Optimized min_std param for different PEQ configurations. Removed ten…

…sorflow from requirements.txt. Added example PEQ config file and improved usage instructions about it.

autoeq.py

@@ -256,23 +256,33 @@ def cli_args():
                                  'no value needed.')
     arg_parser.add_argument('--ten_band_eq', action='store_true',
                             help='Shortcut parameter for activating standard ten band eq optimization.')
-    arg_parser.add_argument('--parametric_eq_config', type=str, default='4_PEAKING_WITH_LOW_SHELF,4_PEAKING_WITH_HIGH_SHELF',
-                            help='Name of parametric equalizer configuration or a path to a configuration file. \n\n'
+    arg_parser.add_argument('--parametric_eq_config', type=str,
+                            default='4_PEAKING_WITH_LOW_SHELF,4_PEAKING_WITH_HIGH_SHELF',
+                            help='Name of parametric equalizer configuration or a path to a configuration file. '
                                  'Available named configurations are "10_PEAKING" for 10 peaking filters, '
                                  '"8_PEAKING_WITH_SHELVES" for 8 peaking filters and a low shelf at 105 Hz for bass '
-                                 'adjustment and high shelf at 10 kHz for treble adjustment, '
+                                 'adjustment and a high shelf at 10 kHz for treble adjustment, '
                                  '"4_PEAKING_WITH_LOW_SHELF" for 4 peaking filters and a low shelf at 105 Hz for bass '
                                  'adjustment, "4_PEAKING_WITH_HIGH_SHELF" for 4 peaking filters and a high shelf '
                                  'at 10 kHz for treble adjustments. You can give multiple named configurations by '
-                                 'separating the names with commas. \n\n'
+                                 'separating the names with commas and filter sets will be built on top of each other. '
                                  'When the value is a file path, the file will be read and used as a configuration. '
                                  'The file needs to be a YAML file with "filters" field as a list of filter '
                                  'configurations, each of which can define "fc", "min_fc", "max_fc", "q", "min_q", '
                                  '"max_q", "gain", "min_gain", "max_gain" and "type" fields. When the fc, q or gain '
                                  'value is given, the parameter won\'t be optimized for the filter. "type" needs to '
                                  'be either "LOW_SHELF", "PEAKING" or "HIGH_SHELF". Also "filter_defaults" field is '
-                                 'supported on the top level and it can have the same fields as the filters do.\n\n'
-                                 'Defaults to "4_PEAKING_WITH_LOW_SHELF,4_PEAKING_WITH_HIGH_SHELF".'),
+                                 'supported on the top level and it can have the same fields as the filters do. '
+                                 'All fields missing from the filters will be read from "filter_defaults". '
+                                 'Defaults to "4_PEAKING_WITH_LOW_SHELF,4_PEAKING_WITH_HIGH_SHELF". '
+                                 'Optimizer behavior can be adjusted by defining "optimizer" field which has fields '
+                                 '"min_f" and "max_f" for lower and upper bounds of the optimization range, "max_time" '
+                                 'for maximum optimization duration in seconds, "target_loss" for RMSE target level '
+                                 'upon reaching which the optimization is ended, "min_change_rate" for minimum rate '
+                                 'of improvement in db/s and "min_std" for minimum standard deviation of the last few '
+                                 'loss values. "min_change_rate" and "min_std" end the optimization when further time '
+                                 'spent optimizing can\'t be expected to improve the results dramatically. See '
+                                 'peq.yaml for an example.'),
     arg_parser.add_argument('--fixed_band_eq_config', type=str, default='10_BAND_GRAPHIC_EQ',
                             help='Path to fixed band equalizer configuration. The file format is the same YAML as '
                                  'for parametric equalizer.')

constants.py

@@ -35,7 +35,7 @@
 DEFAULT_PEQ_OPTIMIZER_MAX_TIME = None
 DEFAULT_PEQ_OPTIMIZER_TARGET_LOSS = None
 DEFAULT_PEQ_OPTIMIZER_MIN_CHANGE_RATE = None
-DEFAULT_PEQ_OPTIMIZER_MIN_STD = 0.1
+DEFAULT_PEQ_OPTIMIZER_MIN_STD = 0.002
 
 DEFAULT_FIXED_BAND_FILTER_MIN_GAIN = -12
 DEFAULT_FIXED_BAND_FILTER_MAX_GAIN = 12
@@ -75,6 +75,9 @@
         'filters': [{'type': 'PEAKING'}] * 10
     },
     '8_PEAKING_WITH_SHELVES': {
+        'optimizer': {
+            'min_std': 0.008
+        },
         'filters': [{
             'type': 'LOW_SHELF',
             'fc': 105,
@@ -87,7 +90,7 @@
     },
     '4_PEAKING_WITH_LOW_SHELF': {
         'optimizer': {
-            'max_f': 10000
+            'max_f': 10000,
         },
         'filters': [{
             'type': 'LOW_SHELF',

frequency_response.py

@@ -1,6 +1,7 @@
 # -*- coding: utf-8 -*_
 
 import os
+from copy import deepcopy
 import matplotlib.pyplot as plt
 import matplotlib.ticker as ticker
 import math
@@ -438,6 +439,8 @@ def write_readme(self, file_path, parametric_eq_peqs=None, fixed_band_eq_peq=Non
                 filter_ranges = ''
                 preamps = ''
                 for i, peq in enumerate(parametric_eq_peqs):
+                    peq = deepcopy(peq)
+                    peq.sort_filters()
                     for filt in peq.filters:
                         compound.add_filter(filt)
                     filter_ranges += f'1-{len(peq.filters) + n}'
@@ -453,6 +456,8 @@ def write_readme(self, file_path, parametric_eq_peqs=None, fixed_band_eq_peq=Non
             else:
                 compound = PEQ(self.frequency.copy(), parametric_eq_peqs[0].fs, [])
                 for peq in parametric_eq_peqs:
+                    peq = deepcopy(peq)
+                    peq.sort_filters()
                     for filt in peq.filters:
                         compound.add_filter(filt)
                 s += f'Apply preamp of -{compound.max_gain + 0.1:.1f} dB when using parametric equalizer.\n\n'

peq.py

@@ -600,7 +600,7 @@ def optimize(self):
         except OptimizationFinished as err:
             # Restore best params
             self._parse_optimizer_params(self.history.params[np.argmin(self.history.loss)])
-        self.sort_filters()
+            #print(err)
 
     def plot(self, fig=None, ax=None):
         if fig is None:

peq.yaml

@@ -0,0 +1,31 @@
+optimizer:
+  min_f: 20
+  max_f: 20000  # Only optimizing up to 10 kHz because of we only have 6 filters
+  max_time: null  # Letting optimizer finish since this is probably for a one-off custom optimization, not for API
+  min_change_rate: null  # Letting optimizer finish
+  min_std: null  # Letting optimizer finish
+filter_defaults:
+  min_fc: 20
+  max_fc: 10000
+  min_q: 0.18248
+  max_q: 6
+  min_gain: -20
+  max_gain: 20
+filters:
+  - type: PEAKING
+  - type: PEAKING
+  - type: PEAKING
+  - type: PEAKING
+  - type: PEAKING
+  - type: PEAKING
+  - type: PEAKING
+  - type: PEAKING
+  - type: LOW_SHELF
+    fc: 105  # Fixed center frequency at 105 Hz
+    q: 0.7  # Fixed quality of 0.7
+  - type: HIGH_SHELF
+    fc: 10000
+    min_fc: 5000  # Letting the optimizer adjust high shelf between 5 kHz and...
+    max_fc: 12000  # ...12 kHz
+    min_q: 0.4  # Shelf filters start to overshoot below 0.4
+    max_q: 0.7  # Shelf filters start to overshoot after 0.7

requirements.txt

@@ -3,7 +3,6 @@ matplotlib~=3.5.2
 pandas~=1.4.3
 scipy~=1.8.1
 numpy~=1.23.1
-tensorflow~=2.9.1
 tabulate~=0.8.10
 soundfile~=0.10.2
 pyyaml~=6.0

research/neo_peq/animation.py

@@ -0,0 +1,61 @@
+import sys
+from pathlib import Path
+from time import time
+from copy import deepcopy
+import yaml
+import numpy as np
+import pandas as pd
+import matplotlib.pyplot as plt
+from matplotlib.animation import FuncAnimation, FFMpegWriter
+ROOT_PATH = Path().absolute().parent.parent
+if str(ROOT_PATH) not in sys.path:
+    sys.path.insert(1, str(ROOT_PATH))
+from frequency_response import FrequencyResponse
+from constants import DEFAULT_FS, PEQ_CONFIGS
+
+with open(ROOT_PATH.joinpath('peq.yaml')) as fh:
+    config = yaml.safe_load(fh)
+fr = FrequencyResponse.read_from_csv(ROOT_PATH.joinpath('my_results/F3/Custom Art FIBAE 3 (CIEM).csv'))
+config = PEQ_CONFIGS['4_PEAKING_WITH_LOW_SHELF']
+config['optimizer']['min_std'] = 0.0
+peqs = fr.optimize_parametric_eq(config, DEFAULT_FS)
+peq = peqs[0]
+#peq.plot()
+#plt.show()
+
+
+def interpolate(current, previous, ratio):
+    return previous + ratio * (current - previous)
+
+
+param_history = []
+time_history = []
+times = []
+for i in range(1, len(peq.history.params)):
+    for j in range(10):
+        param_history.append(interpolate(np.array(peq.history.params[i]), np.array(peq.history.params[i - 1]), j / 10))
+        time_history.append(interpolate(np.array(peq.history.time[i]), np.array(peq.history.time[i - 1]), j / 10))
+
+fig, ax = FrequencyResponse.init_plot()
+fills = [ax.fill_between([], [], []) for _ in peq.filters]
+lines = [ax.plot([], [], linewidth=1, color=f'C{i}')[0] for i in range(len(peq.filters))]
+lines.append(ax.plot([], [], color='black', linestyle='--')[0])
+lines.append(ax.plot([], [], color='black')[0])
+ax.set_ylim([-10, 15])
+
+
+def animate(i):
+    peq._parse_optimizer_params(param_history[i])
+    ax.collections.clear()
+    for j, filt in enumerate(peq.filters):
+        ax.fill_between(filt.f, np.zeros(filt.fr.shape), filt.fr, alpha=0.3, color=f'C{j}')
+        lines[j].set_data(filt.f, filt.fr)
+    lines[-2].set_data(peq.f, peq.target)
+    lines[-1].set_data(peq.f, peq.fr)
+    ax.set_title(f't={time_history[i] * 1000:.0f} ms')
+    return lines
+
+
+ani = FuncAnimation(fig, animate, frames=len(param_history), interval=3, repeat=False)
+ani.save('peq.mp4', writer=FFMpegWriter(fps=60))
+plt.show()