elements

🔥 Elements

Building blocks for productive research.

Installation

pip install elements

Features

Elements aims to provide well thought out solutions to common problems in research code. It is also hackable. If you need to change some of the code, we encourage you to fork the corresponding file into your project directory and make edits.

elements.Logger

A logger for array types that is extensible through backends. Metrics are written in a background thread to not block program execution, which is especially important on cloud services where bucket access is slow.

Provided backends:

• TerminalOutput(pattern): Print scalars to the terminal. Can filter to fewer metrics via regex.
• JSONLOutput(logdir, filename, pattern): Write scalars to JSONL files. For example, can be read directly with pandas.
• TensorBoardOutput(logdir): Scalars, histograms, images, GIFs. Automatically starts new event files when the current one exceeds a size limit to support cloud storage where appding to files requires a full download and reupload.
• WandBOutput(pattern, **init_kwargs): Strings, histograms, images, videos.
• MLFlowOutput(run_name, resume_id): Logs all types of metrics to MLFlow.

step = elements.Counter()
logger = elements.Logger(step, [
    elements.logger.TerminalOutput(),
    elements.logger.JSONLOutput(logdir, 'metrics.jsonl'),
    elements.logger.TensorBoardOutput(logdir),
    elements.logger.WandBOutput(name='name', project='project'),
])

step.increment()
logger.scalar('foo', 42)
logger.scalar('foo', 43)
logger.scalar('foo', 44)
logger.vector('vector', np.zeros(100))
logger.image('image', np.zeros((800, 600, 3, np.uint8)))
logger.video('video', np.zeros((100, 64, 64, 3, np.uint8)))

logger.add({'foo': 42, 'vector': np.zeros(100)}, prefix='scope')

logger.write()

elements.Config

An immutable nested directory to hold configurations. Keys can be accessed via attribute syntax. Values are restricted to primitive types that are supported by JSON. Types are checked when replacing values in the config.

config = elements.Config(
    logdir='path/to/dir',
    foo=dict(bar=42),
)

print(config)                      # Pretty printing
print(config.foo.bar)              # Attribute syntax
print(config['foo']['bar'])        # Dictionary syntax
config.logdir = 'path/to/new/dir'  # Not allowed

# Access a copy of the flattened dictionary underlying the config.
config.flat == {'logdir': 'path/to/dir', 'foo.bar': 42}

# Configs are immutable, so updating them returns a new object.
new_config = config.update({'foo.bar': 43})

# Types are enforced when updating configs, but values of other types are
# allowed as long as they can be converted without loss of information.
new_config = config.update({'foo.bar': float(1e5)})  # Allowed
new_config = config.update({'foo.bar': float(1.5)})  # Not allowed

# Configs can be saved and loaded in JSON and YAML formats.
config.save('config.json')
config = elements.Config.load('config.json')

elements.Flags

A parser for command line flags similar to argparse but faster to use and more flexible. Enforces types and supports nested dictionaries and overwriting multiple flags at once via regex.

A mapping of valid keys and their default values must be provided to infer types. Because there are defaults for all values, there are no required arguments that the user must specify on the command line.

# Create flags parser from default values.
flags = elements.Flags(logdir='path/to/dir', bar=42)

# Create flags parser from config.
flags = elements.Flags(elements.Config({
    'logdir': 'path/to/dir',
    'foo.bar': 42,
    'list': [1, 2, 3],
}))

# Load a config from YAML and overwrite it from it from the command line.
config = elements.Config.load('defaults.yaml')
config = elements.Flags(config).parse()

# Overwrite some of the keys.
config = flags.parse(['--logdir', 'path/to/new/dir', '--foo.bar', '43'])

# Supports syntax with space or equals.
config = flags.parse(['--logdir=path/to/new/dir'])

# Overwrite lists.
config = flags.parse(['--list', '10', '20', '30'])
config = flags.parse(['--list', '10,20,30'])
config = flags.parse(['--list=10,20,30'])

# Set all nested keys that end in 'bar'.
config = flags.parse(['--.*\.bar$', '43'])

# Parse only known flags.
config, remaining = flags.parse_known(['--logdir', 'dir', '--other', '123'])
remaining == ['--other', '123']

# Print a help page and terminate the program.
flags.parse(['--help'])

# Print a help page without terminating the program.
flags = elements.Flags(logdir='path/to/dir', bar=42, help_exits=False)
parsed, remaining = flags.parse_known(['--help', '--other=value'])
remaining == ['--help', '--other=value']
second_parser.parse(remaining)  # Now we exit.

elements.Path

A filesystem abstraction similar to pathlib that is extensible to new filesystems. Comes with support for local filesystems and GCS buckets.

path = elements.Path('gs://bucket/path/to/file.txt')

# String operations
path.parent                           # gs://bucket/path/to
path.name                             # file.txt
path.stem                             # file
path.suffix                           # .txt

# File operations
path.read(mode='r')                   # Content of the file as string
path.read(mode='rb')                  # Content of the file as bytes
path.write(content, mode='w')         # Write string to the file
path.write(content, mode='wb')        # Write bytes to the file
with path.open(mode='r') as f:        # Create a file pointer
  pass

# File system checks
path.parent.glob('*')                 # Get all sibling paths
path.exists()                         # True
path.isdir()                          # False
path.isfile()                         # True

# File system changes
(path.parent / 'foo').mkdir()         # Creates directory including parents
path.remove()                         # Deletes a file or empty directory
path.parent.rmtree()                  # Deletes directory and its content
path.copy(path.parent / 'copy.txt')   # Makes a copy
path.move(path.parent / 'moved.txt')  # Moves the file

elements.Checkpoint

Holds a collection of objects that can be saved to and loaded from disk.

Each object attached to the checkpoint needs to implement save() -> data and load(data) methods, where data is pickleable.

Checkpoints are written in a background thread to not block program execution. New checkpoints are writing to a temporary path first and renamed to the actual path once they are fully written, so that the path always points to a valid name even if the program gets terminated while writing.

step = elements.Counter()

cp = elements.Checkpoint(directory)
# Attach objects to the checkpoint.
cp.step = step
cp.model = model
# After attaching the objects we load the checkpoint from disk if it exists
# and otherwise save an initial checkpoint.
cp.load_or_save()

# Later on, we can change the objects and then save the checkpoint again.
should_save = elements.when.Every(10)
for _ in range(100):
  step.increment()
  if should_save(step):
    cp.save()

# We can also load checkpoints or parts of a checkpoint from a different directory.
cp.load(pretraining_directory, keys=['model'])
print(cp.model)

elements.Timer

Collect timing statistics about the run time of different parts of a program. Measures code inside sections and can wrap methods into sections. Returns execution count, execution time, fraction of overall program time, and more. The resulting statisticse can be added to the logger.

timer = Timer()

timer.section('foo'):
  time.sleep(10)

timer.wrap('name', obj, ['method1', 'method2'])
obj.method1()
obj.method1()
obj.method1()
obj.method2()

stats = timer.stats(reset=True, log=True)
stats == {
    'foo_count': 1,
    'foo_total': 10.0,
    'foo_avg': 10.0,
    'foo_min': 10.0,
    'foo_max': 10.0,
    'foo_frac': 0.92,
    'name.method1_count': 3,
    'name.method1_frac': 0.07,
    # ...
    'name.method2_frac': 0.01,
    # ...
}

elements.when

Helpers for running code at defined times, such as every fixed number of steps or seconds or a certain fraction of the time. The counting is robust, so when you skip a step it will run the code the next time to catch up.

should = elements.when.Every(100)
for step in range(1000):
  if should(step):
    print(step)  # 0, 100, 200, ...

should = elements.when.Ratio(0.3333)
for step in range(100):
  if should(step):
    print(step)  # 0, 4, 7, 10, 13, 16, ...

should = elements.when.Once()
for step in range(100):
  if should(step):
    print(step)  # 0

should = elements.when.Until(5)
for step in range(10):
  if should(step):
    print(step)  # 0, 1, 2, 3, 4

should = elements.when.Clock(1)
for step in range(100):
  if should(step):
    print(step)  # 0, 10, 20, 30, ...
  time.sleep(0.1)

elements.plotting

Tools for storing, loading, and plotting data with sensible defaults.

Data is stored in the run format in gzipped JSON files. Each file contains a list of one or more run. A run is a dictionary with the keys task, method, seed, xs, ys. The task, method, and seed are string fields, whereas xs and ys are lists of equal length containing numbers for the data to plot.

Take a look at plotting.py in the repository to see the list of all available functions, beyond what is used in this snippet.

from elements import plotting

runs = plotting.load('filename.json.gz')
plotting.dump(runs, 'filename.json.gz')

bins = np.linspace(0, 1e6, 100)
tensor, tasks, methods, seeds = plotting.tensor(runs, bins)
tensor.shape == (len(tasks), len(methods), len(seeds), len(bins))

fig, axes = plotting.plots(len(tasks))

for i, task in enumerate(tasks):
  ax = axes[i]
  for j, method in enumerate(methods):
    # Aggregate over seeds.
    mean = np.nanmean(tensor[i, j, :, :], 2)
    std = np.nanstd(tensor[i, j, :, :], 2)
    plotting.curve(ax, bins[1:], mean, std, label=method, order=j)

plotting.legend(fig, adjust=True)

# Saves the figure in both PNG and PDF formats and attempts to crop margins off
# the PDF.
plotting.save(fig, 'path/to/name')

Questions

Please file an issue on Github.