def upload_code():
code_package, files_uploaded = upload_code_package(
".",
excludes=[
"core",
"compression_results",
"output.tmp",
".vscode",
"node_modules",
"scripts",
".git",
"bla.tar",
"*.pyc",
"._*",
"__pycache__",
"*.pdf",
"*.js",
"*.yaml",
".pylintrc",
".gitignore",
".AppleDouble",
".jobignore",
],
)
print("Uploaded {} files.".format(len(files_uploaded)))
return code_package
project = "decentralized_powersgd"
experiment = os.path.splitext(os.path.basename(__file__))[0]
script = "train.py"
description = """
See if we can reproduce Moniqua's results
""".strip()
base_config = {
"n_workers": 8,
"topology": "ring",
"batch_size": 128,
"task_architecture": "ResNet20",
}
code_package, files_uploaded = upload_code_package(".",
excludes=excluded_files +
["gossip_run.py"])
print("Uploaded {} files.".format(len(files_uploaded)))
def schedule(name, config, skip_existing=False):
# Skip pre-existing entries
if (skip_existing and mongo.job.count_documents({
"project": project,
"job": name,
"experiment": experiment
}) > 0):
return
config = {**base_config, **config}
n_workers = config["n_workers"]
job_id = register_job(
".gitignore",
".AppleDouble",
".jobignore",
]
project = "decentralized_powersgd"
experiment = os.path.splitext(os.path.basename(__file__))[0]
script = "gossip_run.py"
description = """
Now using optimal scaling for the uncompressed baseline and PowerGossip.
""".strip()
code_package, files_uploaded = upload_code_package(
".", excludes=excluded_files + ["tasks", "train.py"]
)
print("Uploaded {} files.".format(len(files_uploaded)))
def schedule(name, config):
# Skip pre-existing entries
if mongo.job.count_documents({"project": project, "job": name, "experiment": experiment}) > 0:
return
job_id = register_job(
user="******",
project=project,
experiment=experiment,
job=name,
n_workers=n_workers,
base_config = dict(
task="LSTM",
batch_size=64, # per worker
weight_decay=0,
num_epochs=90,
lr_schedule_milestones=[(60, 0.1), (80, 0.1)],
learning_rate=1.25,
momentum=0.0,
optimizer_diffusion_rate=1,
topology="ring",
use_decentralized_data=True,
n_workers=n_workers,
evaluate_average_model=False,
)
code_package, files_uploaded = upload_code_package(".",
excludes=excluded_files)
print("Uploaded {} files.".format(len(files_uploaded)))
ids = []
def schedule(name, config, skip_existing=True):
# Skip pre-existing entries
config = {**base_config, **config}
if (skip_existing and mongo.job.count_documents(
{
"project": project,
"job": name,
"experiment": experiment,
"config.learning_rate": config["learning_rate"],
}) > 0):
