Commit 6d9f6f2f authored by nilabha's avatar nilabha

Update README

parent b8f59603
......@@ -30,6 +30,15 @@ When we process videos, we sort videos which are more recent (based on the lates
## Sample video uploaded to wandb
![](recording/wandbvideos.png)
## Sample script to save videos
We have provided a simple config which runs 6 iterations and saves the videos to wandb. Change the wandb configs if required before running.
```bash
python train.py -f -f experiments/flatland_sparse/small_v0/tree_obs_fc_net/recording/ppo.yaml
```
Note that videos may not be created for the last iteration for each worker, as the videos for the last iterations are saved only once the process is stopped. RLlib seems to not execute the `on_result` step once the stop condition is achieved. Instead the process is closed finally once the program finishes and only then the video recorder is flushed. Hence, the last video cannot be uploaded directly to wandb during `on_result`.
## Video Folder Names
flatland-1.1250.0.0.11475.video000000.1590223601.6230516
......@@ -50,6 +59,6 @@ If we are running multiple wandb process, caution should be taken to avoid writi
# TODO
- Cannot match exact env details on which video was created and hence defaulting to the env details from when the video was first created. To help identify we must record the video file with the env iteration or/and steps etc.
- Using the env details when the video was created may be useful when recording video during evaluation
where we are more interested in the current training state
- Using the env details when the video was created may be useful when recording video during evaluation. In such cases, we are more interested in the latest training state than the current evaluation run details.
- The last episode videos are not saved due to the reasons mentioned in the [section](#sample-script-to-save-videos)
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment