V2 Model Compression experiment result #4365
Conversation
| @@ -55,6 +55,29 @@ User configuration for Level Pruner | |||
|
|
|||
| .. autoclass:: nni.algorithms.compression.v2.pytorch.pruning.LevelPruner | |||
|
|
|||
| Performance Test | |||
There was a problem hiding this comment.
all these performance numbers are obtained from our example code?
There was a problem hiding this comment.
Yes, the seed was set, and some hyperparameters may change.
There was a problem hiding this comment.
is it better to also provide the hyper-parameters that you use?
There was a problem hiding this comment.
Now this has been provided.
| * - Pruned VGG-16 | ||
| - 93.74% | ||
| - 14.98M | ||
| - 313.46M |
There was a problem hiding this comment.
why parameters and flops do not change in the pruned model?
There was a problem hiding this comment.
This is fine-grained algorithm, it has no speedup procedure.
| - | ||
| * - Pruned VGG-16 | ||
| - 87.24% | ||
| - 0.64M |
There was a problem hiding this comment.
why parameters are reduce so much with 80% sparsity?
There was a problem hiding this comment.
The pruning behavior may differ according to various default mode setting. It has three types: 'normal', 'global' and 'dependency_aware'. In the intermediate config_list we can see it does prune nearly 0.8 in each layer.
| * - Pruned VGG-16 | ||
| - 76.94% | ||
| - | ||
| - |
There was a problem hiding this comment.
why the two numbers are missing?
There was a problem hiding this comment.
Because the result was not saved at that time, now it has been added.
| * - Pruned VGG-16 | ||
| - 33.6% (wo FT) | ||
| - | ||
| - |
There was a problem hiding this comment.
what is the meaning of "wo FT", and there are also two missing numbers
There was a problem hiding this comment.
At first I want the result 'without finetune', but this time it has been normal.
Just like the workflow: prune -> speedup -> finetune
|
These results have already been updated. But the result has little difference, I think maybe we need a more significant benchmark... |
Description
Some simple tests to measure the performance of v2 pruner.
Checklist
How to test