ericjang

"""
To generate dataset, make folder with each video's frames extracted into subdir. the dino codebase has code to do this easily. 

videos = glob.glob(os.path.join(SOURCE_DIR,"*.mp4"))[:num_videos_to_extract]
for i, video in enumerate(videos):
    print(i)
    directory=os.path.join(OUTPUT_DIR, str(i))
    if not os.path.exists(directory):
        os.makedirs(directory)
    _extract_frames_from_video(inp=video, out=directory)

ericjang

"""
A lightweight experiment logbook for Jupyter/Colab-style ad hoc experiments.

Let's say you generate a plot with Matplotlib and want to re-run your notebook with a 
different set of configurations and then compare the resulting plot to the one you saved (to see
if the new configuration is better).

# Saving experiments
f = plt.gcf()
elog.savefig(f, 

ericjang

def _reduce_logmeanexp(x, axis, epsilon):
  """Numerically-stable (?) implementation of log-mean-exp.

  Args:
    x: The tensor to reduce. Should have numeric type.
    axis: The dimensions to reduce. If `None` (the default),
      reduces all dimensions. Must be in the range
      `[-rank(input_tensor), rank(input_tensor)]`.
    epsilon: Floating point scalar to avoid log-underflow.

ericjang

for i in range(num_bijectors):
    bijectors.append(tfb.MaskedAutoregressiveFlow(
      shift_and_log_scale_fn=tfb.masked_autoregressive_default_template(
      hidden_layers=[512, 512])))
    bijectors.append(tfb.Permute(permutation=[1, 0]))
flow_bijector = tfb.Chain(list(reversed(bijectors[:-1])))

ericjang

class BatchNorm(tfb.Bijector):
    def __init__(self, eps=1e-5, decay=0.95, validate_args=False, name="batch_norm"):
        super(BatchNorm, self).__init__(
            event_ndims=1, validate_args=validate_args, name=name)
        self._vars_created = False
        self.eps = eps
        self.decay = decay

    def _create_vars(self, x):
        n = x.get_shape().as_list()[1]

ericjang

loss = -tf.reduce_mean(dist.log_prob(x_samples))
train_op = tf.train.AdamOptimizer(1e-3).minimize(loss)
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
NUM_STEPS = int(1e5)
global_step = []
np_losses = []
for i in range(NUM_STEPS):
    _, np_loss = sess.run([train_op, loss])
    if i % 1000 == 0:

ericjang

d, r = 2, 2
DTYPE = tf.float32
bijectors = []
num_layers = 6
for i in range(num_layers):
    with tf.variable_scope('bijector_%d' % i):
        V = tf.get_variable('V', [d, r], dtype=DTYPE)  # factor loading
        shift = tf.get_variable('shift', [d], dtype=DTYPE)  # affine shift
        L = tf.get_variable('L', [d * (d + 1) / 2],
                            dtype=DTYPE)  # lower triangular

ericjang

# quite easy to interpret - multiplying by alpha causes a contraction in volume.
class LeakyReLU(tfb.Bijector):
    def __init__(self, alpha=0.5, validate_args=False, name="leaky_relu"):
        super(LeakyReLU, self).__init__(
            event_ndims=1, validate_args=validate_args, name=name)
        self.alpha = alpha

    def _forward(self, x):
        return tf.where(tf.greater_equal(x, 0), x, self.alpha * x)

ericjang

import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
tfd = tf.contrib.distributions
tfb = tfd.bijectors

ericjang

base_dist = tfd.MultivariateNormalDiag(loc=tf.zeros([2], tf.float32))