CNN Models for CIFAR-10
Model Architectures
MODEL_ARCH_1
[ (Conv2D → BN)×2 → MaxPool → Dropout(0.3) ]
→ [ (Conv2D → BN)×2 → MaxPool → Dropout(0.4) ]
→ [ (Conv2D → BN)×2 → MaxPool → Dropout(0.5) ]
→ Flatten → Dense(512) → BN → Dropout(0.5) → Dense(10, softmax)
- Number of Parameters:
3,251,018 - Test Accuracy:
90.91%
MODEL_ARCH_2
[ (Conv2D(32) → BN)×2 → MaxPool → Dropout(0.25) ]
→ [ (Conv2D(64) → BN)×2 → MaxPool → Dropout(0.35) ]
→ [ (Conv2D(128) → BN)×2 → MaxPool → Dropout(0.4) ]
→ Flatten → Dense(256) → BN → Dropout(0.5) → Dense(10, softmax)
- Number of Parameters:
815,530 - Test Accuracy:
88.84%
MODEL_ARCH_3
[ Conv2D(32)×2 → MaxPool ]
→ [ Conv2D(64)×2 → MaxPool ]
→ [ Conv2D(96) → MaxPool ]
→ Flatten → Dense(256) → Dense(10, softmax)
- Number of Parameters:
517,002 - Test Accuracy:
85.53%
MODEL_ARCH_4
[ Conv2D(16)×2 → MaxPool ]
→ [ Conv2D(32)×2 → MaxPool ]
→ [ Conv2D(64)×2 → MaxPool ]
→ GAP → Dense(10, softmax)
- Number of Parameters:
72,730 - Test Accuracy:
83.05%
Model Architecture Summary
| Model ID | Number of Parameters | Test Accuracy (%) | Model Size (MB) |
|---|---|---|---|
| MODEL_ARCH_1 | 3,251,018 | 90.91 | 12.40 |
| MODEL_ARCH_2 | 815,530 | 88.84 | 3.11 |
| MODEL_ARCH_3 | 517,002 | 85.53 | 1.97 |
| MODEL_ARCH_4 | 72,730 | 83.05 | 0.28 |
Model size calculated assuming 32-bit floating-point weights (
params × 4 bytes ÷ 1024²).
Verilog Inference Q-Point Results
All results are based on a test set of 100 images (10 from each class). The local test accuracy is therefore considered equal to the number of correct predictions out of 100.
All evaluations use the final model [MODEL_ARCH_4] listed in the table above.
| Format | Python Accuracy (%) | Verilog Accuracy (%) | Notes |
|---|---|---|---|
Float32 (Python) vs Float64 IEEE-754 (Verilog real) | 85 | 84 | Verilog used double precision; the minor accuracy difference is not due to precision loss |
| Q1.31 | 84 | 84 | High-precision fixed-point; results match exactly |
| Q1.15 | 84 | 84 | 16-bit fixed-point; results match exactly |
| Q1.7 | 82 | ~82 | Moderate precision; Verilog value estimated from Python result |
| Q1.3 | 65 | ~65 | Low precision; Verilog value estimated from Python result |