Convis Linear Filters



In [1]:

    
%pylab inline
import convis









    



Populating the interactive namespace from numpy and matplotlib






    



/home/jacob_unencrypted/Environments/py3/lib/python3.6/site-packages/h5py/__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.
  from ._conv import register_converters as _register_converters

One Dimensional Filtering

The Conv1d filter affects only the temporal dimension of the input by convolving it with a temporal weight.



In [2]:

    
model = convis.filters.Conv1d()
model.exponential(0.01)
model.plot_impulse()









    Out[2]:




Output containing a 1x1x500x10x10 Tensor.

Two Dimensional Filtering

The Conv2d filter affects only the spatial dimension of the input by convolving it with a spatial weight.



In [3]:

    
model = convis.filters.Conv2d()
model.gaussian(0.1)
model.plot_impulse((10,21,21))









    Out[3]:




Output containing a 1x1x10x21x21 Tensor.

Thre Dimensional Filtering

The Conv3d filter is the most flexible filter and convolves the input with a spatio-temporal weight.



In [4]:

    
model = convis.filters.Conv3d()
model.set_weight(randn(10,5,5))
model.plot_impulse((100,21,21))









    Out[4]:




Output containing a 1x1x100x21x21 Tensor.

RF: Receptive Field

The RF class has only a 1 pixel output, so different receptive fields have to be created by using additional out_channels that each has a 3d filter. Similar to the Conv3d filter, RF does a 3d convolution, but it removes padding and cuts the input to the exact size of the weight. The result is a spatio-temporal receptive field that will give a single time series as an output.



In [5]:

    
model = convis.filters.RF()
model.set_weight(randn(10,5,5))
model.plot_impulse((100,9,9))









    Out[5]:




Output containing a 1x1x100x1x1 Tensor.



In [6]:

    
model = convis.filters.RF(bias=False)
model.set_weight(randn(3,1,10,20,20))
inp = convis.samples.random_checker_stimulus(10,20,20).repeat(20,0)
o = model.run(inp,dt=200)
o









    Out[6]:




Output containing a 1x3x200x1x1 Tensor.