CIFAR-10 Image Classifier (PyTorch)

A convolutional neural network (CNN) trained on CIFAR-10 with PyTorch.

Information about dataset

Object Recognition in Images - Identify the subject of 60,000 labeled images See more

Architecture Overview

Input: 3×32×32 image (CIFAR-10)
Kernel size: All Conv2d layers use 3×3 kernels

---

Block 1

• Conv2d(3 → 32, 3×3) + SiLU + BatchNorm
• Conv2d(32 → 32, 3×3) + SiLU + BatchNorm
• MaxPooling + Dropout

---

Block 2

• Conv2d(32 → 64, 3×3) + SiLU + BatchNorm
• Conv2d(64 → 64, 3×3) + SiLU + BatchNorm
• MaxPooling + Dropout

---

Block 3

• Conv2d(64 → 128, 3×3) + SiLU + BatchNorm
• Conv2d(128 → 128, 3×3) + SiLU + BatchNorm
• Conv2d(128 → 128, 3×3) + SiLU + BatchNorm
• MaxPooling + Dropout

---

Block 4

• Conv2d(128 → 256, 3×3) + SiLU + BatchNorm
• Conv2d(256 → 256, 3×3) + SiLU + BatchNorm
• Conv2d(256 → 256, 3×3) + SiLU + BatchNorm
• MaxPooling + Dropout

---

Classifier

• Flatten
• Linear(1024 → 512) + SiLU
• Linear(512 → 10)

---

Test Accuracy: ~90% after 100 epochs With:

BATCH_SIZE_TRAIN = 256
BATCH_SIZE_TEST = 64
LEARNING_RATE = 0.001
NUM_EPOCHS = 200
LR_GAMMA = 0.96

Run code

pip install -r requirements.txt

Customize params in config.py

Training the Model

Once dependencies are installed, you can train the model using:

python train.py

Evaluating the Model

After training, you can evaluate the model on the CIFAR-10 test set or your own model:

python eval.py