Upload gtzan_dataset_linear_probe.py
Browse files- gtzan_dataset_linear_probe.py +296 -0
gtzan_dataset_linear_probe.py
ADDED
@@ -0,0 +1,296 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Union, Callable, List, Optional, Dict
|
2 |
+
import torch
|
3 |
+
import torch.nn as nn
|
4 |
+
from torch.utils.data import Dataset, DataLoader
|
5 |
+
from torch.optim import Adam
|
6 |
+
import numpy as np
|
7 |
+
import librosa
|
8 |
+
import miniaudio
|
9 |
+
from pathlib import Path
|
10 |
+
from sklearn.model_selection import train_test_split
|
11 |
+
from tqdm import tqdm
|
12 |
+
from functools import partial
|
13 |
+
import math
|
14 |
+
|
15 |
+
from mae import MaskedAutoencoderViT
|
16 |
+
|
17 |
+
|
18 |
+
def load_audio(
|
19 |
+
path: str,
|
20 |
+
sr: int = 32000,
|
21 |
+
duration: int = 20,
|
22 |
+
) -> (np.ndarray, int):
|
23 |
+
g = miniaudio.stream_file(path, output_format=miniaudio.SampleFormat.FLOAT32, nchannels=1,
|
24 |
+
sample_rate=sr, frames_to_read=sr * duration)
|
25 |
+
signal = np.array(next(g))
|
26 |
+
return signal
|
27 |
+
|
28 |
+
|
29 |
+
def mel_spectrogram(
|
30 |
+
signal: np.ndarray,
|
31 |
+
sr: int = 32000,
|
32 |
+
n_fft: int = 800,
|
33 |
+
hop_length: int = 320,
|
34 |
+
n_mels: int = 128,
|
35 |
+
) -> np.ndarray:
|
36 |
+
mel_spec = librosa.feature.melspectrogram(
|
37 |
+
y=signal, sr=sr, n_fft=n_fft, hop_length=hop_length, n_mels=n_mels,
|
38 |
+
window='hann', pad_mode='constant'
|
39 |
+
)
|
40 |
+
mel_spec = librosa.power_to_db(mel_spec) # (freq, time)
|
41 |
+
return mel_spec.T # (time, freq)
|
42 |
+
|
43 |
+
|
44 |
+
def normalize(arr: np.ndarray, eps: float = 1e-8) -> np.ndarray:
|
45 |
+
return (arr - arr.mean()) / (arr.std() + eps)
|
46 |
+
|
47 |
+
|
48 |
+
device = 'cuda:0'
|
49 |
+
seed = 42
|
50 |
+
train_size = 0.8 # 80% train, 20% test
|
51 |
+
batch_size_train = 10
|
52 |
+
batch_size_test = 32
|
53 |
+
num_workers = 1
|
54 |
+
lr = 1e-3
|
55 |
+
epochs = 200
|
56 |
+
detection_epoch = 20
|
57 |
+
|
58 |
+
sr = 32000
|
59 |
+
n_fft = 800 # 25ms
|
60 |
+
hop_length = 320 # 10ms
|
61 |
+
duration = 10000 # seconds. 10000 ~= Inf for reading the whole audio file
|
62 |
+
|
63 |
+
feature_length = 2048 # length of mel spectrogram (MAE is trained with 2048x128 mel spectrogram)
|
64 |
+
patch_size = 16 # MAE split the mel spectrogram into patches with size 16x16
|
65 |
+
|
66 |
+
feature_padding = True
|
67 |
+
header = 'mean'
|
68 |
+
|
69 |
+
mlp_num_neurons = [768, 10]
|
70 |
+
mlp_activation_layer = nn.ReLU
|
71 |
+
mlp_bias = True
|
72 |
+
|
73 |
+
torch.manual_seed(seed)
|
74 |
+
np.random.seed(seed)
|
75 |
+
torch.cuda.manual_seed_all(seed)
|
76 |
+
torch.backends.cudnn.deterministic = True
|
77 |
+
torch.backends.cudnn.benchmark = False
|
78 |
+
|
79 |
+
# =============================== model ===============================
|
80 |
+
mae = MaskedAutoencoderViT(
|
81 |
+
img_size=(2048, 128),
|
82 |
+
patch_size=16,
|
83 |
+
in_chans=1,
|
84 |
+
embed_dim=768,
|
85 |
+
depth=12,
|
86 |
+
num_heads=12,
|
87 |
+
decoder_mode=1,
|
88 |
+
no_shift=False,
|
89 |
+
decoder_embed_dim=512,
|
90 |
+
norm_layer=partial(nn.LayerNorm, eps=1e-6),
|
91 |
+
norm_pix_loss=False,
|
92 |
+
pos_trainable=False,
|
93 |
+
)
|
94 |
+
|
95 |
+
# Load pre-trained weights
|
96 |
+
ckpt_path = 'music-mae-32kHz.pth.pth'
|
97 |
+
mae.load_state_dict(torch.load(ckpt_path, map_location='cpu'))
|
98 |
+
mae.to(device)
|
99 |
+
mae.eval()
|
100 |
+
|
101 |
+
# =============================== data ===============================
|
102 |
+
fp = Path('GTZAN-dataset/genres_original')
|
103 |
+
audio_data = dict() # {genre: [audio_file1, audio_file2, ...]}
|
104 |
+
|
105 |
+
for d in fp.iterdir():
|
106 |
+
if d.is_dir():
|
107 |
+
for f in d.iterdir():
|
108 |
+
if f.is_file():
|
109 |
+
genres = f.name.split('.')[0]
|
110 |
+
if genres not in audio_data:
|
111 |
+
audio_data[genres] = [str(f)]
|
112 |
+
else:
|
113 |
+
audio_data[genres].append(str(f))
|
114 |
+
|
115 |
+
audio_data_train = dict()
|
116 |
+
audio_data_test = dict()
|
117 |
+
|
118 |
+
for k, v in audio_data.items():
|
119 |
+
train_data, test_data = train_test_split(v, train_size=train_size, random_state=seed, shuffle=True)
|
120 |
+
audio_data_train[k] = train_data
|
121 |
+
audio_data_test[k] = test_data
|
122 |
+
|
123 |
+
|
124 |
+
@torch.no_grad()
|
125 |
+
def infer_mae_embedding(data: Dict) -> Dict:
|
126 |
+
emb_data = dict() # {genre: [embed1, embed2, ...]}
|
127 |
+
|
128 |
+
for k, v in tqdm(data.items(), desc='infer mae embedding', total=len(data)):
|
129 |
+
for f in v:
|
130 |
+
try:
|
131 |
+
mel_spec = mel_spectrogram(load_audio(f, duration=duration), sr=sr, n_fft=n_fft, hop_length=hop_length)
|
132 |
+
except Exception as e:
|
133 |
+
print(e)
|
134 |
+
print(f)
|
135 |
+
continue
|
136 |
+
|
137 |
+
# pad the mel spectrogram to the multiple of patch_size
|
138 |
+
input_length = mel_spec.shape[0]
|
139 |
+
n = math.ceil(input_length / patch_size)
|
140 |
+
if input_length < patch_size * n:
|
141 |
+
pad_length = patch_size * n - input_length
|
142 |
+
mel_spec = np.pad(mel_spec, ((0, pad_length), (0, 0)), mode='constant', constant_values=mel_spec.min())
|
143 |
+
|
144 |
+
# if the length of mel spectrogram after padding is longer than feature_length,
|
145 |
+
# split it into multiple snippets
|
146 |
+
input_length = mel_spec.shape[0]
|
147 |
+
embeds = []
|
148 |
+
for i in range(0, input_length, feature_length):
|
149 |
+
snippet = mel_spec[i:i + feature_length]
|
150 |
+
snippet = normalize(snippet)
|
151 |
+
snippet = snippet[None, None, :, :]
|
152 |
+
x = torch.from_numpy(snippet).to(device)
|
153 |
+
y = mae.forward_encoder_no_mask(x, header=header) # (1, 768)
|
154 |
+
y = y / y.norm(p=2, dim=-1, keepdim=True) # normalize
|
155 |
+
y = y.cpu().numpy().squeeze()
|
156 |
+
embeds.append(y)
|
157 |
+
|
158 |
+
y = np.mean(embeds, axis=0) # (768,)
|
159 |
+
|
160 |
+
if k not in emb_data:
|
161 |
+
emb_data[k] = [y]
|
162 |
+
else:
|
163 |
+
emb_data[k].append(y)
|
164 |
+
|
165 |
+
return emb_data
|
166 |
+
|
167 |
+
|
168 |
+
audio_emb_train = infer_mae_embedding(audio_data_train)
|
169 |
+
audio_emb_test = infer_mae_embedding(audio_data_test)
|
170 |
+
|
171 |
+
label_set = set(audio_emb_train.keys())
|
172 |
+
label_map = {label: i for i, label in enumerate(label_set)}
|
173 |
+
print(label_map)
|
174 |
+
|
175 |
+
|
176 |
+
class MLP(torch.nn.Sequential):
|
177 |
+
def __init__(
|
178 |
+
self,
|
179 |
+
num_neurons: List[int],
|
180 |
+
activation_layer: Optional[Callable[..., torch.nn.Module]] = torch.nn.ReLU,
|
181 |
+
bias: bool = True,
|
182 |
+
dropout: float = 0.0,
|
183 |
+
):
|
184 |
+
layers = []
|
185 |
+
for c_in, c_out in zip(num_neurons[:-1], num_neurons[1:]):
|
186 |
+
layers.append(torch.nn.Linear(c_in, c_out, bias=bias))
|
187 |
+
layers.append(activation_layer())
|
188 |
+
layers.append(torch.nn.Dropout(dropout))
|
189 |
+
|
190 |
+
# remove the last two layers
|
191 |
+
layers.pop()
|
192 |
+
layers.pop()
|
193 |
+
|
194 |
+
super().__init__(*layers)
|
195 |
+
|
196 |
+
|
197 |
+
class SimpleDataset(Dataset):
|
198 |
+
def __init__(self, dict_data: Dict, label_map: Dict):
|
199 |
+
self.embed_with_label = []
|
200 |
+
|
201 |
+
for k, v in dict_data.items():
|
202 |
+
for emb in v:
|
203 |
+
self.embed_with_label.append((emb, label_map[k]))
|
204 |
+
|
205 |
+
def __len__(self):
|
206 |
+
return len(self.embed_with_label)
|
207 |
+
|
208 |
+
def __getitem__(self, idx):
|
209 |
+
return self.embed_with_label[idx]
|
210 |
+
|
211 |
+
|
212 |
+
train_dataset = SimpleDataset(audio_emb_train, label_map)
|
213 |
+
test_dataset = SimpleDataset(audio_emb_test, label_map)
|
214 |
+
print(f"len(train_dataset): {len(train_dataset)}")
|
215 |
+
print(f"len(test_dataset): {len(test_dataset)}")
|
216 |
+
|
217 |
+
|
218 |
+
def train_one_epoch(model, device, dataloader, loss_fn, optimizer):
|
219 |
+
model.train()
|
220 |
+
|
221 |
+
# for batch in tqdm(dataloader, desc='train', total=len(dataloader)):
|
222 |
+
for batch in dataloader:
|
223 |
+
x, y = batch
|
224 |
+
x = x.to(device)
|
225 |
+
y = y.to(device)
|
226 |
+
|
227 |
+
y_logit = model(x)
|
228 |
+
loss = loss_fn(y_logit, y)
|
229 |
+
|
230 |
+
optimizer.zero_grad()
|
231 |
+
loss.backward()
|
232 |
+
optimizer.step()
|
233 |
+
|
234 |
+
|
235 |
+
@torch.no_grad()
|
236 |
+
def eval_one_epoch(model, device, dataloader, loss_fn):
|
237 |
+
model.eval()
|
238 |
+
|
239 |
+
total_loss = 0.0
|
240 |
+
total_correct = 0.0
|
241 |
+
total_num = 0.0
|
242 |
+
|
243 |
+
for batch in dataloader:
|
244 |
+
x, y = batch
|
245 |
+
x = x.to(device)
|
246 |
+
y = y.to(device)
|
247 |
+
|
248 |
+
y_logit = model(x)
|
249 |
+
loss = loss_fn(y_logit, y)
|
250 |
+
|
251 |
+
total_loss += loss.item() * x.shape[0]
|
252 |
+
total_correct += (y_logit.argmax(dim=-1) == y).sum().item()
|
253 |
+
total_num += x.shape[0]
|
254 |
+
|
255 |
+
loss = total_loss / total_num
|
256 |
+
acc = total_correct / total_num
|
257 |
+
|
258 |
+
return loss, acc
|
259 |
+
|
260 |
+
|
261 |
+
mlp = MLP(
|
262 |
+
num_neurons=mlp_num_neurons,
|
263 |
+
activation_layer=mlp_activation_layer,
|
264 |
+
bias=mlp_bias,
|
265 |
+
dropout=0.0
|
266 |
+
)
|
267 |
+
print(MLP)
|
268 |
+
|
269 |
+
mlp.to(device)
|
270 |
+
|
271 |
+
optimizer = Adam(mlp.parameters(), lr=lr)
|
272 |
+
loss_fn = nn.CrossEntropyLoss()
|
273 |
+
|
274 |
+
train_dataloader = DataLoader(train_dataset, batch_size=batch_size_train, shuffle=True, num_workers=num_workers)
|
275 |
+
test_dataloader = DataLoader(test_dataset, batch_size=batch_size_test, shuffle=False, num_workers=num_workers)
|
276 |
+
|
277 |
+
test_loss, test_accuracy = eval_one_epoch(mlp, device, test_dataloader, loss_fn)
|
278 |
+
print(f"init: test loss {test_loss:.4f}, test accuracy {test_accuracy:.4f}")
|
279 |
+
|
280 |
+
best_accuracy = 0.0
|
281 |
+
at = 0
|
282 |
+
|
283 |
+
for epoch in range(epochs):
|
284 |
+
train_one_epoch(mlp, device, train_dataloader, loss_fn, optimizer)
|
285 |
+
test_loss, test_accuracy = eval_one_epoch(mlp, device, test_dataloader, loss_fn)
|
286 |
+
|
287 |
+
print(f"epoch {epoch}: test loss {test_loss:.4f}, test accuracy {test_accuracy:.4f}")
|
288 |
+
|
289 |
+
if test_accuracy > best_accuracy:
|
290 |
+
best_accuracy = test_accuracy
|
291 |
+
at = epoch
|
292 |
+
|
293 |
+
if epoch - at >= detection_epoch:
|
294 |
+
print(f"early stop at epoch {epoch}")
|
295 |
+
print(f"best accuracy: {best_accuracy:.4f} at epoch {at}")
|
296 |
+
break
|