def make_batch_script(trainer_params, model_params, script_params):
# Create LBANN objects
trainer = lbann.Trainer(mini_batch_size=trainer_params.mini_batch_size)
model = make_model(**model_params)
reader = make_data_reader()
# Optimizer with learning rate schedule
# Note: Rough approximation of
# embed_dim^-0.5 * min(step^-0.5, step*warmup^-1.5)
# with embed_dim=512 and warmup=4000.
opt = lbann.Adam(learn_rate=0.0001, beta1=0.9, beta2=0.98, eps=1e-9)
model.callbacks.append(
lbann.CallbackDropFixedLearningRate(
drop_epoch=[1],
amt=2,
))
model.callbacks.append(
lbann.CallbackDropFixedLearningRate(
drop_epoch=[2, 4, 8, 12],
amt=0.75,
))
# Checkpoint after every epoch
trainer.callbacks.append(
lbann.CallbackCheckpoint(
checkpoint_dir=os.path.join(script_params['work_dir'],
'checkpoint'),
checkpoint_epochs=1,
))
# Dump weights after every epoch
model.callbacks.append(
lbann.CallbackDumpWeights(
basename=os.path.join(script_params['work_dir'], 'weights'),
epoch_interval=1,
))
# Create Protobuf file
protobuf_file = os.path.join(script_params['work_dir'],
'experiment.prototext')
lbann.proto.save_prototext(
protobuf_file,
trainer=trainer,
model=model,
data_reader=reader,
optimizer=opt,
)
# Create batch script
script = lbann.contrib.launcher.make_batch_script(**script_params, )
script.add_command('echo "Started training at $(date)"')
script.add_parallel_command([
lbann.lbann_exe(),
f'--prototext={protobuf_file}',
])
script.add_command('status=$?')
script.add_command('echo "Finished training at $(date)"')
script.add_command('exit ${status}')
return script
def setup_experiment(lbann):
"""Construct LBANN experiment.
args:
lbann (module): Module for LBANN Python frontend
"""
trainer = lbann.Trainer(mini_batch_size=mini_batch_size)
callbacks = [
lbann.CallbackPrint(),
lbann.CallbackTimer(),
lbann.CallbackGPUMemoryUsage()
]
model = Sparse_Graph_Trainer.make_model(kernel_type='GatedGraph',
num_epochs=num_epochs,
callbacks=callbacks)
reader = data.PROTEINS.make_data_reader()
# No validation set
optimizer = lbann.Adam(learn_rate=0.01, beta1=0.9, beta2=0.99, eps=1e-8)
return trainer, model, reader, optimizer
def setup_experiment(lbann):
"""Construct LBANN experiment.
args:
lbann (module): Module for LBANN Python frontend
"""
trainer = lbann.Trainer(mini_batch_size=mini_batch_size)
model = make_model(NUM_NODES,
NUM_EDGES,
NUM_NODES_FEATURES,
NUM_EDGE_FEATURES,
EMBEDDING_DIM,
EDGE_EMBEDDING_DIM,
NUM_OUT_FEATURES,
num_epochs)
reader = LSC_PPQM4M.make_data_reader("LSC_100K",
validation_percent=0)
optimizer = lbann.Adam(learn_rate=0.01, beta1=0.9, beta2=0.99, eps=1e-8 )
return trainer, model, reader, optimizer
fldr_name=now.strftime('%Y%m%d_%H%M%S') ## time format
data_pct,val_ratio=1.0,0.1 # Percentage of data to use, % of data for validation
batchsize=args.batchsize
step_interval=args.step_interval
print('Step interval',step_interval)
work_dir="/global/cscratch1/sd/vpa/proj/cosmogan/results_dir/512square/{0}_bsize{1}_{2}".format(fldr_name,batchsize,args.suffix)
#####################
### Run lbann
trainer = lbann.Trainer(mini_batch_size=batchsize,random_seed=random_seed,callbacks=lbann.CallbackCheckpoint(checkpoint_dir='chkpt',
checkpoint_epochs=10))
# checkpoint_steps=step_interval))
spectral_loss=args.spec_loss
print("Spectral loss: ",spectral_loss)
model = construct_model(num_epochs,mcr,spectral_loss=spectral_loss,save_batch_interval=int(step_interval)) #'step_interval*val_ratio' is the step interval for validation set.
# Setup optimizer
opt = lbann.Adam(learn_rate=args.learn_rate,beta1=0.5,beta2=0.99,eps=1e-8)
# Load data reader from prototext
data_reader = construct_data_reader(data_pct,val_ratio)
status = lbann.run(trainer,model, data_reader, opt,
nodes=num_nodes, procs_per_node=num_procs,
work_dir=work_dir,
scheduler='slurm', time_limit=1440, setup_only=False)
print(status)
callbacks = [lbann.CallbackPrint(), lbann.CallbackTimer()]
layers = list(lbann.traverse_layer_graph([images, responses]))
model = lbann.Model(args.num_epochs,
layers=layers,
metrics=metrics,
objective_function=mse,
callbacks=callbacks)
# Load data reader from prototext
data_reader_proto = lbann.lbann_pb2.LbannPB()
with open(data_reader_prototext, 'r') as f:
txtf.Merge(f.read(), data_reader_proto)
data_reader_proto = data_reader_proto.data_reader
# Setup trainer
trainer = lbann.Trainer(mini_batch_size=args.mini_batch_size)
# Setup optimizer
opt = lbann.Adam(learn_rate=0.0002, beta1=0.9, beta2=0.99, eps=1e-8)
# Run experiment
kwargs = lbann.contrib.args.get_scheduler_kwargs(args)
lbann.contrib.launcher.run(trainer,
model,
data_reader_proto,
opt,
lbann_args=" --use_data_store --preload_data_store",
job_name=args.job_name,
**kwargs)
def main():
run_args = construct_lc_launcher_args()
# add data_config data
# and do not overwrite args if data_reader_prototext is enabled
if os.path.isfile(
run_args.data_config) and not run_args.data_reader_prototext:
with open(run_args.data_config, "r") as f:
config = json.load(f)
for k, v in config.items():
setattr(run_args, k, v)
trainer = lbann.Trainer(run_args.batch_size,
#name=None,
)
# define data_reader
if run_args.data_reader_prototext:
print("Using data_reader_prototext")
assert run_args.sequence_length is not None
assert run_args.vocab is not None
data_reader_proto = lbann.lbann_pb2.LbannPB()
with open(run_args.data_reader_prototext, "r") as f:
txtf.Merge(f.read(), data_reader_proto)
data_reader = data_reader_proto.data_reader
else:
data_reader = construct_data_reader(run_args)
if "LBANN_EXPERIMENT_DIR" in os.environ:
work_dir = os.environ["LBANN_EXPERIMENT_DIR"]
else:
work_dir = os.path.join(os.getcwd())
timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
experiment_dir = os.path.join(work_dir,
"{}_{}".format(timestamp, run_args.job_name))
if not os.path.exists(experiment_dir):
os.makedirs(experiment_dir)
# model and optimizer
model = construct_model(run_args)
opt = lbann.Adam(learn_rate=run_args.lr, beta1=0.9, beta2=0.99, eps=1e-8)
# dump the config to the experiment_dir so that it can be used to load the model in pytorch (moses codebase)
ppn = 4 if run_args.scheduler == "lsf" else 2
print("args:\n" + str(run_args))
if (run_args.scheduler == 'slurm'):
import torch
torch.save(run_args, "{}/{}_config.pt".format(experiment_dir,
run_args.job_name))
m_lbann_args = f"--load_model_weights_dir_is_complete --load_model_weights_dir={run_args.dump_model_dir} --vocab={run_args.vocab} --num_samples={run_args.num_samples} --sequence_length={run_args.sequence_length} --num_io_threads={run_args.num_io_threads} --no_header={run_args.no_header} --delimiter={run_args.delimiter}"
if (run_args.data_reader_prototext):
m_lbann_args = " ".join(
(m_lbann_args, " --use_data_store --preload_data_store "))
if (run_args.procs_per_trainer):
m_lbann_args = " ".join(
(m_lbann_args,
f"--procs_per_trainer={run_args.procs_per_trainer}"))
status = lbann.contrib.launcher.run(
trainer,
model,
data_reader,
opt,
lbann_exe,
partition=run_args.partition,
scheduler=run_args.scheduler,
account=run_args.account,
time_limit=run_args.time_limit,
nodes=run_args.nodes,
procs_per_node=ppn,
batch_job=True,
#setup_only = True,
job_name=run_args.job_name,
experiment_dir=experiment_dir,
lbann_args=m_lbann_args,
#turn on for tensor core
environment={
'LBANN_USE_CUBLAS_TENSOR_OPS': 1,
'LBANN_USE_CUDNN_TENSOR_OPS': 1,
},
)
print("LBANN launcher status:\n" + str(status))
def create_unet3d_optimizer(learn_rate):
# TODO: This is a temporal optimizer copied from CosomoFlow.
adam = lbann.Adam(learn_rate=learn_rate, beta1=0.9, beta2=0.999, eps=1e-8)
return adam
objective_function=obj,
metrics=metrics,
callbacks=[
lbann.CallbackPrint(),
lbann.CallbackTimer(),
],
)
# Setup trainer, optimizer, data_reader
trainer = lbann.Trainer(
mini_batch_size=lbann_params.mini_batch_size,
num_parallel_readers=1,
)
optimizer = lbann.Adam(
learn_rate=0.01,
beta1=0.9,
beta2=0.99,
eps=1e-8,
)
data_reader = make_data_reader()
# Launch LBANN
kwargs = lbann.contrib.args.get_scheduler_kwargs(lbann_params)
kwargs["environment"] = {}
lbann.contrib.launcher.run(
trainer,
model,
data_reader,
optimizer,
work_dir=lbann_params.work_dir,
job_name=lbann_params.job_name,
lbann_args=["--num_io_threads=1"],
checkpoint_dir='chkpt',
# checkpoint_epochs=10))
checkpoint_steps=gdict['step_interval']))
spectral_loss = gdict['lambda_spec']
if spectral_loss: print("Using Spectral loss with coupling", spectral_loss)
model = construct_model(
num_epochs,
gdict['mcr'],
spectral_loss=spectral_loss,
save_batch_interval=int(gdict['step_interval'])
) #'step_interval*val_ratio' is the step interval for validation set.
# Setup optimizer
opt = lbann.Adam(learn_rate=gdict['learn_rate'],
beta1=gdict['beta1'],
beta2=gdict['beta2'],
eps=float(gdict['eps']))
# Load data reader from prototext
data_reader = construct_data_reader(data_pct, val_ratio)
status = lbann.run(trainer,
model,
data_reader,
opt,
nodes=num_nodes,
procs_per_node=num_procs,
work_dir=work_dir,
scheduler='slurm',
time_limit=1440,
setup_only=False)
