CS6140 NN Setup

Setup for the Neural-Network Homeworks (HW5–HW7)

HW5–HW7 use PyTorch plus the helper package of the Dive into Deep Learning (d2l) book. You set it up once, before HW5, in about 20 minutes. You do not need a GPU. Every homework runs on a normal laptop; a GPU just makes HW6–HW7 faster.

1. Install

In a terminal (macOS/Linux):

python3 -m venv ~/cs6140_env
source ~/cs6140_env/bin/activate        # repeat this line in every new terminal
pip install --upgrade pip
pip install torch torchvision numpy scipy pandas matplotlib scikit-learn jupyterlab requests tqdm sacrebleu gensim
pip install d2l==1.0.3 --no-deps        # keep --no-deps (see below)
jupyter lab

Windows (Command Prompt): the same, except create and activate the environment with py -m venv %USERPROFILE%\cs6140_env and %USERPROFILE%\cs6140_env\Scripts\activate. If you have an NVIDIA GPU, see "NVIDIA GPU" below before installing torch.

Why --no-deps? The d2l package pins very old versions of numpy/matplotlib that won't install on current Python. d2l itself works fine with current versions, so we install it without its pins.

2. Check that it works

Paste into a notebook cell and run:

import torch
from d2l import torch as d2l
device = 'cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu'
print('torch', torch.__version__, '| device:', device)
data = d2l.FashionMNIST(batch_size=256)          # downloads ~30 MB the first time
print(len(data.train), 'training images')

Expected: no errors, a device name (cuda = NVIDIA GPU, mps = Apple Silicon GPU, cpu = no GPU, which is fine), and 60000 training images. That's it: you're set for HW5–HW7. Each starter notebook begins with a cell that picks your device automatically.

3. Which machine do I have?

Your machine	What happens
Mac with Apple Silicon (M1–M4)	Uses the Mac's GPU (`mps`) automatically. Nothing extra to install.
Linux or Windows with an NVIDIA GPU	Uses the GPU (`cuda`). On Windows this needs a special torch install (below).
Anything else (Intel Mac, laptop without NVIDIA GPU)	Runs on the CPU. HW5 is fast; for HW6–HW7 use the starters' `FAST` setting, or Colab (below).

Optional details

Open these only if they apply to you.

NVIDIA GPU (Windows, Linux, WSL2)

You need an up-to-date NVIDIA driver (nvidia-smi shows it). You do not need to install CUDA separately; PyTorch includes it.
Linux: the plain pip install torch torchvision above already has GPU support.
Windows: plain pip install torch is CPU-only. Instead, use the command from the PyTorch install page (choose Pip / Windows / CUDA). It looks like pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128.
Check: python -c "import torch; print(torch.cuda.is_available())" prints True.

Apple Silicon notes

The Mac GPU supports float32 only. When you turn numpy arrays (float64 by default) into tensors, use torch.tensor(X, dtype=torch.float32). Otherwise you get Cannot convert a MPS Tensor to float64.
If an operation reports not currently implemented for the MPS device, start Jupyter with the environment variable PYTORCH_ENABLE_MPS_FALLBACK=1 set, so that operation runs on the CPU.
Very small models (e.g. HW5's first problems) can be faster on the CPU. That's normal; the starters choose accordingly.
The d2l book's own GPU helpers only know NVIDIA GPUs. The starters' first cell patches them, so d2l.Trainer(..., num_gpus=1) uses the Mac GPU.

Google Colab (free GPU in the browser)

Upload the starter notebook to Colab, then choose Runtime → Change runtime type → T4 GPU.
First cell: !pip install d2l==1.0.3 --no-deps sacrebleu gensim (everything else is preinstalled).
Sessions reset after a while. Save results to Google Drive, especially HW6's results_hw6.json, which HW7 reads. Course data files (e.g. the translation sentence pairs) must be uploaded to the session or to Drive.

Using plain PyTorch instead of d2l's Trainer

The starters use d2l's Module / Trainer classes, which match the book. You may use a plain PyTorch training loop instead (the most common style in practice), as long as you evaluate exactly as the starter does. This is what d2l.Trainer does for you:

model = model.to(device)
opt = torch.optim.SGD(model.parameters(), lr=0.1)
loss_fn = torch.nn.CrossEntropyLoss()
for epoch in range(num_epochs):
    model.train()                                  # dropout / batch norm in training mode
    for X, y in train_loader:
        X, y = X.to(device), y.to(device)
        loss = loss_fn(model(X), y)
        opt.zero_grad(); loss.backward(); opt.step()
    model.eval()                                   # evaluation mode
    with torch.no_grad():
        val_acc = sum((model(X.to(device)).argmax(1) == y.to(device)).float().mean().item()
                      for X, y in val_loader) / len(val_loader)
    print(f'epoch {epoch}: val acc {val_acc:.3f}')

No other frameworks (Lightning, Keras, JAX, Hugging Face) are needed for this course.

conda users, and picking the right Jupyter kernel

conda works too: conda create -n cs6140 python=3.12, conda activate cs6140, then the same pip install lines. Install torch with pip, not conda.
If Jupyter can't find torch although you installed it, it's using a different Python. Start jupyter lab from the activated environment, or register it once as a kernel: pip install ipykernel then python -m ipykernel install --user --name cs6140, and pick "cs6140" in Jupyter.

Troubleshooting

Symptom	Fix
Installing d2l tries to build old numpy and fails	You left out `--no-deps`. Run `pip install --upgrade numpy matplotlib pandas scipy`, then the d2l line again.
`ModuleNotFoundError: torch` in Jupyter	Wrong kernel; see the section above.
Out of memory, or the machine freezes during training	Halve `batch_size`, use the `FAST` setting, and restart the kernel to free old models.
`DataLoader worker ... exited unexpectedly` (usually in `.py` scripts)	Put the script's code under `if __name__ == "__main__":`, or use `num_workers=0`.
`OMP: Error #15 ... libomp already initialized` (conda on Mac)	Use the plain venv from Step 1 instead of conda.
Numbers differ slightly between runs or machines	Normal (random initialization, floating point). Compare with tolerances.

What each homework downloads

HW5: Fashion-MNIST, MNIST (~30 MB each, automatic); Wine and the 2-D Gaussian files are in the course data folder.
HW6: CIFAR-10 (~170 MB), pretrained ResNet-18 weights (~45 MB), Penn Treebank (~2 MB), all automatic; Spanish–English sentence pairs (course data folder, ~24 MB).
HW7: the same Penn Treebank and sentence pairs, plus your saved HW6 results.

d2l stores downloads in a ../data folder next to the notebook's folder. Keep your HW notebooks in sibling folders (e.g. cs6140/hw5, cs6140/hw6) so they share one download cache.

Still stuck? Post on Piazza with your OS, the output of the Step 2 check, and the full error message (as text, not a screenshot).