top_filters.py

#!/usr/bin/env python
"""
Usage:
./top_filters <path_to_a_saved_DBM.pkl> <optional: output path prefix>

Displays the matrix product of the layer 1 and layer 2 weights.
Also displays a grid visualization the connections in more detail.
Row i of the grid corresponds to the second layer hidden unit
with the ith largest filter norm.
Grid cell (i,j) shows the filter for the first layer unit with the
jth largest weight going into the second layer unit for this row.
The cells is surrounded by a colored box.
Its brightness indicates the relative strength of the connection between
the first layer unit and second layer unit, and its color indicates
the sign of that connection (yellow = positive / excitatory,
magenta = negative / inhibitory).

Optionally saves these images as png files prefixed with
the given output path name instead of displaying them.
This can be useful when working over ssh.
"""

from __future__ import print_function

__authors__ = "Ian Goodfellow"
__copyright__ = "Copyright 2012, Universite de Montreal"
__credits__ = ["Ian Goodfellow"]
__license__ = "3-clause BSD"
__maintainer__ = "LISA Lab"

import numpy as np
import sys
from theano.compat.six.moves import xrange
from pylearn2.config import yaml_parse
from pylearn2.gui.patch_viewer import PatchViewer
from pylearn2.gui.patch_viewer import make_viewer
from pylearn2.utils import serial


def sort_layer2(W2):
    """
    Sort weights of the a layer.

    Parameters
    ----------
    W2: list
        The hidden layer to sort.
    """
    print('Sorting so largest-norm layer 2 weights are plotted at the top')
    norms = np.square(W2).sum(axis=0)
    idxs = [elem[1] for elem in sorted(zip(-norms, range(norms.shape[0])))]

    new = W2.copy()

    for i in xrange(len(idxs)):
        new[:, i] = W2[:, idxs[i]]
    W2 = new

    return new


def get_mat_product_viewer(W1, W2):
    """
    Show the matrix product of 2 layers.

    Parameters
    ----------
    W1: list
        First hidden layer.
    W2: list
        Second hidden layer.
    out_prefix: str
        Path where to save image.
    """
    prod = np.dot(W1, W2)
    pv = make_viewer(prod.T)

    return pv


def get_connections_viewer(imgs, W1, W2):
    """
    Show connections between 2 hidden layers.

    Parameters
    ----------
    imgs: ndarray
        Images of weights from the first layer.
    W1: list
        First hidden layer.
    W2: list
        Second hidden layer.
    """
    W2 = sort_layer2(W2)

    N1 = W1.shape[1]
    N = W2.shape[1]
    N = min(N, 100)

    count = get_elements_count(N, N1, W2)

    pv = create_connect_viewer(N, N1, imgs, count, W2)

    return pv


def create_connect_viewer(N, N1, imgs, count, W2):
    """
    Create the patch to show connections between layers.

    Parameters
    ----------
    N: int
        Number of rows.
    N1: int
        Number of elements in the first layer.
    imgs: ndarray
        Images of weights from the first layer.
    count: int
        Number of elements to show.
    W2: list
        Second hidden layer.
    """
    pv = PatchViewer((N, count), imgs.shape[1:3], is_color=imgs.shape[3] == 3)

    for i in xrange(N):
        w = W2[:, i]

        wneg = w[w < 0.]
        wpos = w[w > 0.]

        w /= np.abs(w).max()

        wa = np.abs(w)

        to_sort = zip(wa, range(N1), w)

        s = sorted(to_sort)

        for j in xrange(count):

            idx = s[N1-j-1][1]
            mag = s[N1-j-1][2]

            if mag > 0:
                act = (mag, 0)
            else:
                act = (0, -mag)

            pv.add_patch(imgs[idx, ...], rescale=True, activation=act)

    return pv


def get_elements_count(N, N1, W2):
    """
    Retrieve the number of elements to show.

    Parameters
    ----------
    N: int
        Number of rows.
    N1: int
        Number of elements in the first layer.
    W2: list
        Second hidden layer.
    """
    thresh = .9
    max_count = 0
    total_counts = 0.
    for i in xrange(N):
        w = W2[:, i]

        wa = np.abs(w)

        total = wa.sum()

        s = np.asarray(sorted(wa))

        count = 1

        while s[-count:].sum() < thresh * total:
            count += 1

        if count > max_count:
            max_count = count

        total_counts += count
    ave = total_counts / float(N)

    print('average needed filters', ave)

    count = max_count

    print('It takes', count, 'of', N1, 'elements to account for ',
          (thresh*100.), '\% of the weight in at least one filter')

    lim = 10
    if count > lim:
        count = lim
        print('Only displaying ', count, ' elements though.')

    if count > N1:
        count = N1

    return count

if __name__ == '__main__':
    if len(sys.argv) == 2:
        _, model_path = sys.argv
        out_prefix = None
    else:
        _, model_path, out_prefix = sys.argv

    model = serial.load(model_path)

    layer_1, layer_2 = model.hidden_layers[0:2]

    W1 = layer_1.get_weights()
    W2 = layer_2.get_weights()
    print(W1.shape)
    print(W2.shape)

    mat_v = get_mat_product_viewer(W1, W2)

    if out_prefix is None:
        mat_v.show()
    else:
        mat_v.save(out_prefix+"_prod.png")

    dataset_yaml_src = model.dataset_yaml_src
    dataset = yaml_parse.load(dataset_yaml_src)
    imgs = dataset.get_weights_view(W1.T)

    conn_v = get_connections_viewer(imgs, W1, W2)
    if out_prefix is None:
        conn_v.show()
    else:
        conn_v.save(out_prefix+".png")