob.glob(f'../data/processed/{target}*.npz')[0]
+save_path = '../data/inv/noise_env/'
+
+# ANALYSIS SETTINGS:
+scales = np.geomspace(0.1, 10000, 200)
+sd_inputs = np.arange(10.9, 11.1, 0.01)
+n_trials = 10
+tol_to_one = 0.1
+
+# EXECUTION:
+
+# Load signal data:
+data, config = load_data(data_path, files='filt')
+signal, rate = data['filt'], config['rate']
+
+# Reduce to song segment and normalize:
+time = np.arange(signal.shape[0]) / rate
+start, end = data['songs_0'].ravel()
+segment = (time >= start) & (time <= end)
+signal /= signal[segment].std()
+
+# Get rescaled signals (time, scale):
+signal = signal[:, None] * scales[None, :]
+
+# Prepare storage:
+current_match = 0
+storage = dict(
+    scales=scales,
+    n_trials=n_trials,
+    sd_factor=np.array([0.]),
+    trials=np.zeros((scales.size, n_trials), dtype=float),
+    mean=np.zeros(scales.size, dtype=float),
+    spread=np.zeros(scales.size, dtype=float),
+)
+
+# Analyze piece-wise:
+rng = np.random.default_rng()
+for i, sigma in enumerate(sd_inputs):
+    print(f'Testing SD: {sigma:.3f} ...')
+
+    # Add Gaussian noise of given SD to rescaled signals (time, scale, trial):
+    mix = signal[..., None] + rng.normal(0, sigma, (*signal.shape, n_trials))
+
+    # Get mixture envelopes (time, scale, trial):
+    mix = sosfilter(np.abs(mix), rate, config['env_fcut'], 'lp',
+                    padtype='even', padlen=config['padlen'])[segment, ...]
+
+    # Get noise remainders of mean over trials:
+    mix -= mix.mean(axis=-1, keepdims=True)
+
+    # Estimate noise SD:
+    sd = mix.std(axis=0)
+    mean_sd = sd.mean(axis=-1)
+    n_match = (np.abs(1 - mean_sd) <= tol_to_one).sum()
+    if n_match > current_match:
+        print(f'Found better SD: {sigma:.3f} with {n_match} matches (previous: {current_match})')
+        storage['sd_factor'][0] = sigma
+        storage['trials'][:, :] = sd
+        storage['mean'][:] = mean_sd
+        storage['spread'][:] = sd.std(axis=-1)
+        current_match = n_match
+    del mix
+del signal
+if save_path is not None:
+    np.savez(save_path + 'sd_conversion.npz', **storage)
+
+plt.plot(scales, storage['mean'], 'k')
+plt.show()
+embed()
+
+print('Done.')
+embed()
+