Multi-label neural networks can be challenging to make. With many multilabel datasets being text, difficult to find, or having subjective questionable labels, I created this generator to eliminate one point of error when making multi-label NNs; the dataset.

This generator makes it easy to confirm that a neural network is functioning as expected.

The recommended use case is to create an easy dataset that any neural network should do well on as a basic verification that everything is working as expected; however, if robustness is desired, the difficulty can be increased.

• Images and labels are easily verifiable.
• Dataset difficulty can be easily adjusted using generation parameters.

generate_multilabel_toy_dataset(sample_number=10000, label_count=3, x_res=256, y_res=256, channels=3, v_min=0, v_max=1, size=[10, 40], frequency=[2, 20], label_count=3, label_frequency=0.5, opacity=1, path="", export_folder="ShapesDataset" export_type="image_folder", verbose=True, random_seed=0, random_channel_classes=False)

• Number of samples
• X_res, R_res, Channels
• Number of labels
• Shape Sizes (random between bounds or hard-set)
• Frequency of shapes in image (random between bounds or hard-set)
• Frequency of images to have a label (linear space between bounds or hard-set)
• If shapes should be generated to random channels
• Opacity of generated shapes

• Exporting the dataset into an image-folder/pickle file
• Setting a seed for random numbers
• If messages should be shown (progress bar, comments)

• Implement an opacity option. (Added 9/4, Done 9/6/23)
• Detect files in dataset folder and replace them if generating a new dataset. (Added 8/30, Done 9/2/23)
• Add more warnings, tracebacks, and comments. (Added 8/27, Ongoing)
• Add an option to get rid of no-label images (all zeros). (Added 8/27, Removed 8/30/23)
• Add a progress bar. (Added 8/27, Done 8/30/23)
• Add a set seed for random numbers. (Added 8/27, Done 8/30/23)
• Add channel-specific classes/shapes. (Added 8/27, Done 8/30/23)
• Make an option to export a pickle file. (Added 8/27, Done 8/29/23)
• Allow different classes to have different frequencies (currently only a single-hard-set value). (Added 8/27, Done 8/29/23)

3-labels: (circles, lines, triangles)

5-labels: (circles, lines, triangles, squares, pentagons)

256x256x3 3-label w/ all 3 classes:	128x256x3 3-label w/ all 3 classes:

256x256x3 5-label w/ all 5 classes:	256x256x3 5-label w/ all & random channels:	256x256x3 5-label w/ all & Opacity=0.5: