Custom data at test time

[eljdos]

Hello,

Thank you so very much for the impressive work and for making it open source.

Out of curiosity I was wondering if you are going to explain how one could get it to work with custom datasets (by that i mean not from the list of supported datasets).

Thank you so very much and best regards,

[david-berthelot]

One way is to look at scripts/create_datasets.py, we convert datasets into a TFRecord file of tf.Example with two fields:

• image is a string containing the png encoded image.
• label is an int64 (unused, you can set it to 0)

Once you created two record files for your dataset (train and test), you can simply modify libml/data.py to add your dataset to the dictionary. And then when calling lag.py you would need to add --dataset=your_dataset_name.

[eljdos]

Thank you very much for the answer :)