Johannes Fischer
88b4466e57
MInor change in value dice loss, activate print statements, only do EITHER value OR policy update for each batch
2021-08-06 18:52:17 +02:00
Johannes Fischer
025c71767f
Change final value network activation to identity
2021-08-06 18:48:53 +02:00
Johannes Fischer
66bfba3986
Minor formatting
2021-08-05 18:37:54 +02:00
Johannes Fischer
6afb112277
Bugfix in value dice
...
FIRST backward() has to be called on both, policy and value, before step() is called for either of them
2021-08-05 18:35:14 +02:00
Johannes Fischer
bf4c19a4d0
Add value dice ray config
2021-08-05 18:33:34 +02:00
Johannes Fischer
40c55478f3
Bugfixes in valuedice
2021-08-04 20:53:03 +02:00
Johannes Fischer
2224e2cd14
Merge branch 'main' of github.com:sisl/InteractionImitation
2021-08-04 19:44:55 +02:00
Johannes Fischer
5c40de66fa
Implement ValueDICE and some restructuring
2021-08-04 19:34:49 +02:00
Arec
f9729b0a9d
making expert data save s, a, sp. making dataloader also load batches thisway. renaming state to ego_state. converting path_x and path_y to single path variable. making number of samples for ray an argument. adjusting metrics, policy, and other functions to be able to handle this
2021-08-04 09:45:36 -07:00
Johannes Fischer
cba42c6e4d
Set default divergence to histogram based
2021-08-03 17:18:30 +02:00
Johannes Fischer
5919a4e439
Use JS divergence in metrics
2021-08-03 17:04:47 +02:00
Johannes Fischer
1ee46214a7
Implement jenson shannon divergence
2021-08-03 17:03:37 +02:00
Johannes Fischer
1916a8fe69
Implement metrics and write to tensorboard summary at test time
2021-08-03 15:24:29 +02:00
Arec
98294e0c95
Merge branch 'main' of https://github.com/sisl/InteractionImitation into main
2021-08-03 03:28:23 -07:00
Arec
6e524cf4b5
adding options for regularization and relative state masking via interaction graphs during data processing and experiment running. found 0.002 regularization on actions gives up to 3m of deviation with no collisions. added shell script to run ray experiments overnight
2021-08-02 14:38:06 -07:00
Johannes Fischer
b869597717
extend comment on divergence
2021-08-02 19:15:02 +02:00
Johannes Fischer
f468b3b7a4
Implement histogram based kl divergence computation
2021-08-02 19:14:40 +02:00
Johannes Fischer
6177b1f7e1
Add kl_cat
2021-08-02 11:12:24 +02:00
Johannes Fischer
4ea4d42df7
Merge branch 'main' of github.com:sisl/InteractionImitation
2021-08-02 11:08:35 +02:00
Johannes Fischer
281f7773c4
Implement kl divergence methods and tests
2021-08-02 11:04:03 +02:00
Johannes Fischer
99f7df2e7c
Add comment
2021-07-30 18:19:29 +02:00
Arec
fb91ee1a62
Merge branch 'main' of https://github.com/sisl/InteractionImitation into main
2021-07-30 05:17:43 -07:00
Johannes Fischer
bd5e854720
Add collision and avg velocity metrics
2021-07-30 13:35:42 +02:00
Arec
e30d0ab1ba
removing outliers from expert tracks
2021-07-29 08:20:40 -07:00
Arec
3b4ef6ffb5
adding tool for visualizing acceleration distributions, and making nframes an arg
2021-07-29 07:35:21 -07:00
Johannes Fischer
afc3719ab9
Fix bug with wrong argument order
2021-07-29 15:30:55 +02:00
Johannes Fischer
3b623a1467
Add testing script for raytune experiments
2021-07-29 15:30:41 +02:00
Johannes Fischer
35634fd2eb
Separate experiment from main.py
2021-07-29 14:40:19 +02:00
Arec
e47d69dbc1
changing relative state dim to 6
2021-07-29 02:54:41 -07:00
Arec
fdbdc7f9f0
making data processing happen on front end, not on data loader. saving a ton of time
2021-07-29 02:53:20 -07:00
Arec
09f77e0587
changing how tune reporting works so the scheduler doesnt break if itcant find cv loss. also fixed bug in config structure that was rendering impossible policies
2021-07-28 09:55:11 -07:00
Arec
87aa19b86b
getting hyperparameter tunning with ray tune working. updating default network with optimization and general parameters.
2021-07-27 14:30:28 -07:00
Arec
d9436daaba
changing default network, making deepsets network choose output dimension appropriately, making bc config to be called by ray
2021-07-26 07:52:11 -07:00
Arec
69359b5af3
periodically savingin out model and adding functionality to make identity Phi networks (for 0-dim NNs)
2021-07-26 05:51:08 -07:00
Arec
7b2ca6edc7
adding tensorboard writer for training loss and cv loss
2021-07-26 02:38:50 -07:00
Arec
5b09374c77
getting training and testing loop working, adding tqdm to simulator, and reduced number of frames, updating readme
2021-07-23 04:44:50 -07:00
Arec
563a2cfbd4
making a differentiable transform for use for pytorch, making sure the fitting function treats nans properly while fitting. next issue: forward pass is returning nans
2021-07-22 12:37:36 -07:00
Arec
08eb898812
adding dtypes and fixing matrix indexing
2021-07-22 12:06:42 -07:00
Johannes Fischer
1c22bd6111
bugfix in deepsets
2021-07-22 18:44:12 +02:00
Arec
91d052445e
making test case for typing bug and fixing some small typing errors in bc
2021-07-22 08:23:46 -07:00
Johannes Fischer
827a8e7172
bugfix in Phi module
...
nn.ModuleList has to be used in order to register layer parameters as module parameters (similar to add_module)
2021-07-22 14:33:05 +02:00
Arec
1ca9914bf9
fixing bugs in transform, expert demo processing, main train function, and behavior cloning class. need to get bc class parameters to return nonempty list
2021-07-21 09:44:20 -07:00
Arec
5758af5dd8
Merge branch 'main' of https://github.com/sisl/InteractionImitation into main
2021-07-21 08:22:15 -07:00
Arec
18b8e0c58f
updating main testing function to use configs and seeds, finishing first pass at behavior cloning policy and training loop. not yet tested
2021-07-21 08:22:08 -07:00
Arec
4350cf8cd5
defining transform class to do tensor size manipulation before and after transform
2021-07-21 08:20:55 -07:00
Arec
3422e9c9ef
exporting policy class and letting default final activation do nothing
2021-07-21 08:20:15 -07:00
Johannes Fischer
ca0f520c89
Merge branch 'main' of github.com:sisl/InteractionImitation
2021-07-20 18:44:11 +02:00
Johannes Fischer
61f06c95c2
Improve deepsets module
...
module can now deal with nan values for nonexisting relative states
in case all relative states are nan, the latent representation is zeroed, which is consitent with an empty sum
2021-07-20 18:44:07 +02:00
Johannes Fischer
ffdff12ccb
Update deepsets to deal with nans (first version)
2021-07-20 18:23:51 +02:00
Arec
6794b4cad8
making saving and loading functions class requirements, working on behavior cloning policy class and training function
2021-07-20 08:29:00 -07:00
