wassname/ray
1
0
Fork 0
mirror of https://github.com/wassname/ray.git synced 2026-09-12 12:51:15 +08:00
Code Issues Packages Projects Releases Wiki Activity

[RLlib] TorchPolicies: Accessing "infos" dict in train_batch causes TypeError. (#13039)

Browse Source
This commit is contained in:
Sven Mika
2020-12-23 11:30:50 -05:00
committed by GitHub
parent 670d083a56
commit 1e74187179
1 changed files with 13 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rllib/utils/torch_ops.py
+13 -4
View File
@@ -63,11 +63,20 @@ def convert_to_torch_tensor(x, device=None):
return RepeatedValues(
tree.map_structure(mapping, item.values), item.lengths,
item.max_len)
# Non-writable numpy-arrays will cause PyTorch warning.
if isinstance(item, np.ndarray) and item.flags.writeable is False:
with warnings.catch_warnings():
warnings.simplefilter("ignore")
# Numpy arrays.
if isinstance(item, np.ndarray):
# np.object_ type (e.g. info dicts in train batch): leave as-is.
if item.dtype == np.object_:
return item
# Non-writable numpy-arrays will cause PyTorch warning.
elif item.flags.writeable is False:
with warnings.catch_warnings():
warnings.simplefilter("ignore")
tensor = torch.from_numpy(item)
# Already numpy: Wrap as torch tensor.
else:
tensor = torch.from_numpy(item)
# Everything else: Convert to numpy, then wrap as torch tensor.
else:
tensor = torch.from_numpy(np.asarray(item))
# Floatify all float64 tensors.