[Performance] Accelerate GAE #1142
Conversation
An optimized vecotrized version for the generalized advantage estimation is used in case gamma and lambda are scalars. When handling consecutive trajectories of the form ``` reward = [r00, r01, r02, r03, r10, r11] done = [False, False, False, True, False, False] ``` , `vec_generalized_advantage_estimate` needs to build a giant famma tensor of size [Batch, T, T] with a decayed gamma tensor that suits each trajectory. Thus it needs to allocate a big tensor `[B, T, T]` and do a heavy matrix multiplication. In case gamma and lambda are scalars, this can be optimized by building a single tensor of the form ``` r_transformed = [[r00, r01, r02, r03] [r10, r11, 0, 0]] ``` and applying the gamma filter `[r00 + gamma r01 + gamma ** 2 r02 + ..., ro1 + gamma r02 + gamma ** 2 r03 + ...,]` to calculate the GAE.
* move helper methods to util * reuse existing helper methods * remove wip file
facebook-github-bot
added
the
CLA Signed
In case gamma and lmbda are scalars, `fast_vec_gae` should be always faster than `vec_generalized_advantage_estimate` if len(T) is large enough.
…lementation" This reverts commit 245e68f.
in case there is only one split, _inv_pad_sequence can skip its calculation.
| gamma = torch.full(size, gamma) | ||
| lmbda = 0.95 | ||
|
|
||
| benchmark( |
There was a problem hiding this comment.
Why not using benchmark.pedantic to get some extra options?
I'm open to use plain benchmark if you think it's a better fit
There was a problem hiding this comment.
My guess was to use the automatic calibration is better here, when I have no need of the fine-grained control of benchmark.pedantic. The pytest-benchmark docu says roughly: "don't use pedantic if you don't need it".
| if reward.ndim > 2: | ||
| done = done.transpose(-2, -1) | ||
| reward = reward.transpose(-2, -1) | ||
| state_value = state_value.transpose(-2, -1) | ||
| next_state_value = next_state_value.transpose(-2, -1) |
There was a problem hiding this comment.
what happens if the reward has 2 dimensions? Don't we want to swap them>
There was a problem hiding this comment.
You are right, that does not make sense. In particular, there is the following line in vec_generalized_advantage_estimate
*batch_size, time_steps, lastdim = not_done.shape
which ensures that reward and the other tensors have at least 3 dimensions, so the checks are never executed.
I will remove the checks, although I need to remember why I introduced them in the first place.
Description
An optimized vecotrized version for the generalized advantage estimation is used in case gamma and lambda are scalars.
Motivation and Context
When handling consecutive trajectories of the form
,
vec_generalized_advantage_estimateneeds to build a giant gamma tensor of size [Batch, T, T] with a decayed gamma tensor that suits each trajectory. Thus it needs to allocate a big tensor [B, T, T] and do a heavy matrix multiplication. In case gamma and lambda are scalars, this can be optimized by building a single tensor of the form
and applying the gamma filter
[r00 + gamma r01 + gamma ** 2 r02 + ..., ro1 + gamma r02 + gamma ** 2 r03 + ...,]to calculate the GAE.close #1052Types of changes
What types of changes does your code introduce? Remove all that do not apply:
Checklist
Go over all the following points, and put an
xin all the boxes that apply.If you are unsure about any of these, don't hesitate to ask. We are here to help!