Skip to main content

A Simple and Scalable Logging Utility With a Beautiful Visualization Dashboard

Project description


ML-Logger makes it easy to:

  • save data locally and remotely, as binary, in a transparent pickle file, with the same API and zero configuration.
  • write from 500+ worker containers to a single instrumentation server
  • visualize matplotlib.pyplot figures from a remote server locally with logger.savefig('my_figure.png?raw=true')

And ml-logger does all of these with minimal configuration — you can use the same logging code-block both locally and remotely with no code-block change.

ML-logger is highly performant – the remote writes are asynchronous. For this reason it doesn’t slow down your training even with 100+ metric keys.

Why did we built this, you might ask? Because we want to make it easy for people in ML to use the same logging code-block in all of they projects, so that it is easy to get started with someone else’s baseline.

Getting Started!

To install ml_logger, do:

pip install ml-logger

Now you can rock!

from ml_logger import logger
# ~> logging data to /tmp/ml-logger-debug

Log key/value pairs, and metrics:

for i in range(1):
    logger.log(metrics={'some_val/smooth': 10, 'status': f"step ({i})"}, reward=20, timestep=i)
    ## flush the data, otherwise the value would be overwritten with new values in the next iteration.
# outputs ~>
# ╒════════════════════╤════════════════════════════╕
# │       reward       │             20             │
# ├────────────────────┼────────────────────────────┤
# │      timestep      │             0              │
# ├────────────────────┼────────────────────────────┤
# │  some val/smooth   │             10             │
# ├────────────────────┼────────────────────────────┤
# │       status       │          step (0)          │
# ├────────────────────┼────────────────────────────┤
# │      timestamp     │'2018-11-04T11:37:03.324824'│
# ╘════════════════════╧════════════════════════════╛

Logging to a Server

Skip this if you just want to log locally. When training in parallel, you want to kickstart an logging server (Instrument Server). To do so, run:

python -m ml_logger.server --log-dir /home/yourname/ml-logger-debug --host --port 8081

Use ssh tunnel if you are running on a managed cluster.

Allowing Non-local Requests

The default host is set to This would prevent external requests from being accepted. To allow requests from a non-localhost client, set host to

Full Readme on the Server

ml-logger uses params-proto to declaratively define the cli interface. To view the help document, you can simply type

python -m ml_logger.server --help

Asynchronously log the summary of LOTs of training metrics

A common scenario is you only want to upload averaged statistics of your metrics. A pattern that @jachiam uses is the following: store_metrics(), peak_stored_metrics(), and log_metrics_summary()

# You log lots of metrics during training.
for i in range(100):
    logger.store_metrics(metrics={'some_val/smooth': 10}, some=20, timestep=i)
# you can peak what's inside the cache and print out a table like this:
# outputs ~>
#      some      |   timestep    |some_val/smooth
# ━━━━━━━━━━━━━━━┿━━━━━━━━━━━━━━━┿━━━━━━━━━━━━━━━
#       20       |       0       |      10
#       20       |       1       |      10
#       20       |       2       |      10
#       20       |       3       |      10

# The metrics are stored in-memory. Now we need to actually log the summaries:
# outputs ~> . (data is now logged to the server)

Table of Contents

  • logging matplotlib.pyplot figures on an headless server
  • [documentation under construction]

How to Develop

First clone repo, install dev dependencies, and install the module under evaluation mode.

git clone
cd ml_logger && cd ml_logger && pip install -r requirements-dev.txt
pip install -e .

Testing local-mode (without a server)

You should be inside ml_logger/ml_logger folder

pwd # ~> ml_logger/ml_logger
make test

Testing with a server (You need to do both for an PR)

To test with a live server, first run (in a separate console)

python -m ml_logger.server --log-dir /tmp/ml-logger-debug

or do:

make start-test-server

Then run this test script with the option:

python -m pytest tests --capture=no --log-dir

or do

make test-with-server

Your PR should have both of these two tests working. ToDo: add CI to this repo.

To Publish

You need twine, rst-lint etc, which are included in the requirements-dev.txt file.

Logging Matplotlib pyplots

Configuring The Experiment Folder

from ml_logger import logger, Color, percent
from datetime import datetime

now =
logger.configure(log_directory="/tmp/ml-logger-demo", f"deep_Q_learning/{now:%Y%m%d-%H%M%S}")

This is a singleton pattern similar to matplotlib.pyplot. However, you could also use the logger constructor

from ml_logger import ML_Logger

logger = ML_Logger(root_dir="/tmp/ml-logger-demo", f"deep_Q_learning/{now:%Y%m%d-%H%M%S}")

Logging Text, and Metrics

logger.log({"some_var/smooth": 10}, some=Color(0.85, 'yellow', percent), step=3)

colored output: (where the values are yellow)

│  some var/smooth   │         10         │
│        some        │       85.0%        │

Logging Matplotlib Figures

We have optimized ML-Logger, so it supports any format that pyplot supports. To save a figure locally or remotely,

import matplotlib.pyplot as plt
import numpy as np

xs = np.linspace(-5, 5)

plt.plot(xs, np.cos(xs), label='Cosine Func')

Logging Videos

It is especially hard to visualize RL training sessions on a remote computer. With ML-Logger this is easy, and super fast. We optimized the serialization and transport process, so that a large stack of video tensor gets first compressed by ffmepg before getting sent over the wire.

The compression rate (and speed boost) can be 2000:1.

import numpy as np

def im(x, y):
    canvas = np.zeros((200, 200))
    for i in range(200):
        for j in range(200):
            if x - 5 < i < x + 5 and y - 5 < j < y + 5:
                canvas[i, j] = 1
    return canvas

frames = [im(100 + i, 80) for i in range(20)]

logger.log_video(frames, "test_video.mp4")

Saving PyTorch Modules

PyTorch has a very nice module saving and loading API that has inspired the one in Keras. We make it easy to save this state dictionary (state_dict) to a server, and load it. This way you can load from 100+ of your previous experiments, without having to download those weights to your code-block repository.

# save a module

# load a module
state_dict, = logger.load_pkl(f"modules/{0:04d}_Test.pkl")

Saving Tensorflow Models

The format tensorflow uses to save the models is opaque. I prefer to save model weights in pickle as a dictionary. This way the weight files are transparent. ML_Logger offers easy helper functions to save and load from checkpoints saved in this format:

## To save checkpoint
from ml_logger import logger
import tensorflow as tf


x = tf.get_variable('x', shape=[], initializer=tf.constant_initializer(0.0))
y = tf.get_variable('y', shape=[], initializer=tf.constant_initializer(10.0))
c = tf.Variable(1000)

sess = tf.InteractiveSession()

trainables = tf.trainable_variables()
logger.save_variables(trainables, path="variables.pkl", namespace="checkpoints")

which creates a file checkpoints/variables.pkl under /tmp/ml-logger-demos.


An idea visualization dashboard would be 1. Fast, instantaneous. On an AWS headless server? View the plots as if they are on your local computer. 2. Searchable, performantly. So that you don’t have to remember where an experiment is from last week. 3. Answer Questions, from 100+ Experiments. We make available Google’s internal hyperparameter visualization tool, on your own computer.

Searching for Hyper Parameters

Experiments are identified by the metrics.pkl file. You can log multiple times to the same metrics.pkl file, and the later parameter values overwrites earlier ones with the same key. We enforce namespace in this file, so each key/value argument you pass into the logger.log_parameters function call has to be a dictionary.

Args = dict(

How to launch the Vis App

This requires node.js and yarn dev environment at the moment. We will streamline this process without these requirements soon.

  1. download this repository
  2. go to ml-vis-app folder
  3. Install the dev dependencies
    1. install node: Installation
    2. install yarn: Installation
    3. install the dependencies of this visualization app:
      1. yarn install
  4. in that folder, run yarn.

The IP address of the server is currently hard code-blockdhere. To use this with your own instrumentation server, over-write this line. I’m planning on making this configuration more accessible.

Full Logging API

from ml_logger import logger, Color, percent

logger.log(some=Color(0.1, 'yellow'), step=0)
logger.log(some=Color(0.28571, 'yellow', lambda v: "{:.5f}%".format(v * 100)), step=1)
logger.log(some=Color(0.85, 'yellow', percent), step=2)
logger.log({"some_var/smooth": 10}, some=Color(0.85, 'yellow', percent), step=3)
logger.log(some=Color(10, 'yellow'), step=4)

colored output: (where the values are yellow)

│        some        │        0.1         │
│        some        │     28.57100%      │
│        some        │       85.0%        │
│  some var/smooth   │         10         │
│        some        │       85.0%        │

In your project files, do:

from params_proto import cli_parse
from ml_logger import logger

class Args:
    seed = 1
    D_lr = 5e-4
    G_lr = 1e-4
    Q_lr = 1e-4
    T_lr = 1e-4
    plot_interval = 10
    log_dir = ""
    log_prefix = ""

logger.configure(log_directory="", prefix="your-experiment-prefix!")

for epoch in range(10):
    logger.log(step=epoch, D_loss=0.2, G_loss=0.1, mutual_information=0.01)
    logger.log_key_value(epoch, 'some string key', 0.0012)
    # when the step index updates, logger flushes all of the key-value pairs to file system/logging server


# Images
face = scipy.misc.face()
face_bw = scipy.misc.face(gray=True)
logger.log_image(index=4, color_image=face, black_white=face_bw)
image_bw = np.zeros((64, 64, 1))
image_bw_2 = scipy.misc.face(gray=True)[::4, ::4]

logger.log_image(i, animation=[face] * 5)

This version of logger also prints out a tabular printout of the data you are logging to your stdout. - can silence stdout per key (per logger.log call) - can print with color: logger.log(timestep, some_key=green(some_data)) - can print with custom formatting: logger.log(timestep, some_key=green(some_data, percent)) where percent - uses the correct unix table characters (please stop using | and +. Use ``│``, ``┼`` instead)

A typical print out of this logger look like the following:

from ml_logger import ML_Logger

logger = ML_Logger(root_dir=f"/mnt/bucket/deep_Q_learning/{}")

logger.log_params(G=vars(G), RUN=vars(RUN), Reporting=vars(Reporting))

outputs the following

           env_name            │ MountainCar-v0
             seed              │ None
      stochastic_action        │ True
         conv_params           │ None
         value_params          │ (64,)
        use_layer_norm         │ True
         buffer_size           │ 50000
      replay_batch_size        │ 32
      prioritized_replay       │ True
            alpha              │ 0.6
          beta_start           │ 0.4
           beta_end            │ 1.0
    prioritized_replay_eps     │ 1e-06
      grad_norm_clipping       │ 10
           double_q            │ True
         use_dueling           │ False
     exploration_fraction      │ 0.1
          final_eps            │ 0.1
         n_timesteps           │ 100000
        learning_rate          │ 0.001
            gamma              │ 1.0
        learning_start         │ 1000
        learn_interval         │ 1
target_network_update_interval │ 500
        log_directory          │ /mnt/slab/krypton/machine_learning/ge_dqn/2017-11-20/162048.353909-MountainCar-v0-prioritized_replay(True)
          checkpoint           │ checkpoint.cp
           log_file            │ output.log
     checkpoint_interval       │ 10000
        reward_average         │ 100
        print_interval         │ 10
│      timestep      │        1999        │
│      episode       │         10         │
│    total reward    │       -200.0       │
│ total reward/mean  │       -200.0       │
│  total reward/max  │       -200.0       │
│time spent exploring│       82.0%        │
│    replay beta     │        0.41        │


Visualization (Preview):boom:

In addition, ml-logger also comes with a powerful visualization dashboard that beats tensorboard in every aspect.

ml visualization dashboard

ml visualization dashboard

An Example Log from ML-Logger

A common pain that comes after getting to launch ML training jobs on AWS is a lack of a good way to manage and visualize your data. So far, a common practice is to upload your experiment data to aws s3 or google cloud buckets. Then one quickly realizes that downloading data from s3 can be slow. s3 does

not offer diffsync like gcloud-cli’s g rsync. This makes it hard to sync a large collection of data that is constantly appended to.

So far the best way we have found for organizing experimental data is to have a centralized instrumentation server. Compared with managing your data on S3, a centralized instrumentation server makes it much easier to move experiments around, run analysis that is co-located with your data, and hosting visualization dashboards on the same machine. To download data locally, you can use sshfs, smba, rsync or a variety of remote disks. All faster than s3.

ML-Logger is the logging utility that allows you to do this. To make ML_logger easy to use, we made it so that you can use ml-logger with zero configuration, logging to your local hard-drive by default. When the logging directory field logger.configure(log_directory= <your directory>) is an http end point, the logger will instantiate a fast, future based logging client that launches http requests in a separate thread. We optimized the client so that it won’t slow down your training code-block.

API wise, ML-logger makes it easy for you to log textual printouts, simple scalars, numpy tensors, image tensors, and pyplot figures. Because you might also want to read data from the instrumentation server, we also made it possible to load numpy, pickle, text and binary files remotely.

In the future, we will start building an integrated dashboard with fast search, live figure update and markdown-based reporting/dashboarding to go with ml-logger.

Now give this a try, and profit!

Project details

Release history Release notifications | RSS feed

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for ml-logger, version 0.6.17
Filename, size File type Python version Upload date Hashes
Filename, size ml_logger-0.6.17-py3-none-any.whl (51.2 kB) File type Wheel Python version py3 Upload date Hashes View

Supported by

Pingdom Pingdom Monitoring Google Google Object Storage and Download Analytics Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN DigiCert DigiCert EV certificate StatusPage StatusPage Status page