[thesis] updated pipelines

.gitignore

@@ -1,12 +1,19 @@
 
+## Logs
+src/logs/*
+
+## Trained Models
+src/models/trained_models/separation/*
+src/models/trained_models/transcription/*
+
 ## Datasets
 # Ignore all files in these specific directories
-/src/data/processed/test/model1/*
-/src/data/processed/test/model2/*
-/src/data/processed/train/model1/*
-/src/data/processed/train/model2/*
-/src/data/raw/model1/*
-/src/data/raw/model2/*
+src/data/processed/test/model1/*
+src/data/processed/test/model2/*
+src/data/processed/train/model1/*
+src/data/processed/train/model2/*
+src/data/raw/model1/*
+src/data/raw/model2/*
 
 # Except the directory structures
 !/src/data/processed/test/model1/.gitkeep

src/config/constants.py

@@ -8,7 +8,7 @@
 # ----------
 
 # Model 1 Constants
-MODEL1_SUBSET = "V2"
+MODEL1_SUBSET = "V5"
 MODEL1_BASE_PATH = f"../data/raw/model1/{MODEL1_SUBSET}"
 MODEL1_TRAIN_FOLDER_PATH = "../data/processed/train/model1"
 MODEL1_TRAIN_FILE_NAME = f"mix_bass_train_data_{MODEL1_SUBSET}-TRAIN"
@@ -38,7 +38,11 @@
 TRAINED_MODEL2_SAVE_PATH = "../models/trained_models/transcription/mark1.pt"
 
 # Visualisation and Testing constants
-TRAINED_AUDIO_FILE_PATH = "../visualization/audio/trained_audio"
-PRED_AUDIO_FILE_PATH = "../visualization/audio/predicted_audio"
-VISUALIZATION_SAVE_PATH = "../visualization/spectrograms/outputs"
+TRAINED_AUDIO_FILE_PATH = (
+    "../../../../Bachelor Thesis/thesis/audio outputs/trained audio"
+)
+PRED_AUDIO_FILE_PATH = (
+    "../../../../Bachelor Thesis/thesis/audio outputs/predicted audio"
+)
+VISUALIZATION_SAVE_PATH = "../../../../Bachelor Thesis/thesis/spectrograms"
 PRED_MIDI_FILE_PATH = "../visualization/midi/predicted_midi"

src/config/fourierparameters.json

@@ -1,8 +1,8 @@
 {
   "n_fft": 2048,
   "n_mels": 128,
-  "hop_length": 3308,
+  "hop_length": 512,
   "sample_rate": 44100,
   "track_seconds_considered": 30,
-  "audio_amplitude_threshold": 0.0001
+  "audio_amplitude_threshold": 0.0000001
 }

src/config/hyperparameters_separation.json

@@ -1,7 +1,7 @@
 {
   "learning_rate": 0.0001,
-  "n_epochs": 10,
-  "hidden_dim": 512,
+  "n_epochs": 250,
+  "hidden_dim": 1024,
   "n_layers": 3,
-  "dropout_rate": 0.3
+  "dropout_rate": 0.1
 }

src/data_manipulation/__helpers__/normalization/decibel_normalizer.py

@@ -0,0 +1,38 @@
+import numpy as np
+
+
+class DecibelNormalizer:
+    def __init__(self, array):
+        self.array = array
+        self.min_val = None
+        self.max_val = None
+
+    def normalize(self):
+        # Convert spectrogram magnitudes to decibels
+        epsilon = 1e-10  # small constant to avoid log(0)
+        self.array = 20 * np.log10(np.maximum(self.array, epsilon))
+
+        # Find min and max values in the dB-scaled array
+        self.min_val, self.max_val = np.min(self.array), np.max(self.array)
+
+        # Avoid division-by-zero error in case of all same elements
+        if self.min_val == self.max_val:
+            self.array[:] = 0
+        else:
+            self.array = (self.array - self.min_val) / (self.max_val - self.min_val)
+
+        return self.array
+
+    def get_min_max(self):
+        return self.min_val, self.max_val
+
+    def denormalize(self, min_val, max_val):
+        if min_val is not None and max_val is not None:
+
+            # Apply the inverse of min-max normalization
+            self.array = (self.array * (max_val - min_val)) + min_val
+
+            # Convert decibels back to linear scale
+            self.array = 10 ** (self.array / 20.0)
+
+        return self.array

src/data_manipulation/for_production/production_audio_to_dict.py

@@ -3,8 +3,8 @@
     convert_t_dict_key_to_numpy_arrays,
     convert_to_recarray,
 )
-from src.data_manipulation.__helpers__.normalization.mix_bass_data_normalizer import (
-    Normalizer,
+from src.data_manipulation.__helpers__.normalization.decibel_normalizer import (
+    DecibelNormalizer,
 )
 from src.transformers.audio_to_freq_time_analysis import audio_to_freq_time_analysis
 
@@ -36,8 +36,8 @@ def production_audio_to_dict(file_name, input_path):
         t_dict = convert_t_dict_key_to_numpy_arrays(dictionary=t_dict, keys=["x", "y"])
 
         # Normalize the data
-        norm_x = Normalizer(t_dict["x"])
-        t_dict["x"], t_dict["min_max_amplitudes"] = (
+        norm_x = DecibelNormalizer(t_dict["x"])
+        t_dict["x"], t_dict["x_min_max_amplitudes"] = (
             norm_x.normalize(),
             norm_x.get_min_max(),
         )

src/data_manipulation/for_training/audio_spectrograms/mixed_signal_to_dict.py

@@ -6,8 +6,8 @@
     savez_numpy_data,
     convert_to_recarray,
 )
-from src.data_manipulation.__helpers__.normalization.mix_bass_data_normalizer import (
-    Normalizer,
+from src.data_manipulation.__helpers__.normalization.decibel_normalizer import (
+    DecibelNormalizer,
 )
 from src.data_manipulation.__helpers__.truncator.mix_bass_data_truncator import (
     data_truncator,
@@ -111,16 +111,12 @@ def mixed_signal_to_dict(base_path, files_to_transform, save_file_path, pause=Fa
         t_dict = convert_t_dict_key_to_numpy_arrays(dictionary=t_dict, keys=["x", "y"])
 
         # Normalize the data
-        norm_x = Normalizer(t_dict["x"])
-        t_dict["x"], t_dict["min_max_amplitudes"] = (
-            norm_x.normalize(),
-            norm_x.get_min_max(),
-        )
-        norm_y = Normalizer(t_dict["y"])
-        t_dict["y"], t_dict["min_max_amplitudes"] = (
-            norm_y.normalize(),
-            norm_y.get_min_max(),
-        )
+        norm_x = DecibelNormalizer(t_dict["x"])
+        t_dict["x"] = norm_x.normalize()
+        t_dict["x_min_max_amplitudes"] = norm_x.get_min_max()
+        norm_y = DecibelNormalizer(t_dict["y"])
+        t_dict["y"] = norm_y.normalize()
+        t_dict["y_min_max_amplitudes"] = norm_y.get_min_max()
 
         # Transform to recarray
         t_dict_recarray = convert_to_recarray(data_dict=t_dict)

src/data_manipulation/for_training/tab_transcriptions/bass_and_midi_to_dict.py

@@ -5,8 +5,8 @@
     convert_to_recarray,
     savez_numpy_data,
 )
-from src.data_manipulation.__helpers__.normalization.mix_bass_data_normalizer import (
-    Normalizer,
+from src.data_manipulation.__helpers__.normalization.decibel_normalizer import (
+    DecibelNormalizer,
 )
 from src.data_manipulation.__helpers__.truncator.mix_bass_data_truncator import (
     data_truncator,
@@ -95,8 +95,8 @@ def bass_and_midi_to_dict(base_path, files_to_transform, save_file_path, pause=F
         t_dict = convert_t_dict_key_to_numpy_arrays(dictionary=t_dict, keys=["x", "y"])
 
         # Normalize the data
-        norm_x = Normalizer(t_dict["x"])
-        t_dict["x"], t_dict["min_max_amplitudes"] = (
+        norm_x = DecibelNormalizer(t_dict["x"])
+        t_dict["x"], t_dict["x_min_max_amplitudes"] = (
             norm_x.normalize(),
             norm_x.get_min_max(),
         )

src/main.py

@@ -15,80 +15,11 @@
 
 
 def main():
-    while True:
-        choice = (
-            input(
-                "@@@@@@ Would you like to create new models before processing your audio? @@@@@@\n"
-                '@@@@@@ WARNING: Selecting "yes" requires training/testing datasets to be present in the '
-                "data/raw/model1 and data/raw/model2 folders. @@@@@@\n"
-                '@@@@@@  Selecting "no" will use the pre-existing mark1 models instead. [Y/N]:'
-            )
-            .strip()
-            .lower()
-        )
 
-        if choice in ["yes", "y"]:
-            # Trains and tests both models separately
-            print()
-            print("@@@@@@ SEPARATION TRAINING/TESTING START @@@@@@")
-            separation_manager()
-            print()
-            print("@@@@@@ TRANSCRIPTION TRAINING/TESTING START @@@@@@")
-            transcription_manager()
-            print()
-            print("@@@@@@ MODELS SUCCESSFULLY CREATED @@@@@@")
-            break
-
-        elif choice in ["no", "n"]:
-            break
-
-        else:
-            print("Please enter a valid input. Either [Y/N] or [Yes/No].")
-
-    try:
-        input_folder = f"{PATH_TO_AUDIO}/{PRODUCTION_INPUT_FOLDER_PATH}"
-        if not os.listdir(input_folder):
-            print("No input files present in input folder. Programm terminated.")
-            return
-        for track in os.listdir(f"{input_folder}"):
-            if track.endswith(".wav"):
-                print()
-                print(f"@@@@ USING MODELS ON SELECTED TRACK: {track} @@@@")
-                input_file_path = (
-                    f"{PATH_TO_AUDIO}/{PRODUCTION_INPUT_FOLDER_PATH}/{track}"
-                )
-                output_file_path = (
-                    f"{PATH_TO_AUDIO}/{PRODUCTION_OUTPUT_FOLDER_PATH}/{track}"
-                )
-
-                print("@@ Pre-Processing... @@")
-                # Pre-process the audio into correct spectrogram format
-                processed_input = transform_data(
-                    flag="production input",
-                    file_name=track,
-                    input_file_path=input_file_path,
-                )
-
-                print("@@ Separating and transcribing... @@")
-                # Pass processed audio through both models
-                output = use_models_on_audio(processed_input)
-
-                print("@@ Post-Processing... @@")
-                # Post-process the midi-spectrogram into pure midi format and save it
-                piano_roll_to_midi(
-                    piano_roll=output,
-                    output_file_path=output_file_path,
-                    tag="",
-                    mix_names=["mix_name"],
-                    flag="production",
-                )
-
-            else:
-                print(f"@@@@ {track} is not a .wav file. Skipping... @@@@")
-    except Exception as e:
-        raise Exception(
-            f"An error occurred while processing the production audio files: {e}"
-        )
+    print("@@@@@@ SEPARATION TRAINING/TESTING START @@@@@@")
+    separation_manager()
+    print()
+    print("@@@@@@ MODEL SUCCESSFULLY CREATED @@@@@@")
 
 
 if __name__ == "__main__":

src/models/__helpers__/visualize_for_testing_evaluation.py

@@ -0,0 +1,31 @@
+from src.config.constants import VISUALIZATION_SAVE_PATH
+from src.visualization.spectrograms.visualize_spectrogram import visualize_spectrogram
+
+
+def visualize_for_testing_evaluation(y_pred, y_test, x_test, data, tag):
+    # Convert tensor back into numpy array
+    y_pred_for_visualization = y_pred.detach().cpu().numpy()
+    x_test_for_visualization = x_test.detach().cpu().numpy()
+    y_test_for_visualization = y_test.detach().cpu().numpy()
+
+    visualize_spectrogram(
+        save_folder_path=VISUALIZATION_SAVE_PATH,
+        mel_spectrogram=x_test_for_visualization[0],
+        mix_name=data["mix_name"],
+        tag=tag,
+        label="x_test",
+    )
+    visualize_spectrogram(
+        save_folder_path=VISUALIZATION_SAVE_PATH,
+        mel_spectrogram=y_pred_for_visualization[0],
+        mix_name=data["mix_name"],
+        tag=tag,
+        label="y_pred",
+    )
+    visualize_spectrogram(
+        save_folder_path=VISUALIZATION_SAVE_PATH,
+        mel_spectrogram=y_test_for_visualization[0],
+        mix_name=data["mix_name"],
+        tag=tag,
+        label="y_test",
+    )

src/models/__helpers__/visualize_for_training_evaluation.py

@@ -0,0 +1,41 @@
+from src.config.constants import VISUALIZATION_SAVE_PATH, TRAINED_AUDIO_FILE_PATH
+from src.transformers.freq_time_analysis_to_audio import freq_time_analysis_to_audio
+from src.visualization.spectrograms.visualize_spectrogram import visualize_spectrogram
+
+
+def visualize_for_training_evaluation(outputs, x, y, data, flag, tag):
+    # Convert tensor back into numpy array
+    outputs_for_visualization = outputs.detach().cpu().numpy()
+    x_train_for_visualization = x.detach().cpu().numpy()
+    y_train_for_visualization = y.detach().cpu().numpy()
+
+    visualize_spectrogram(
+        save_folder_path=VISUALIZATION_SAVE_PATH,
+        mel_spectrogram=x_train_for_visualization[0],
+        mix_name=data["mix_name"],
+        tag=tag,
+        label="x_train",
+    )
+    visualize_spectrogram(
+        save_folder_path=VISUALIZATION_SAVE_PATH,
+        mel_spectrogram=y_train_for_visualization[0],
+        mix_name=data["mix_name"],
+        tag=tag,
+        label="y_train",
+    )
+    visualize_spectrogram(
+        save_folder_path=VISUALIZATION_SAVE_PATH,
+        mel_spectrogram=outputs_for_visualization[0],
+        mix_name=data["mix_name"],
+        tag=tag,
+        label="y_output",
+    )
+
+    if flag:
+        freq_time_analysis_to_audio(
+            mel_spectrogram_array=outputs_for_visualization[:3],
+            output_file_path=TRAINED_AUDIO_FILE_PATH,
+            mix_names=data["mix_name"],
+            min_max_amplitudes=data["y_min_max_amplitudes"],
+            tag=tag,
+        )