Skip to content

Appendices

Appendix A. PyTorch vs TensorFlow Cheatsheet

This quick reference is designed to help you switch between frameworks with confidence. It highlights the syntax, structure, and behavioral differences that you need to remember during development.

Task PyTorch TensorFlow
Import import torch import tensorflow as tf
Tensor Creation torch.tensor([[1., 2.], [3., 4.]]) tf.constant([[1., 2.], [3., 4.]])
Shape Format [C, H, W] [H, W, C]
Reshape tensor.view(-1, 3, 32, 32) tf.reshape(tensor, [-1, 32, 32, 3])
Permute/Transpose x.permute(0, 2, 3, 1) tf.transpose(x, [0, 2, 3, 1])
Device Handling x.to("cuda") Automatically uses GPU if available
Model Definition Subclass nn.Module Subclass tf.keras.Model
Forward Pass def forward(self, x): def call(self, inputs):
Loss Functions nn.CrossEntropyLoss() tf.keras.losses.CategoricalCrossentropy()
Optimizers torch.optim.Adam() tf.keras.optimizers.Adam()
Training Mode model.train() / model.eval() training=True / False
Disable Gradients with torch.no_grad(): @tf.function or with tf.GradientTape():
Checkpointing torch.save(model.state_dict()) model.save_weights()

Use this table anytime you're translating code or debugging between frameworks.

Appendix B. Troubleshooting Image Model Failures

Training an image classifier but it keeps failing? Here’s a diagnostic checklist.

⚠️ Symptom: Model Predicts Only One Class

  • Check normalization: are pixel values normalized properly (0–1 or mean-std)?
  • Inspect label imbalance—does one class dominate the training data?
  • Try a confusion matrix to confirm.

⚠️ Symptom: Shape Mismatch Errors

  • Inspect input shape: is [C, H, W] (PyTorch) or [H, W, C] (TensorFlow)?
  • Add .unsqueeze(0) or expand_dims() if batch dim is missing.
  • Is the model expecting 3 channels? Grayscale images may only have 1.

⚠️ Symptom: Loss Not Decreasing

  • Use model.train() (PyTorch) or training=True (TensorFlow).
  • Check learning rate: too high causes instability, too low slows learning.
  • Are labels correctly encoded (integers for CrossEntropy, one-hot for categorical)?

⚠️ Symptom: Good Train Accuracy, Poor Val Accuracy

  • Overfitting? Add dropout or L2 regularization.
  • Add more data or augmentations.
  • Match preprocessing between training and inference.

Tip: Always Test with a Single Image First

# PyTorch single image test
model.eval()
with torch.no_grad():
    out = model(image.unsqueeze(0).to(device))

Appendix C. Glossary of Key Terms

A quick dictionary of the most important terms in CNN-based vision.

Term Definition
Tensor A multi-dimensional array (e.g., 4D tensor for image batches).
Convolution A sliding window operation to extract spatial patterns.
Kernel / Filter The small matrix (e.g., 3×3) used in convolutions.
Stride Step size the filter moves each time.
Padding Adding borders to preserve spatial size after convolution.
Channel Color or feature dimension (e.g., RGB = 3 channels).
Batch Size Number of images processed in one forward/backward pass.
Pooling Downsampling layer (MaxPool, AvgPool) to reduce dimensions.
Activation Non-linearity applied after each layer (ReLU, Sigmoid).
Feature Map Output of a convolutional layer representing features.
Fine-Tuning Adapting a pretrained model on a new dataset.
Freeze Layers Prevent layers from updating during training.
Transfer Learning Reusing a trained model’s knowledge in a new task.

Appendix D. Pretrained Model Reference Table

Use this table to find the right pretrained model for your project and the correct preprocessing function.

Model Framework Import Preprocess Function Notes
ResNet50 PyTorch torchvision.models.resnet50() transforms.Normalize(mean, std) Default 224×224
TensorFlow tf.keras.applications.ResNet50 preprocess_input
MobileNetV2 PyTorch torchvision.models.mobilenet_v2() Normalize(mean, std) Lightweight
TensorFlow tf.keras.applications.MobileNetV2 preprocess_input
EfficientNetB0 PyTorch torchvision.models.efficientnet_b0() Normalize(mean, std) Good accuracy/speed
TensorFlow tf.keras.applications.EfficientNetB0 preprocess_input
VGG16 PyTorch torchvision.models.vgg16() Normalize(mean, std) Large and old
TensorFlow tf.keras.applications.VGG16 preprocess_input

🔗 All models available at:

Appendix E. Sample Projects and Mini-Exercises per Chapter

Each chapter includes a practical checkpoint or challenge. Here’s a recap to revisit later:

Chapter Exercise
1. Image Input Convert an image (JPG) to a 3D tensor and visualize its shape.
2. Tensors Practice reshaping, permuting, and broadcasting tensors.
3. Input Pipeline Write a full image → tensor → model input pipeline.
4. Preprocessing Visualize how normalization and resize affect your image.
5. Pretrained Models Try ResNet50 on a custom image and print predictions.
6. Datasets Load a folder of images using ImageFolder or image_dataset_from_directory.
7. Augmentation Apply at least 3 different augmentations and view the effects.
8. CNN Layers Build a custom Conv2d → ReLU → MaxPool pipeline manually.
9. CNN Terms Draw out how kernel, stride, and padding affect feature map size.
10. Forward/Call Implement forward() in PyTorch or call() in TensorFlow.
11. Summary Inspect model parameters and freeze layers selectively.
12. Build CNN Build and train a LeNet-style CNN from scratch.
13. Training Write a full training loop or use model.fit().
14. Fine-Tuning Load MobileNetV2, freeze base, and train a new classifier head.
15. Generalization Implement early stopping and visualize overfitting.
16. Eval Mode Compare predictions in train() vs eval() mode.
17. Feature Maps Visualize intermediate layer outputs with hooks or submodels.
18. Inference Pipeline Build a reusable inference pipeline that handles image → model prediction.
19. Debug Checklist Pick a broken model and fix it using the debugging checklist.

Final Words

These appendices complete your toolkit. Whether you're writing training code from scratch, deploying models, or debugging strange edge cases, this section will bring you back on track.

Next Steps?

  • Bookmark this cheatsheet.
  • Try the mini-exercises if you skipped any.
  • Build and deploy a real project using one of the pretrained models.

If you’ve reached this far—you’re not just reading a book. You’ve trained your own mind like a neural net: layer by layer, with careful activation.

Let’s keep building.