Custom CNN for CIFAR-10
Model Description
Model ini adalah arsitektur Convolutional Neural Network (CNN) kustom yang dilatih dari awal pada dataset CIFAR-10 untuk tugas klasifikasi gambar menjadi 10 kelas berbeda.
Arsitektur model terdiri dari:
- 5 convolutional blocks dengan Batch Normalization
- Global Average Pooling
- 2 Fully Connected layers dengan Dropout (0.5 dan 0.3)
10 Kelas yang Dikenali
| ID |
Kelas |
| 0 |
airplane |
| 1 |
automobile |
| 2 |
bird |
| 3 |
cat |
| 4 |
deer |
| 5 |
dog |
| 6 |
frog |
| 7 |
horse |
| 8 |
ship |
| 9 |
truck |
Performa Model
| Set |
Akurasi |
| Train |
82.75% |
| Validation |
80.73% |
| Test |
84.78% |
Detail Training
| Parameter |
Nilai |
| Dataset |
CIFAR-10 (50,000 training, 10,000 test) |
| Epochs |
50 (dengan early stopping, patience=8) |
| Batch Size |
64 |
| Optimizer |
Adam (lr=0.001, weight_decay=5e-4) |
| Loss Function |
CrossEntropyLoss |
| Learning Rate Scheduler |
ReduceLROnPlateau (patience=3, factor=0.5) |
| Best Model Epoch |
Epoch 49 (Val Acc: 80.73%) |
Data Augmentasi
- RandomHorizontalFlip (p=0.5)
- RandomRotation (20 derajat)
- ColorJitter (brightness=0.3, contrast=0.3, saturation=0.3)
- RandomAffine (translate=0.15)
Hasil Pengujian
| Metrik |
Nilai |
| Test Accuracy |
84.78% |
| Correct Predictions |
8,478 / 10,000 |
| Test Loss |
0.4654 |
Penulis
- Nama: [Daniel Wuliutomo]
- Batch: batch-11
