Skip to content

Fix model inference issue with Barracuda 1.2.1 #4766

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

Merged
merged 3 commits into from
Dec 17, 2020
Merged

Fix model inference issue with Barracuda 1.2.1 #4766

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
64871e3
fix export
Dec 16, 2020
24c00a7
typo
Dec 16, 2020
457a344
Update ml-agents/mlagents/trainers/torch/distributions.py
Dec 16, 2020
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
8 changes: 5 additions & 3 deletions ml-agents/mlagents/trainers/torch/distributions.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -173,9 +173,11 @@ def forward(self, inputs: torch.Tensor) -> List[DistInstance]:
log_sigma = torch.clamp(self.log_sigma(inputs), min=-20, max=2)
else:
# Expand so that entropy matches batch size. Note that we're using
# torch.cat here instead of torch.expand() becuase it is not supported in the
# verified version of Barracuda (1.0.2).
log_sigma = torch.cat([self.log_sigma] * inputs.shape[0], axis=0)
# mu*0 here to get the batch size implicitly since Barracuda 1.2.1
# throws error on runtime broadcasting due to unknown reason. We
# use this to replace torch.expand() becuase it is not supported in
# the verified version of Barracuda (1.0.X).
log_sigma = mu * 0 + self.log_sigma
if self.tanh_squash:
return TanhGaussianDistInstance(mu, torch.exp(log_sigma))
else:
Expand Down
14 changes: 8 additions & 6 deletions ml-agents/mlagents/trainers/torch/networks.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -258,9 +258,11 @@ def __init__(
):
super().__init__()
self.action_spec = action_spec
self.version_number = torch.nn.Parameter(torch.Tensor([2.0]))
self.version_number = torch.nn.Parameter(
torch.Tensor([2.0]), requires_grad=False
)
self.is_continuous_int_deprecated = torch.nn.Parameter(
torch.Tensor([int(self.action_spec.is_continuous())])
torch.Tensor([int(self.action_spec.is_continuous())]), requires_grad=False
)
self.continuous_act_size_vector = torch.nn.Parameter(
torch.Tensor([int(self.action_spec.continuous_size)]), requires_grad=False
Expand All @@ -283,6 +285,9 @@ def __init__(
self.encoding_size = network_settings.memory.memory_size // 2
else:
self.encoding_size = network_settings.hidden_units
self.memory_size_vector = torch.nn.Parameter(
torch.Tensor([int(self.network_body.memory_size)]), requires_grad=False
)

self.action_model = ActionModel(
self.encoding_size,
Expand Down Expand Up @@ -335,10 +340,7 @@ def forward(
disc_action_out,
action_out_deprecated,
) = self.action_model.get_action_out(encoding, masks)
export_out = [
self.version_number,
torch.Tensor([self.network_body.memory_size]),
]
export_out = [self.version_number, self.memory_size_vector]
if self.action_spec.continuous_size > 0:
export_out += [cont_action_out, self.continuous_act_size_vector]
if self.action_spec.discrete_size > 0:
Expand Down