Skip to content

muqiaoy/mimir

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

54 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Mímir

When training machine learning models there are often many things we want to log: training error, validation error, gradient and weights norms, samples, etc. There are a few considerations:

  • For long-running experiments the log should be streamed to disk so that memory-usage doesn't grow.
  • The log should be stored in a format that is portable, easy to analyze, and space-efficient.
  • We want to be able to plot and analyze the log while the experiment is still running, potentially over the network.

Mímir stores logs as line-delimited JSON data and can stream them to disk as a gzipped files. It can also publish new entries over TCP sockets using ZeroMQ, enabling things such as live plotting.

Quickstart

To use Mímir, simply create a logger object:

import mimir
import time

logger = mimir.Logger()
for i in range(100):
    logger.log({'iteration': i, 'training_error': 1. / (i + 1)})
    time.sleep(1)

By default the log will just print the entry to standard output and then discard it. If you don't want to print anything, pass formatter=None, or pass a custom formatter to change the way the data is printed.

Accessing entries

If you want to keep entries in memory so that you can access past entries, pass a nonzero maxlen argument, which determines the maximum number of entries kept in memory. This is done so that long-running experiments don't run out of memory.

logger = mimir.Logger(maxlen=10)
logger.log({'iteration': 0, 'training_error': 10})
assert logger[-1]['training_error'] == 10

If you're sure that you won't run out of memory you can use maxlen=None for unlimited memory.

Saving to disk

We often want to save the log to disk to analyze it afterwards. Mímir allows you to save the log as line-delimited JSON files.

logger = mimir.Logger(filename='log.jsonl.gz')
for i in range(100):
    logger.log({'iteration': i, 'training_error': 1. / (i + 1)})
    time.sleep(1)

If the filename ends with .gz the log will be compressed in a streaming manner using gzlog.

Loading logs

If you want to load a log that was saved to disk so that its entries can be accessed in memory, use the load method. Any keyword arguments passed to this method will be passed on to json.loads, which can be useful for the deserialization of non-basic types. By default, NumPy objects are deserialized using mimir.serialization.deserialize_numpy.

logger = mimir.Logger('log.jsonl.gz')
logger.log({'iteration': 12})
logger.close()

new_logger = mimir.Logger('log.jsonl.gz', maxlen=10)
new_logger.load('log.jsonl.gz')
assert new_logger[-1]['iteration'] == 12

Streaming

Mímir can stream log entries over a TCP socket which clients can connect to, both locally as well as over a network. This allows you to do things like live-plotting your experiments. To enable this, pass stream=True. By default the data is streamed, which means that clients only get the entries from after when they joined. If you want clients to receive past log entries as well, there is a stream_maxlen argument similar to the maxlen argument.

logger = mimir.Logger(stream=True, stream_maxlen=50)
for i in range(100):
    logger.log({'iteration': i, 'training_error': 1. / (i + 1)})
    time.sleep(1)

To see a live plot of your log, open up a Jupyter notebook and type the following. Note that this requires a Bokeh server to be running (start it using the command bokeh serve). It will plot the last 50 datapoints, and then live plot every entry as it comes in.

import mimir.plot
mimir.plot.notebook_plot('iteration', 'training_error')

Context manager

The logger object can be used as a context manager, in which case all file objects are closed when the runtime context is exited.

with Logger(filename='log.jsonl') as logger:
    logger.log({'iteration': 0, 'training_error': 10})

Log analysis

To analyze the training logs jq is recommended. Most operations can be done easily on the command line.

# Get all training errors
cat log.jsonl | jq '.training_error'

# For compressed logs
gunzip -c log.jsonl.gz | jq '.training_error'

# Equivalently
zcat log.jsonl.gz | jq '.training_error'

To operate on the entire log as one array use the -s (slurp) flag.

cat log.json | jq -s 'min_by(.training_error)'

If your log entries have an irregular set of keys (e.g. if you only draw samples every n iterations) you use the select function to filter these out.

{"iteration": 0, "training_error": 1.2}
{"iteration": 1, "training_error": 0.7, "sample": 0.2}
{"iteration": 2, "training_error": 0.3}
cat log.jsonl | jq 'select(.sample)'

If you want to write the log back to a file after operating on it use the -c flag for compact output.

# Sorting the log by a timestamp
cat log.json | jq -s -c 'sort_by(.timestamp)[]' > sorted_log.json

# Subsampling the log
cat log.json | jq 'select(.iteration % 100 == 0).training_error' | less

JSON serialization

For streaming log entries over TCP sockets and saving logs to disk, Mímir uses JSON. To serialize non-basic types you need to pass a custom serialization function. Any keyword arguments passed to the Logger class will be passed to json.dumps. By default Mímir will pass default=serialize_numpy, which enables the serialization of NumPy arrays and scalars (numpy.ndarray and numpy.generic). Below is an example of how to go about serializing other objects:

import numpy
import mimir
from mimir.serialization import serialize_numpy, deserialize_numpy

def serialize_set(obj):
    if isinstance(obj, set):
        return tuple(obj)
    return serialize_numpy(obj)

logger = mimir.Logger(filename='log.jsonl.gz', default=serialize_set)
logger.log({'foo': set([1, 2]), 'bar': numpy.random.rand(10, 10)})

# In legacy Python use codecs.getreader('utf-8')(gzip.open(fn))
with gzip.open('log.jsonl.gz', 'rt') as f:
    entry = json.loads(f.readline(), obj_hook=deserialize_numpy)

About

JSON-based logging framework

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • C 65.8%
  • Assembly 8.0%
  • Ada 5.0%
  • Pascal 4.2%
  • C++ 3.7%
  • C# 3.1%
  • Other 10.2%