New mixin-based template structure #1092

HamedHemati · 2022-07-12T15:18:34Z

I'm creating this draft PR for the new mixin-based template structure.
For now, the main goal is to implement working version of the templates that we want and get feedback to improve the structure's simplicity and fix potential bugs.

Current status and next step:

Current changes are completely separated from the main templates and do not affect existing strategies.
I've tested the current changes for Naive and OnlineNaive strategies and both work fine.
Next step: add meta-learning template.

Here is a summary of the changes that I've made:

I added a new template called BaseGeneralSGDTemplate that combines the callbacks from the existingBaseSGDTemplate and SupervisedTemplate in order to create a more general template.
As discussed in ADD Meta Learning Template #1079 a new template can be implemented with the following structure:
NewTemplate(ObservationType, ProblemType, UpdateType, BaseGeneralSGDTemplate)
In common_templates.py I've added the templates for supervised and online supervised strategies and inside strategy_wrappers_temp.py you can see Naive and OnlineNaive strategies.

coveralls · 2022-07-12T15:28:26Z

Pull Request Test Coverage Report for Build 3284238644

221 of 278 (79.5%) changed or added relevant lines in 44 files are covered.
2 unchanged lines in 2 files lost coverage.
Overall coverage decreased (-0.3%) to 73.124%

Changes Missing Coverage	Changed/Added Lines	%
avalanche/evaluation/metric_definitions.py	1	0.0%
avalanche/evaluation/metric_utils.py	1	0.0%
avalanche/evaluation/metrics/checkpoint.py	1	0.0%
avalanche/evaluation/metrics/confusion_matrix.py	1	0.0%
avalanche/evaluation/metrics/forgetting_bwt.py	1	0.0%
avalanche/evaluation/metrics/forward_transfer.py	1	0.0%
avalanche/evaluation/metrics/gpu_usage.py	1	0.0%
avalanche/evaluation/metrics/images_samples.py	1	0.0%
avalanche/evaluation/metrics/labels_repartition.py	1	0.0%
avalanche/evaluation/metrics/mean_scores.py	1	0.0%

Files with Coverage Reduction	New Missed Lines	%
avalanche/benchmarks/utils/data_loader.py	1	81.17%
avalanche/training/templates/base_sgd.py	1	89.19%

Totals
Change from base Build 3282871086:	-0.3%
Covered Lines:	12932
Relevant Lines:	17685

💛 - Coveralls

HamedHemati · 2022-07-14T08:49:17Z

@AntonioCarta There is also another point regarding plugins:
I thought about implementing an existing meta-learning-based strategy as a plugin for SupervisedMetaLearningTemplate and realized that we also need to include inner/outer callbacks to the base plugin class.

We can either have one general base plugin class that consists of all possible callbacks that trigger plugins, or we can extend BaseSGDPlugin as below:
BaseSGDPlugin->SupervisedPlugin *which already exists
BaseSGDPlugin->MetaLearningPlugin *this needs to be added

What do you think about the extension of plugins? we should also consider that we might need to update it for every new "common" template as well.

AntonioCarta · 2022-07-18T15:53:40Z

We can either have one general base plugin class that consists of all possible callbacks

This doesn't make much sense from an inheritance perspective. Childs can add callbacks but they cannot remove them, i.e. they need to support all the callbacks of the parents.

Notice that this is what we do now, where Base -> BaseSGD -> Supervised progressively add new callbacks.

AntonioCarta · 2022-07-12T16:00:28Z

avalanche/training/templates/base_general_sgd.py

+
+    # ==================================================================> NEW
+
+    def maybe_adapt_model_and_make_optimizer(self):


We don't need this additional method. It's already covered by _before_training_exp. In a subclass you would do something like:

def _before_training_exp(...): # do stuff like model adapt and optimizer init # can also define new callbacks here if needed super()._before_training_exp(...)

AntonioCarta · 2022-07-12T16:05:16Z

avalanche/training/templates/observation_type/online_observation.py

+
+
+class OnlineObservation:
+    def _train_exp(


doesn't look specific to OCL. Can we move it to BaseSGD or up the hierarchy?

Here we don't have a loop over train_epochs unlike batch_observations.

AntonioCarta · 2022-07-12T16:08:08Z

avalanche/training/templates/update_type/meta_update.py

+            self.loss = 0
+
+            # Fast updates
+            self._before_fast_update(**kwargs)


you also need the fast_update method between the before/after calls. Same problem for slow update

AntonioCarta · 2022-07-18T15:55:17Z

avalanche/training/templates/base_general_sgd.py

+from avalanche.training.utils import trigger_plugins
+
+
+class BaseGeneralSGDTemplate(BaseTemplate):


Why not call it just SGDTemplate? it's taking an optimizer as argument, so it must be some kind of sgd-based training.

alternatively, you can call it IterativeTemplate. In this case, I would consider removing the optimizer from this class and pushing it down the hierarchy

…o new_templates

HamedHemati · 2022-07-24T18:51:56Z

The meta-update class is updated.
@AntonioCarta I applied some of your comments, for the other comments I need to discuss them with you in the next meeting.
Working examples for Naive, OnlineNaive, and LaMAML (as a plugin) are added.
We also need to discuss potential conflicts that can happen with the distributed training PR from Lorenzo

** I've added the prefix NEW_ for the files related to this PR.

AntonioCarta · 2022-08-01T15:57:02Z

avalanche/NEW_core.py

+
+    # ====================================================================> NEW
+
+    def before_inner_updates(


why do we have inner/outer updates here? SGD doesn't have an inner/outer loop

This is what I discussed previously. The new base SGD template is supposed to be the base class for all possible sorts of SGD-based strategies, and the meta-learning template is one of them (as defined in common_tamplates.py). So the idea is to add all callbacks that can be triggered by any of the SGD-based templates. Or are you suggesting splitting it into a separate plugin class for meta-learning-based strategies?

AntonioCarta · 2022-08-01T15:58:21Z

avalanche/training/plugins/NEW_lamaml.py

+        self.rough_sz = math.ceil(bsize_data / self.n_inner_updates)
+        self.meta_losses = [0 for _ in range(self.n_inner_updates)]
+
+    def single_inner_update(self, x, y, t, criterion):


plugins usually implent only before/after methods.

I explained this in the new plugin comments

AntonioCarta · 2022-08-01T15:58:50Z

avalanche/NEW_core.py

+        """Called before `_inner_updates` by the `BaseTemplate`."""
+        pass
+
+    def inner_updates(


this is not a before/after method, shouldn't be here.

That's true, but the reason I added the inner_updates callback as a plugin trigger is that, unlike supervised strategies, we don't have any "Naive" type of updates that I can set as default for meta-learning-based strategies.
More precisely, in supervised strategies, we have the training_epoch function that is implemented in its most basic form (which is Naive fine-tuning), and you can augment it by adding new plugins. We don't have such a general structure similar totraining_epoch for inner updates in meta-learning, and it can be completely different from method to method. That's why I added it as a plugin trigger that has to be implemented by the user. Do you have other suggestions?

AntonioCarta · 2022-08-01T15:59:06Z

avalanche/NEW_core.py

+        """Called before `_outer_updates` by the `BaseTemplate`."""
+        pass
+
+    def outer_update(


same as inner_update. Shouldn't be here

Same answer as the previous one.

AntonioCarta · 2022-08-01T16:01:11Z

avalanche/training/templates/NEW_base_sgd.py

+
+    # ==================================================================> NEW
+
+    def maybe_adapt_model_and_make_optimizer(self):


why do we need this method? strategies can override _before_training_exp for the same behavior

if you need a call that can be implemented by observation_type I would use a more general name, something like prepare_training_exp or similar.

AntonioCarta · 2022-08-01T16:04:58Z

avalanche/training/templates/observation_type/batch_observation.py

+            Use [] if you do not want to evaluate during training.
+        :param kwargs: custom arguments.
+        """
+        if eval_streams is None:


L19-L26 should be factored out of this method since it's common for all strategies

The loop over training_epochs and the plugin triggers can be different in different observations, that's why I explicitly defined them here.

Then it's not clear to me what is the difference between observation type and templates. I expected templates to modify the loop and observation types to only add new attributes/methods that are called by the loop.

Do we have a clear definition of template/observation/problem/update? For example a clear interface that they must specify

…o new_templates

HamedHemati

As mentioned in the comments, let me know your opinion about how to create a more "general" plugin class that would cover all possible callbacks. Thanks!

AntonioCarta · 2022-08-23T11:21:13Z

let me know your opinion about how to create a more "general" plugin class that would cover all possible callbacks

I think you are breaking the inheritance hierarchy with this approach. For example, in your general template you have inner and outer loops, but these do not make sense for some of the child classes. This is undesirable for many reasons.

To give a more concrete example: imagine that you want to create a hierarchy of animals (BaseAnimal is the parent, FlyingAnimal and SeaAnimal the childs). You are saying that all the animals should implement both swim and fly (methods in BaseAnimal), even though SeaAnimal will never fly.

…o new_templates

HamedHemati · 2022-08-31T16:51:41Z

@AntonioCarta the updates are made according to our last discussion.

AntonioCarta · 2022-09-06T14:33:45Z

avalanche/training/templates/base_sgd.py

-
-            self._after_training_iteration(**kwargs)
+        # Should be implemented in Update Type
+        raise NotADirectoryError()


should be NotImplementedError

AntonioCarta · 2022-09-06T14:35:16Z

avalanche/training/templates/base_sgd.py

@@ -271,8 +216,142 @@ def eval_epoch(self, **kwargs):

            self._after_eval_iteration(**kwargs)

+    # ==================================================================> NEW
+
+    def maybe_adapt_model_and_make_optimizer(self):


what is this method? why do we need it? we already have before_training_exp which covers model adaptation and optimizer init.

maybe change the name to prepare_for_observation, or something similar. It doesn't have to be about model adaptation or optimizers

AntonioCarta · 2022-09-06T14:40:37Z

avalanche/training/templates/observation_type/batch_observation.py

+            Use [] if you do not want to evaluate during training.
+        :param kwargs: custom arguments.
+        """
+        if eval_streams is None:


Then it's not clear to me what is the difference between observation type and templates. I expected templates to modify the loop and observation types to only add new attributes/methods that are called by the loop.

AntonioCarta · 2022-09-06T14:42:09Z

avalanche/training/templates/observation_type/batch_observation.py

+            Use [] if you do not want to evaluate during training.
+        :param kwargs: custom arguments.
+        """
+        if eval_streams is None:


Do we have a clear definition of template/observation/problem/update? For example a clear interface that they must specify

AntonioCarta · 2022-09-06T14:43:11Z

avalanche/training/templates/update_type/meta_update.py

+
+
+class MetaUpdate:
+    def training_epoch(self, **kwargs):


Can we use the same training_epoch method for all and define only the iteration here?

because the epoch is the same, what changes is the single iteration

AntonioCarta · 2022-09-06T14:46:53Z

@HamedHemati can you explain the difference between a template and an observation type? update and problem type are clear to me. I think there isn't a clear separation between template and observation.

…o new_templates

HamedHemati · 2022-09-30T15:00:46Z

Update:

Old template files are removed, and all imports are updated
_train_exp is removed from observation_typenow it's a part of the BasedSGDTemplate again
observation_type now only checks the model adaptation and optimizer resetting. The behavior is currently different for batch and online updates because of the experience boundary checks (currently specific to OnlineCLExperience)

…o new_templates � Conflicts: � examples/lamaml_cifar100.py

…o new_templates � Conflicts: � avalanche/evaluation/metrics/images_samples.py � avalanche/training/supervised/lamaml.py � avalanche/training/templates/base_online_sgd.py � avalanche/training/templates/base_sgd.py � avalanche/training/templates/online_supervised.py � avalanche/training/templates/supervised.py � examples/online_naive.py � tests/training/test_online_strategies.py

HamedHemati added 2 commits July 12, 2022 16:47

Initialize new template structure

136c69e

Syntax fix

38e0cef

AntonioCarta reviewed Jul 18, 2022

View reviewed changes

HamedHemati added 3 commits July 24, 2022 20:36

Update meta-learning template and add examples

346017b

Merge branch 'master' of https://github.com/ContinualAI/avalanche int…

28d8156

…o new_templates

Fix syntax

6c5bfb0

AntonioCarta reviewed Aug 1, 2022

View reviewed changes

Merge branch 'master' of https://github.com/ContinualAI/avalanche int…

23f7af1

…o new_templates

HamedHemati commented Aug 11, 2022

View reviewed changes

HamedHemati added 4 commits August 31, 2022 15:50

Merge branch 'master' of https://github.com/ContinualAI/avalanche int…

02b9267

…o new_templates

Remove NEW_ prefixes and split plugin templates

3e0c54a

Update meta-learning example w.r.t the new template

50f2683

Add model_adaptation to observation_type implementation

a66a6a3

AntonioCarta reviewed Sep 6, 2022

View reviewed changes

HamedHemati added 3 commits September 30, 2022 14:32

Merge branch 'master' of https://github.com/ContinualAI/avalanche int…

5efeefc

…o new_templates

Update imports and delete old template files

1cd1924

Update observation_type and base_sgd

aee1d2e

HamedHemati added 3 commits October 3, 2022 16:27

Update observation_type and base_sgd

678a234

Merge branch 'master' of https://github.com/ContinualAI/avalanche int…

f00fd4d

…o new_templates � Conflicts: � examples/lamaml_cifar100.py

AntonioCarta marked this pull request as ready for review October 20, 2022 12:24

AntonioCarta merged commit 4a54a2a into ContinualAI:master Oct 20, 2022

HamedHemati deleted the new_templates branch March 16, 2023 11:01


		# ==================================================================> NEW

		def maybe_adapt_model_and_make_optimizer(self):

		from avalanche.training.utils import trigger_plugins


		class BaseGeneralSGDTemplate(BaseTemplate):


		# ====================================================================> NEW

		def before_inner_updates(

New mixin-based template structure #1092

New mixin-based template structure #1092

Conversation

HamedHemati commented Jul 12, 2022

coveralls commented Jul 12, 2022 • edited

Pull Request Test Coverage Report for Build 3284238644

💛 - Coveralls

HamedHemati commented Jul 14, 2022 • edited

AntonioCarta commented Jul 18, 2022

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

HamedHemati commented Jul 24, 2022

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

HamedHemati Aug 11, 2022 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

HamedHemati left a comment

Choose a reason for hiding this comment

AntonioCarta commented Aug 23, 2022 • edited

HamedHemati commented Aug 31, 2022

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

AntonioCarta commented Sep 6, 2022

HamedHemati commented Sep 30, 2022

coveralls commented Jul 12, 2022 •

edited

HamedHemati commented Jul 14, 2022 •

edited

HamedHemati Aug 11, 2022 •

edited

AntonioCarta commented Aug 23, 2022 •

edited