dustinvtran

### Keybase proof

I hereby claim:

  * I am dustinvtran on github.
  * I am dvt (https://keybase.io/dvt) on keybase.
  * I have a public key whose fingerprint is 91B5 FBA6 2E1D 2938 4866  BBB8 3571 FE5A 0F13 2DFE

To claim this, I am signing this object:

dustinvtran

plot.risk <- function(data, est) {
  # Plot estimated biases of the 4 optimization routines performed.
  #
  # Args:
  #   data: List of x, y, A, theta in particular form
  #   est: A list of matrix estimates, one for each set of optimization
  #        methods done on the data.
  #
  # Returns:
  #   A log-log plot with 4 curves, showing excess risk over training size.

dustinvtran

Summary Portion

1. Minjae: Brief overview of the contributions of paper, summary of decomposition of the Q into k basis functions, quick derivation of the weight.
2. Dustin: LSQ, asymptotic, incremental update, LSPI.

Some things to reinforce:

• That it's linear with respect to the parameters and not necessarily the data.
• That it calculates parameters using the least squares approach (in other words, decision theory), rather than doing MLE or MAP or whatever.
• That its primary advantages are: intuition, simplicity, good theoretical

dustinvtran

# This note is to show that TensorFlow objects are shallow copied:
# a deep copy is not made when a class contains another class.
#
# This is relevant for variational parameters. For example, Inference
# holds Variational, we train variational parameters via
# inference.run(). We can use the original variational object as it will
# have the trained parameters, and not just resort to using
# inference.variational.
# (This is probably a feature in Python in general.)
from __future__ import print_function

dustinvtran

# TensorFlow r0.10
# 
# Building Graph
#
add_to_collection,tf.add_to_collection
as_dtype,tf.as_dtype
bytes,tf.bytes
container,tf.container
control_dependencies,tf.control_dependencies
convert_to_tensor,tf.convert_to_tensor

dustinvtran

from edward.models import RandomVariable
from tensorflow.contrib.distributions import Distribution

class Concrete(RandomVariable, Distribution):
  def __init__(self, *args, **kwargs):
    kwargs['is_continuous'] = True
    kwargs['is_reparameterized'] = True
    super(Concrete, self).__init__(*args, **kwargs)
    
  def _log_prob(self, value):

dustinvtran

#!/usr/bin/env python
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import edward as ed
import matplotlib.cm as cm
import numpy as np
import seaborn as sns
import tensorflow as tf

dustinvtran

dustinvtran

import inspect

class LocalsCapture(object):
  def __enter__(self):
    caller_frame = inspect.currentframe().f_back
    self.local_names = set(caller_frame.f_locals)
    return self
  
  def __exit__(self, exc_type, exc_val, exc_tb):
    caller_frame = inspect.currentframe().f_back

dustinvtran

import operator

def toposort(start_node, children=operator.attrgetter('children')):
  """Generate nodes in DAG's topological order. This guarantees
  for any edge U -> V, we always visit U before visiting V.

  This lets us play the tape via the "push" dataflow model.
  https://stackoverflow.com/questions/981027/what-are-pro-cons-of-push-pull-data-flow-models
  """
  parent_counts = {}