Summary

# ===== Mental model =====
# Tensor: typed n-dim array on CPU / GPU / TPU
# Operation: pure function on tensors; builds a graph (in tf.function)
# Variable: trainable state with autograd
# Model: stack of layers + a forward pass + (optionally) a training step
# Optimizer: apply gradients to variables

# ===== Daily reflexes =====
# - shape, dtype, device on every tensor in new code
# - tf.data: cache + shuffle + prefetch
# - Keras Sequential / Functional / Subclass — pick by shape
# - compile + fit + EarlyStopping + ModelCheckpoint + ReduceLROnPlateau
# - SavedModel for serving; TFLite for mobile; TFJS for web

# ===== Production patterns =====
# - Mixed precision on modern GPUs
# - tf.data pipelines for fast input
# - tf.function on hot paths
# - TensorBoard for live metrics
# - SavedModel as the deploy unit

# ===== Where TF wins vs PyTorch =====
# TF: production deploys to mobile / edge (TFLite); TPU; Keras DX
# PyTorch: research; HuggingFace + community models; debug ergonomics
# In 2026, many teams use both — PyTorch for training, ONNX / TFLite for deploy.

# ===== The broader ecosystem =====
# - Keras 3 (multi-backend: TF, JAX, PyTorch)
# - TensorFlow Probability (probabilistic models)
# - TensorFlow Hub (pretrained models)
# - TensorFlow Recommenders
# - TensorFlow.js (JS / browser)
# - TFLite Micro (embedded)

# ===== Next steps =====
# - Practise: build + ship 2-3 small models end-to-end (data + train + serve)
# - Read: 'Deep Learning' by Goodfellow et al for the foundations
# - Compete: Kaggle for evaluated benchmarks
# - Specialise: vision / NLP / RL / recommender
# - Cross-pollinate: try PyTorch + JAX
# - Deploy: TFLite + TFJS to see your model on real devices

# ===== Patterns to internalise =====
# - tf.data + mixed precision + tf.function = perf trinity
# - EarlyStopping + ModelCheckpoint + TensorBoard every run
# - SavedModel as the artifact contract
# - Pin TF + Keras versions; mind their multi-backend story

# ===== Pitfalls =====
# - CUDA / cuDNN mismatch -> 'Could not load library' errors
# - Mixing eager + graph carelessly -> retracing on every call
# - Forgetting validation_data + EarlyStopping -> overfitting
# - Heavy preprocessing in Python loops (use tf.data)

# ===== Where to publish =====
# - Kaggle for benchmarked + open data work
# - HuggingFace Hub for pretrained models
# - Personal blog with rendered notebooks (Quarto)

# ===== Closing thought =====
# Models are dumb; data is everything; the loop is the skill.
# Master the input pipeline + the training loop + the deployment story, and the
# specific architecture matters less than people think. Start small, ship often.

# Next: tf.function, custom training loops, JAX comparisons, KerasNLP / KerasCV.