File size: 6,371 Bytes
97dae9c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1eebc2d
97dae9c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
"""
This script implements a deep learning pipeline for audio classification using a pre-trained MobileNetV2 model.
The pipeline includes data loading, model training, evaluation, and emissions tracking.
"""

import os
import torch
import torch.nn as nn
import torchaudio
from torch.utils.data import Dataset, DataLoader
import numpy as np
from transformers import AutoModelForImageClassification
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score
from tqdm import tqdm
import logging
from datasets import load_dataset
from accelerate import Accelerator
from codecarbon import EmissionsTracker
import time


class Config:
    """
    Configuration class to store hyperparameters and model settings.
    """
    SAMPLE_RATE = 16000
    N_FFT = 800
    N_MELS = 128
    HOP_LENGTH = None
    SIZE = (96, 96)
    SCALING_DIM = (1, 2)
    LEARNING_RATE = 0.0005
    BATCH_SIZE = 32
    NUM_WORKERS = 4
    NUM_EPOCHS = 1
    MODEL_NAME = "google/mobilenet_v2_0.35_96"
    MODEL_PATH = "scaled_model_800_128_96x96_mobilenet_small_unscaled_submission.pth"

config = Config()

class AudioDataset(Dataset):
    """
    Custom Dataset class for loading and processing audio data.

    Args:
        data (list): List of audio data samples.
        sample_rate (int, optional): Target sample rate for audio resampling. Defaults to 16000.
        audio_target_length (float, optional): Target length of audio in seconds. Defaults to 4.5.
    """
    def __init__(self, data, sample_rate=16000, audio_target_length=4.5):
        self.data = data
        self.sample_rate = sample_rate
        self.audio_target_length = audio_target_length
        
    def __len__(self):
        return len(self.data)

    def __getitem__(self, index):
        # 1. Cache the resampler
        if not hasattr(self, '_resampler_cache'):
            self._resampler_cache = {}
        
        # 2. Get data efficiently
        data_item = self.data[index]
        waveform = torch.FloatTensor(data_item["audio"]["array"]) if len(data_item["audio"]["array"]) > 0 else torch.ones(36000)*1E-5

        # 4. Cached resampler creation
        orig_freq = waveform.shape[-1]
        target_freq = self.audio_target_length * self.sample_rate
        resampler_key = (orig_freq, target_freq)
        
        if resampler_key not in self._resampler_cache:
            self._resampler_cache[resampler_key] = torchaudio.transforms.Resample(
                orig_freq=orig_freq,
                new_freq=target_freq
            )
        # 5. Apply resampling and return
        return self._resampler_cache[resampler_key](waveform), data_item["label"]


def collate_fn(batch):
    """
    Collate function to stack inputs and labels into batches.
    Args:
        batch (list): List of tuples containing inputs and labels.
    Returns:
        tuple: Stacked inputs and labels.
    """
    return torch.stack([inputs for inputs, _ in batch]), torch.tensor([label for _, label in batch])


class AudioClassifier(nn.Module):
    """
    Audio classification model using a pre-trained MobileNetV2.
    Args:
        model_name (str): Name of the pre-trained model.
        model_path (str): Path to save/load the model.
        new (bool, optional): Whether to load a new model or an existing one. Defaults to True.
    """
    def __init__(self, model_name, model_path, new=True):
        super().__init__()
        self.model = self.load_model(model_name, model_path, new)
        self.num_classes = 2
        self.mel_spectrogram = torchaudio.transforms.MelSpectrogram(
            sample_rate=config.SAMPLE_RATE,
            n_fft=config.N_FFT,
            n_mels=config.N_MELS,
            hop_length=config.HOP_LENGTH
        )
        self.amplitude_to_db = torchaudio.transforms.AmplitudeToDB()

    def load_model(self, model_name, model_path, new=False):
        """
        Load the pre-trained model and modify the classifier.

        Args:
            model_name (str): Name of the pre-trained model.
            model_path (str): Path to save/load the model.
            new (bool, optional): Whether to load a new model or an existing one. Defaults to False.
        Returns:
            nn.Module: Loaded model.
        """
        model = AutoModelForImageClassification.from_pretrained(model_name)
        model.classifier = torch.nn.Sequential(
            nn.Linear(in_features=1280, out_features=2))
        
        for param in model.parameters():
            param.requires_grad = True
        state_dict = torch.load(model_path)
        model.load_state_dict(state_dict)
        return model

    def forward(self, waveforms):
        """
        Forward pass through the model.
        Args:
            waveforms (torch.Tensor): Input audio waveforms.
        Returns:
            torch.Tensor: Model output.
        """
        melspectrogram = self.mel_spectrogram(waveforms)
        melspectrogram = nn.functional.interpolate(melspectrogram.unsqueeze(1),
                                                   size=config.SIZE,
                                                   mode="bilinear",
                                                   align_corners=False).squeeze(1)
        db_melspectrogram = self.amplitude_to_db(melspectrogram)
        delta = torchaudio.functional.compute_deltas(melspectrogram)
        x = torch.stack([melspectrogram, db_melspectrogram, delta], dim=1)
        return self.model(x)


class Evaluator:
    def __init__(self, model, dataloader, device):
        self.model = model
        self.dataloader = dataloader
        self.device = device
        
    @torch.no_grad()
    def evaluate(self):
        self.model.eval()
        all_predictions = []
        all_labels = []

        idx = 0
        for waveforms, labels in self.dataloader:
            waveforms = waveforms.to(self.device)
            outputs = self.model(waveforms).logits
            predictions = torch.argmax(outputs, dim=1)
            all_predictions.extend(predictions.cpu().numpy())
            all_labels.extend(labels.cpu().numpy())
            idx += 1
            if idx % 10 == 0:
                torch.cuda.empty_cache()
                
        all_predictions = np.array(all_predictions)
        all_labels = np.array(all_labels)

        # return self.compute_metrics(all_predictions, all_labels)
        return all_predictions