hongshaorou/dreamerv3-torch
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

set default replay buffer size as 1M

Browse Source
This commit is contained in:
NM512 2023-04-05 21:38:51 +09:00
parent 57ac1c11d3
commit cd935b7dd9
2 changed files with 4 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
configs.yaml
View File

@ -80,7 +80,7 @@ defaults:
ac_opt_eps: 1e-5 ac_opt_eps: 1e-5
value_grad_clip: 100 value_grad_clip: 100
actor_grad_clip: 100 actor_grad_clip: 100
dataset_size: 0 dataset_size: 1000000
oversample_ends: False oversample_ends: False
slow_value_target: True slow_value_target: True
slow_target_update: 1 slow_target_update: 1

6
dreamer.py
View File

@ -252,14 +252,14 @@ class ProcessEpisodeWrap:
cls.eval_lengths = [] cls.eval_lengths = []
cache.clear() cache.clear()
if mode == "train" and config.dataset_size: if mode == "train":
total = 0 total = 0
for key, ep in reversed(sorted(cache.items(), key=lambda x: x[0])): for key, ep in reversed(sorted(cache.items(), key=lambda x: x[0])):
if total <= config.dataset_size - length: if not config.dataset_size or total <= config.dataset_size - length:
total += len(ep["reward"]) - 1 total += len(ep["reward"]) - 1
else: else:
del cache[key] del cache[key]
logger.scalar("dataset_size", total + length) logger.scalar("dataset_size", total)
print(f"{mode.title()} episode has {length} steps and return {score:.1f}.") print(f"{mode.title()} episode has {length} steps and return {score:.1f}.")
logger.scalar(f"{mode}_return", score) logger.scalar(f"{mode}_return", score)
logger.scalar(f"{mode}_length", length) logger.scalar(f"{mode}_length", length)