Part 3: End‑to‑End CNN on MNIST (Keras • scikit‑learn • TensorFlow • PyTorch)

Author: Suyog Garg, Dated: 2025/11/18

Accompanying notebooks:

• Notebook A - Keras: Part3A_Keras_MNIST_CNN.ipynb
• Notebook B - scikit‑learn (via SciKeras): Part3B_ScikitLearn_SciKeras_MNIST_CNN.ipynb
• Notebook C - TensorFlow Core (GradientTape): Part3C_TensorFlowCore_MNIST_CNN.ipynb
• Notebook D - PyTorch: Part3D_PyTorch_MNIST_CNN.ipynb
• Comparison notebook: Part3E_Compare_Frameworks.ipynb (reads artifacts/*_metrics.json)

All notebooks save plots and metrics into ./artifacts/ for later comparison.

0) Why four mechanisms?

• Keras: shortest code and fastest path to a solid CNN; batteries included.
• scikit‑learn (+ SciKeras): lets you use sklearn’s fit/score, CV, and grid‑search around a Keras CNN.
• TensorFlow Core: full control over training (for custom loops, research, mixed precision, etc.).
• PyTorch: explicit, pythonic training loop; great research ergonomics and ecosystem.

The comparison notebook summarizes test accuracy, train time, parameter count, and an approximate effective LOC (lines of code, measured by introspecting key functions).

1) Quick start (local or Colab)

Colab: upload any notebook, run all.
Local conda env (CPU baseline):

conda create -n mnist-cnn python=3.11 -c conda-forge -y
conda activate mnist-cnn
conda install numpy matplotlib scikit-learn scikeras tensorflow pytorch torchvision -c conda-forge -y

GPU users should install vendor wheels:

• PyTorch (CUDA): see https://pytorch.org/get-started/locally/ and pick the --index-url for your CUDA version.
• TensorFlow (CUDA): install a build matching your CUDA/cuDNN.
• Apple M‑series (MPS): recent PyTorch supports mps; TensorFlow uses the metal plugin on macOS.

2) Running on a remote server (CPU / MPS / CUDA)

2.1 Headless Jupyter over SSH

# On remote
conda activate mnist-cnn
jupyter lab --no-browser --port 8890

# On your laptop
ssh -L 8890:localhost:8890 youruser@remote.host
# Then open http://localhost:8890

2.2 Device selection

• PyTorch: the notebooks auto-detect cuda → mps → cpu:

  device = torch.device("cuda" if torch.cuda.is_available() else ("mps" if torch.backends.mps.is_available() else "cpu"))

• TensorFlow / Keras:

  import tensorflow as tf
  tf.config.list_physical_devices()   # shows CPU/GPU/MPS

• If multiple GPUs exist, you can select visibility by CUDA_VISIBLE_DEVICES=0 python ... (PyTorch & TF).

2.3 CUDA vs CPU vs MPS

• CUDA (NVIDIA): fastest training, ensure the installation wheel matches driver/CUDA runtime.
• MPS (Apple): good speedups vs CPU; feature coverage still maturing.
• CPU: slower but perfectly fine for MNIST demos.

3) Outline of the session

1. Start with A (Keras): quick win, show loss/accuracy plots.
2. Continue B (scikit‑learn): fit/score API + tiny grid search.
3. Then use C (TF Core): how to customize training with GradientTape etc.?
4. Finally with D (PyTorch): point out explicit training loops and device placement.
5. Run E (Compare) to produce the summary plots/table from the artifacts/*_metrics.json files.

4) Tips

• Keep the same batch size and epochs when comparing speed.
• For reproducibility, set numpy and framework seeds if you need bitwise‑stable results.
• To try your CNN on Fashion‑MNIST dataset, change the dataset imports, the rest of the code stays the same.

Enjoy!