Add the implementation for downloading mnist

[eyumboo]

This PR adds download_mnist.py to examples/singa_peft/examples/data.

[eyumboo]

Add the implementation for downloading mnist

[eyumboo]

Add the implementation for downloading mnist

[moazreyad]

download_dir is currently hard-coded to /tmp/, but the PEFT MNIST loader defaults to /tmp/mnist and expects the dataset files under that directory:

• Default dataset path: https://github.com/eyumboo/singa/blob/dev-postgresql/examples/singa_peft/examples/data/mnist.py#L85
• File lookup logic assumes /tmp/mnist: https://github.com/eyumboo/singa/blob/dev-postgresql/examples/singa_peft/examples/data/mnist.py#L38

In addition, the example runner explicitly passes -dir /tmp/mnist:
https://github.com/eyumboo/singa/blob/dev-postgresql/examples/singa_peft/examples/run.sh#L21

Because of this mismatch, even after running the new downloader, training fails since it looks for files like /tmp/mnist/train-images-idx3-ubyte.gz, which are not present.

It would be better if the downloader:

• Writes directly to /tmp/mnist (and creates the directory if it doesn’t exist)
• Ideally supports the same --dir-path argument used by the training pipeline for consistency

[moazreyad]

Suggestion: improve download logging clarity

The current download log is helpful, but it would be even better if we also print the resolved local path (or target directory) where files are stored. This makes it easier to debug storage issues and verify where artifacts end up, especially since the download location may depend on defaults or environment variables (e.g. /tmp/mnist).

In addition, it would improve clarity to include the local filename/path in both download states.

Proposed changes:

if not os.path.isfile(filename):
    print("Downloading %s -> %s" % (url, filename))
    urllib.request.urlretrieve(url, filename)
else:
    print("Already Downloaded: %s -> %s" % (url, filename))

[moazreyad]

One remaining integration suggestion for a future PR: examples/singa_peft/examples/autograd/mnist_cnn.py currently expects the MNIST files to be located under /tmp, while the downloader now defaults to /tmp/mnist.

As a result, running the downloader with its default settings is not sufficient for training to work out of the box.

A future improvement could be to either:

• update mnist_cnn.py to use /tmp/mnist (or the same default as the downloader), or
• document that users should invoke the downloader with --dir-path /tmp when using mnist_cnn.py.

Ideally, both scripts should share the same default directory to avoid this mismatch.

Another non-blocking suggestion for a future cleanup is the directory structure under examples/singa_peft/examples, which feels a bit redundant since everything under examples/ is already example code. It might be worth reorganizing this in a follow-up PR by removing the second examples directory and simplifying the layout. A flatter, more consistent structure would make the project easier to navigate and reduce unnecessary nesting.

Neither of these is a blocking issue for merging this PR, but addressing them in a follow-up PR would improve the out-of-the-box experience, project organization, and overall clarity for future users.