File size: 4,397 Bytes
c84f2c4
fc7b4a9
 
 
 
c84f2c4
fc7b4a9
 
75d43d2
fc7b4a9
 
c84f2c4
fc7b4a9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
75d43d2
fc7b4a9
 
75d43d2
 
fc7b4a9
75d43d2
fc7b4a9
91f3c16
fc7b4a9
 
75d43d2
 
 
 
fc7b4a9
 
75d43d2
fc7b4a9
 
 
 
 
 
2b37a16
75d43d2
fc7b4a9
 
 
75d43d2
344eaf8
 
 
 
 
 
 
 
 
 
fc7b4a9
 
c84f2c4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fc7b4a9
 
75d43d2
 
 
 
 
c84f2c4
75d43d2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
from src.preprocessing.preprocessor import single_preprocessing, single_audio_preprocessing
from src.spectttra.spectttra_trainer import spectttra_predict
from src.llm2vectrain.model import load_llm2vec_model
from src.llm2vectrain.llm2vec_trainer import l2vec_single_train, load_pca_model
from src.models.mlp import build_mlp, load_config
from src.utils.dataset import instance_scaler, audio_instance_scaler

import numpy as np
import pandas as pd


def predict_multimodal(audio_file, lyrics):
    """
    Predict script which includes preprocessing, feature extraction, and
    training the MLP model for a single data sample.

    Parameters
    ----------
    audio : audio_object
        Audio object file

    lyric : string
        Lyric string

    Returns
    -------
    prediction : str
        A string result of the prediction

    label : int
        A numerical representation of the prediction
    """
    # 1.) Instantiate LLM2Vec Model
    llm2vec_model = load_llm2vec_model()

    # 2.) Preprocess both audio and lyrics
    audio, lyrics = single_preprocessing(audio_file, lyrics)

    # 3.) Call the train method for both models
    audio_features = spectttra_predict(audio)
    audio_features = audio_features.reshape(1, -1)
    lyrics_features = l2vec_single_train(llm2vec_model, lyrics)

    # 4.) Scale the vectors using Z-Score
    audio_features, lyrics_features = instance_scaler(audio_features, lyrics_features)

    # 5.) Reduce the lyrics using saved PCA model
    reduced_lyrics = load_pca_model(lyrics_features)

    # 6.) Concatenate the vectors of audio_features + lyrics_features
    results = np.concatenate([audio_features, reduced_lyrics], axis=1)

    # ---- Load MLP Classifier ----
    config = load_config("config/model_config.yml")
    classifier = build_mlp(input_dim=results.shape[1], config=config)

    # 7.) Load trained weights
    model_path = "models/mlp/mlp_best.pth"
    classifier.load_model(model_path)
    classifier.model.eval()

    # 8.) Run prediction
    confidence, prediction, label, probability = classifier.predict_single(
        results.flatten()
    )

    return {
        "confidence": confidence,
        "prediction": prediction,
        "label": label,
        "probability": probability,
    }


def predict_unimodal(audio_file):
    """
    Predict script of AUDIO only which includes preprocessing, feature extraction, and
    training the MLP model for a single data sample.

    Parameters
    ----------
    audio : audio_object
        Audio object file

    Returns
    -------
    prediction : str
        A string result of the prediction

    label : int
        A numerical representation of the prediction
    """

    # 1.) Preprocess the audio
    audio = single_audio_preprocessing(audio_file)

    # 2.) Call the inference method from SpecTTTra
    audio_features = spectttra_predict(audio)
    audio_features = audio_features.reshape(1, -1)

    # 4.) Scale the vector using Z-Score
    audio_features = audio_instance_scaler(audio_features)

    # 5.) Load MLP Classifier
    config = load_config("config/model_config.yml")
    classifier = build_mlp(input_dim=audio_features.shape[1], config=config)

    # 6.) Load trained weights
    model_path = "models/spectttra/mlp_best.pth"
    classifier.load_model(model_path)
    classifier.model.eval()

    # 8.) Run prediction
    confidence, prediction, label, probability = classifier.predict_single(
        audio_features.flatten()
    )

    return {
        "confidence": confidence,
        "prediction": prediction,
        "label": label,
        "probability": probability,
    }


if __name__ == "__main__":
    # Example usage (replace with real inputs, place song inside data/raw.)
    data = pd.read_csv("data/raw/predict_data_final.csv")

    result = []
    label = []
    for row in data.itertuples():
        prediction = predict_multimodal(row.song, row.lyrics)
        result.append(
            {
                "song": row.song,
                "label": row.label,
                "predicted_label": prediction["label"],
                "probability": prediction["probability"],
            }
        )

    for r in result:
        print(f"Song: {r['song']}")
        print(f"Actual Label: {r['label']}")
        print(f"Predicted: {r['predicted_label']}")
        print(f"Confidence: {r['probability']: .8f}%")
        print("-" * 50)