tf-benchmark.py

import datetime
import pytz
from tensorflow.contrib.session_bundle import exporter
from tensorflow.python.client import timeline
import glob
import json
import time
import math
import numpy as np
import os
import tensorflow as tf

steps_to_validate = 20
epoch_number = 2
thread_number = 2
batch_size = 100
min_after_dequeue = 1000
capacity = thread_number * batch_size + min_after_dequeue

filename = "20161227/*.zlib"

filename_queue = tf.train.string_input_producer(
      tf.train.match_filenames_once(filename),
      shuffle=True,
      seed = int(time.time()),
      num_epochs=epoch_number)

def read_and_decode(filename_queue):
    reader = tf.TFRecordReader(options = tf.python_io.TFRecordOptions(tf.python_io.TFRecordCompressionType.ZLIB))
    _, serialized_example = reader.read(filename_queue)
    return serialized_example

serialized_example = read_and_decode(filename_queue)
batch_serialized_example = tf.train.shuffle_batch(
    [serialized_example],
    batch_size=batch_size,
    num_threads=thread_number,
    capacity=capacity,
    min_after_dequeue=min_after_dequeue)
features = tf.parse_example(
    batch_serialized_example,
    features={
        "label": tf.FixedLenFeature([], tf.float32),
        "ids": tf.VarLenFeature(tf.int64),
        "values": tf.VarLenFeature(tf.float32),
    })

batch_labels = features["label"]
batch_ids = features["ids"]
batch_values = features["values"]

init_op = tf.global_variables_initializer()

with tf.Session() as sess:
    run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
    run_metadata = tf.RunMetadata()
    sess.run(init_op,options=run_options, run_metadata=run_metadata)
    sess.run(tf.local_variables_initializer(),options=run_options, run_metadata=run_metadata)

    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(coord=coord, sess=sess)
    start_time = datetime.datetime.now(pytz.timezone('Asia/Shanghai'))
    step = 1
    try:
        while not coord.should_stop():
            f1,f2,f3 = sess.run([batch_ids,batch_values,batch_labels],options=run_options, run_metadata=run_metadata)
            step +=1
            if step % steps_to_validate == 0:
                end_time = datetime.datetime.now(pytz.timezone('Asia/Shanghai'))
                sec = (end_time - start_time).total_seconds()
                print("[{}] time[{:6.2f}] step[{:10d}] speed[{:6d}]".format(
                    str(end_time).split(".")[0],sec, step,
                    int((steps_to_validate*batch_size)/sec)
                    ))
                start_time = end_time
            if step > 2000:
              break


    except tf.errors.OutOfRangeError:
        print("Done training after reading all data")
    finally:
        coord.request_stop()
        print("coord stopped")

    coord.join(threads)

    tl = timeline.Timeline(run_metadata.step_stats)
    ctf = tl.generate_chrome_trace_format()
    with open('timeline.json', 'w') as f:
           f.write(ctf)
    print "all done"

timeline.json

{
    "traceEvents": [
        {
            "ph": "M",
            "args": {
                "name": "Allocators"
            },
            "pid": 0,
            "name": "process_name"
        },
        {
            "ph": "M",
            "args": {
                "name": "/job:localhost/replica:0/task:0/cpu:0 Compute"
            },
            "pid": 1,
            "name": "process_name"
        },
        {
            "ph": "M",
            "args": {
                "name": "/job:localhost/replica:0/task:0/cpu:0 Tensors"
            },
            "pid": 2,
            "name": "process_name"
        },
        {
            "name": "NoOp",
            "args": {
                "name": "_SOURCE",
                "op": "NoOp"
            },
            "pid": 1,
            "ts": 1484406792814677,
            "cat": "Op",
            "tid": 0,
            "ph": "X",
            "dur": 4
        },
        {
            "name": "Const",
            "args": {
                "name": "shuffle_batch/n",
                "op": "Const"
            },
            "pid": 1,
            "ts": 1484406792814687,
            "cat": "Op",
            "tid": 0,
            "ph": "X",
            "dur": 6
        },
        {
            "name": "RandomShuffleQueue",
            "args": {
                "name": "shuffle_batch/random_shuffle_queue",
                "op": "RandomShuffleQueue"
            },
            "pid": 1,
            "ts": 1484406792814689,
            "cat": "Op",
            "tid": 1,
            "ph": "X",
            "dur": 7
        },
        {
            "name": "Const",
            "args": {
                "name": "ParseExample/Const",
                "op": "Const"
            },
            "pid": 1,
            "ts": 1484406792814695,
            "cat": "Op",
            "tid": 0,
            "ph": "X",
            "dur": 4
        },
        {
            "name": "Const",
            "args": {
                "name": "ParseExample/ParseExample/names",
                "op": "Const"
            },
            "pid": 1,
            "ts": 1484406792814701,
            "cat": "Op",
            "tid": 0,
            "ph": "X",
            "dur": 4
        },
        {
            "name": "Const",
            "args": {
                "name": "ParseExample/ParseExample/sparse_keys_0",
                "op": "Const"
            },
            "pid": 1,
            "ts": 1484406792814708,
            "cat": "Op",
            "tid": 0,
            "ph": "X",
            "dur": 4
        },
        {
            "name": "Const",
            "args": {
                "name": "ParseExample/ParseExample/sparse_keys_1",
                "op": "Const"
            },
            "pid": 1,
            "ts": 1484406792814714,
            "cat": "Op",
            "tid": 0,
            "ph": "X",
            "dur": 5
        },
        {
            "name": "Const",
            "args": {
                "name": "ParseExample/ParseExample/dense_keys_0",
                "op": "Const"
            },
            "pid": 1,
            "ts": 1484406792814721,
            "cat": "Op",
            "tid": 0,
            "ph": "X",
            "dur": 4
        },
        {
            "name": "QueueDequeueMany",
            "args": {
                "input0": "shuffle_batch/random_shuffle_queue",
                "input1": "shuffle_batch/n",
                "name": "shuffle_batch",
                "op": "QueueDequeueMany"
            },
            "pid": 1,
            "ts": 1484406792814699,
            "cat": "Op",
            "tid": 1,
            "ph": "X",
            "dur": 8889
        },
        {
            "name": "shuffle_batch/n",
            "pid": 1,
            "ts": 1484406792814693,
            "cat": "DataFlow",
            "tid": 0,
            "ph": "s",
            "id": 0
        },
        {
            "name": "shuffle_batch/n",
            "pid": 1,
            "ts": 1484406792814699,
            "cat": "DataFlow",
            "tid": 1,
            "ph": "t",
            "id": 0
        },
        {
            "name": "ParseExample",
            "args": {
                "input2": "ParseExample/ParseExample/sparse_keys_0",
                "input3": "ParseExample/ParseExample/sparse_keys_1",
                "input0": "shuffle_batch",
                "input1": "ParseExample/ParseExample/names",
                "name": "ParseExample/ParseExample",
                "input4": "ParseExample/ParseExample/dense_keys_0",
                "input5": "ParseExample/Const",
                "op": "ParseExample"
            },
            "pid": 1,
            "ts": 1484406792823598,
            "cat": "Op",
            "tid": 0,
            "ph": "X",
            "dur": 233
        },
        {
            "name": "shuffle_batch",
            "pid": 1,
            "ts": 1484406792823588,
            "cat": "DataFlow",
            "tid": 1,
            "ph": "s",
            "id": 1
        },
        {
            "name": "shuffle_batch",
            "pid": 1,
            "ts": 1484406792823598,
            "cat": "DataFlow",
            "tid": 0,
            "ph": "t",
            "id": 1
        }
    ]
}