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

Jump to bottom

Rl from scratch debugging #124

Open
wants to merge 4 commits into
base: qedcartographer
Choose a base branch
from
Open

Rl from scratch debugging #124

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
745e91b
debug subprocess run output
sahilx13 Mar 11, 2024
de582dc
fix divide by zero
sahilx13 Mar 21, 2024
751ba03
fix torch log nan
sahilx13 Mar 21, 2024
d89c101
Revert changes to distributed_rl.py related to output
sahilx13 Apr 9, 2024
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
10 changes: 6 additions & 4 deletions src/distributed_rl_learning_server.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,6 +33,8 @@
import torch.distributed as dist
import torch.optim.lr_scheduler as scheduler

EPSILON = 1e-7

print("Finished torch imports", file=sys.stderr)
# pylint: disable=wrong-import-position
sys.path.append(str(Path(os.getcwd()) / "src"))
Expand Down Expand Up @@ -179,7 +181,7 @@ def serve_parameters(args: argparse.Namespace, backend='mpi') -> None:
loss_buffer.append(loss)
iters_trained += 1
if len(loss_buffer) == args.loss_smoothing:
smoothed_loss = sum(loss_buffer) / args.loss_smoothing
smoothed_loss = sum(loss_buffer) / (args.loss_smoothing + EPSILON)
lr = optimizer.param_groups[0]['lr']
eprint(f"Loss: {smoothed_loss} (learning rate {lr/20:.3e})")
if vsample_changed and args.reset_on_updated_sample:
Expand Down Expand Up @@ -392,7 +394,7 @@ def train(args: argparse.Namespace, v_model: VModel,
loss = F.mse_loss(actual_values, target_values)
else:
assert args.loss == "log"
loss = F.mse_loss(torch.log(actual_values), torch.log(target_values))
loss = F.mse_loss(torch.log(F.relu(actual_values) + EPSILON), torch.log(F.relu(target_values) + EPSILON))
if args.verbose >= 1:
eprint("Training obligations to values:")
for idx, (context, prev_tactic, output, actual_value) \
Expand Down Expand Up @@ -785,15 +787,15 @@ def print_vvalue_errors(gamma: float, vnetwork: nn.Module,
predicted_v_values = vnetwork(
torch.cat([obl.local_context.view(1, -1) for obl, _ in items], dim=0),
torch.LongTensor([obl.previous_tactic for obl, _ in items]).to(device)).view(-1)
predicted_steps = torch.log(predicted_v_values) / math.log(gamma)
predicted_steps = torch.log(F.relu(predicted_v_values) + EPSILON) / (math.log(gamma) + EPSILON)
num_predicted_zeros = torch.count_nonzero(predicted_steps == float("inf"))
target_steps: FloatTensor = torch.tensor([steps for _, steps in items]).to(device) #type: ignore
step_errors = torch.abs(predicted_steps - target_steps)

total_error = torch.sum(torch.where(predicted_steps == float("inf"),
torch.zeros_like(step_errors),
step_errors)).item()
avg_error = total_error / (len(items) - num_predicted_zeros)
avg_error = total_error / (len(items) - num_predicted_zeros + EPSILON)
eprint(f"Average step error across {len(items) - num_predicted_zeros} "
f"initial states with finite predictions: {avg_error:.6f}")
eprint(f"{num_predicted_zeros} predicted as infinite steps (impossible)")
Expand Down