Commit Graph

65 Commits

Author SHA1 Message Date
Alexander Engelsberger
7b9b767113 fix: training loss is a zero dimensional tensor
Should fix the problem with EarlyStopping callback.
2021-06-25 17:07:06 +02:00
Jensun Ravichandran
72af03b991 refactor: use LinearTransform instead of torch.nn.Linear 2021-06-25 17:07:06 +02:00
Jensun Ravichandran
d42693a441 refactor(api)!: merge the new api changes into dev 2021-06-20 19:00:12 +02:00
danielstaps
0a2da9ae50
Added Vis for GMLVQ with more then 2 dims using PCA (#11)
* Added Vis for GMLVQ with more then 2 dims using PCA

* Added initialization possibility to GMlVQ with PCA and one example with omega init + PCA vis of 3 dims

* test(githooks): Add githooks for automatic commit checks

Co-authored-by: staps@hs-mittweida.de <staps@hs-mittweida.de>
Co-authored-by: Alexander Engelsberger <alexanderengelsberger@gmail.com>
2021-06-18 13:28:11 +02:00
Alexander Engelsberger
8956ee75ad
test(githooks): Add githooks for automatic commit checks 2021-06-16 16:16:34 +02:00
Jensun Ravichandran
a37095409b [BUGFIX] examples/cbc_iris.py works again 2021-06-15 15:59:47 +02:00
Jensun Ravichandran
68034d56f6 [BUGFIX] examples/glvq_iris.py works again 2021-06-14 20:13:25 +02:00
Jensun Ravichandran
69e5ff3243 Import from the newly cleaned-up prototorch namespace 2021-06-14 20:08:08 +02:00
Jensun Ravichandran
d558fa6a4a [REFACTOR] Clean up GLVQ-types 2021-06-07 17:00:38 +02:00
Jensun Ravichandran
016fcb4060 [REFACTOR] Major cleanup 2021-06-04 22:20:32 +02:00
Jensun Ravichandran
42d974e08c No implicit learning rate scheduling 2021-06-04 15:55:06 +02:00
Alexander Engelsberger
47db1965ee [BUGFIX] GNG Example 2021-06-03 15:42:54 +02:00
Alexander Engelsberger
358f27257d [REFACTOR] Remove prototype_initializer function from GLVQ
Fixes #9
2021-06-03 15:15:22 +02:00
Alexander Engelsberger
e209bf73d5 [BUGFIX] Pruning example works on GPU now 2021-06-03 14:35:24 +02:00
Jensun Ravichandran
8f7deb75dd [FEATURE] Log prototype win ratios over all training batches 2021-06-02 02:32:54 +02:00
Jensun Ravichandran
757f4e980d Add Local-Matrix LVQ
Also remove the use of `self.distance_fn` in favor of `self.distance_layer`.
2021-06-01 23:44:16 +02:00
Jensun Ravichandran
27eccf44d4 Use LambdaLayer from ProtoTorch 2021-05-31 16:53:04 +02:00
Alexander Engelsberger
2a218c0ede Add example for dynamic components in callbacks 2021-05-31 11:39:24 +02:00
Alexander Engelsberger
db064b5af1 Improvement of model __repr__ 2021-05-31 11:19:06 +02:00
Alexander Engelsberger
0ac4ced85d [refactor] Use functional variant of accuracy
Prevents Accuracy in `__repr__` of the models.
2021-05-31 11:12:27 +02:00
Jensun Ravichandran
cd73f6c427 Add examples/dynamic_components.py 2021-05-31 00:32:27 +02:00
Jensun Ravichandran
b7edee02c3 [WIP] Add CELVQ
TODO Ensure that the distances/probs corresponding to the plabels are sorted
like the target labels.
2021-05-27 17:40:16 +02:00
Jensun Ravichandran
d411e52be4 Refactor non-gradient-lvq models into lvq.py 2021-05-25 20:37:34 +02:00
Jensun Ravichandran
139109804f [BUGFIX] Use _forward in LVQ1 and LVQ21 2021-05-25 17:43:37 +02:00
Alexander Engelsberger
8ce18f83ce Add prototype_initializer function to GLVQ
This allows overwriting it inside subclasses.
2021-05-21 17:13:10 +02:00
Alexander Engelsberger
7b4f7d84e0 Update Documentation
Clean up project
2021-05-21 15:42:45 +02:00
Jensun Ravichandran
a5e086ce0d Refactor code 2021-05-21 13:33:57 +02:00
Jensun Ravichandran
88a34a06ef [WIP] Update CBC implementation to use SiameseGLVQ 2021-05-20 17:36:00 +02:00
Jensun Ravichandran
df061cc2ff Refactor code 2021-05-20 14:40:02 +02:00
Alexander Engelsberger
969fb34cc3 Accumulate test loss 2021-05-20 14:20:23 +02:00
Alexander Engelsberger
0204f5eab6 Log test accuracy. 2021-05-20 14:03:31 +02:00
Alexander Engelsberger
b7fc5df386 Log test loss. 2021-05-20 13:47:20 +02:00
Alexander Engelsberger
faf1a88f99 [Bugfix] Remove optimzer_idx from validation and test. 2021-05-20 13:17:27 +02:00
Jensun Ravichandran
5ffbd43a7c Refactor into shared_step 2021-05-19 16:57:51 +02:00
Jensun Ravichandran
fdf9443a2c Add validation and test logic 2021-05-19 16:30:19 +02:00
Jensun Ravichandran
eefec19c9b Custom non-gradient training 2021-05-18 19:49:16 +02:00
Jensun Ravichandran
81346785bd Cleanup models
Siamese architectures no longer accept a `backbone_module`. They have to be
initialized with an pre-initialized backbone object instead. This is so that the
visualization callbacks could use the very same object for visualization
purposes. Also, there's no longer a dependent copy of the backbone. It is
managed simply with `requires_grad` instead.
2021-05-17 17:00:23 +02:00
Jensun Ravichandran
b7684ae512 predict_latent no longer returns numpy 2021-05-15 12:52:16 +02:00
Alexander Engelsberger
0eac2ce326 Examples use GPUs if available. 2021-05-13 15:22:01 +02:00
Jensun Ravichandran
8f9c29bd2b [BUGFIX] Remove incorrect import statement 2021-05-12 16:45:01 +02:00
Jensun Ravichandran
ca39aa00d5 Stop passing component initializers as hparams
Pass the component initializer as an hparam slows down the script very much. The
API has now been changed to pass it as a kwarg to the models instead.

The example scripts have also been updated to reflect the new changes.

Also, ImageGMLVQ and an example script `gmlvq_mnist.py` that uses it have also
been added.
2021-05-12 16:36:22 +02:00
Jensun Ravichandran
eab1ec72c2 Change optimizer using kwargs 2021-05-11 16:13:00 +02:00
Alexander Engelsberger
3fa6378c4d Add LVQ1 and LVQ2.1 Models. 2021-05-11 13:26:13 +02:00
Jensun Ravichandran
ed03ab168e [BUGFIX] Fix lambda_matrix property in GMLVQ 2021-05-10 14:09:25 +02:00
Jensun Ravichandran
c6e06ceaa4 Properly initialize prototypes in LVQMLN 2021-05-09 20:55:28 +02:00
Jensun Ravichandran
ca4c9da10a Add the namespace hook for GMLVQ in the model class 2021-05-09 20:53:31 +02:00
Jensun Ravichandran
d7972a69e8 Update GMLVQ model 2021-05-07 15:24:47 +02:00
Alexander Engelsberger
4bbe73e3a9 Add GRLVQ with examples. 2021-05-06 18:42:06 +02:00
Alexander Engelsberger
5a2f4f6170 Revert deletion of training accuracy. 2021-05-06 18:02:01 +02:00
Alexander Engelsberger
1c3613019b Update Examples to new initializer architecture.
Visualization still borken for some examples.
2021-05-06 14:10:09 +02:00