Batch size once again
Designs
- Show closed items
Relates to
Activity
-
Newest first Oldest first
-
Show all activity Show comments only Show history only
- Daniel Lukats added Bug label
added Bug label
- Daniel Lukats assigned to @dl337788
assigned to @dl337788
- Daniel Lukats changed milestone to %Official dates
changed milestone to %Official dates
- Author Owner
My implementation of batch handling might deviate from the OpenAI implementation. Mainly the number of updates performed and the size of mini batches might differ.
Edited by Daniel Lukats - Daniel Lukats marked this issue as related to #72 (closed)
marked this issue as related to #72 (closed)
- Daniel Lukats added In Progress label
added In Progress label
- Author Owner
Property OpenAI Me Kostrikov time steps 10_000_000 10_000_000 10_000_000 rollout length 128 128 128 number of envs 8 8 8 total rollout items (batch size) 128 * 8 = 1024 same same number of epochs 4 3 4 number of mini batches 4 3 32 mini batches per epoch 4 1 32 number of gradient updates 4 * 4 = 16 3 number of epochs * number of mini batches = 128 data used per mini batch 25% 25% 32 mini batch size batch size / number of mini batches = 256 rollout length * data used per batch * num_envs = 32 * 8 = 256 32 number of updates time steps / total rollout items = 9765 time steps / rollout length = 78_125 time steps / rollout length / number of envs = 9765 Edited by Daniel Lukats Collapse replies - Author Owner
Both OpenAI and Kostrikov use the full dataset in each PPO epoch.
- Daniel Lukats closed
closed
- Daniel Lukats reopened
reopened
- Daniel Lukats mentioned in commit d48816c6
mentioned in commit d48816c6
- Daniel Lukats mentioned in commit 0c458e36
mentioned in commit 0c458e36
- Daniel Lukats added In Review label and removed In Progress label
added In Review label and removed In Progress label
- Daniel Lukats closed
closed
- Daniel Lukats removed In Review label
removed In Review label
Please register or sign in to reply