MFCC (Mel-Frequency Cepstral Coefficients)

157 Views

MFCC (Mel-Frequency Cepstral Coefficients) adalah fitur yang paling umum digunakan dalam pemrosesan sinyal audio, terutama untuk pengenalan suara dan klasifikasi suara. MFCC meniru cara manusia mendengar dan menangkap karakteristik suara berdasarkan persepsi frekuensi.

Langkah-langkah Ekstraksi MFCC dari Sinyal Audio

1. Pre-emphasis (opsional)
Meningkatkan energi frekuensi tinggi:

$y[n] = x[n] - \alpha x[n-1]$

Biasanya, $\alpha \approx 0.95$

2. Framing
Sinyal dibagi menjadi frame kecil (misalnya 25ms) untuk menangkap dinamika lokal.

3. Windowing
Untuk menghindari efek diskontinuitas di tepi frame. Umumnya digunakan jendela Hamming:

$w[n] = 0.54 - 0.46 \cos\left(\frac{2\pi n}{N-1}\right)$

4. FFT (Fast Fourier Transform)
Mengubah sinyal dari domain waktu ke domain frekuensi.

5. Mel Filter Bank
Energi spektrum dikalikan dengan filter bank mel (biasanya 20-40 filter). Skala Mel meniru persepsi manusia terhadap frekuensi:

$m = 2595 \log_{10}\left(1 + \frac{f}{700}\right)$

6. Log Energi
Ambil log dari hasil setiap filter untuk mensimulasikan persepsi logaritmik manusia terhadap suara.

7. DCT (Discrete Cosine Transform)
DCT diterapkan pada log-energi dari filter bank untuk mendapatkan koefisien MFCC (biasanya ambil 12-13 pertama + 1 energi total):

$\text{MFCC}[k] = \sum_{n=1}^{N} \log(S[n]) \cos\left[\frac{\pi k}{N}(n-0.5)\right]$

Contoh menggunakan librosa

import librosa
import librosa.display
import matplotlib.pyplot as plt

# Load audio
y, sr = librosa.load('audio.wav', sr=None)

# Ekstraksi MFCC
mfcc = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=13)

# Visualisasi
plt.figure(figsize=(10, 4))
librosa.display.specshow(mfcc, x_axis='time')
plt.colorbar()
plt.title('MFCC')
plt.tight_layout()
plt.show()

Contoh menggunakan torchaudio

import torch
import torchaudio
import matplotlib.pyplot as plt

# Load audio (contoh: file .wav)
waveform, sample_rate = torchaudio.load("audio.wav")  # waveform shape: [channel, time]

# Pastikan audio mono (gunakan channel pertama jika stereo)
if waveform.shape[0] > 1:
    waveform = waveform[0:1, :]  # ambil channel pertama

# Ekstraksi MFCC
mfcc_transform = torchaudio.transforms.MFCC(
    sample_rate=sample_rate,
    n_mfcc=13,  # Jumlah koefisien MFCC
    melkwargs={
        'n_fft': 400,        # ukuran FFT
        'hop_length': 160,   # jarak antar frame (biasanya 10ms = 160 jika sr=16k)
        'n_mels': 40,        # jumlah filter mel
        'center': False
    }
)

mfcc = mfcc_transform(waveform)  # shape: [channel, n_mfcc, time]

# Visualisasi
plt.figure(figsize=(10, 4))
plt.imshow(mfcc[0].detach().numpy(), aspect='auto', origin='lower')
plt.title("MFCC (PyTorch)")
plt.xlabel("Frame Index")
plt.ylabel("MFCC Coefficients")
plt.colorbar()
plt.tight_layout()
plt.show()

Contoh bila menggunakan numpy

import numpy as np
from scipy.fftpack import dct

def mfcc_numpy(signal, sample_rate=16000, 
               pre_emphasis_coef=0.97, 
               frame_size=0.025,  # 25 ms
               frame_stride=0.01, # 10 ms
               num_filters=26, 
               nfft=512, 
               num_ceps=13):
    # 1. Pre-emphasis
    emphasized_signal = np.append(signal[0], signal[1:] - pre_emphasis_coef * signal[:-1])

    # 2. Framing
    frame_length = int(round(frame_size * sample_rate))
    frame_step = int(round(frame_stride * sample_rate))
    signal_length = len(emphasized_signal)
    num_frames = int(np.ceil(float(np.abs(signal_length - frame_length)) / frame_step)) + 1

    pad_signal_length = num_frames * frame_step + frame_length
    z = np.zeros((pad_signal_length - signal_length))
    pad_signal = np.append(emphasized_signal, z)

    indices = np.tile(np.arange(0, frame_length), (num_frames,1)) + \
              np.tile(np.arange(0, num_frames * frame_step, frame_step), (frame_length,1)).T
    frames = pad_signal[indices.astype(np.int32, copy=False)]

    # 3. Windowing - Hamming
    frames *= np.hamming(frame_length)

    # 4. FFT and Power Spectrum
    mag_frames = np.absolute(np.fft.rfft(frames, nfft))  # Magnitude of the FFT
    pow_frames = ((1.0 / nfft) * (mag_frames ** 2))      # Power Spectrum

    # 5. Filter Banks
    low_freq_mel = 0
    high_freq_mel = 2595 * np.log10(1 + (sample_rate / 2) / 700)  # Convert Hz to Mel
    mel_points = np.linspace(low_freq_mel, high_freq_mel, num_filters + 2)  # Equally spaced in Mel scale
    hz_points = 700 * (10**(mel_points / 2595) - 1)  # Convert Mel to Hz
    bin = np.floor((nfft + 1) * hz_points / sample_rate).astype(np.int32)

    fbank = np.zeros((num_filters, int(np.floor(nfft / 2 + 1))))
    for m in range(1, num_filters + 1):
        f_m_minus = bin[m - 1]   # left
        f_m = bin[m]             # center
        f_m_plus = bin[m + 1]    # right

        for k in range(f_m_minus, f_m):
            fbank[m - 1, k] = (k - bin[m - 1]) / (bin[m] - bin[m - 1])
        for k in range(f_m, f_m_plus):
            fbank[m - 1, k] = (bin[m + 1] - k) / (bin[m + 1] - bin[m])

    filter_banks = np.dot(pow_frames, fbank.T)
    # Numerical stability
    filter_banks = np.where(filter_banks == 0, np.finfo(float).eps, filter_banks)

    # 6. Log Energies
    filter_banks = 20 * np.log10(filter_banks)

    # 7. DCT to get MFCC
    mfcc = dct(filter_banks, type=2, axis=1, norm='ortho')[:, :num_ceps]

    return mfcc

# Contoh penggunaan:
if __name__ == "__main__":
    # Generate contoh sinyal sinus 440 Hz selama 1 detik
    fs = 16000
    t = np.linspace(0, 1, fs)
    freq = 440
    signal = 0.5 * np.sin(2 * np.pi * freq * t)

    mfcc_features = mfcc_numpy(signal, sample_rate=fs)
    print(mfcc_features.shape)  # (num_frames, num_ceps)

Pengenalan kata Yes dan No menggunakan MFCC

Contents

0.1 Pengenalan kata Yes dan No menggunakan MFCC
0.2 Penerapan MFCC
0.3 CNN untuk klasifikasi suara sederhana
- 0.3.1 Dataset Kustom: YESNO Segmen per Kata
- 0.3.2 Model CNN nya sebagai berikut
  - 0.3.2.1 solusi 1: Pad MFCC agar Semua Sama Panjang
  - 0.3.2.2 solusi 2: Pakai collate_fn custom (jika ingin flexible padding)

1 Catatan Penting untuk Model

Kita akan menggunakan dataset yang sudah ada di torchaudio. torchaudio.datasets.YESNO adalah dataset built-in dari torchaudio yang berisi rekaman suara seseorang yang mengucapkan kata “yes” atau “no” dalam bahasa Inggris sebanyak 8 kata per rekaman.

| Atribut         | Keterangan                               |
| --------------- | ---------------------------------------- |
| Asal            | CMU Arctic corpus                        |
| Isi             | 60 rekaman audio WAV                     |
| Panjang rekaman | \~2 detik                                |
| Ucapan per file | 8 kata (kombinasi "yes" dan "no")        |
| Jumlah label    | 8 angka biner per file (1 = yes, 0 = no) |
| Jumlah file     | 60 file WAV (dengan label)               |
| Format          | WAV 16kHz mono                           |

mari kita loading dataset nya agar lebih jelas

import torchaudio

# Download dan load dataset YESNO
dataset = torchaudio.datasets.YESNO(root="data", download=True)

# Ambil satu sampel
waveform, sample_rate, labels = dataset[0]

print("Waveform shape:", waveform.shape)
print("Sample rate:", sample_rate)
print("Labels:", labels)  # Contoh: [1, 1, 1, 1, 1, 1, 1, 1] → semua "yes"

mari kita putar saja audionya

import matplotlib.pyplot as plt
import IPython.display as ipd

# Tampilkan waveform
plt.figure(figsize=(10, 3))
plt.plot(waveform[0].numpy())
plt.title(f"Labels: {labels}")
plt.show()

# Putar audio (di notebook)
ipd.Audio(waveform.numpy(), rate=sample_rate)

Kita buat spliting 8 segment diatas

Berikut adalah fungsi otomatis memotong sinyal audio dari torchaudio.datasets.YESNO menjadi 8 segmen kata (masing-masing mewakili satu kata: “yes” atau “no”).

def split_yesno_segments(waveform, sample_rate):
    """
    Memotong satu waveform YESNO menjadi 8 segmen kata (yes/no).
    
    Args:
        waveform (Tensor): Tensor audio [1, T]
        sample_rate (int): Sampel rate (biasanya 16000 Hz)

    Returns:
        List[Tensor]: List 8 segmen waveform, masing-masing [1, segment_length]
    """
    total_duration_sec = waveform.shape[1] / sample_rate
    segment_duration = total_duration_sec / 8  # 8 kata per audio
    
    segment_samples = int(segment_duration * sample_rate)

    segments = []
    for i in range(8):
        start = i * segment_samples
        end = (i + 1) * segment_samples
        segment = waveform[:, start:end]
        segments.append(segment)
    
    return segments

contoh penggunaan

import torchaudio
import matplotlib.pyplot as plt

# Load dataset YESNO
dataset = torchaudio.datasets.YESNO(root="data", download=True)
waveform, sample_rate, labels = dataset[0]

# Potong jadi 8 segmen
segments = split_yesno_segments(waveform, sample_rate)

# Visualisasi salah satu segmen (misalnya segmen ke-3)
plt.plot(segments[2][0].numpy())
plt.title(f"Segment 3 - Label: {'yes' if labels[2] else 'no'}")
plt.show()

Penerapan MFCC

Berikut versi lengkap fungsi yang langsung memotong audio menjadi 8 segmen dan mengekstrak MFCC dari masing-masing segmen.

import torch
import torchaudio

def extract_mfcc_segments(waveform, sample_rate, n_mfcc=13, n_mels=40):
    """
    Memotong waveform menjadi 8 segmen dan mengekstrak MFCC dari tiap segmen.
    
    Args:
        waveform (Tensor): Tensor audio [1, T]
        sample_rate (int): Sampel rate, default YESNO = 16000
        n_mfcc (int): Jumlah koefisien MFCC
        n_mels (int): Jumlah filter Mel

    Returns:
        List[Tensor]: List 8 MFCC tensors, masing-masing shape [n_mfcc, time]
    """
    total_duration_sec = waveform.shape[1] / sample_rate
    segment_duration = total_duration_sec / 8
    segment_samples = int(segment_duration * sample_rate)

    # Inisialisasi transformasi MFCC
    mfcc_transform = torchaudio.transforms.MFCC(
        sample_rate=sample_rate,
        n_mfcc=n_mfcc,
        melkwargs={
            'n_fft': 400,
            'hop_length': 160,
            'n_mels': n_mels,
            'center': False
        }
    )

    mfcc_segments = []
    for i in range(8):
        start = i * segment_samples
        end = (i + 1) * segment_samples
        segment = waveform[:, start:end]

        mfcc = mfcc_transform(segment)  # [1, n_mfcc, time]
        mfcc_segments.append(mfcc.squeeze(0))  # [n_mfcc, time]
    
    return mfcc_segments

cara penggunaan

waveform, sample_rate, labels = yesno_data[0]

# Ekstraksi MFCC per segmen
mfcc_segments = extract_mfcc_segments(waveform, sample_rate)

# Contoh visualisasi segmen ke-1
import matplotlib.pyplot as plt
for i in range(0, len(mfcc_segments)):
    if labels[i]==1:
        label = 'yes'
    else:
        label = 'no'
    plt.imshow(mfcc_segments[i].numpy(), aspect='auto', origin='lower')
    plt.title("MFCC Segment "+str(i)+" - Label: "+label)
    plt.colorbar()
    plt.tight_layout()
    plt.show()

CNN untuk klasifikasi suara sederhana

Kita bisa menggunakan MFCC sebagai input di CNN. Mengingat ukuran MFCC yang dihasilkan bisa berbeda-beda ukurannya, kita akan banyak modifikasi terlebih dahulu agar model CNN bisa menerima input dengan beragam ukuran.

Dataset Kustom: YESNO Segmen per Kata

Misalkan menggunakan class sebagai berikut sebagai rencana awal nya

from torch.utils.data import Dataset

class YesNoSegmentDataset(Dataset):
    def __init__(self, root="data", n_mfcc=13):
        self.dataset = torchaudio.datasets.YESNO(root=root, download=True)
        self.sample_rate = 16000
        self.n_mfcc = n_mfcc
        self.mfcc_transform = torchaudio.transforms.MFCC(
            sample_rate=self.sample_rate,
            n_mfcc=n_mfcc,
            melkwargs={'n_fft': 400, 'hop_length': 160, 'n_mels': 40, 'center': False}
        )

        self.samples = []  # list of (mfcc_tensor, label)

        for waveform, sr, labels in self.dataset:
            segment_mfccs = self._extract_mfcc_segments(waveform)
            for mfcc, label in zip(segment_mfccs, labels):
                self.samples.append((mfcc, torch.tensor(label, dtype=torch.long)))

    def _extract_mfcc_segments(self, waveform):
        segment_len = waveform.shape[1] // 8
        mfccs = []
        for i in range(8):
            seg = waveform[:, i * segment_len : (i + 1) * segment_len]
            mfcc = self.mfcc_transform(seg).squeeze(0)  # [n_mfcc, time]
            mfccs.append(mfcc)
        return mfccs

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

    def __getitem__(self, idx):
        x, y = self.samples[idx]
        return x.unsqueeze(0), y  # [1, n_mfcc, time], label

Model CNN nya sebagai berikut

import torch.nn as nn
import torch.nn.functional as F

class MFCC_CNN(nn.Module):
    def __init__(self):
        super().__init__()
        self.net = nn.Sequential(
            nn.Conv2d(1, 16, kernel_size=3, padding=1),  # [B, 16, n_mfcc, time]
            nn.ReLU(),
            nn.MaxPool2d(2),  # [B, 16, n_mfcc/2, time/2]

            nn.Conv2d(16, 32, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(2),  # [B, 32, n_mfcc/4, time/4]

            nn.Flatten(),
            nn.Linear(32 * (13 // 4) * (40 // 4), 64),
            nn.ReLU(),
            nn.Linear(64, 2)  # 2 kelas: yes / no
        )

    def forward(self, x):
        return self.net(x)

bila kita run pelatihannya

from torch.utils.data import DataLoader
import torch.optim as optim

# Setup
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
dataset = YesNoSegmentDataset()
loader = DataLoader(dataset, batch_size=16, shuffle=True)
model = MFCC_CNN().to(device)
optimizer = optim.Adam(model.parameters(), lr=1e-3)
loss_fn = nn.CrossEntropyLoss()

# Training
for epoch in range(10):
    total_loss = 0
    correct = 0
    for x, y in loader:
        x, y = x.to(device), y.to(device)
        logits = model(x)
        loss = loss_fn(logits, y)

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

        total_loss += loss.item()
        correct += (logits.argmax(1) == y).sum().item()
    
    acc = correct / len(dataset)
    print(f"Epoch {epoch+1} | Loss: {total_loss:.4f} | Accuracy: {acc:.2%}")

hemm akan ada error karena ukuran MFCC nya berbeda-beda

RuntimeError: stack expects each tensor to be equal size, but got [1, 13, 40] at entry 0 and [1, 13, 37] at entry 1

Error ini muncul karena panjang dimensi waktu (time axis) dari MFCC berbeda-beda antar segmen, contohnya:

Segment 0: [1, 13, 40] (13 MFCC, 40 frame waktu)
Segment 1: [1, 13, 37] (13 MFCC, 37 frame waktu)

DataLoader mencoba stack semua batch jadi 1 tensor (seperti torch.stack()), tapi tidak bisa jika shape-nya beda. Solusinya

solusi 1: Pad MFCC agar Semua Sama Panjang

Buat fungsi padding ke panjang waktu tetap, misalnya 40 frame:

import torch.nn.functional as F

def pad_mfcc(mfcc, target_length=40):
    """
    Pad atau potong mfcc ke panjang waktu tetap.
    Args:
        mfcc: Tensor [n_mfcc, time]
        target_length: jumlah frame waktu yang diinginkan

    Returns:
        Tensor [n_mfcc, target_length]
    """
    current_length = mfcc.shape[1]
    if current_length < target_length:
        pad_amt = target_length - current_length
        return F.pad(mfcc, (0, pad_amt))  # pad di akhir time axis
    else:
        return mfcc[:, :target_length]  # potong jika terlalu panjang

tambahkan ke dalam YesNoSegmentDataset:

Ubah bagian __getitem__():

def __getitem__(self, idx):
    x, y = self.samples[idx]
    x = pad_mfcc(x, target_length=40)  # pastikan panjang waktu tetap
    return x.unsqueeze(0), y  # [1, 13, 40], label

solusi 2: Pakai `collate_fn` custom (jika ingin flexible padding)

Kalau kamu ingin padding dinamis dalam DataLoader, kamu bisa juga buat collate_fn, tapi solusi di atas lebih sederhana dan cukup untuk model CNN yang butuh ukuran tetap.

Setelah padding semua MFCC ke panjang tetap ([1, 13, 40]), DataLoader akan bisa stack batch dengan lancar.

Tujuan collate_fn ini:

Menerima list [mfcc, label]
Menyamakan panjang dimensi waktu semua MFCC dalam batch
Padding agar bisa dibentuk jadi batch tensor [B, 1, 13, max_time]

mari kita buat fungsi collate_fn

import torch
import torch.nn.functional as F

def collate_fn_pad_mfcc(batch):
    """
    Custom collate function untuk padding MFCC ke panjang waktu sama.
    Args:
        batch: list of (mfcc_tensor [1, 13, T], label)

    Returns:
        padded_mfccs: Tensor [B, 1, 13, max_time]
        labels: Tensor [B]
    """
    mfccs, labels = zip(*batch)  # unzip

    # Temukan panjang maksimum time axis
    max_len = max(mfcc.shape[-1] for mfcc in mfccs)

    padded_mfccs = []
    for mfcc in mfccs:
        # mfcc: [1, 13, T]
        pad_amt = max_len - mfcc.shape[-1]
        padded = F.pad(mfcc, (0, pad_amt))  # pad hanya di time axis (dim -1)
        padded_mfccs.append(padded)

    # Stack jadi batch tensor
    padded_mfccs = torch.stack(padded_mfccs)  # [B, 1, 13, max_len]
    labels = torch.stack(labels)  # [B]

    return padded_mfccs, labels

pasangkan ke dataloader

from torch.utils.data import DataLoader

dataset = YesNoSegmentDataset()
loader = DataLoader(dataset, batch_size=16, shuffle=True, collate_fn=collate_fn_pad_mfcc)

Catatan Penting untuk Model

Karena dimensi time bisa bervariasi di setiap batch (misal 13×37, 13×40, dst), maka:

Pastikan CNN kamu bisa menerima input dengan dimensi waktu fleksibel.
Biasanya, Conv2d dan MaxPool2d aman asalkan resolusi minimum tetap cukup besar.

See also Mengenal spectogram

Maka kita ubah dulu modelnya sebagai berikut

import torch.nn as nn

class MFCC_CNN(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv_layers = nn.Sequential(
            nn.Conv2d(1, 16, 3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(2),

            nn.Conv2d(16, 32, 3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(2),

            nn.AdaptiveAvgPool2d((4, 4))
        )
        self.classifier = nn.Sequential(
            nn.Flatten(),
            nn.Linear(32 * 4 * 4, 64),
            nn.ReLU(),
            nn.Linear(64, 2)
        )

    def forward(self, x):
        x = self.conv_layers(x)
        x = self.classifier(x)
        return x

Berikut final dari pelatihannya

from torch.utils.data import DataLoader
import torch.optim as optim

# Setup
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
dataset = YesNoSegmentDataset()
# loader = DataLoader(dataset, batch_size=16, shuffle=True)
loader = DataLoader(dataset, batch_size=16, shuffle=True, collate_fn=collate_fn_pad_mfcc)
model = MFCC_CNN().to(device)
optimizer = optim.Adam(model.parameters(), lr=1e-3)
loss_fn = nn.CrossEntropyLoss()

# Training
for epoch in range(1000):
    total_loss = 0
    correct = 0
    for x, y in loader:
        x, y = x.to(device), y.to(device)
        logits = model(x)
        loss = loss_fn(logits, y)

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

        total_loss += loss.item()
        correct += (logits.argmax(1) == y).sum().item()
    
    acc = correct / len(dataset)
    print(f"Epoch {epoch+1} | Loss: {total_loss:.4f} | Accuracy: {acc:.2%}")

dengan hasil sebagai berikut

Epoch 1 | Loss: 26.2292 | Accuracy: 52.71%
Epoch 2 | Loss: 18.2889 | Accuracy: 65.83%
Epoch 3 | Loss: 17.0057 | Accuracy: 70.42%
Epoch 4 | Loss: 16.0658 | Accuracy: 72.29%
Epoch 5 | Loss: 15.2807 | Accuracy: 71.25%
Epoch 6 | Loss: 15.3737 | Accuracy: 72.92%
Epoch 7 | Loss: 16.1613 | Accuracy: 70.21%
Epoch 8 | Loss: 14.7362 | Accuracy: 76.04%
Epoch 9 | Loss: 14.5108 | Accuracy: 74.79%
Epoch 10 | Loss: 14.6439 | Accuracy: 75.21%
Epoch 11 | Loss: 14.2847 | Accuracy: 75.00%
Epoch 12 | Loss: 14.3857 | Accuracy: 75.00%
Epoch 13 | Loss: 12.5071 | Accuracy: 78.96%
Epoch 14 | Loss: 12.1144 | Accuracy: 81.46%
Epoch 15 | Loss: 12.0223 | Accuracy: 80.83%
Epoch 16 | Loss: 11.5861 | Accuracy: 82.08%
Epoch 17 | Loss: 11.7297 | Accuracy: 80.83%
Epoch 18 | Loss: 10.8024 | Accuracy: 84.17%
Epoch 19 | Loss: 10.9587 | Accuracy: 83.75%
Epoch 20 | Loss: 9.7609 | Accuracy: 86.04%
Epoch 21 | Loss: 9.5121 | Accuracy: 85.21%
Epoch 22 | Loss: 9.5792 | Accuracy: 85.83%
Epoch 23 | Loss: 7.7602 | Accuracy: 88.54%
Epoch 24 | Loss: 7.1596 | Accuracy: 90.42%
Epoch 25 | Loss: 6.7793 | Accuracy: 90.62%
Epoch 26 | Loss: 6.3053 | Accuracy: 91.88%
Epoch 27 | Loss: 5.8734 | Accuracy: 92.50%
Epoch 28 | Loss: 5.3829 | Accuracy: 92.50%
Epoch 29 | Loss: 5.0102 | Accuracy: 94.17%
Epoch 30 | Loss: 4.2111 | Accuracy: 95.00%
Epoch 31 | Loss: 4.1987 | Accuracy: 94.58%
Epoch 32 | Loss: 3.9876 | Accuracy: 95.00%
Epoch 33 | Loss: 4.4293 | Accuracy: 94.17%
Epoch 34 | Loss: 3.0900 | Accuracy: 96.46%
Epoch 35 | Loss: 3.9723 | Accuracy: 94.58%
Epoch 36 | Loss: 4.0678 | Accuracy: 95.00%
Epoch 37 | Loss: 3.4426 | Accuracy: 96.25%
Epoch 38 | Loss: 2.2301 | Accuracy: 97.50%
Epoch 39 | Loss: 3.2947 | Accuracy: 95.83%
Epoch 40 | Loss: 1.8023 | Accuracy: 98.54%
Epoch 41 | Loss: 1.6340 | Accuracy: 98.12%
Epoch 42 | Loss: 1.2069 | Accuracy: 99.17%
Epoch 43 | Loss: 0.9231 | Accuracy: 98.96%
Epoch 44 | Loss: 1.0076 | Accuracy: 99.38%
Epoch 45 | Loss: 1.0945 | Accuracy: 98.75%
Epoch 46 | Loss: 1.5428 | Accuracy: 98.54%
Epoch 47 | Loss: 2.8091 | Accuracy: 96.88%
Epoch 48 | Loss: 4.7232 | Accuracy: 92.92%
Epoch 49 | Loss: 4.0821 | Accuracy: 95.42%
Epoch 50 | Loss: 2.3349 | Accuracy: 97.50%
Epoch 51 | Loss: 0.8448 | Accuracy: 99.58%
Epoch 52 | Loss: 0.3800 | Accuracy: 100.00%
Epoch 53 | Loss: 0.1901 | Accuracy: 100.00%
Epoch 54 | Loss: 0.1935 | Accuracy: 100.00%
Epoch 55 | Loss: 0.5328 | Accuracy: 99.58%
Epoch 56 | Loss: 0.4569 | Accuracy: 99.58%
Epoch 57 | Loss: 0.5992 | Accuracy: 99.79%
Epoch 58 | Loss: 0.2574 | Accuracy: 100.00%
Epoch 59 | Loss: 0.2273 | Accuracy: 100.00%
Epoch 60 | Loss: 0.2254 | Accuracy: 99.79%
Epoch 61 | Loss: 0.6182 | Accuracy: 99.38%
Epoch 62 | Loss: 0.3135 | Accuracy: 100.00%
Epoch 63 | Loss: 0.1344 | Accuracy: 100.00%
Epoch 64 | Loss: 0.0817 | Accuracy: 100.00%
Epoch 65 | Loss: 0.0730 | Accuracy: 100.00%
Epoch 66 | Loss: 0.0894 | Accuracy: 100.00%
Epoch 67 | Loss: 0.0578 | Accuracy: 100.00%
Epoch 68 | Loss: 0.0539 | Accuracy: 100.00%
Epoch 69 | Loss: 0.0707 | Accuracy: 100.00%
Epoch 70 | Loss: 0.0524 | Accuracy: 100.00%
Epoch 71 | Loss: 0.0570 | Accuracy: 100.00%
Epoch 72 | Loss: 0.0418 | Accuracy: 100.00%
Epoch 73 | Loss: 0.0482 | Accuracy: 100.00%
Epoch 74 | Loss: 0.0705 | Accuracy: 100.00%
Epoch 75 | Loss: 0.0430 | Accuracy: 100.00%
Epoch 76 | Loss: 0.0659 | Accuracy: 100.00%
Epoch 77 | Loss: 0.0547 | Accuracy: 100.00%
Epoch 78 | Loss: 0.0297 | Accuracy: 100.00%
Epoch 79 | Loss: 0.0347 | Accuracy: 100.00%
Epoch 80 | Loss: 0.0340 | Accuracy: 100.00%
Epoch 81 | Loss: 0.0280 | Accuracy: 100.00%
Epoch 82 | Loss: 0.0241 | Accuracy: 100.00%
Epoch 83 | Loss: 0.0241 | Accuracy: 100.00%
Epoch 84 | Loss: 0.0185 | Accuracy: 100.00%
Epoch 85 | Loss: 0.0235 | Accuracy: 100.00%
Epoch 86 | Loss: 0.0199 | Accuracy: 100.00%
Epoch 87 | Loss: 0.0255 | Accuracy: 100.00%
Epoch 88 | Loss: 0.0178 | Accuracy: 100.00%
Epoch 89 | Loss: 0.0179 | Accuracy: 100.00%
Epoch 90 | Loss: 0.0211 | Accuracy: 100.00%
Epoch 91 | Loss: 0.0166 | Accuracy: 100.00%
Epoch 92 | Loss: 0.0189 | Accuracy: 100.00%
Epoch 93 | Loss: 0.0155 | Accuracy: 100.00%
Epoch 94 | Loss: 0.0165 | Accuracy: 100.00%
Epoch 95 | Loss: 0.0177 | Accuracy: 100.00%
Epoch 96 | Loss: 0.0117 | Accuracy: 100.00%
Epoch 97 | Loss: 0.0142 | Accuracy: 100.00%
Epoch 98 | Loss: 0.0136 | Accuracy: 100.00%
Epoch 99 | Loss: 0.0128 | Accuracy: 100.00%
Epoch 100 | Loss: 0.0109 | Accuracy: 100.00%
Epoch 101 | Loss: 0.0115 | Accuracy: 100.00%
Epoch 102 | Loss: 0.0110 | Accuracy: 100.00%
Epoch 103 | Loss: 0.0131 | Accuracy: 100.00%
Epoch 104 | Loss: 0.0121 | Accuracy: 100.00%
Epoch 105 | Loss: 0.0139 | Accuracy: 100.00%
Epoch 106 | Loss: 0.0088 | Accuracy: 100.00%
Epoch 107 | Loss: 0.0096 | Accuracy: 100.00%
Epoch 108 | Loss: 0.0096 | Accuracy: 100.00%
Epoch 109 | Loss: 0.0085 | Accuracy: 100.00%
Epoch 110 | Loss: 0.0099 | Accuracy: 100.00%
Epoch 111 | Loss: 0.0100 | Accuracy: 100.00%
Epoch 112 | Loss: 0.0108 | Accuracy: 100.00%
Epoch 113 | Loss: 0.0079 | Accuracy: 100.00%
Epoch 114 | Loss: 0.0075 | Accuracy: 100.00%
Epoch 115 | Loss: 0.0064 | Accuracy: 100.00%
Epoch 116 | Loss: 0.0127 | Accuracy: 100.00%
Epoch 117 | Loss: 0.0095 | Accuracy: 100.00%
Epoch 118 | Loss: 0.0077 | Accuracy: 100.00%
Epoch 119 | Loss: 0.0079 | Accuracy: 100.00%
Epoch 120 | Loss: 0.0100 | Accuracy: 100.00%
Epoch 121 | Loss: 0.0095 | Accuracy: 100.00%
Epoch 122 | Loss: 0.0073 | Accuracy: 100.00%
Epoch 123 | Loss: 0.0102 | Accuracy: 100.00%
Epoch 124 | Loss: 0.0071 | Accuracy: 100.00%
Epoch 125 | Loss: 0.0078 | Accuracy: 100.00%
Epoch 126 | Loss: 0.0062 | Accuracy: 100.00%
Epoch 127 | Loss: 0.0080 | Accuracy: 100.00%
Epoch 128 | Loss: 0.0054 | Accuracy: 100.00%
Epoch 129 | Loss: 0.0055 | Accuracy: 100.00%
Epoch 130 | Loss: 0.0062 | Accuracy: 100.00%
Epoch 131 | Loss: 0.0052 | Accuracy: 100.00%
Epoch 132 | Loss: 0.0096 | Accuracy: 100.00%
Epoch 133 | Loss: 0.0071 | Accuracy: 100.00%
Epoch 134 | Loss: 0.0051 | Accuracy: 100.00%
Epoch 135 | Loss: 0.0050 | Accuracy: 100.00%
Epoch 136 | Loss: 0.0046 | Accuracy: 100.00%
Epoch 137 | Loss: 0.0054 | Accuracy: 100.00%
Epoch 138 | Loss: 0.0039 | Accuracy: 100.00%
Epoch 139 | Loss: 0.0040 | Accuracy: 100.00%
Epoch 140 | Loss: 0.0185 | Accuracy: 100.00%
Epoch 141 | Loss: 0.0071 | Accuracy: 100.00%
Epoch 142 | Loss: 0.0045 | Accuracy: 100.00%
Epoch 143 | Loss: 0.0054 | Accuracy: 100.00%
Epoch 144 | Loss: 0.0057 | Accuracy: 100.00%
Epoch 145 | Loss: 0.0044 | Accuracy: 100.00%
Epoch 146 | Loss: 0.0040 | Accuracy: 100.00%
Epoch 147 | Loss: 0.0054 | Accuracy: 100.00%
Epoch 148 | Loss: 0.0038 | Accuracy: 100.00%
Epoch 149 | Loss: 0.0043 | Accuracy: 100.00%
Epoch 150 | Loss: 0.0055 | Accuracy: 100.00%
Epoch 151 | Loss: 0.0030 | Accuracy: 100.00%
Epoch 152 | Loss: 0.0047 | Accuracy: 100.00%
Epoch 153 | Loss: 0.0032 | Accuracy: 100.00%
Epoch 154 | Loss: 0.0046 | Accuracy: 100.00%
Epoch 155 | Loss: 0.0037 | Accuracy: 100.00%
Epoch 156 | Loss: 0.0034 | Accuracy: 100.00%
Epoch 157 | Loss: 0.0033 | Accuracy: 100.00%
Epoch 158 | Loss: 0.0037 | Accuracy: 100.00%
Epoch 159 | Loss: 0.0034 | Accuracy: 100.00%
Epoch 160 | Loss: 0.0032 | Accuracy: 100.00%
Epoch 161 | Loss: 0.0049 | Accuracy: 100.00%
Epoch 162 | Loss: 0.0029 | Accuracy: 100.00%
Epoch 163 | Loss: 0.0032 | Accuracy: 100.00%
Epoch 164 | Loss: 0.0027 | Accuracy: 100.00%
Epoch 165 | Loss: 0.0029 | Accuracy: 100.00%
Epoch 166 | Loss: 0.0036 | Accuracy: 100.00%
Epoch 167 | Loss: 0.0028 | Accuracy: 100.00%
Epoch 168 | Loss: 0.0031 | Accuracy: 100.00%
Epoch 169 | Loss: 0.0023 | Accuracy: 100.00%
Epoch 170 | Loss: 0.0023 | Accuracy: 100.00%
Epoch 171 | Loss: 0.0025 | Accuracy: 100.00%
Epoch 172 | Loss: 0.0023 | Accuracy: 100.00%
Epoch 173 | Loss: 0.0021 | Accuracy: 100.00%
Epoch 174 | Loss: 0.0024 | Accuracy: 100.00%
Epoch 175 | Loss: 0.0022 | Accuracy: 100.00%
Epoch 176 | Loss: 0.0019 | Accuracy: 100.00%
Epoch 177 | Loss: 0.0021 | Accuracy: 100.00%
Epoch 178 | Loss: 0.0025 | Accuracy: 100.00%
Epoch 179 | Loss: 0.0017 | Accuracy: 100.00%
Epoch 180 | Loss: 0.0018 | Accuracy: 100.00%
Epoch 181 | Loss: 0.0015 | Accuracy: 100.00%
Epoch 182 | Loss: 0.0019 | Accuracy: 100.00%
Epoch 183 | Loss: 0.0021 | Accuracy: 100.00%
Epoch 184 | Loss: 0.0018 | Accuracy: 100.00%
Epoch 185 | Loss: 0.0023 | Accuracy: 100.00%
Epoch 186 | Loss: 0.0024 | Accuracy: 100.00%
Epoch 187 | Loss: 0.0019 | Accuracy: 100.00%
Epoch 188 | Loss: 0.0023 | Accuracy: 100.00%
Epoch 189 | Loss: 0.0017 | Accuracy: 100.00%
Epoch 190 | Loss: 0.0024 | Accuracy: 100.00%
Epoch 191 | Loss: 0.0023 | Accuracy: 100.00%
Epoch 192 | Loss: 0.0390 | Accuracy: 100.00%
Epoch 193 | Loss: 18.6600 | Accuracy: 80.00%
Epoch 194 | Loss: 12.5987 | Accuracy: 84.17%
Epoch 195 | Loss: 8.5316 | Accuracy: 88.96%
Epoch 196 | Loss: 7.4993 | Accuracy: 88.75%
Epoch 197 | Loss: 5.5558 | Accuracy: 92.08%
Epoch 198 | Loss: 4.6345 | Accuracy: 94.17%
Epoch 199 | Loss: 2.7514 | Accuracy: 96.88%
Epoch 200 | Loss: 2.9323 | Accuracy: 96.88%
Epoch 201 | Loss: 2.0268 | Accuracy: 97.08%
Epoch 202 | Loss: 1.3389 | Accuracy: 98.33%
Epoch 203 | Loss: 2.7342 | Accuracy: 96.04%
Epoch 204 | Loss: 1.2913 | Accuracy: 98.54%
Epoch 205 | Loss: 0.8766 | Accuracy: 99.38%
Epoch 206 | Loss: 0.6008 | Accuracy: 99.58%
Epoch 207 | Loss: 1.1676 | Accuracy: 98.33%
Epoch 208 | Loss: 1.1118 | Accuracy: 98.96%
Epoch 209 | Loss: 0.7183 | Accuracy: 99.17%
Epoch 210 | Loss: 0.2214 | Accuracy: 100.00%
Epoch 211 | Loss: 0.1458 | Accuracy: 100.00%
Epoch 212 | Loss: 0.1117 | Accuracy: 100.00%
Epoch 213 | Loss: 0.0873 | Accuracy: 100.00%
Epoch 214 | Loss: 0.1275 | Accuracy: 100.00%
Epoch 215 | Loss: 0.1159 | Accuracy: 100.00%
Epoch 216 | Loss: 0.1644 | Accuracy: 100.00%
Epoch 217 | Loss: 0.0570 | Accuracy: 100.00%
Epoch 218 | Loss: 0.0635 | Accuracy: 100.00%
Epoch 219 | Loss: 0.0301 | Accuracy: 100.00%
Epoch 220 | Loss: 0.0714 | Accuracy: 100.00%
Epoch 221 | Loss: 0.1437 | Accuracy: 100.00%
Epoch 222 | Loss: 0.0681 | Accuracy: 100.00%
Epoch 223 | Loss: 0.0375 | Accuracy: 100.00%
Epoch 224 | Loss: 0.0515 | Accuracy: 100.00%
Epoch 225 | Loss: 0.0328 | Accuracy: 100.00%
Epoch 226 | Loss: 0.0202 | Accuracy: 100.00%
Epoch 227 | Loss: 0.0182 | Accuracy: 100.00%
Epoch 228 | Loss: 0.0255 | Accuracy: 100.00%
Epoch 229 | Loss: 0.0148 | Accuracy: 100.00%
Epoch 230 | Loss: 0.0191 | Accuracy: 100.00%
Epoch 231 | Loss: 0.0184 | Accuracy: 100.00%
Epoch 232 | Loss: 0.0217 | Accuracy: 100.00%
Epoch 233 | Loss: 0.0240 | Accuracy: 100.00%
Epoch 234 | Loss: 0.0141 | Accuracy: 100.00%
Epoch 235 | Loss: 0.0123 | Accuracy: 100.00%
Epoch 236 | Loss: 0.0130 | Accuracy: 100.00%
Epoch 237 | Loss: 0.0128 | Accuracy: 100.00%
Epoch 238 | Loss: 0.0123 | Accuracy: 100.00%
Epoch 239 | Loss: 0.0143 | Accuracy: 100.00%
Epoch 240 | Loss: 0.0094 | Accuracy: 100.00%
Epoch 241 | Loss: 0.0146 | Accuracy: 100.00%
Epoch 242 | Loss: 0.0080 | Accuracy: 100.00%
Epoch 243 | Loss: 0.0089 | Accuracy: 100.00%
Epoch 244 | Loss: 0.0201 | Accuracy: 100.00%
Epoch 245 | Loss: 0.0141 | Accuracy: 100.00%
Epoch 246 | Loss: 0.0101 | Accuracy: 100.00%
Epoch 247 | Loss: 0.0092 | Accuracy: 100.00%
Epoch 248 | Loss: 0.0102 | Accuracy: 100.00%
Epoch 249 | Loss: 0.0104 | Accuracy: 100.00%
Epoch 250 | Loss: 0.0083 | Accuracy: 100.00%
Epoch 251 | Loss: 0.0106 | Accuracy: 100.00%
Epoch 252 | Loss: 0.0102 | Accuracy: 100.00%
Epoch 253 | Loss: 0.0084 | Accuracy: 100.00%
Epoch 254 | Loss: 0.0069 | Accuracy: 100.00%
Epoch 255 | Loss: 0.0064 | Accuracy: 100.00%
Epoch 256 | Loss: 0.0075 | Accuracy: 100.00%
Epoch 257 | Loss: 0.0055 | Accuracy: 100.00%
Epoch 258 | Loss: 0.0058 | Accuracy: 100.00%
Epoch 259 | Loss: 0.0089 | Accuracy: 100.00%
Epoch 260 | Loss: 0.0072 | Accuracy: 100.00%
Epoch 261 | Loss: 0.0058 | Accuracy: 100.00%
Epoch 262 | Loss: 0.0055 | Accuracy: 100.00%
Epoch 263 | Loss: 0.0058 | Accuracy: 100.00%
Epoch 264 | Loss: 0.0045 | Accuracy: 100.00%
Epoch 265 | Loss: 0.0053 | Accuracy: 100.00%
Epoch 266 | Loss: 0.0049 | Accuracy: 100.00%
Epoch 267 | Loss: 0.0043 | Accuracy: 100.00%
Epoch 268 | Loss: 0.0049 | Accuracy: 100.00%
Epoch 269 | Loss: 0.0043 | Accuracy: 100.00%
Epoch 270 | Loss: 0.0058 | Accuracy: 100.00%
Epoch 271 | Loss: 0.0038 | Accuracy: 100.00%
Epoch 272 | Loss: 0.0045 | Accuracy: 100.00%
Epoch 273 | Loss: 0.0055 | Accuracy: 100.00%
Epoch 274 | Loss: 0.0041 | Accuracy: 100.00%
Epoch 275 | Loss: 0.0040 | Accuracy: 100.00%
Epoch 276 | Loss: 0.0034 | Accuracy: 100.00%
Epoch 277 | Loss: 0.0048 | Accuracy: 100.00%
Epoch 278 | Loss: 0.0048 | Accuracy: 100.00%
Epoch 279 | Loss: 0.0048 | Accuracy: 100.00%
Epoch 280 | Loss: 0.0035 | Accuracy: 100.00%
Epoch 281 | Loss: 0.0036 | Accuracy: 100.00%
Epoch 282 | Loss: 0.0044 | Accuracy: 100.00%
Epoch 283 | Loss: 0.0028 | Accuracy: 100.00%
Epoch 284 | Loss: 0.0028 | Accuracy: 100.00%
Epoch 285 | Loss: 0.0034 | Accuracy: 100.00%
Epoch 286 | Loss: 0.0027 | Accuracy: 100.00%
Epoch 287 | Loss: 0.0035 | Accuracy: 100.00%
Epoch 288 | Loss: 0.0033 | Accuracy: 100.00%
Epoch 289 | Loss: 0.0023 | Accuracy: 100.00%
Epoch 290 | Loss: 0.0028 | Accuracy: 100.00%
Epoch 291 | Loss: 0.0024 | Accuracy: 100.00%
Epoch 292 | Loss: 0.0035 | Accuracy: 100.00%
Epoch 293 | Loss: 0.0023 | Accuracy: 100.00%
Epoch 294 | Loss: 0.0024 | Accuracy: 100.00%
Epoch 295 | Loss: 0.0023 | Accuracy: 100.00%
Epoch 296 | Loss: 0.0021 | Accuracy: 100.00%
Epoch 297 | Loss: 0.0023 | Accuracy: 100.00%
Epoch 298 | Loss: 0.0034 | Accuracy: 100.00%
Epoch 299 | Loss: 0.0027 | Accuracy: 100.00%
Epoch 300 | Loss: 0.0041 | Accuracy: 100.00%