gtzan_dataset_linear_probe.py

from typing import Union, Callable, List, Optional, Dict
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
from torch.optim import Adam
import numpy as np
import librosa
import miniaudio
from pathlib import Path
from sklearn.model_selection import train_test_split
from tqdm import tqdm
from functools import partial
import math

from mae import MaskedAutoencoderViT


def load_audio(
        path: str,
        sr: int = 32000,
        duration: int = 20,
) -> (np.ndarray, int):
    g = miniaudio.stream_file(path, output_format=miniaudio.SampleFormat.FLOAT32, nchannels=1,
                              sample_rate=sr, frames_to_read=sr * duration)
    signal = np.array(next(g))
    return signal


def mel_spectrogram(
        signal: np.ndarray,
        sr: int = 32000,
        n_fft: int = 800,
        hop_length: int = 320,
        n_mels: int = 128,
) -> np.ndarray:
    mel_spec = librosa.feature.melspectrogram(
        y=signal, sr=sr, n_fft=n_fft, hop_length=hop_length, n_mels=n_mels,
        window='hann', pad_mode='constant'
    )
    mel_spec = librosa.power_to_db(mel_spec)  # (freq, time)
    return mel_spec.T  # (time, freq)


def normalize(arr: np.ndarray, eps: float = 1e-8) -> np.ndarray:
    return (arr - arr.mean()) / (arr.std() + eps)


device = 'cuda:0'
seed = 42
train_size = 0.8  # 80% train, 20% test
batch_size_train = 10
batch_size_test = 32
num_workers = 1
lr = 1e-3
epochs = 200
detection_epoch = 20

sr = 32000
n_fft = 800  # 25ms
hop_length = 320  # 10ms
duration = 10000  # seconds. 10000 ~= Inf for reading the whole audio file

feature_length = 2048  # length of mel spectrogram (MAE is trained with 2048x128 mel spectrogram)
patch_size = 16  # MAE split the mel spectrogram into patches with size 16x16

feature_padding = True
header = 'mean'

mlp_num_neurons = [768, 10]
mlp_activation_layer = nn.ReLU
mlp_bias = True

torch.manual_seed(seed)
np.random.seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False

# =============================== model ===============================
mae = MaskedAutoencoderViT(
    img_size=(2048, 128),
    patch_size=16,
    in_chans=1,
    embed_dim=768,
    depth=12,
    num_heads=12,
    decoder_mode=1,
    no_shift=False,
    decoder_embed_dim=512,
    norm_layer=partial(nn.LayerNorm, eps=1e-6),
    norm_pix_loss=False,
    pos_trainable=False,
)

# Load pre-trained weights
ckpt_path = 'music-mae-32kHz.pth.pth'
mae.load_state_dict(torch.load(ckpt_path, map_location='cpu'))
mae.to(device)
mae.eval()

# =============================== data ===============================
fp = Path('GTZAN-dataset/genres_original')
audio_data = dict()  # {genre: [audio_file1, audio_file2, ...]}

for d in fp.iterdir():
    if d.is_dir():
        for f in d.iterdir():
            if f.is_file():
                genres = f.name.split('.')[0]
                if genres not in audio_data:
                    audio_data[genres] = [str(f)]
                else:
                    audio_data[genres].append(str(f))

audio_data_train = dict()
audio_data_test = dict()

for k, v in audio_data.items():
    train_data, test_data = train_test_split(v, train_size=train_size, random_state=seed, shuffle=True)
    audio_data_train[k] = train_data
    audio_data_test[k] = test_data


@torch.no_grad()
def infer_mae_embedding(data: Dict) -> Dict:
    emb_data = dict()  # {genre: [embed1, embed2, ...]}

    for k, v in tqdm(data.items(), desc='infer mae embedding', total=len(data)):
        for f in v:
            try:
                mel_spec = mel_spectrogram(load_audio(f, duration=duration), sr=sr, n_fft=n_fft, hop_length=hop_length)
            except Exception as e:
                print(e)
                print(f)
                continue

            # pad the mel spectrogram to the multiple of patch_size
            input_length = mel_spec.shape[0]
            n = math.ceil(input_length / patch_size)
            if input_length < patch_size * n:
                pad_length = patch_size * n - input_length
                mel_spec = np.pad(mel_spec, ((0, pad_length), (0, 0)), mode='constant', constant_values=mel_spec.min())

            # if the length of mel spectrogram after padding is longer than feature_length,
            # split it into multiple snippets
            input_length = mel_spec.shape[0]
            embeds = []
            for i in range(0, input_length, feature_length):
                snippet = mel_spec[i:i + feature_length]
                snippet = normalize(snippet)
                snippet = snippet[None, None, :, :]
                x = torch.from_numpy(snippet).to(device)
                y = mae.forward_encoder_no_mask(x, header=header)  # (1, 768)
                y = y / y.norm(p=2, dim=-1, keepdim=True)  # normalize
                y = y.cpu().numpy().squeeze()
                embeds.append(y)

            y = np.mean(embeds, axis=0)  # (768,)

            if k not in emb_data:
                emb_data[k] = [y]
            else:
                emb_data[k].append(y)

    return emb_data


audio_emb_train = infer_mae_embedding(audio_data_train)
audio_emb_test = infer_mae_embedding(audio_data_test)

label_set = set(audio_emb_train.keys())
label_map = {label: i for i, label in enumerate(label_set)}
print(label_map)


class MLP(torch.nn.Sequential):
    def __init__(
            self,
            num_neurons: List[int],
            activation_layer: Optional[Callable[..., torch.nn.Module]] = torch.nn.ReLU,
            bias: bool = True,
            dropout: float = 0.0,
    ):
        layers = []
        for c_in, c_out in zip(num_neurons[:-1], num_neurons[1:]):
            layers.append(torch.nn.Linear(c_in, c_out, bias=bias))
            layers.append(activation_layer())
            layers.append(torch.nn.Dropout(dropout))

        # remove the last two layers
        layers.pop()
        layers.pop()

        super().__init__(*layers)


class SimpleDataset(Dataset):
    def __init__(self, dict_data: Dict, label_map: Dict):
        self.embed_with_label = []

        for k, v in dict_data.items():
            for emb in v:
                self.embed_with_label.append((emb, label_map[k]))

    def __len__(self):
        return len(self.embed_with_label)

    def __getitem__(self, idx):
        return self.embed_with_label[idx]


train_dataset = SimpleDataset(audio_emb_train, label_map)
test_dataset = SimpleDataset(audio_emb_test, label_map)
print(f"len(train_dataset): {len(train_dataset)}")
print(f"len(test_dataset): {len(test_dataset)}")


def train_one_epoch(model, device, dataloader, loss_fn, optimizer):
    model.train()

    # for batch in tqdm(dataloader, desc='train', total=len(dataloader)):
    for batch in dataloader:
        x, y = batch
        x = x.to(device)
        y = y.to(device)

        y_logit = model(x)
        loss = loss_fn(y_logit, y)

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()


@torch.no_grad()
def eval_one_epoch(model, device, dataloader, loss_fn):
    model.eval()

    total_loss = 0.0
    total_correct = 0.0
    total_num = 0.0

    for batch in dataloader:
        x, y = batch
        x = x.to(device)
        y = y.to(device)

        y_logit = model(x)
        loss = loss_fn(y_logit, y)

        total_loss += loss.item() * x.shape[0]
        total_correct += (y_logit.argmax(dim=-1) == y).sum().item()
        total_num += x.shape[0]

    loss = total_loss / total_num
    acc = total_correct / total_num

    return loss, acc


mlp = MLP(
    num_neurons=mlp_num_neurons,
    activation_layer=mlp_activation_layer,
    bias=mlp_bias,
    dropout=0.0
)
print(MLP)

mlp.to(device)

optimizer = Adam(mlp.parameters(), lr=lr)
loss_fn = nn.CrossEntropyLoss()

train_dataloader = DataLoader(train_dataset, batch_size=batch_size_train, shuffle=True, num_workers=num_workers)
test_dataloader = DataLoader(test_dataset, batch_size=batch_size_test, shuffle=False, num_workers=num_workers)

test_loss, test_accuracy = eval_one_epoch(mlp, device, test_dataloader, loss_fn)
print(f"init: test loss {test_loss:.4f}, test accuracy {test_accuracy:.4f}")

best_accuracy = 0.0
at = 0

for epoch in range(epochs):
    train_one_epoch(mlp, device, train_dataloader, loss_fn, optimizer)
    test_loss, test_accuracy = eval_one_epoch(mlp, device, test_dataloader, loss_fn)

    print(f"epoch {epoch}: test loss {test_loss:.4f}, test accuracy {test_accuracy:.4f}")

    if test_accuracy > best_accuracy:
        best_accuracy = test_accuracy
        at = epoch

    if epoch - at >= detection_epoch:
        print(f"early stop at epoch {epoch}")
        print(f"best accuracy: {best_accuracy:.4f} at epoch {at}")
        break