Include transform to change spectrograms into images.

[jcr0269]

Would love a way to save the datasets as .png files or other file types. I have run into a few issues when trying to decode the lmdb file and pull out the data instead of just passing it into torch. Such as by trying to unpickle the folder like: ############################################################################# def extract_images(lmdb_path, output_dir): env = lmdb.open(lmdb_path, readonly=True)

if not os.path.exists(output_dir):
    os.makedirs(output_dir)

with env.begin() as txn:
    cursor = txn.cursor()
    for i, (key, value) in enumerate(cursor):
        key = int.from_bytes(key, 'big')  # Convert from bytes to int
        try:
            img_data = pickle.loads(value)  # Convert from bytes to data
            # Convert byte data to PIL Image
            img = Image.open(io.BytesIO(img_data))

            img.save(os.path.join(output_dir, f'{key}.jpg'))

        except Exception as e:
            print(f'Error processing key {key}: {str(e)}')
            print(value)
        # Stop after checking 5 images
        if i >= 5:
            break

############################################################################## If there is an easier way to do this though, please let me know.

[gvanhoy]

I think the best way here is to have a ToPNG or ToJPEG transform that can be used in tandem with the DatasetCreator. You just add this transform at the end of the transforms pipeline and then hand the dataset over to the DatasetCreator.

On the training/reading-from-disk end, you'd have to have a FromPNG or FromJPEG transform when you read it from the disk. I think this is the best because then we can keep WidebandSig53 as it is and just add a few transforms to the codebase.

[jcr0269]

awe okay, ill spend some time today working on that and try to post a solution if I get one working. Thank you for the quick reply

[gvanhoy]

Related to #181 in a way.

[ereoh]

Hello! Just to update you that we are currently rewriting a lot of the code for the v1.0.0 release in a few months, and this feature will be much easier to include then. Thanks for you suggestion and patience.