dora commanddora run: Running XP locallydora launch: Launching XP remotelydora info: Inspecting an XPdora grid: Managing a grid search# For bleeding edge
pip install -U git+https://github.com/facebookincubator/submitit@main#egg=submitit
pip install -U git+https://git@github.com/facebookresearch/dora#egg=dora-search
# For stable release
pip install -U dora-search
See the changelog for details on releases.
git_save option. This will ensure that the project git is clean
and make a clone from which the experiment will run. This does not apply to dora run for easier
debugging (but you can force it with --git_save).(FB Only) If you are using Dora and want to receive updates on bug fixes and new versions, ping me (@defossez) on Workchat.
Dora is an experiment launching tool which provides the following features:
Some Dora concepts:
explorer function, wrapped in a dora.Explorer.
The explorer function takes a dora.Launcher as argument. Call repeatidly
the dora.Launcher with a set of
hyper-parameters to schedule different experiments.In order to derive the XP signature, Dora must know about the configuration schema your project is following, as well as
the parsed arguments for a run.
Dora supports two backends for that : argparse, and hydra. On top of that, Dora provides a smooth integration
with Pytorch Lightning for projects that uses it.
In all cases, you must have a specific python package (which we will call here myproj),
with a train module in it, (i.e. myproj.train module, stored in the myproj/train.py file.)
The train.py file must contain a main function that is properly decorated, as explained hereafter.
Here is a template for the train.py file:
import argparse
from dora import argparse_main, get_xp
parser = argparse.ArgumentParser("mycode.train")
...
@argparse_main(
dir="./where_to_store_logs_and_checkpoints",
parser=parser,
exclude=["list_of_args_to_ignore_in_signature, e.g.", "num_workers",
"can_be_pattern_*", "log_*"],
use_underscore=True, # flags are --batch_size vs. --batch-size
git_save=False, # if True, scheduled experiments will run from a separate clone of the repo.
)
def main():
# No need to reparse args, you can directly access them from the current XP
# object.
xp = get_xp()
xp.cfg # parsed arguments
xp.sig # signature for the current run
xp.folder # folder for the current run, please put your checkpoint relative
# to this folder, so that it is automatically resumed!
xp.link # link object, can send back metrics to Dora
# If you load a previous checkpoint, you should always make sure
# That the Dora Link is consistent with what is in the checkpoint with
# history = checkpoint['history']
# xp.link.update_history(history)
for t in range(10):
xp.link.push_metrics({"loss": 1/(t + 1)})
...
The template for train.py:
from dora import hydra_main, get_xp
@hydra_main(
config_path="./conf", # path where the config is stored, relative to the parent of `mycode`.
config_name="config" # a file `config.yaml` should exist there.
)
def main(cfg):
xp = get_xp()
xp.cfg # parsed configuration
xp.sig # signature for the current run
# Hydra run folder will automatically be set to xp.folder!
xp.link # link object, can send back metrics to Dora
# If you load a previous checkpoint, you should always make sure
# That the Dora Link is consistent with what is in the checkpoint with
# history = checkpoint['history']
# xp.link.update_history(history)
for t in range(10):
xp.link.push_metrics({"loss": 1/(t + 1)})
...
You can customize dora behavior from the config.yaml file, e.g.
my_config: plop
num_workers: 40
logs:
interval: 10
level: info
dora:
exclude: ["num_workers", "logs.*"]
dir: "./outputs"
git_save: true # set git_save option for the project.
Dora supports PyTorch Lightning (PL) out of the box. Dora will automatically
capture logged metrics (make sure to use per_epoch=True), and handles distribution
(you should not pass gpus=... or num_nodes=... to PL).
import dora.lightning
@dora.argparse_main(...)
def main():
xp = dora.get_xp()
args = xp.cfg
# Replace Pytorch lightning `Trainer(...)` with the following:
trainer = dora.lightning.get_trainer(...)
# Or when using argparse parsing:
trainer = dora.lightning.trainer_from_argparse_args(args)
See examples/pl/train.py for a full example including automatic reloading of the last checkpoint, logging etc.
Important: Dora deactivates the default PL behavior of dumping a mid-epoch
checkpoint upon preemption, as this lead to non deterministic behavior
(as PL would skip this epoch upon restart). Dora assumes you save checkpoints
from time to time (e.g. every epoch). To get back the old behavior,
pass no_unfinished_epochs=False to get_trainer. See examples/pl/train.py
for an example of how to implement checkpointing in a reliable manner.
Dora supports distributed training, and makes a few assumptions for you. You should initialize distributed training through Dora, by calling in your main function:
import dora.distrib
dora.distrib.init()
Note: This is not required for Pytorch Lightning users, see the PL section hereafter, everything will be setup automatically for you :)
You can set the git_save option on your project, see hereafter on how to do it for either argparse or Hydra
based projects.
When this option is set, Dora makes individual clones of your project repository for each experiment that is scheduled.
The job will then run from that clean clone. This allows both to keep track of the exact
code that was used for an experiment, as well as preventing code changes to impact pending, or requeued
jobs.
If you reschedule a failed or cancelled job, the clone will however be updated with the current code.
In order to use this option, your code should be able to run from a fresh clone of the repository.
If you need to access to resources that are specified with a path relative to the original
repo, use dora.to_absolute_path(). Note that this is similar to hydra.utils.to_absolute_path(). In fact, you can safely replace the Hydra version with this one,
as even when git_save is not set, the Dora one automatically falls back to the Hydra one (if Hydra is used).
The repository must be completely clean before scheduling remote jobs, and all files should be either
tracked or git ignored. This is very restricive, but this makes implementing this feature
much simpler and safe. Also this forces good practice :)
Only the dora run command can be used on a dirty repository, to allow
for easy debugging. For the dora launch and dora grid command, you can also use the --no_git_save
option to temporarily deactivate this feature.
The clone for each experiment is located inside the code/ subfolder inside the XP folder (which you can get with the dora info command for instance).
dora commandDora will install a dora command that is the main way to interact with it.
The dora command defines 4 sub-commands, detailed in the following sections:
dora run: run training code locally (e.g. for debugging).dora launch: launch remote jobs, useful for one-off experiments.dora info: get information on a specific job/XP, logs etc.dora grid: launch an entire grid search defined in a grid file. Only missing XP will be scheduled.
Will also reports status and latest metrics.In order for Dora to find your code, you must pass your training package
(i.e. mycode) as dora -P mycode [run|launch|grid|info].
This flag can be skipped if mycode is in the current working directory and is the only folder with a train.py file in it, in which
case Dora will find it automatically.
You can also export DORA_PACKAGE=mycode to avoid having to give the -P flag explicitely.
You can use a different script name than train.py with -M, --main_module,
or setting the DORA_MAIN_MODULE env variable.
Attention: those flags should be specified BEFORE the run | launch |info |grid part: dora -P mycode run, not dora run -P mycode.
See the examples folder for a few examples using argparse, Hydra
and Pytorch Lightning, in order to test the commands described here.
To play with them, first install Dora (pip install . from the top-level of the repo), then cd examples, and use dora -P example_name ...
to let Dora know which example to use!
dora run: Running XP locallyYou can run an XP locally with
dora run [TRAINING_ARGS ...]
Warning: for the argparse backend, you must insert -- between the dora args and your own training args, i.e.:
dora run -- [TRAINING_ARGS ...]
dora run supports two flags:
-d: distributed training using all available gpus. The master worker output will be to the shell, and other workers will be redirected to a log file in the XP folder.-f sig: this will inject the hyper-parameters from the XP with the given sig on top of the one provided on the command line. Useful to resume locally a remote job that failed.--git_save: clone the repo inside the XP folder and execute from there. This is mostly for debugging,
and in general is not needed.dora launch: Launching XP remotelyDora supports scheduling experiments on Slurm. If you need to schedule many of them, then a grid file is properly better.
dora launch [--dev] [-g NUMBER_OF_GPUS] [TRAINING_ARGS ...]
Dora will automatically select the appropriate number of nodes and tasks per nodes based on the number of GPUs required, as well as scale required memory.
For instance, if -g 16, Dora will schedule on 2 nodes with 8 gpus each.
This command will launch the command, and immediately tail its log and monitor its progress, just like if it were running in locally.
If you want to kill the command if you kill the local process, you can add the -a, --attach flag.
To avoid tailing the log, just pass --no_tail.
If a job already exist for the given XP, Dora will not schedule a second one, but reuse the existing job.
If a previous run has failed or was canceled, Dora will not automatically start a new one, to give you a chance to inspect the logs.
If you want to reschedule a run, use the -r, --retry flag.
Other flags:
-f SIG: injects the arguments from the XP matching this signature, on top of the one provided on the command line.-R, --replace: replace any running job (i.e. cancels, and schedules a new one).-D, --replace_done: also reschedule a job even if a previous one completed successfully.-p, --partition PARTITION: partition to use.-c, --comment COMMENT: comment for the job (e.g. if priority is used).--clear: cancel any previous job, clear the XP folder (i.e. delete checkpoints) and reschedule.dora info: Inspecting an XPYou can get information on an XP with the dora info command:
dora info [TRAINING_ARGS ...]
dora info -f SIGNATURE
dora info -j SLURM_ID
You can either specify the XP by listing all of its training arguments, by passing its signature, or even the latest Slurm id associated with it. The info command supports a number of flags:
-l: print the entire log for the main task (this only work for remote jobs, not XP ran locally with dora run)-t: tail the log for the main task.dora grid: Managing a grid searchThe main benefit from Dora is the ability to handle arbitarily complex grid searches.
Each grid is defined by a grid file, inside a grids package (i.e. mycode.grids.my_grid).
The grid file defines an explorer function, decorated by an Explorer class.
The Explorer class defines various metadata, in particular on which metrics
to display when calling the grid command.
The explorer function takes a dora.Launcher as an argument, and
should repeatidly call it to schedule experiments.
Here is an example of grid search file, for instance mycode.grids.mygrid.
from itertools import product
from dora import Explorer, Launcher
@Explorer
def explorer(launcher: Launcher):
launcher(batch_size=128) # Schedule an experiments with the given batch size.
# For an argparse based project, this will get converted to the `--batch_size=128` flag
# You can pass `use_underscore=False` to `argparse_main` to get instead `--batch-size=128`.
sub = launcher.bind(lr=0.01) # bind some parameter value, in a new launcher
sub.slurm_(gpus=8) # all jobs scheduled with `sub` will use 8 gpus.
sub() # Job with lr=0.01 and 8 gpus.
sub.bind_(epochs=40) # in-place version of bind()
sub.slurm(partition="dev")(batch_size=64) # lr=0.01, 8 gpus, dev, bs=64 and epochs=40.
launcher.slurm_(gpus=16) # Now using 16 gpus per job, i.e. 2 full nodes.
# Nice thing of native python, you can define arbitrary set of XP!
for lr, bs in product([0.1, 0.01, 0.001], [16, 32, 64]):
if bs > 32 and lr < 0.01:
# this is just too extreme, let's skip
continue
launcher(lr=lr, batch_size=bs)
# Job arrays are also supported.
# The only limitation is that all jobs in an array must use exactly
# the same slurm config.
with launcher.job_array():
for seed in range(1, 100):
launcher(seed=seed)
You can then call
dora grid mygrid
This will do 3 thing:
explorer function will be scheduled, if not already running
or completed.Here is a more comprehensive description of what Launcher object can do.
launcher.bind_(...): remember the given parameters (command line option for argparse based
project, or overrides for Hydra based ones) for future scheduling, i.e. all experiments
later scheduled with that launcher will have those parameters set.sub = launcher.bind(...): same as bind, but returns a new "sub" launcher, i.e. the object
launcher is not changed, only experiments scheduled with sub will use the given params.
sub also inherits from all the params already bound to its parent launcher (i.e. previous call to launcher.bind_).
Creating a sub-launcher is especially recommended inside loops, to avoid leaking params to the next loop iteration.launcher(...): schedules an experiment with the given params, plus all the ones that have
been aggregated through the various calls to bind_ and to bind. This is equivalent to
launcher.bind(...)().launcher.slurm_(key=value, ...) and launcher.slurm(key=value, ...): same as bind_ and bind
but for the slurm config (nb of GPUs etc). For a list of possible options, checkout
SlurmConf.Now let us describe the format for passing parameters overrides or command line flags to
launcher.bind_(), launcher.bind() or launcher():
--batch_size flag, you can
do launcher.bind(batch_size=64).launcher.bind(['--lr=1e-4']).launcher.bind({'batch_size': 64}). Note that this
also allows for nested keys in Hydra: launcher.bind({'model.channels': 256}). With Hydra, you can
also define new keys with {'+model.activation': 'relu'}. You must not remove keys though.launcher.bind(['optim.lr=1e-4'], {'model.channels': 256, 'seed': 42}, {'+model.activation': 'relu'}, batch_size=64)
The dora grid command supports the following flags:
-r, --retry: failed or cancelled XP within one grid file will
be rescheduled.
-R, --replace: any running XP will be replaced by a new job.
-D, --replace_done: any XP in the grid that previously completed will be rescheduled.
-C, --cancel: cancel all XPs in a grid.
--clear: cancel any previous jobs, clear all XP folders (i.e. delete checkpoints) and reschedule. This will ask confirmation first, because this is quite dangerous.
-i, --interval INTERVAL: the table monitoring all jobs will be updated every INTERVAL
minutes, until all jobs are finished or failed.
-T, --trim IDX: trim all the metrics to the number of epochs of the XP
with the given index inside the grid, i.e. pretend that all XPs have at most
as many epochs as the XP with the given index.
-L, --trim_last: trim all XPs to the least advanced XP i.e. if the least
advanced XP has only 3 epochs, show the metrics at epoch 3 for all XPs.
-f, --folder IDX: only print the folder of the XP with the given idnex.
-l, --log IDX: print the full log of the XP with the given index.
-t, --tail IDX: tail the log of the XP with the given index.
--no_monitoring: only show the table once and return.
--dry_run: only simulate actions.
You can also pass patterns to the grid command, for instance
dora grid mygrid bs=64
will only show XPs which have bs=64 in their name. You can see the name by launching
the grid normally. Names are heavily shorten to avoid running out of space, in particular
nested structure will have all their components but the leaf be shorten. You
can negate a query with !, for instance dora grid mygrid '!bs=64' (use quotes
because ! will be interpreted by the shell otherwise).
Multiple patterns are interpreted as logical AND between them.
Note that with the latest version (be sure to update), the --clear, or -C, --cancel flags
will only apply to the XP matching the pattern. Similarly, only XP matching those patterns
will be scheduled.
The Explorer class allows to customize which metric to report, and with what precision. It also gives you a chance to reorganize metrics or further post process them (for instance, extracting max, min etc.).
See Customize metrics displayed hereafter for more explanation.
By convention, files starting with _ inside the grids folder are ignored by Dora, and are a good place to put utility code such as your custom Explorer classes.
For an example with detailed comments, go checkout the Explorer classes for BrainMagick.
Dora supports HiPlot out of the box. Make sure it is installed (pip install hiplot), then
you can run
python -m hiplot dora.hiplot.load --port=XXXX
In the prompt, you can type any number of grid names or XP sig, separated by spaces.
You can customize the metrics displayed by inheriting dora.hiplot.HiPlotExplorer
in a class inside yourproject.grids._hiplot. Then, select your explorer with
the command explorer=MyExplorer inside the HiPlot prompt (along the grid names and XP
sigs, in any order).
Dora provides some API, including the possibility to run grid searches directly from an IPython notebook. See the Dora API.
The most useful class is the DecoratedMain, which is the decorated main function in your project. You can use it to retrieve an XP object from a list of argv, or a signature:
from myproj.train import main
xp = main.get_xp_from_sig('ae43f645')
xp2 = main.get_xp_from_argv(xp.argv + ['batch_size=32'])
xp2_name = main.get_name(xp2) # See API in dora.names.NameMixin
with xp2.enter():
# You can pretend to be in an XP with this.
...
If you want to do some advance customization of the behavior of DecoratedMain
(e.g. custom naming for XP, or weird parsing of flags), feel free
to inherit dora.main.ArgparseMain or dora.hydra.HydraMain and use your
custom class instead.
You can schedule and manage grids from the Dora API rather than the command-line. This is useful to manage XPs from a notebook for instance! See the dora.grid.run_grid. Flags are passed with as an instance of dora.grid.RunGridArgs. Submission rules (e.g. cancel, retry etc.) are passed as a dora.conf.SubmitRules.
import dora
import dora.grid
from myproj.train import main
@dora.Explorer
def explorer(launcher):
launcher.slurm_(gpus=2, partition='learnlab')
launcher.bind_({
'epochs': 200,
'model.depth': 10,
'batch_size': 128
})
launcher()
for p in [0.005, 0.01, 0.05, 0.1]:
sub = launcher.bind({'task.penalty': p})
sub()
sub(lr=1e-4)
args = dora.grid.RunGridArgs(jupyter=True, monitor=False, interval=1)
rules = dora.conf.SubmitRules(retry=True) # Should we reschedule failed jobs?
# The run_grid function returns a list of sheeps
# each sheep as 2 attributues: sheep.xp and sheep.job_id.
sheeps = dora.grid.run_grid(main, explorer, grid_name='jupy', rules=rules, args=args)
args.monitor = True
args.jupyter = True
# The jupyter flag will make the grid API use the display API to clear the cell
# output and update it regularly. This one will not return until all jobs
# are done or failed.
# In the following, `grid_name` should be unique. It will be used
# to determine which experiments were previously scheduled with that grid
# and should potentially be cancelled if no longer needed.
dora.grid.run_grid(main, explorer, grid_name='jupy', rules=rules, args=args)
# You can retrieve the short names by using `main.get_names()`
short_names, ref_name = main.get_names([sheep.xp for sheep in sheeps])
At the moment, checkpoints and metrics cannot be directly shared (you can always copy manually an XP folder in someone else XPs folder). However, there are now two ways to share easily an XP hyper-params using its signature. This is useful if you want someone else to quickly reproduce your XP!
import/export commandGiven a list of signatures, you can export its hyper-params to a compact textual format with dora export:
dora export SIG1 [OTHER_SIG ...]
Copy paste the given string and your teammate can import it with
dora import
# now paste to stdin
The command will show you the list of imported XPs. Once an XP is imported, you can simply run it or query hyper params
with dora run -f SIG, dora info -f SIG etc. From a grid file, you can programmatically retrieve the hyper-params from that XP, e.g.
from myproject.train import main
xp = main.get_xp_from_sig(SIG)
launcher.bind_(xp.argv)
You can now configure a secondary shared XPs repository where only mappings from SIG -> hyper params are stored. With Hydra you can add
dora:
dir: outputs
shared: /shared_folder/shared_xps
Then other teammates can reference any SIG from an XP launched by other team members within the Dora commands.
Slurm configuration is detailed in dora/conf.py. It is a bit different from the usual Slurm config, as it tries to make it as easy as possible to change the number of GPUs without requiring to manually compute the number of nodes, tasks per nodes etc.
gpus (int): number of GPUs to schedule. Number of nodes
and tasks per nodes will be automatically inferred.mem_per_gpu (float): amount of memory in GB to schedule per gpus.time (int): maximum duration for the job in minutes.cpus_per_gpu (int): number of cpus per gpu, this will set
the cpus_per_task automatically, based on the
number of gpus and one_task_per_node, unless cpus_per_task
is explicitely provided.cpus_per_task (int or None): number of cpus per task.partition (str): partition name.comment (str): comment for the job.setup (List[str]): list of shell commands to execute
before the actual command. Use it for module load.max_num_timeout (int): maximum number of requeue.one_task_per_node (bool): if True, schedules a single task
per node, otherwise, will schedule one task per gpu (default is False).You can pass an instance of SlurmConfig to argparse_main that will be used as the default
config for all dora launch commands and grid files.
Grid files can override any field defined on SlurmConfig with the launcher.slurm (return new launcher) and
launcher.slurm_ (in-place) methods.
For Hydra, the default slurm configuration is taken from the slurm entry in the yaml file, for instance, you can
have their:
my_param: whatever
batch_size: 42
dora:
dir: outputs
slurm:
partition: devlab
mem_per_gpu: 30 # this is in GB
Metrics are formatted with the treetable, which is not heavily documented, but it should be easy enough to get how it works with this example:
from dora import Explorer
import treetable as tt
class MyExplorer(Explorer):
test_metrics = ['sisnr', 'pesq']
def get_grid_metrics(self):
"""Return the structure of the metrics that should be displayed in the tracking table.
"""
# This will return a list of `tt.group`, each group will
# be in separate parts of the table.
return [
tt.group("train", [
tt.leaf("epoch"),
tt.leaf("loss", ".3f"), # The second argument of tt.leaf is a formatting string.
], align=">"), # Align can be left (<) or right (>) and will apply on all
# leaves inside the group.
tt.group("valid", [
tt.leaf("best", ".3f"),
tt.leaf("loss", ".3f"),
], align=">"),
tt.group("test", [
tt.leaf(name, ".3f")
for name in self.test_metrics
], align=">")
]
# In practice you can even have deeply nested groups etc. but honestly
# you probably don't need that.
def process_history(self, history):
"""This process the history obtained from the Dora Link
into the right format for the `get_grid_metrics()` layout.
This should return a dictionary, with one key per group, each
being a sub-dict with one key per metric.
It is fine for a key to be absent, things won't crash, and it will
just not show anything there in the table.
"""
train = {
'epoch': len(history),
}
valid = {}
test = {}
best = float('inf')
for metrics in history:
train.update(metrics['train'])
valid.update(metrics['valid'])
# Let say you forgot to compute best valid loss, you can
# fill in it here on the fly.
best = min(valid['loss'], best)
valid['best'] = best
# The reason for having this for loop is also if some metrics
# are not defined with every epoch. Let say you compute test metrics
# only every 10 epochs, then this will automatically fill the
# test metrics of the last epoch which evaluated them.
if 'test' in metrics:
test.update(metrics['test'])
return {"train": train, "valid": valid, "test": test}
Dora is released under the MIT license as found in the LICENSE file.
Before submitting any change, please run make to run unit tests and code linting.
FAQs
Easy grid searches for ML.
We found that dora-search demonstrated a healthy version release cadence and project activity because the last version was released less than a year ago. It has 1 open source maintainer collaborating on the project.
Did you know?
Socket for GitHub automatically highlights issues in each pull request and monitors the health of all your open source dependencies. Discover the contents of your packages and block harmful activity before you install or update your dependencies.
Security News
Research
The Socket Research Team breaks down a malicious wrapper package that uses obfuscation to harvest credentials and exfiltrate sensitive data.
Research
Security News
Attackers used a malicious npm package typosquatting a popular ESLint plugin to steal sensitive data, execute commands, and exploit developer systems.
Security News
The Ultralytics' PyPI Package was compromised four times in one weekend through GitHub Actions cache poisoning and failure to rotate previously compromised API tokens.