This is the first part in what will hopefully be a series of notes 1 on my exploration of the recently released Google QuickDraw dataset, using the concurrently released SketchRNN model.
The QuickDraw dataset is curated from the millions of drawings contributed by over 15 million people around the world who participated in the "Quick, Draw!" A.I. Experiment, in which they were given the challenge of drawing objects belonging to a particular class (such as "cat") in under 20 seconds.
SketchRNN is a very impressive generative model that was trained to produce vector drawings using this dataset. It was of particular interest to me because it cleverly combines many of the latest tools and techniques recently developed in machine learning, such as Variational Autoencoders, HyperLSTMs (a HyperNetwork for LSTM), Autoregressive models, Layer Normalization, Recurrent Dropout, the Adam optimizer, and others.
This notebook was based on the notebook included with the code release. I've made significant stylistic changes and some minor changes to ensure Python 3 compatibility as Magenta only supports Python 2 currently.
These notes will likely be quite hasty and unpolished , as it has been written more for myself than anyone else. While I've always tried to avoid cluttering up my blog with notebooks on in-progress work, I've decided to just make a habit of posting them everytime I complete a session of work on something for the day. The aim is to start creating series of short and succinct notebooks, and avoid further accumulating a collection of long and disorganized notebooks that I never want to touch again, because the sheer effort of going through, cleaning up the experimental code, and articulating it effectively in a well-thought-out blog article just requires too much willpower. Doing it this way will hopefully make it much easier for me to create and share useful content quickly.↩
In [1]:
%matplotlib inline
%config InlineBackend.figure_format = 'svg'
%load_ext autoreload
%autoreload 2
In [2]:
import matplotlib.pyplot as plt
import matplotlib.patches as patches
import numpy as np
import tensorflow as tf
from matplotlib.path import Path
In [3]:
from magenta.models.sketch_rnn.sketch_rnn_train import \
(load_env,
load_checkpoint,
reset_graph,
download_pretrained_models,
PRETRAINED_MODELS_URL)
from magenta.models.sketch_rnn.model import Model, sample
from magenta.models.sketch_rnn.utils import (get_bounds,
to_big_strokes,
to_normal_strokes)
In [4]:
# set numpy output to something sensible
np.set_printoptions(precision=8,
edgeitems=6,
linewidth=200,
suppress=True)
In [5]:
tf.logging.info("TensorFlow Version: {}".format(tf.__version__))
In [6]:
DATA_DIR = ('http://github.com/hardmaru/sketch-rnn-datasets/'
'raw/master/aaron_sheep/')
MODELS_ROOT_DIR = '/tmp/sketch_rnn/models'
In [7]:
DATA_DIR
Out[7]:
In [8]:
PRETRAINED_MODELS_URL
Out[8]:
In [9]:
download_pretrained_models(
models_root_dir=MODELS_ROOT_DIR,
pretrained_models_url=PRETRAINED_MODELS_URL)
The directory tree looks like this. There are a few pretrained models for us to explore.
In [10]:
!tree -L 3 /tmp/sketch_rnn/
We look at the layer normalized model trained on the aaron_sheep dataset for now.
In [11]:
MODEL_DIR = MODELS_ROOT_DIR + '/aaron_sheep/layer_norm'
In [12]:
(train_set,
valid_set,
test_set,
hps_model,
eval_hps_model,
sample_hps_model) = load_env(DATA_DIR, MODEL_DIR)
The strokes object variable is the list of data points, which are sequences of strokes, represented as a 2D NumPy array of xy-offsets and the pen state.
In [13]:
len(train_set.strokes)
Out[13]:
We can get a random sample from the dataset like so
In [122]:
a = train_set.random_sample()
a.shape
Out[122]:
In the original notebook, the authors implemented their own function to iterate through a stroke sequence and write out a SVG Path string. I found this a bit inelegant and cumbersome to work with.
Here we simply subclass Path, which underpins all matplotlib.patch classes. This is almost perfect for our data format since it "supports the standard set of moveto, lineto, curveto commands to draw simple and compound outlines consisting of line segments and splines. The Path is instantiated with a (N,2) array of (x,y) vertices, and a N-length array of path codes." We just need to normalize the data slightly as we shall explain later.
In [133]:
class StrokesPath(Path):
def __init__(self, data, factor=.2, *args, **kwargs):
vertices = np.cumsum(data[::, :-1], axis=0) / factor
codes = np.roll(self.to_code(data[::,-1].astype(int)),
shift=1)
super(StrokesPath, self).__init__(vertices,
codes,
*args,
**kwargs)
@staticmethod
def to_code(cmd):
# if cmd == 0, the code is LINETO
# if cmd == 1, the code is MOVETO (which is LINETO - 1)
return Path.LINETO - cmd
Now drawing the strokes becomes as simple as
In [134]:
fig, ax = plt.subplots(figsize=(3, 3))
strokes = StrokesPath(a)
patch = patches.PathPatch(strokes, facecolor='none')
ax.add_patch(patch)
x_min, x_max, y_min, y_max = get_bounds(data=a, factor=.2)
ax.set_xlim(x_min-5, x_max+5)
ax.set_ylim(y_max+5, y_min-5)
ax.axis('off')
plt.show()
We define this as a function to maximize modularity and reusability
In [135]:
def draw(stroke, factor=.2, pad=(10, 10), ax=None):
if ax is None:
ax = plt.gca()
x_pad, y_pad = pad
x_pad //= 2
y_pad //= 2
x_min, x_max, y_min, y_max = get_bounds(data=stroke,
factor=factor)
ax.set_xlim(x_min-x_pad, x_max+x_pad)
ax.set_ylim(y_max+y_pad, y_min-y_pad)
strokes = StrokesPath(stroke)
patch = patches.PathPatch(strokes, facecolor='none')
ax.add_patch(patch)
ax.axis('off')
Now it is easy to fully take advantage of the functionality provided by Matplotlib to create more complex plots. For example, to draw the sketches in a grid, we just call our draw function on the grid of axes created with subplots.
In [136]:
fig, ax_arr = plt.subplots(nrows=5,
ncols=10,
figsize=(8, 4),
subplot_kw=dict(xticks=[],
yticks=[],
frame_on=False))
fig.tight_layout()
for ax_row in ax_arr:
for ax in ax_row:
strokes = train_set.random_sample()
draw(strokes, ax=ax)
plt.show()
In [137]:
a[::,-1].astype(int)
Out[137]:
A 1 means the pen is to be lifted from the current point before being moved to the next point. In terms of SVG Path commands or Matplotlib Path codes, this corresponds to the next point being preceded by the m command or having the MOVETO code, respectively. Otherwise, the pen just draws a line to the next point, which means it is preceded by the l command or has the LINETO code.
Our to_code static method above simply turns the pen action for the next state to the corresponding Matplotlib Path code.
In [138]:
{c: getattr(Path, c) for c in dir(Path) if c.isupper()}
Out[138]:
In [139]:
StrokesPath.to_code(a[::,-1].astype(int))
Out[139]:
As explained, the $i$th code is meant for the $i+1$th vertex, so we are not done just yet. We must right shift the array by 1. But what do we do with the first and last elements of the sequence?
Note that the pen action for all stroke sequences are terminated with the pen being lifted. This corresponds to the next vertex having a MOVETO code. However, there are no vertices left, so this is superfluous and may be discarded.
In [140]:
all(StrokesPath.to_code(a[-1,-1]) == Path.MOVETO \
for a in train_set.strokes)
Out[140]:
On the other hand, the first vertex is always required to have a MOVETO code (otherwise, where the drawing actually begins is not well-defined). Therefore, the most simple and elegant solution is just to np.roll the array to right by 1, so the first code takes on the value of the last code, which is always a MOVETO.
Lastly, while SVG Paths supports both absolute positions and relative offsets (with M and m commands respectively), Matplotlib only supports absolute positions. As the datasets gives the points as relative offsets, we convert it to absolute positions simply by calling taking the cumulative sum with np.cumsum.