Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Fix memory issues in multitask generator #57

Merged
merged 3 commits into from
Aug 18, 2022
Merged

Fix memory issues in multitask generator #57

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
82d65ba
Dynamically adapt `init_batch_limit`
AngelFP Jul 21, 2022
636459c
Merge branch 'run_ax_on_gpu' into fix_memory_issues
AngelFP Jul 21, 2022
cc1a570
Improve prints
AngelFP Jul 23, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
60 changes: 49 additions & 11 deletions libe_opt/persistent_gp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -589,6 +589,9 @@ def persistent_gp_mt_ax_gen_f(H, persis_info, gen_specs, libE_info):
else:
torch_device = 'cpu'

# Batch limit for the initialization of the aquisition function optimization.
init_batch_limit = 1000

# Receive information from the manager (or a STOP_TAG)
tag = None
model_iteration = 0
Expand All @@ -612,7 +615,7 @@ def persistent_gp_mt_ax_gen_f(H, persis_info, gen_specs, libE_info):
else:
# Run multi-task BO.

# Fit the MTGP.
# 1) Fit the MTGP.
m = get_MTGP(
experiment=exp,
data=exp.fetch_data(),
Expand All @@ -621,23 +624,58 @@ def persistent_gp_mt_ax_gen_f(H, persis_info, gen_specs, libE_info):
device=torch.device(torch_device)
)

# Find the best points for the high fidelity task.
gr = m.gen(
n=n_opt_lofi,
optimization_config=exp.optimization_config,
fixed_features=ObservationFeatures(
parameters={}, trial_index=hifi_trials[-1]),
)
# 2) Find the best points for the high fidelity task.

# Too large initialization batches can lead to out-of-memory errors.
# The loop below tries to generate the next points to evaluate
# using `init_batch_limit`. If an RuntimeError is raised during
# generation (namely an out-of-memory error), `init_batch_limit`
# is divided by two and a new attempt is made. This is repeated
# until the generation runs successfully.
generator_success = True
while True:
try:
# Try to generate the new points.
gr = m.gen(
n=n_opt_lofi,
optimization_config=exp.optimization_config,
fixed_features=ObservationFeatures(
parameters={}, trial_index=hifi_trials[-1]),
model_gen_options={
'optimizer_kwargs': {
'init_batch_limit': init_batch_limit
}
}
)
# When successful, break loop.
break
except RuntimeError as e:
# Print exception.
print('RuntimeError: {}'.format(e), flush=True)
# Divide batch size by 2.
init_batch_limit //= 2
print('Retrying with `init_batch_limit={}`'.format(
init_batch_limit), flush=True)
finally:
# If all attempts have failed (even for batch size of 1),
# mark generation as failed and break loop.
if init_batch_limit == 0:
generator_success = False
break
# If generation failed, stop optimization.
if not generator_success:
tag = STOP_TAG
break

# But launch them at low fidelity.
# 3) But launch them at low fidelity.
tr = exp.new_batch_trial(trial_type=lofi_task, generator_run=gr)
tr.run()
tr.mark_completed()
tag = tr.run_metadata['tag']
if tag in [STOP_TAG, PERSIS_STOP]:
break

# Update the model.
# 4) Update the model.
m = get_MTGP(
experiment=exp,
data=exp.fetch_data(),
Expand All @@ -646,7 +684,7 @@ def persistent_gp_mt_ax_gen_f(H, persis_info, gen_specs, libE_info):
device=torch.device(torch_device)
)

# Select max-utility points from the low fidelity batch to generate a high fidelity batch.
# 5) Select max-utility points from the low fidelity batch to generate a high fidelity batch.
gr = max_utility_from_GP(
n=n_opt_hifi,
m=m,
Expand Down