Is there a way to perform batch mode active learning ?

[lironesamoun]

Hi,

Instead of having of having unlabeled data which come as a stream, I would like to know if there is a way with libact to perform batch mode active learning meaning that the users can select multiples images at once (positive and negatives) ?

thank you in advance

[yangarbiter]

We haven't officially support batch mode active learning yet.

Though some algorithms with slight modification can achieve this feature. Take uncertainty sampling for example, you may change the following line

https://github.com/ntucllab/libact/blob/master/libact/query_strategies/uncertainty_sampling.py#L111

by replacing np.argmin with something like n_smallest to return the index of n most uncertain data.

[lironesamoun]

Thank you, I will try. When do you think you will support officially batch mode active learning ?

My question may be silly but I have made some changes inside the uncertainty_sampling file, then I have made sudo python setup.py build and sudo python setup.py install and my changes haven't been taken in consideration...I can do whatever I want, where I run a python example, nothing has changed. Do you know why ? Or maybe did I miss something ?

[yangarbiter]

You may want to check your environment variables like PYTHONPATH or PATH. Or maybe you are inside a virtualenv where you don't need sudo to install the package.

[lironesamoun]

Yes, the problem was because of the virtual environment. Thank for your help !

[lironesamoun]

In order to not open a new thread, is there a way to associate an image to each feature ? I mean If I don't use pixel as a feature, I cannot use InteractiveLabel. So for example, If I use an abstract feature like SIFT, how can I associate each feature to each image. I tried to use a dictionary but it doesn't work.. thank in advance

[yangarbiter]

Do you mean that the feature is separated with the pixel array?

If this is what you mean, I think you can try putting the corresponding pixel array to the label function of InteractiveLabeler.

For the label_digits example: https://github.com/ntucllab/libact/blob/master/examples/label_digits.py#L90

Instead of lb = lbr.label(trn_ds.data[ask_id][0].reshape(8, 8)), you may put lb = lbr.label(feature_to_image(trn_ds.data[ask_id][0]))

If your image is not in a form of pixel, you may need to modify the image rendering part of InteractiveLabeler. https://github.com/ntucllab/libact/blob/master/libact/labelers/interactive_labeler.py#L32

[lironesamoun]

Yes. For example, I have an array of images (not in a form of pixel) and I have an array of associated features. images = [image1,..,imageN] and features = [features1,...,featuresN]

So yes I would like to create a Dataset object but after doing that, I lose track of my corresponding images as I have only in Dataset the pool X and label Y. There is no way to know which X corresponds to which image, if you see what I mean. I will try to do like you advise me.

[yangarbiter]

The most simply way of doing it is to take featureX and search through [features1,...,featuresN] one by one for the index and then go back to the imageX.

Though if I remembered correctly, the index of features in the dataset won't change during the process. This means that if you create dataset like this Dataset(features, Y). The index returned by i = qs.make_query() should be the same as the original one. which means trn_ds.data[i][0] shoud be the same as features[i], and you will show image[i].

If the first method is too slow for your application, you can double check with the second method by running some small samples.