[PTQ][Torch][Native] Experimental tensor statistics migration

[daniil-lyakhov]

Changes

• Torch Init range moved to experimental TensorCollector
• Torch PTQ migrated to experimental TensorCollector
• Threesigma/MAD aggregators are implemented as experimental tensor collectors for Torch/OV backends
• no_outlier_map function refactored to not depend on shape of input tensor to make it possible to unify Torch and OV backends as they are using different shape statistics

Reason for changes

• To use all experimental reducers/aggregators (tensor collectors) in Torch/Torch PTQ backend

Related tickets

109665

Tests

• Tests are updated
• Test register_unnamed_inputs
• Test MedianAbsoluteDeviationAggregator/PrecentileAggregator/PostAggregateHook
• Test mixed min max TensorCollector creation function

TODO:

• Confirm that new no_outlier_map performs as good as previous implementation

[daniil-lyakhov]

@vshampor, @AlexanderDokuchaev, this PR is conflicting with PRs #2032, #2058. We should decide how we will be merging as all three PRs are huge

[vshampor]

@vshampor, @AlexanderDokuchaev, this PR is conflicting with PRs #2032, #2058. We should decide how we will be merging as all three PRs are huge

I closed the #2032, but there will be another one with the algo unification. You should proceed with merging this PR disregarding mine, though - I will rebase my changes accordingly.

[codecov]

Codecov Report

Merging #2117 (60c2804) into develop (b16d4a1) will decrease coverage by 0.11%.
Report is 2 commits behind head on develop.
The diff coverage is 39.16%.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #2117      +/-   ##
===========================================
- Coverage    36.24%   36.13%   -0.11%     
===========================================
  Files          477      480       +3     
  Lines        42570    42998     +428     
===========================================
+ Hits         15428    15539     +111     
- Misses       27142    27459     +317     

Files	Coverage Δ
nncf/common/tensor_statistics/collectors.py	77.18% <100.00%> (+1.18%)	⬆️
nncf/common/tensor_statistics/statistics.py	72.72% <100.00%> (+2.13%)	⬆️
nncf/onnx/graph/node_utils.py	97.77% <100.00%> (ø)
...antization/algorithms/bias_correction/algorithm.py	93.88% <ø> (ø)
...quantization/algorithms/bias_correction/backend.py	100.00% <ø> (ø)
...ization/algorithms/bias_correction/onnx_backend.py	97.72% <100.00%> (-0.03%)	⬇️
...antization/algorithms/channel_alignment/backend.py	100.00% <ø> (ø)
...ation/algorithms/fast_bias_correction/algorithm.py	90.83% <ø> (ø)
...ization/algorithms/fast_bias_correction/backend.py	100.00% <ø> (ø)
...on/algorithms/fast_bias_correction/onnx_backend.py	97.64% <100.00%> (-0.03%)	⬇️
... and 25 more

... and 21 files with indirect coverage changes

[daniil-lyakhov]

@vshampor, I've renamed percentile and have resolved all relative comments, thanks!

[vshampor]

Better now, still some changes required.

[alexsu52]

LGTM

[alexsu52]

Could you provide onnx e2e results?

[daniil-lyakhov]

Could you provide onnx e2e results?

onnx nightly e2e build 455 finished successfully