ericjang

#!/bin/env python

# uses tf-slim from release 0.10.0

import tensorflow as tf
slim = tf.contrib.slim

batch = 13
in_height, in_width, in_channels = 7, 7, 512

ericjang

r = 0.0
for n in range(1000000):
  r += (-1.0)**n/(2.0*n+1.0)
print(4*r)

ericjang

# Author: Eric Jang
# 2013-Jul-17

# The problem with using audioWave + time nodes + audioWave bonus tool in Maya
# to drive animation is that we can't really do spectral analysis (high pass/low pass filters)
# short of implementing it using hypershade nodes, so we can't get really fine-tuned animation
# ... but Python is good for this!

# caveat: this was coded up in one night so it may be unstable.  Use with caution.
# future work: instead of just the simple amplitude, we can perform FFT/spectral analysis to extract

ericjang

#
# Qt qmake integration with Google Protocol Buffers compiler protoc
#
# To compile protocol buffers with qt qmake, specify PROTOS variable and
# include this file
#
# Example:
# LIBS += /usr/lib/libprotobuf.so
# PROTOS = a.proto b.proto
# include(protobuf.pri)

ericjang

#Brown CS Potions Textbook:

• List available printers: lpstat -p -d
• removing something from gnome menubar: windows + alt + right click -> option to remove
• Screen Recording: Get window id: xwininfo -display :0, then use recordmydesktop
• Find out who's logged in: who
• GTX970 GPUs: cslab6a-6h, gemini, pastorale, pokemon
• penny has a 550 GTX Ti GPU
• Makes application full screen: F11 key

ericjang

-- torch implementation of KL between Gaussians
self.output = -1/2( 
  + torch.cmul(iqv, p.sigma):sum() -- trace term, tr(\Sigma_q^{-1} * \Sigma_p)
  + torch.cmul(diff:clone():pow(2), iqv):sum()  --  difference in means, (\mu_q-\mu_p)^T\Sigma_q^{-1}(\mu_q-\mu_p)
  - ndim -- k
  + torch.log(q.sigma):sum() - torch.log(p.sigma):sum() -- ratio of determinants log |\Sigma_q| / |\Sigma_p|
)

ericjang

def binary_crossentropy(t,o):
    return -(t*tf.log(o+eps) + (1.0-t)*tf.log(1.0-o+eps))

# reconstruction term appears to have been collapsed down to a single scalar value (rather than one per item in minibatch)
x_recons=tf.nn.sigmoid(cs[-1])

# after computing binary cross entropy, sum across features then take the mean of those sums across minibatches
Lx=tf.reduce_sum(binary_crossentropy(x,x_recons),1) # reconstruction term
Lx=tf.reduce_mean(Lx)

ericjang

def write_no_attn(h_dec):
    with tf.variable_scope("write",reuse=DO_SHARE):
        return linear(h_dec,img_size)

def write_attn(h_dec):
    with tf.variable_scope("writeW",reuse=DO_SHARE):
        w=linear(h_dec,write_size) # batch x (write_n*write_n)
    N=write_n
    w=tf.reshape(w,[batch_size,N,N])
    Fx,Fy,gamma=attn_window("write",h_dec,write_n)

ericjang

def decode(state,input):
    with tf.variable_scope("decoder",reuse=DO_SHARE):
        return lstm_dec(input, state)

ericjang

def sampleQ(h_enc):
    """
    Samples Zt ~ normrnd(mu,sigma) via reparameterization trick for normal dist
    mu is (batch,z_size)
    """
    with tf.variable_scope("mu",reuse=DO_SHARE):
        mu=linear(h_enc,z_size)
    with tf.variable_scope("sigma",reuse=DO_SHARE):
        logsigma=linear(h_enc,z_size)
        sigma=tf.exp(logsigma)