Change Log
Unreleased
- Fix: Multiple
stack_compress
(#143)
- Improve: PyPI page
- Fix: Some typos at docstring
- Fix: Replace
assert
with ValueError
- Fix: Remove deprecated
__getslice__
/ __setslice__
custom methods
- Improve:
MPReplayBuffer
and MPPrioritizedReplayBuffer
works with Ray (Request, 141) (Thanks @jamartinh)
- Add: Context parameter (
ctx
) at constructor. SyncManager
can be used for Lock
and Event
.
- Add: Backend parameter (
backend
) at constructor. SharedMemory
can be used.
- Fix: Remove unused method from internal C++ sampler
- Improve: Speed up Segment Tree implementation.
- Fix: Pass nextstate (instead of state) to reward function in
HindsightReplayBuffer
- Fix: Take
goal_func
to convert state to goal in HindsightReplayBuffer
- Add:
HindsightReplayBuffer
- Fix: Nstep accepts JAX array
- Fix: GitHub Action binary build
- Add: Support JAX array (aka.
jaxlib.xla_extension.DeviceArray
) input.
- Add: Binary distibution for macOS x8664, and Linux AArch64
- Add:
LaBERmean
, LaBERlazy
, and LaBERmax
- Fix: Numpy version at build from source
- Add:
save_transitions
/ load_transitions
(138)
- Breaking Change: Nstep no longer returns discounts (137)
- Fix: stackcompress cache (135)
- Fix: NumPy ABI change (131)
- Fix:
PrioritizedReplayBuffer
, MPPrioritizedReplayBuffer
support non C-convertible dtype (e.g. np.float16
) (130)
- Fix:
PrioritizedReplayBuffer
, MPPrioritizedReplayBuffer
adds read only priority (128)
- Fix:
MPPrioritizedReplayBuffer
lock mechanism, which makes faster
- Fix:
MPReplayBuffer
and MPPrioritizedReplayBuffer
with spawn multiprocessing
- Add:
MPReplayBuffer
and MPPrioritizedReplayBuffer
- Fix: Walkaround
next_of
at Nstep
- Fix: Input
dict
modification (116)
- Fix: CodeQL reported warnings
- Unused import (1,2)
- Empty except (1)
- Fix: cache for nextof and stackcompress (114)
- Add:
get_all_trainsitions(shuffle=True)
for shuffle transitions
- Fix: Windows binary distribution
- Add: Large buffer exceeding memory using mmap
- Fix: Logger writer multiple same massages
- Fix: buffer dtypes were mis-converted (112)
- Fix: Add Warning to
train
- Fix: Embed signature at
train
- Add: New free function
train
for simple train loop (beta)
- Fix: Nstep
PrioritizedReplayBuffer
raises exception at on_episode_end
(111)
Fix: Nstep PrioritizedReplayBuffer
raises exception at on_episode_end
(111) This was not fixed correctly.
- Fix: Cache for
"next_of"
when episode is smaller than stack (110)
- Fix: Cache for
"next_of"
when "stack_compress"
at last item (108)
- Remove: author email address from setup.py, which is not necessary for PyPI
- Fix: Cache for
"next_of"
with "stack_compress"
at non last item (108)
- Remove: cpprb.gym sub-package, which is not cpprb main purpose
- Fix: Support PyVirtualDisplay 1.0 above.
- Fix: Try to acquire physical memory in ReplayBuffer initialization (#105)
- Add: Pruning update chain in Segment Tree (#103)
- Fix: Windows binary build on GitHub Actions
- Fix: Cache non last transitions. (#97)
- Fix: Clear caches when
ReplayBuffer.clear
(#98)
- Fix: PER priority shape mismatch (#96)
- Fix: Unrequire done (#61)
- Add:
ReplayBuffer.get_all_transitions
- Fix: Add gym into requirements, since
create_env_dict
, create_before_func
use
- Fix: GitHub Action settings (Recover Windows binary)
- Add: Hepler functions
create_env_dict
, create_before_func
- Fix: GitHub Actions syntax
- Fix: README on PyPI (overwritten upload from GitHub)
- Fix: Binary upload on Github Actions
- Add: Distribute binary of Windows and Linux
- Fix: gym module
is
comparison with literal
- Fix: Build without pre-install cython
- Add: debug build with environment
DEBUG_CPPRB
- Fix: compile argument at setup.py for Windows.
- Fix: Missing <numeric> header at ReplayBuffer.hh
- Fix: Accept float shape at PrioritizedReplayBuffer.unchangesincesample
- Add:
check_for_update
for PrioritizedReplayBuffer