jessstringham

from collections import defaultdict
import random


WINDOW_SIZE = 2


with open("alice.txt") as f:
    full_text = f.read()

jessstringham

jessstringham

from bokeh.models import HoverTool
from bokeh.plotting import figure, show, output_file, ColumnDataSource
from bokeh.layouts import gridplot
from bokeh.palettes import magma

def plot(u, descs=index_to_interesting_subject):
    source = ColumnDataSource(data=dict(
        x1=u[:, 0],
        x2=u[:, 1],
        x3=u[:, 2],

jessstringham

'''This is hacky code I use to convert Jupyter notebooks into Jekyll posts.

Notebooks' first line should be
    # Title

and `Title` will be used as the post's title.

I convert LaTeX into the form MathJax needs.

To make plots work, see `get_maybe_image_code` comment.

jessstringham

image = mpimg.imread(IMAGE_LOCATION)

# pasta the first image
h, w, c = image.shape
image = image.reshape(h // 2, 2, w, c)
image = image.transpose(1, 0, 2, 3)
image = np.hstack(image)

# rotate the image
image = image.transpose(1, 0, 2)

jessstringham

import numpy as np
import matplotlib.image as mpimg
import matplotlib.pyplot as plt


PASTA_SIZE = 16
NUM_IMAGE_SPLITS = 2
IMAGE_LOCATION = 'something.png'

image = mpimg.imread(IMAGE_LOCATION)

jessstringham

np.einsum

I had this gross reshape/tensor product/transpose stuff on huge matrices, and I knew it was making intermediate copies of the matrices that I didn't want to. So I tried out np.einsum, and I think it actually turned out simpler than thinking through the other matrix manipulation.

Here are some quick notes.

Real blogs/documentation

jessstringham

import numpy as np
from numpy.lib.stride_tricks import as_strided

# Makes a new matrix like this:
# input: [0, 1, 2, 3]
# output: [[0, 1], [1, 2], [2, 3]]
# But, does it all in memory using `as_strided`, so it's speedy.
# And if you make a mistake, you can see fake numbers

jessstringham

Srry, this is a bit of an awkward format. I might move this and a few other demos to a repo.

This is code one could drop into Jupyter [1] notebook cells and draw some wicked-looking graphs. I wrote it to understand Bayesian linear regression (for this classMLPR, specifically this and this). No promises that I did it perfectly.

The graphs show samples from the posterior weights in pink.