Part V – Inference, Evaluation, and Visual Debugging¶
“Once the model is trained, the real-world test begins. Inference is the moment your model meets reality.”
Why This Part Matters¶
Training a CNN isn’t the final goal. What really matters is what happens after training:
- Does it make correct predictions on unseen data?
- Does it behave consistently across platforms?
- Can you debug it when it fails?
Most real-world issues occur after training:
- Input mismatch during inference
- Dropout or BatchNorm behaving incorrectly
- Unexplainable results that need visualization
This part of the book equips you to evaluate models confidently and debug them visually.
What You’ll Master in This Part¶
- Understand and use the correct training/evaluation modes
- Handle Dropout and BatchNorm properly during inference
- Perform accurate predictions in real-time systems
- Visualize feature maps and filters to interpret what the model is learning
- Build intuition for CNN internals using visual tools
Chapter Breakdown¶
| Chapter | Title | What You’ll Learn |
|---|---|---|
| 16 | Train vs Eval Mode | Understand mode switching, dropout behavior, and inference consistency |
| 17 | Visualizing Feature Maps and Filters | Peek inside CNNs to see what filters are activating and when |
Key Problems This Part Solves¶
| Problem | Solution You'll Learn |
|---|---|
| Model behaves differently at inference | Use model.eval() or training=False properly |
| Accuracy drops on deployment | Check normalization, dropout, batchnorm behavior |
| Can't explain model predictions | Visualize filters and activations |
| Confusion between logits and probabilities | Confirm mode, output, and softmax usage |
| Debugging performance bottlenecks | Test inference-only mode with no gradients |
Tools You’ll Be Using¶
| Tool | Description |
|---|---|
model.eval() / .train() |
Control dropout/batchnorm in PyTorch |
training=True/False |
Control training/inference in TensorFlow |
torch.no_grad() |
Disable gradients for inference |
Keras Model(inputs, outputs) |
Create sub-models for visualization |
| PyTorch forward hooks | Inspect layer outputs during forward pass |
matplotlib, seaborn |
Plot activation maps and filters |
What You’ll Be Able To Do After This Part¶
- Switch between training and inference with complete confidence
- Ensure consistent behavior between training and deployment environments
- Analyze intermediate activations in CNNs
- Debug failing models visually
- Understand how convolution filters “light up” for different image regions
Real-World Relevance¶
In production, bugs often stem not from your training pipeline—but from incorrect inference behavior or misunderstood activations.
You’ll encounter scenarios like:
- A model doing well in training but failing in deployment due to
Dropoutstill being active - A medical image classifier that seems to focus on text labels in X-rays—visible through feature maps
- A classifier always choosing one class—fixable only after visualizing activation saturation
This part helps you trust, verify, and explain what your CNN is doing under the hood.