Chapter 11: Model Summary and Parameter Inspection

“Understanding your model is like reading blueprints before construction. Don’t train it blind—inspect, analyze, and optimize.”

---

Why This Chapter Matters

You’ve built a CNN and defined its forward flow. But now you need to:

• Check if the architecture matches your expectations
• Inspect layer shapes and total parameter counts
• Freeze or unfreeze layers for transfer learning
• Load/save weights or extract specific layer outputs

This chapter helps you understand:

• How to summarize your model
• How to access weights and parameters
• How to manage layers for training or inference

You’ll learn to treat models not as magic boxes—but as transparent, inspectable systems.

---

Conceptual Breakdown

🔹 What to Inspect in a Model

Property	Why it Matters
Layer names/types	Ensure architecture is correct
Output shapes	Catch shape mismatches early
Total parameters	Know model size and overfitting risk
Trainable vs frozen	Required for transfer learning/fine-tuning
Weight values	For debugging, initialization checks

---

🔹 Freezing vs Unfreezing Layers

Freezing a layer means its weights won’t update during training (used in transfer learning). Unfreezing means allowing gradient flow again.

📌 Freeze base layers → train top layers only → unfreeze gradually.

---

PyTorch Implementation

🔸 Model Summary

Use the torchsummary package (or print manually):

pip install torchsummary

from torchsummary import summary
import torch

model = ConvClassifier()
summary(model, input_size=(3, 224, 224))

🔸 Inspect Parameters

for name, param in model.named_parameters():
    print(f"{name}: {param.shape}, requires_grad={param.requires_grad}")

🔸 Freeze Layers

for param in model.features.parameters():
    param.requires_grad = False

To unfreeze later:

for param in model.features.parameters():
    param.requires_grad = True

🔸 Save/Load Weights

torch.save(model.state_dict(), "model_weights.pth")

model.load_state_dict(torch.load("model_weights.pth"))
model.eval()  # Set to inference mode

---

TensorFlow Implementation

🔸 Model Summary

model = ConvClassifier()
model.build(input_shape=(None, 224, 224, 3))
model.summary()

🔸 Inspect Weights

for layer in model.layers:
    print(layer.name, layer.trainable)
    for weight in layer.weights:
        print(f"  {weight.name} - shape: {weight.shape}")

🔸 Freeze Layers

for layer in model.layers:
    layer.trainable = False  # freeze

Unfreeze:

for layer in model.layers:
    layer.trainable = True

🔸 Save/Load Weights

model.save_weights("model_checkpoint.h5")

# Reload weights into the same architecture
model.load_weights("model_checkpoint.h5")

---

PyTorch vs TensorFlow Parameter Access

Task	PyTorch	TensorFlow
Get all weights	model.parameters()	model.weights
Get named weights	model.named_parameters()	layer.weights per layer
Freeze training	param.requires_grad = False	layer.trainable = False
Save weights	torch.save(state_dict)	model.save_weights()
Load weights	model.load_state_dict(...)	model.load_weights()

---

Model Modes: Train vs Eval

Mode	PyTorch	TensorFlow
Training	model.train()	training=True in call()
Inference	model.eval()	training=False in call()
Dropout/BNorm	Behave differently in modes	Same applies in both frameworks

Always remember:

• Use model.eval() in PyTorch during inference (turns off dropout, uses running stats in BatchNorm).
• In TensorFlow, use training=False explicitly in call().

---

Mini-Exercise

1. Build a small CNN for 3-class classification.
2. Print full model summary.
3. Freeze all convolutional layers.
4. Confirm only Linear / Dense layers are trainable.
5. Save and reload weights.
6. Switch between train/eval modes and print dropout effect.

Bonus: Write a utility function that counts:

• Total parameters
• Trainable parameters
• Frozen parameters

---