aggregation of filters/activities/etas into unified MultiBind instance

[ilibarra]

into mubind/tl/aggregation.py

• [x] function mb.read_h5 to read multibind model from file, in case not implemented.
• [x] Function mb.tl.combine_h5 to load and merge BindingModes from h5 files into single MultiBind classes with all those concatenated.
  • [ ] Unit test for this, by taking one of the current models and placing it in _data.
• [x] The activities and log_etas can be unified into a single value. This assumes the model will be merged into a new dataset.
• [x] Function that measures the distance between two binding modes, by correlation of weights across all positions (this can be reported from other internal functions. Ask for input). Next necessary features
• [x] function mb.get.binding_modes() that takes the already loaded from h5 files, and results in a list of binding modes. The input can be a h5_path/multibind instance, or list of h5_paths/multibind instances. h5 and multibind instances can be mixed.
• [ ] function mb.tl.concatenate, that returns a single multibind instance from the combine_h5 and get.binding_modes.
• [ ] assert after concatenation that log_activities/etas make sense.
• [ ] assert that after concatenation that conv_mono.shape[1] == conv_di.shape[1] + 1, for each filter.

[ilibarra]

@ege-erdogan may you please indicate to me where mb.tl.combine_h5 is and/or merge it to development if ready? The last task can be separated, I suppose.

[ege-erdogan]

@ilibarra Sorry for the late response; pushed to the model-aggregation branch now (https://github.com/theislab/mubind/blob/model-aggregation/mubind/tl/aggregation.py).

Model layers are appended to a new (empty) model and activations/etas are summed right now, but changing what is done during aggregation is straightforward in the implementation.

[ilibarra]

Next necessary features

• [ ] function mb.get.binding_modes() that takes the already loaded from h5 files, and results in a list of binding modes. The input can be a h5_path/multibind instance, or list of h5_paths/multibind instances. h5 and multibind instances can be mixed.
• [ ] function mb.tl.concatenate, that returns a single multibind instance from the combine_h5 and get.binding_modes.
• [ ] assert after concatenation that log_activities/etas make sense.
• [ ] assert that after concatenation that conv_mono.shape[1] == conv_di.shape[1] + 1, for each filter.

[ilibarra]

@ege-erdogan please feel free to use snippets from these scripts for faster understanding/implementation of intermediate steps. https://github.com/theislab/mubind/blob/development/notebooks/concatenation/01_load_and_concatenate.ipynb

[ilibarra]

majorly solved by #108

[ilibarra]

Reopening due to test_aggregation.py not working. It needs inspection.