Support nested Flux Chains in LRP #119

adrhill · 2023-08-21T13:24:04Z

Closes #90.

This PR adds support for Flux models that are nested Chains containing other Chains.
It is a large step towards #10 and a v0.6 release that supports ResNets.

Changes:

refactors LRP rules, composites, model checks to support Chains containing other Chains.
introduces "functional"-like utilities on Flux Chains:
- wrapper types ChainTuple and ParallelTuple to maintain model structure on outputs without type piracy
- internal functions chainmap, chainzip, chainall, chainflatten, chainkeys in src/flux_chain_utils.jl
drop LinearAlgebra and PrettyTables.jl dependencies

Breaking changes:

LRP keyword argument is_flat=false is now called flatten=true
removed check_model (replaced by non-exported check_lrp_compat)
lrp! calls now take an extra argument layer, avoiding copies of unmodified layers
removed composite LastNTypeRule
removed internal functions head_tail, collect_activations, has_activation_fn

removes `LastNTypeRule`, which is of limited use

flattens both model and rules by default.

This is required for chains of chains. It also allows for small performance optimizations, e.g. defining `modify_layer(::Rule, layer) = nothing` for rules without layer modifications.

Remove duplicate include of `test/vgg11.jl`

- also supports Parallel layers - add tests

codecov · 2023-08-21T13:35:52Z

Codecov Report

Merging #119 (a871756) into master (ee31b37) will increase coverage by 0.89%.
Report is 2 commits behind head on master.
The diff coverage is 88.84%.

@@            Coverage Diff             @@
##           master     #119      +/-   ##
==========================================
+ Coverage   92.29%   93.18%   +0.89%     
==========================================
  Files          17       18       +1     
  Lines         558      646      +88     
==========================================
+ Hits          515      602      +87     
- Misses         43       44       +1

Files Changed	Coverage Δ
src/ExplainableAI.jl	`100.00% <ø> (ø)`
src/analyze_api.jl	`100.00% <ø> (ø)`
src/lrp/composite_presets.jl	`100.00% <ø> (ø)`
src/lrp/checks.jl	`80.00% <73.33%> (+33.84%)`	⬆️
src/flux_chain_utils.jl	`82.95% <82.95%> (ø)`
src/flux_layer_utils.jl	`86.66% <86.66%> (ø)`
src/lrp/rules.jl	`93.19% <87.50%> (-2.67%)`	⬇️
src/lrp/lrp.jl	`96.42% <94.44%> (-3.58%)`	⬇️
src/lrp/show.jl	`97.80% <95.83%> (-2.20%)`	⬇️
src/lrp/canonize.jl	`96.42% <100.00%> (-0.13%)`	⬇️
... and 1 more

Fixes `MethodError: no method matching iterate(::Base.Iterators.Reverse{ChainTuple{Tuple{ZeroRule}}})`

github-actions · 2023-08-21T14:51:42Z

Benchmark result

Judge result

Benchmark Report for /home/runner/work/ExplainableAI.jl/ExplainableAI.jl

Job Properties

Time of benchmarks:
- Target: 21 Aug 2023 - 14:45
- Baseline: 21 Aug 2023 - 14:51
Package commits:
- Target: 42bded
- Baseline: 10a5e0
Julia commits:
- Target: e4ee48
- Baseline: e4ee48
Julia command flags:
- Target: None
- Baseline: None
Environment variables:
- Target: None
- Baseline: None

Results

A ratio greater than 1.0 denotes a possible regression (marked with ❌), while a ratio less
than 1.0 denotes a possible improvement (marked with ✅). Only significant results - results
that indicate possible regressions or improvements - are shown below (thus, an empty table means that all
benchmark results remained invariant between builds).

ID	time ratio	memory ratio
`["VGG11", "Gradient", "construct analyzer"]`	0.77 (5%) ✅	0.70 (1%) ✅
`["VGG11", "InputTimesGradient", "analyze"]`	0.05 (5%) ✅	0.71 (1%) ✅
`["VGG11", "InputTimesGradient", "construct analyzer"]`	0.83 (5%) ✅	0.70 (1%) ✅
`["VGG11", "IntegratedGradients", "construct analyzer"]`	0.83 (5%) ✅	0.64 (1%) ✅
`["VGG11", "LREpsilonPlusFlat", "analyze"]`	0.55 (5%) ✅	0.84 (1%) ✅
`["VGG11", "LREpsilonPlusFlat", "construct analyzer"]`	984.46 (5%) ❌	1176.24 (1%) ❌
`["VGG11", "LRP", "analyze"]`	1.01 (5%)	1.08 (1%) ❌
`["VGG11", "LRP", "construct analyzer"]`	12.72 (5%) ❌	2.89 (1%) ❌
`["VGG11", "SmoothGrad", "construct analyzer"]`	0.86 (5%) ✅	0.64 (1%) ✅
`["apply rule", "AlphaBetaRule", "Conv"]`	1.11 (5%) ❌	1.00 (1%)
`["apply rule", "AlphaBetaRule", "Dense"]`	0.70 (5%) ✅	1.00 (1%)
`["apply rule", "EpsilonRule", "Conv"]`	1.14 (5%) ❌	1.00 (1%)
`["apply rule", "FlatRule", "Conv"]`	1.06 (5%) ❌	1.00 (1%)
`["apply rule", "GammaRule", "Conv"]`	1.10 (5%) ❌	1.00 (1%)
`["apply rule", "WSquareRule", "Conv"]`	1.08 (5%) ❌	1.00 (1%)
`["apply rule", "WSquareRule", "Dense"]`	1.08 (5%) ❌	1.00 (1%)
`["apply rule", "ZBoxRule", "Conv"]`	1.06 (5%) ❌	1.00 (1%)
`["apply rule", "ZPlusRule", "Conv"]`	1.08 (5%) ❌	1.00 (1%)
`["apply rule", "ZPlusRule", "Dense"]`	1.08 (5%) ❌	1.00 (1%)
`["apply rule", "ZeroRule", "Conv"]`	1.09 (5%) ❌	1.00 (1%)
`["apply rule", "ZeroRule", "Dense"]`	1.11 (5%) ❌	1.00 (1%)
`["modify layer", "AlphaBetaRule", "Dense"]`	0.90 (5%) ✅	1.00 (1%)
`["modify layer", "EpsilonRule", "Conv"]`	0.47 (5%) ✅	1.00 (1%)
`["modify layer", "EpsilonRule", "Dense"]`	0.11 (5%) ✅	1.00 (1%)
`["modify layer", "FlatRule", "Conv"]`	1.05 (5%) ❌	1.00 (1%)
`["modify layer", "FlatRule", "Dense"]`	1.39 (5%) ❌	1.00 (1%)
`["modify layer", "GammaRule", "Conv"]`	1.18 (5%) ❌	1.00 (1%)
`["modify layer", "GammaRule", "Dense"]`	1.73 (5%) ❌	1.00 (1%)
`["modify layer", "WSquareRule", "Conv"]`	1.06 (5%) ❌	1.00 (1%)
`["modify layer", "WSquareRule", "Dense"]`	1.23 (5%) ❌	1.00 (1%)
`["modify layer", "ZBoxRule", "Conv"]`	1.06 (5%) ❌	1.00 (1%)
`["modify layer", "ZBoxRule", "Dense"]`	2.01 (5%) ❌	1.00 (1%)
`["modify layer", "ZPlusRule", "Dense"]`	1.80 (5%) ❌	1.00 (1%)
`["modify layer", "ZeroRule", "Conv"]`	0.43 (5%) ✅	1.00 (1%)
`["modify layer", "ZeroRule", "Dense"]`	0.12 (5%) ✅	1.00 (1%)

Benchmark Group List

Here's a list of all the benchmark groups executed by this job:

["VGG11", "Gradient"]
["VGG11", "InputTimesGradient"]
["VGG11", "IntegratedGradients"]
["VGG11", "LREpsilonPlusFlat"]
["VGG11", "LRP"]
["VGG11", "SmoothGrad"]
["apply rule", "AlphaBetaRule"]
["apply rule", "EpsilonRule"]
["apply rule", "FlatRule"]
["apply rule", "GammaRule"]
["apply rule", "WSquareRule"]
["apply rule", "ZBoxRule"]
["apply rule", "ZPlusRule"]
["apply rule", "ZeroRule"]
["modify layer", "AlphaBetaRule"]
["modify layer", "EpsilonRule"]
["modify layer", "FlatRule"]
["modify layer", "GammaRule"]
["modify layer", "WSquareRule"]
["modify layer", "ZBoxRule"]
["modify layer", "ZPlusRule"]
["modify layer", "ZeroRule"]

Julia versioninfo

Target

Julia Version 1.9.2
Commit e4ee485e909 (2023-07-05 09:39 UTC)
Platform Info:
  OS: Linux (x86_64-linux-gnu)
      Ubuntu 22.04.3 LTS
  uname: Linux 5.15.0-1041-azure #48-Ubuntu SMP Tue Jun 20 20:34:08 UTC 2023 x86_64 x86_64
  CPU: Intel(R) Xeon(R) CPU E5-2673 v4 @ 2.30GHz: 
              speed         user         nice          sys         idle          irq
       #1  2294 MHz       9868 s          0 s        585 s        776 s          0 s
       #2  2294 MHz       2967 s          0 s        506 s       7670 s          0 s
  Memory: 6.7694854736328125 GB (2927.0390625 MB free)
  Uptime: 1132.17 sec
  Load Avg:  1.02  1.13  1.0
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-14.0.6 (ORCJIT, broadwell)
  Threads: 1 on 2 virtual cores

Baseline

Julia Version 1.9.2
Commit e4ee485e909 (2023-07-05 09:39 UTC)
Platform Info:
  OS: Linux (x86_64-linux-gnu)
      Ubuntu 22.04.3 LTS
  uname: Linux 5.15.0-1041-azure #48-Ubuntu SMP Tue Jun 20 20:34:08 UTC 2023 x86_64 x86_64
  CPU: Intel(R) Xeon(R) CPU E5-2673 v4 @ 2.30GHz: 
              speed         user         nice          sys         idle          irq
       #1  2294 MHz      13270 s          0 s        658 s       1169 s          0 s
       #2  2294 MHz       3388 s          0 s        589 s      11028 s          0 s
  Memory: 6.7694854736328125 GB (3445.84765625 MB free)
  Uptime: 1519.92 sec
  Load Avg:  1.16  1.09  1.01
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-14.0.6 (ORCJIT, broadwell)
  Threads: 1 on 2 virtual cores

Target result

Benchmark Report for /home/runner/work/ExplainableAI.jl/ExplainableAI.jl

Job Properties

Time of benchmark: 21 Aug 2023 - 14:45
Package commit: 42bded
Julia commit: e4ee48
Julia command flags: None
Environment variables: None

Results

Below is a table of this job's results, obtained by running the benchmarks.
The values listed in the ID column have the structure [parent_group, child_group, ..., key], and can be used to
index into the BaseBenchmarks suite to retrieve the corresponding benchmarks.
The percentages accompanying time and memory values in the below table are noise tolerances. The "true"
time/memory value for a given benchmark is expected to fall within this percentage of the reported value.
An empty cell means that the value was zero.

ID	time	GC time	memory	allocations
`["VGG11", "Gradient", "analyze"]`	2.501 s (5%)	10.476 ms	1.10 GiB (1%)	1758
`["VGG11", "Gradient", "construct analyzer"]`	10.200 μs (5%)		8.02 KiB (1%)	157
`["VGG11", "InputTimesGradient", "analyze"]`	2.508 s (5%)	10.666 ms	1.11 GiB (1%)	1761
`["VGG11", "InputTimesGradient", "construct analyzer"]`	10.000 μs (5%)		8.02 KiB (1%)	157
`["VGG11", "IntegratedGradients", "analyze"]`	12.318 s (5%)	74.489 ms	5.30 GiB (1%)	7508
`["VGG11", "IntegratedGradients", "construct analyzer"]`	9.800 μs (5%)		6.05 KiB (1%)	155
`["VGG11", "LREpsilonPlusFlat", "analyze"]`	3.274 s (5%)	16.755 ms	1.08 GiB (1%)	2740
`["VGG11", "LREpsilonPlusFlat", "construct analyzer"]`	36.328 ms (5%)		70.52 MiB (1%)	1257
`["VGG11", "LRP", "analyze"]`	2.274 s (5%)	17.626 ms	761.91 MiB (1%)	1881
`["VGG11", "LRP", "construct analyzer"]`	183.102 μs (5%)		124.66 KiB (1%)	868
`["VGG11", "SmoothGrad", "analyze"]`	12.265 s (5%)	77.189 ms	5.30 GiB (1%)	7463
`["VGG11", "SmoothGrad", "construct analyzer"]`	9.000 μs (5%)		6.05 KiB (1%)	155
`["apply rule", "AlphaBetaRule", "Conv"]`	3.821 ms (5%)		5.48 MiB (1%)	506
`["apply rule", "AlphaBetaRule", "Dense"]`	644.507 μs (5%)		800.94 KiB (1%)	38
`["apply rule", "EpsilonRule", "Conv"]`	930.909 μs (5%)		1.36 MiB (1%)	127
`["apply rule", "EpsilonRule", "Dense"]`	78.600 μs (5%)		1.45 KiB (1%)	3
`["apply rule", "FlatRule", "Conv"]`	958.709 μs (5%)		1.41 MiB (1%)	130
`["apply rule", "FlatRule", "Dense"]`	93.306 ns (5%)
`["apply rule", "GammaRule", "Conv"]`	949.710 μs (5%)		1.36 MiB (1%)	127
`["apply rule", "GammaRule", "Dense"]`	71.000 μs (5%)		1.45 KiB (1%)	3
`["apply rule", "WSquareRule", "Conv"]`	964.306 μs (5%)		1.41 MiB (1%)	130
`["apply rule", "WSquareRule", "Dense"]`	25.300 μs (5%)		576 bytes (1%)	5
`["apply rule", "ZBoxRule", "Conv"]`	2.835 ms (5%)		4.12 MiB (1%)	381
`["apply rule", "ZBoxRule", "Dense"]`	482.305 μs (5%)		602.48 KiB (1%)	32
`["apply rule", "ZPlusRule", "Conv"]`	1.876 ms (5%)		2.79 MiB (1%)	257
`["apply rule", "ZPlusRule", "Dense"]`	334.404 μs (5%)		402.55 KiB (1%)	21
`["apply rule", "ZeroRule", "Conv"]`	949.009 μs (5%)		1.36 MiB (1%)	127
`["apply rule", "ZeroRule", "Dense"]`	80.200 μs (5%)		1.45 KiB (1%)	3
`["modify layer", "AlphaBetaRule", "Conv"]`	823.077 ns (5%)		1.56 KiB (1%)	12
`["modify layer", "AlphaBetaRule", "Dense"]`	116.801 μs (5%)		783.38 KiB (1%)	12
`["modify layer", "EpsilonRule", "Conv"]`	1.600 ns (5%)
`["modify layer", "EpsilonRule", "Dense"]`	1.400 ns (5%)
`["modify layer", "FlatRule", "Conv"]`	128.447 ns (5%)		400 bytes (1%)	3
`["modify layer", "FlatRule", "Dense"]`	27.600 μs (5%)		195.84 KiB (1%)	3
`["modify layer", "GammaRule", "Conv"]`	314.889 ns (5%)		400 bytes (1%)	3
`["modify layer", "GammaRule", "Dense"]`	34.001 μs (5%)		195.84 KiB (1%)	3
`["modify layer", "WSquareRule", "Conv"]`	236.809 ns (5%)		400 bytes (1%)	3
`["modify layer", "WSquareRule", "Dense"]`	33.001 μs (5%)		195.84 KiB (1%)	3
`["modify layer", "ZBoxRule", "Conv"]`	504.597 ns (5%)		800 bytes (1%)	6
`["modify layer", "ZBoxRule", "Dense"]`	63.800 μs (5%)		391.69 KiB (1%)	6
`["modify layer", "ZPlusRule", "Conv"]`	505.533 ns (5%)		800 bytes (1%)	6
`["modify layer", "ZPlusRule", "Dense"]`	63.201 μs (5%)		391.69 KiB (1%)	6
`["modify layer", "ZeroRule", "Conv"]`	1.600 ns (5%)
`["modify layer", "ZeroRule", "Dense"]`	1.600 ns (5%)

Benchmark Group List

Here's a list of all the benchmark groups executed by this job:

["VGG11", "Gradient"]
["VGG11", "InputTimesGradient"]
["VGG11", "IntegratedGradients"]
["VGG11", "LREpsilonPlusFlat"]
["VGG11", "LRP"]
["VGG11", "SmoothGrad"]
["apply rule", "AlphaBetaRule"]
["apply rule", "EpsilonRule"]
["apply rule", "FlatRule"]
["apply rule", "GammaRule"]
["apply rule", "WSquareRule"]
["apply rule", "ZBoxRule"]
["apply rule", "ZPlusRule"]
["apply rule", "ZeroRule"]
["modify layer", "AlphaBetaRule"]
["modify layer", "EpsilonRule"]
["modify layer", "FlatRule"]
["modify layer", "GammaRule"]
["modify layer", "WSquareRule"]
["modify layer", "ZBoxRule"]
["modify layer", "ZPlusRule"]
["modify layer", "ZeroRule"]

Julia versioninfo

Julia Version 1.9.2
Commit e4ee485e909 (2023-07-05 09:39 UTC)
Platform Info:
  OS: Linux (x86_64-linux-gnu)
      Ubuntu 22.04.3 LTS
  uname: Linux 5.15.0-1041-azure #48-Ubuntu SMP Tue Jun 20 20:34:08 UTC 2023 x86_64 x86_64
  CPU: Intel(R) Xeon(R) CPU E5-2673 v4 @ 2.30GHz: 
              speed         user         nice          sys         idle          irq
       #1  2294 MHz       9868 s          0 s        585 s        776 s          0 s
       #2  2294 MHz       2967 s          0 s        506 s       7670 s          0 s
  Memory: 6.7694854736328125 GB (2927.0390625 MB free)
  Uptime: 1132.17 sec
  Load Avg:  1.02  1.13  1.0
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-14.0.6 (ORCJIT, broadwell)
  Threads: 1 on 2 virtual cores

Baseline result

Benchmark Report for /home/runner/work/ExplainableAI.jl/ExplainableAI.jl

Job Properties

Time of benchmark: 21 Aug 2023 - 14:51
Package commit: 10a5e0
Julia commit: e4ee48
Julia command flags: None
Environment variables: None

Results

Below is a table of this job's results, obtained by running the benchmarks.
The values listed in the ID column have the structure [parent_group, child_group, ..., key], and can be used to
index into the BaseBenchmarks suite to retrieve the corresponding benchmarks.
The percentages accompanying time and memory values in the below table are noise tolerances. The "true"
time/memory value for a given benchmark is expected to fall within this percentage of the reported value.
An empty cell means that the value was zero.

ID	time	GC time	memory	allocations
`["VGG11", "Gradient", "analyze"]`	2.563 s (5%)	47.594 ms	1.10 GiB (1%)	1950
`["VGG11", "Gradient", "construct analyzer"]`	13.200 μs (5%)		11.45 KiB (1%)	200
`["VGG11", "InputTimesGradient", "analyze"]`	46.080 s (5%)	476.754 ms	1.55 GiB (1%)	8315209
`["VGG11", "InputTimesGradient", "construct analyzer"]`	12.100 μs (5%)		11.45 KiB (1%)	200
`["VGG11", "IntegratedGradients", "analyze"]`	12.544 s (5%)	134.417 ms	5.30 GiB (1%)	8468
`["VGG11", "IntegratedGradients", "construct analyzer"]`	11.800 μs (5%)		9.48 KiB (1%)	198
`["VGG11", "LREpsilonPlusFlat", "analyze"]`	5.965 s (5%)	167.849 ms	1.29 GiB (1%)	2272404
`["VGG11", "LREpsilonPlusFlat", "construct analyzer"]`	36.901 μs (5%)		61.39 KiB (1%)	123
`["VGG11", "LRP", "analyze"]`	2.247 s (5%)	20.995 ms	704.91 MiB (1%)	1821
`["VGG11", "LRP", "construct analyzer"]`	14.400 μs (5%)		43.09 KiB (1%)	82
`["VGG11", "SmoothGrad", "analyze"]`	12.818 s (5%)	221.714 ms	5.30 GiB (1%)	55967
`["VGG11", "SmoothGrad", "construct analyzer"]`	10.500 μs (5%)		9.48 KiB (1%)	198
`["apply rule", "AlphaBetaRule", "Conv"]`	3.438 ms (5%)		5.48 MiB (1%)	506
`["apply rule", "AlphaBetaRule", "Dense"]`	919.208 μs (5%)		800.94 KiB (1%)	38
`["apply rule", "EpsilonRule", "Conv"]`	816.002 μs (5%)		1.36 MiB (1%)	127
`["apply rule", "EpsilonRule", "Dense"]`	76.701 μs (5%)		1.45 KiB (1%)	3
`["apply rule", "FlatRule", "Conv"]`	907.208 μs (5%)		1.41 MiB (1%)	130
`["apply rule", "FlatRule", "Dense"]`	96.445 ns (5%)
`["apply rule", "GammaRule", "Conv"]`	859.508 μs (5%)		1.36 MiB (1%)	127
`["apply rule", "GammaRule", "Dense"]`	71.401 μs (5%)		1.45 KiB (1%)	3
`["apply rule", "WSquareRule", "Conv"]`	891.508 μs (5%)		1.41 MiB (1%)	130
`["apply rule", "WSquareRule", "Dense"]`	23.500 μs (5%)		576 bytes (1%)	5
`["apply rule", "ZBoxRule", "Conv"]`	2.674 ms (5%)		4.12 MiB (1%)	381
`["apply rule", "ZBoxRule", "Dense"]`	500.405 μs (5%)		601.52 KiB (1%)	30
`["apply rule", "ZPlusRule", "Conv"]`	1.745 ms (5%)		2.79 MiB (1%)	257
`["apply rule", "ZPlusRule", "Dense"]`	309.003 μs (5%)		402.55 KiB (1%)	21
`["apply rule", "ZeroRule", "Conv"]`	870.507 μs (5%)		1.36 MiB (1%)	127
`["apply rule", "ZeroRule", "Dense"]`	72.501 μs (5%)		1.45 KiB (1%)	3
`["modify layer", "AlphaBetaRule", "Conv"]`	844.231 ns (5%)		1.56 KiB (1%)	12
`["modify layer", "AlphaBetaRule", "Dense"]`	130.301 μs (5%)		783.38 KiB (1%)	12
`["modify layer", "EpsilonRule", "Conv"]`	3.400 ns (5%)
`["modify layer", "EpsilonRule", "Dense"]`	13.100 ns (5%)
`["modify layer", "FlatRule", "Conv"]`	122.272 ns (5%)		400 bytes (1%)	3
`["modify layer", "FlatRule", "Dense"]`	19.801 μs (5%)		195.84 KiB (1%)	3
`["modify layer", "GammaRule", "Conv"]`	266.416 ns (5%)		400 bytes (1%)	3
`["modify layer", "GammaRule", "Dense"]`	19.601 μs (5%)		195.84 KiB (1%)	3
`["modify layer", "WSquareRule", "Conv"]`	222.619 ns (5%)		400 bytes (1%)	3
`["modify layer", "WSquareRule", "Dense"]`	26.900 μs (5%)		195.84 KiB (1%)	3
`["modify layer", "ZBoxRule", "Conv"]`	474.495 ns (5%)		800 bytes (1%)	6
`["modify layer", "ZBoxRule", "Dense"]`	31.700 μs (5%)		391.69 KiB (1%)	6
`["modify layer", "ZPlusRule", "Conv"]`	493.472 ns (5%)		800 bytes (1%)	6
`["modify layer", "ZPlusRule", "Dense"]`	35.200 μs (5%)		391.69 KiB (1%)	6
`["modify layer", "ZeroRule", "Conv"]`	3.700 ns (5%)
`["modify layer", "ZeroRule", "Dense"]`	13.100 ns (5%)

Benchmark Group List

Here's a list of all the benchmark groups executed by this job:

["VGG11", "Gradient"]
["VGG11", "InputTimesGradient"]
["VGG11", "IntegratedGradients"]
["VGG11", "LREpsilonPlusFlat"]
["VGG11", "LRP"]
["VGG11", "SmoothGrad"]
["apply rule", "AlphaBetaRule"]
["apply rule", "EpsilonRule"]
["apply rule", "FlatRule"]
["apply rule", "GammaRule"]
["apply rule", "WSquareRule"]
["apply rule", "ZBoxRule"]
["apply rule", "ZPlusRule"]
["apply rule", "ZeroRule"]
["modify layer", "AlphaBetaRule"]
["modify layer", "EpsilonRule"]
["modify layer", "FlatRule"]
["modify layer", "GammaRule"]
["modify layer", "WSquareRule"]
["modify layer", "ZBoxRule"]
["modify layer", "ZPlusRule"]
["modify layer", "ZeroRule"]

Julia versioninfo

Julia Version 1.9.2
Commit e4ee485e909 (2023-07-05 09:39 UTC)
Platform Info:
  OS: Linux (x86_64-linux-gnu)
      Ubuntu 22.04.3 LTS
  uname: Linux 5.15.0-1041-azure #48-Ubuntu SMP Tue Jun 20 20:34:08 UTC 2023 x86_64 x86_64
  CPU: Intel(R) Xeon(R) CPU E5-2673 v4 @ 2.30GHz: 
              speed         user         nice          sys         idle          irq
       #1  2294 MHz      13270 s          0 s        658 s       1169 s          0 s
       #2  2294 MHz       3388 s          0 s        589 s      11028 s          0 s
  Memory: 6.7694854736328125 GB (3445.84765625 MB free)
  Uptime: 1519.92 sec
  Load Avg:  1.16  1.09  1.01
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-14.0.6 (ORCJIT, broadwell)
  Threads: 1 on 2 virtual cores

Runtime information

Runtime Info
BLAS #threads	1
`BLAS.vendor()`	`lbt`
`Sys.CPU_THREADS`	2

lscpu output:

Architecture:                    x86_64
CPU op-mode(s):                  32-bit, 64-bit
Address sizes:                   46 bits physical, 48 bits virtual
Byte Order:                      Little Endian
CPU(s):                          2
On-line CPU(s) list:             0,1
Vendor ID:                       GenuineIntel
Model name:                      Intel(R) Xeon(R) CPU E5-2673 v4 @ 2.30GHz
CPU family:                      6
Model:                           79
Thread(s) per core:              1
Core(s) per socket:              2
Socket(s):                       1
Stepping:                        1
BogoMIPS:                        4589.37
Flags:                           fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology cpuid pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch invpcid_single pti fsgsbase bmi1 hle avx2 smep bmi2 erms invpcid rtm rdseed adx smap xsaveopt md_clear
Hypervisor vendor:               Microsoft
Virtualization type:             full
L1d cache:                       64 KiB (2 instances)
L1i cache:                       64 KiB (2 instances)
L2 cache:                        512 KiB (2 instances)
L3 cache:                        50 MiB (1 instance)
NUMA node(s):                    1
NUMA node0 CPU(s):               0,1
Vulnerability Itlb multihit:     KVM: Mitigation: VMX unsupported
Vulnerability L1tf:              Mitigation; PTE Inversion
Vulnerability Mds:               Mitigation; Clear CPU buffers; SMT Host state unknown
Vulnerability Meltdown:          Mitigation; PTI
Vulnerability Mmio stale data:   Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown
Vulnerability Retbleed:          Not affected
Vulnerability Spec store bypass: Vulnerable
Vulnerability Spectre v1:        Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:        Mitigation; Retpolines, STIBP disabled, RSB filling, PBRSB-eIBRS Not affected
Vulnerability Srbds:             Not affected
Vulnerability Tsx async abort:   Mitigation; Clear CPU buffers; SMT Host state unknown

Cpu Property	Value
Brand	Intel(R) Xeon(R) CPU E5-2673 v4 @ 2.30GHz
Vendor	:Intel
Architecture	:Broadwell
Model	Family: 0x06, Model: 0x4f, Stepping: 0x01, Type: 0x00
Cores	2 physical cores, 2 logical cores (on executing CPU)
	No Hyperthreading hardware capability detected
Clock Frequencies	Not supported by CPU
Data Cache	Level 1:3 : (32, 256, 51200) kbytes
	64 byte cache line size
Address Size	48 bits virtual, 46 bits physical
SIMD	256 bit = 32 byte max. SIMD vector size
Time Stamp Counter	TSC is accessible via `rdtsc`
	TSC increased at every clock cycle (non-invariant TSC)
Perf. Monitoring	Performance Monitoring Counters (PMC) are not supported
Hypervisor	Yes, Microsoft

adrhill added 30 commits January 11, 2023 21:07

Draft adding activation collection on non-sequential Flux models

b5b1354

Add ChainTuple, ParallelTuple and chainmap

8d7e89e

Add connection to ParallelTuple, update chainmap

aeab0c7

Add collect_activations

50c6629

Add and update tests

1353c1a

Fix for Chains containing Arrays of layers

82c127f

Restructure code, fix strip_softmax

1bd31e1

Remove flatten_model

1401e6c

Merge branch 'master' into ah/chain-of-chains

4ecaa61

Fix activation_fn

1207d8c

Simplify chainmap

2e4efbd

Add chainzip, remove head_tail, collect_activations

b838792

Add back better flatten_model

27f9783

Refactor composites using objectid based lambda functions

778a2c1

Refactor composite application, get_modified_layers

df97657

Fix composites

d633217

Add multi argument chainzip

d1bf185

Add lrp_rules

f06fe7b

Add temporary TODOs before release

ac2b095

Add chainkeys

943f053

Refactor composite primitives (again)

c6253a6

removes `LastNTypeRule`, which is of limited use

Add pretty printing for LRP analyzers with Chains / Parallel layers

cfdd8fa

Prepare Parallel layer support, add identity to composite presets

412430a

Help LSP / linter by not generating ChainTuple/ParallelTuple structs

e060ff7

Delete id_list as chainkeys replaces it

341908a

Add chainflatten, generalizing flatten_model to Tuple structs

ed5378a

Add kwarg flatten=true to LRP analyzer constructor

d8c0e1e

flattens both model and rules by default.

Linting and formatting fixes

637856a

Add modified_layers field to LRP analyzer

a6405ba

Refactor lrp!, adding layer argument

633b37e

This is required for chains of chains. It also allows for small performance optimizations, e.g. defining `modify_layer(::Rule, layer) = nothing` for rules without layer modifications.

adrhill added 10 commits August 16, 2023 19:34

Support chains of chains

4b107c0

Update tests, set models and dropout layers to testmode!

b9c0c45

Remove duplicate include of `test/vgg11.jl`

Fix chainflatten on singleton chains

b50525f

Formatting fixes

adb6280

Test LRP on chain of chains

9e8fc70

Refactor LRP model tests, deprecate check_model

32e3e77

Add chainall

fbd9bfa

Remove unusedhas_activation_fn

b495509

Refactor model checks for chains of chains

3c05d81

- also supports Parallel layers - add tests

Add LRP supported activations functions to supported layers

f7a826e

adrhill added 3 commits August 21, 2023 16:01

Fix for Julia 1.6 compat

0cd4d30

Fixes `MethodError: no method matching iterate(::Base.Iterators.Reverse{ChainTuple{Tuple{ZeroRule}}})`

Increase code coverage

430ca3b

Drop LinearAlgebra and PrettyTables dependencies

a871756

adrhill added the run benchmark Run benchmark CI action label Aug 21, 2023

adrhill merged commit 6e12cd2 into master Aug 21, 2023
7 of 8 checks passed

adrhill deleted the ah/chain-of-chains branch August 21, 2023 14:49

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Support nested Flux Chains in LRP #119

Support nested Flux Chains in LRP #119

adrhill commented Aug 21, 2023 •

edited

codecov bot commented Aug 21, 2023 •

edited

github-actions bot commented Aug 21, 2023

Judge result

Benchmark Report for /home/runner/work/ExplainableAI.jl/ExplainableAI.jl

Job Properties

Results

Benchmark Group List

Julia versioninfo

Target

Baseline

Target result

Benchmark Report for /home/runner/work/ExplainableAI.jl/ExplainableAI.jl

Job Properties

Results

Benchmark Group List

Julia versioninfo

Baseline result

Benchmark Report for /home/runner/work/ExplainableAI.jl/ExplainableAI.jl

Job Properties

Results

Benchmark Group List

Julia versioninfo

Runtime information

Support nested Flux Chains in LRP #119

Support nested Flux Chains in LRP #119

Conversation

adrhill commented Aug 21, 2023 • edited

codecov bot commented Aug 21, 2023 • edited

Codecov Report

github-actions bot commented Aug 21, 2023

Judge result

Benchmark Report for /home/runner/work/ExplainableAI.jl/ExplainableAI.jl

Job Properties

Results

Benchmark Group List

Julia versioninfo

Target

Baseline

Target result

Benchmark Report for /home/runner/work/ExplainableAI.jl/ExplainableAI.jl

Job Properties

Results

Benchmark Group List

Julia versioninfo

Baseline result

Benchmark Report for /home/runner/work/ExplainableAI.jl/ExplainableAI.jl

Job Properties

Results

Benchmark Group List

Julia versioninfo

Runtime information

adrhill commented Aug 21, 2023 •

edited

codecov bot commented Aug 21, 2023 •

edited