This repository has been archived by the owner on Sep 18, 2024. It is now read-only.
fix(speedup): refactor the execution logic of functions in speedup #5107
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
J-shang
reviewed
Sep 5, 2022
docs/source/compression/pruning.rst
Outdated
| @@ -105,4 +105,5 @@ In addition, for the convolutional layers that have more than one filter group, | |||
| ``dependency-aware pruner`` will also try to prune the same number of the channels for each filter group. | |||
| Overall, this pruner will prune the model according to the L1 norm of each filter and try to meet the topological constrains (channel dependency, etc) to improve the final speed gain after the speedup process. | |||
|
|
|||
| Operations that will be recognized as having channel dependencies: add/sub/mul/div, addcmul/addcdiv, logical_and/or/xor | |||
J-shang
reviewed
Sep 20, 2022
| @@ -725,8 +717,6 @@ def _build_graph(self): | |||
|
|
|||
| # associate module name with their trace graph nodes | |||
| for node in graph.nodes(): | |||
| if node.kind() == CONSTANT_KIND: | |||
| continue | |||
There was a problem hiding this comment.
when we should skip the CONSTANT_KIND and when we should not? I found you did not remove all the if node.kind() == CONSTANT_KIND: in the code
J-shang
reviewed
Sep 23, 2022
| if node.type == 'module': | ||
| inputs_name = node.inputs | ||
| else: | ||
| inputs_name = [val_node.debugName() for val_node in node.key_node.inputs()] |
There was a problem hiding this comment.
what is key_node? maybe we need a doc to explain the important attrs in a node
J-shang
reviewed
Sep 29, 2022
| elif isinstance(obj, dict): | ||
| return {k: recr_detacher(v) for k, v in obj.items()} | ||
| else: | ||
| return obj |
There was a problem hiding this comment.
can obj a customized data type with detach function?
try:
return obj.detach()
except AttributeError:
return obj
| @@ -297,16 +318,15 @@ def update_indirect_sparsity(self, node): | |||
| debug_name = auto_infer.input_debugname[in_id] | |||
|
|
|||
| last_output = self.internal_result[debug_name] | |||
| # if isinstance(last_output, torch.Tensor): | |||
| # TODO what if last output is tuple/list of tensor | |||
There was a problem hiding this comment.
so what if last output is tuple/list of tensor?
|
totally refactored old speedup in https://github.com/Louis-J/nni/tree/refactor_speedup |
|
the speedup v2 now is runnable. so better to reassess that refactor the old speedup is suitable or not. |
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
In speedup, we use a non-general trick to treat the prim ops. We analyze the relationship between the prim ops, then delete all the prim ops, finally connect the non-prim op directly. We also assume that all the intermediate variables are tensors, tuples of tensors, or lists of tensors. But it'll cause many problems in various models.
So, it's better to execute all the prim ops for real. In this pr, I remove the deletion of prim ops, and execute the prim recursively. And the assumption of variable types is also removed. Now it's closer to the original execution.
But it changed a lot and is untested now. More work is needed.
#5097
Test Options
Checklist
How to test