-
Notifications
You must be signed in to change notification settings - Fork 1.4k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
How to use the model in inference mode #95
Comments
If you load the agent but run it in eval mode, I think that will do what you want (combined to #71). |
How is the eval mode used? Setting the training steps to 0 causes a division by zero exception. |
That sounds like a bug -- I think probably just missing a check that if training_steps = 0, we shouldn't run the train phase. We already do this for the eval phase (when eval_steps = 0). |
For the eval mode, are all checkpoints files necessary? |
You should only need the last checkpoint -- the code keeps a few (3) around, in case the most recent checkpoint is corrupted. |
I mean these kind of checkpoint files:
Are all relevant for the eval mode? Especially the $store$_observation_ckpt can comprise many gigabytes. |
The tf_ckpt* files are for restoring the network parameters, so you
definitely want those.
The other files are for restoring the replay buffer to the state it was at
when the checkpoint was saved.
If you're only doing inference for evaluation (i.e. not sampling batches
from the replay buffer to learn from) then you don't need those files.
…On Mon, Apr 8, 2019, 12:21 AM Marco Pleines, ***@***.***> wrote:
I mean these kind of checkpoint files:
add_count_ckpt.19.gz
ckpt.19
invalid_range_ckpt.19.gz
sentinel_checkpoint_complete.19
$store$_action_ckpt.19.gz
$store$_observation_ckpt.19.gz
$store$_reward_ckpt.19.gz
sum_tree_ckpt.19.gz
tf_ckpt-19.data-00000-of-00001
tf_ckpt-19.index
tf_ckpt-19.meta
Are all relevant for the eval mode? Especially the
$store$_observation_ckpt can comprise many gigabytes.
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#95 (comment)>, or mute
the thread
<https://github.com/notifications/unsubscribe-auth/ATYhMfe3XwytyJiin-OoW0mQbsIklc8Mks5vesPbgaJpZM4cQ3m6>
.
|
Hello,
how does one load the model for execution only (not training) to watch the trained behavior?
This is related to #71
The text was updated successfully, but these errors were encountered: