yaroslavvb

import torch
import numpy as np
import time

def benchmark():
  
  SIZE_GB = 1
  num_iters = 100
  def compute():
    a = torch.ones(int(SIZE_GB*1000*250000)).cuda()

yaroslavvb

import tensorflow as tf
from tensorflow.contrib.eager.python import tfe
tfe.enable_eager_execution()
context = tf.device('/gpu:0')
context.__enter__()

# download resnet_model
import sys, os, urllib.request
resnet_model_url="https://raw.githubusercontent.com/tensorflow/models/master/official/resnet/resnet_model.py"
response = urllib.request.urlopen(resnet_model_url)

yaroslavvb

import argparse
import os
import sys
from six.moves import shlex_quote

parser = argparse.ArgumentParser(description="Run commands")
parser.add_argument('-w', '--num-workers', default=1, type=int,
                    help="Number of workers")
parser.add_argument('-r', '--remotes', default=None,
                    help='The address of pre-existing VNC servers and '

yaroslavvb

import numpy as np
import tensorflow as tf
import scipy
from tensorflow.contrib.eager.python import tfe
tfe.enable_eager_execution()

# manual numpy example
# X = np.array(([[0., 1], [2, 3]]))
# W0 = X
# W1 = np.array(([[0., 1], [2, 3]]))/10

yaroslavvb

# Times: min: 440.60, median: 452.39, mean: 453.87


import util as u
u.check_mkl()

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim

yaroslavvb

import tensorflow as tf
from tensorflow.python.ops import control_flow_ops
from tensorflow.contrib import graph_editor as ge


def make_conditional_initializer(v):
    """Makes initializer of variable var lazy, returns new conditional init
    op."""

    cond = tf.is_variable_initialized(v)

yaroslavvb

def copy_initializer(from_var, to_var):
    from tensorflow.contrib import graph_editor as ge
    assert ge.reroute_a2b_ts(from_var.initial_value, to_var.initial_value), "No copy took place"

yaroslavvb

1. Making  Pull Request

First, click "Fork" on github tensorflow page

Then in Terminal:

export user=yaroslavvb
export branch_name
git clone https://github.com/$user/tensorflow.git
cd tensorflow

yaroslavvb

# Example of profiling session.run overhead
# for python profiling
# python -m cProfile -o session-run-benchmark-feed.prof session-run-benchmark.py feed_dict
# python -m cProfile -o session-run-benchmark-variable.prof session-run-benchmark.py variable
# pip install snakeviz
# snakeviz  session-run-benchmark-feed.prof
# snakeviz  session-run-benchmark.prof
#
#
# Feed_dict: 147 usec, no feed dict, 71 usec

yaroslavvb

# GPU picking
# http://stackoverflow.com/a/41638727/419116
# Nvidia-smi GPU memory parsing.
# must set
# CUDA_DEVICE_ORDER=PCI_BUS_ID
# see https://github.com/tensorflow/tensorflow/issues/152#issuecomment-273663277

# Tested on nvidia-smi 370.23

def run_command(cmd):