Benchmark 2D filter

[alessandrofelder]

In order to make a documented and informed decision on a good long-term solution to balance maintainability and performance #85 #73, we should compare the 2d filtering between

• original cython implementation
• numba implementation
• pure python implementation

both in terms of memory usage and speed.

[alessandrofelder]

memory usage

original Cython implementation

numba

pure python

[alessandrofelder]

Speed

original Cython implementation

13.611 <module>  filter_2d.py:1
└─ 13.601 TileProcessor.get_tile_mask  cellfinder_core/detect/filters/plane/plane_filter.py:19
   ├─ 12.600 enhance_peaks  cellfinder_core/detect/filters/plane/classical_filter.py:6
   │  ├─ 8.883 medfilt2d  scipy/signal/_signaltools.py:1828
   │  │  └─ 8.883 _medfilt2d  None
   │  ├─ 2.087 gaussian_filter  scipy/ndimage/_filters.py:279
   │  │  └─ 2.087 gaussian_filter1d  scipy/ndimage/_filters.py:211
   │  │     └─ 2.087 correlate1d  scipy/ndimage/_filters.py:84
   │  │        └─ 2.087 correlate1d  None
   │  ├─ 1.184 laplace  scipy/ndimage/_filters.py:492
   │  │  └─ 1.164 generic_laplace  scipy/ndimage/_filters.py:450
   │  │     └─ 1.062 derivative2  scipy/ndimage/_filters.py:517
   │  │        └─ 1.062 correlate1d  scipy/ndimage/_filters.py:84
   │  │           └─ 1.062 correlate1d  None
   │  ├─ 0.227 ndarray.astype  None
   │  └─ 0.166 [self]  None
   ├─ 0.291 _std  numpy/core/_methods.py:263
   │  └─ 0.277 _var  numpy/core/_methods.py:196
   │     └─ 0.214 [self]  None
   ├─ 0.245 TileWalker.__init__  cellfinder_core/detect/filters/plane/tile_walker.py:11
   │  └─ 0.245 ndarray.copy  None
   └─ 0.225 TileWalker.walk_out_of_brain_only  cellfinder_core/detect/filters/plane/tile_walker.py:56

numba

Program: benchmarks/filter_2d.py

14.688 <module>  filter_2d.py:1
└─ 14.688 TileProcessor.get_tile_mask  cellfinder_core/detect/filters/plane/plane_filter.py:19
   ├─ 12.728 enhance_peaks  cellfinder_core/detect/filters/plane/classical_filter.py:6
   │  ├─ 8.966 medfilt2d  scipy/signal/_signaltools.py:1828
   │  │  └─ 8.966 _medfilt2d  None
   │  ├─ 2.113 gaussian_filter  scipy/ndimage/_filters.py:279
   │  │  └─ 2.113 gaussian_filter1d  scipy/ndimage/_filters.py:211
   │  │     └─ 2.113 correlate1d  scipy/ndimage/_filters.py:84
   │  │        └─ 2.113 correlate1d  None
   │  ├─ 1.195 laplace  scipy/ndimage/_filters.py:492
   │  │  └─ 1.174 generic_laplace  scipy/ndimage/_filters.py:450
   │  │     └─ 1.083 derivative2  scipy/ndimage/_filters.py:517
   │  │        └─ 1.083 correlate1d  scipy/ndimage/_filters.py:84
   │  │           └─ 1.083 correlate1d  None
   │  ├─ 0.231 ndarray.astype  None
   │  └─ 0.169 [self]  None
   ├─ 1.187 TileWalker.walk_out_of_brain_only  cellfinder_core/detect/filters/plane/tile_walker.py:56
   │  ├─ 0.425 CPUDispatcher.typeof_pyval  numba/core/dispatcher.py:724
   │  │  └─ 0.366 typeof  numba/core/typing/typeof.py:32
   │  │     └─ 0.264 wrapper  functools.py:884
   │  │        └─ 0.191 typeof_impl  numba/core/typing/typeof.py:46
   │  ├─ 0.297 is_low_average  cellfinder_core/detect/filters/plane/base_tile_filter.py:37
   │  ├─ 0.218 CPUDispatcher._compile_for_args  numba/core/dispatcher.py:388
   │  │  └─ 0.213 CPUDispatcher.compile  numba/core/dispatcher.py:915
   │  │     └─ 0.213 _FunctionCompiler.compile  numba/core/dispatcher.py:124
   │  │        └─ 0.213 _FunctionCompiler._compile_cached  numba/core/dispatcher.py:131
   │  │           └─ 0.213 _FunctionCompiler._compile_core  numba/core/dispatcher.py:146
   │  │              └─ 0.213 compile_extra  numba/core/compiler.py:690
   │  └─ 0.149 [self]  None
   ├─ 0.293 _std  numpy/core/_methods.py:263
   │  └─ 0.277 _var  numpy/core/_methods.py:196
   │     └─ 0.214 [self]  None
   └─ 0.247 TileWalker.__init__  cellfinder_core/detect/filters/plane/tile_walker.py:11
      └─ 0.247 ndarray.copy  None

pure python

Program: benchmarks/filter_2d.py

34.908 <module>  filter_2d.py:1
└─ 34.897 TileProcessor.get_tile_mask  cellfinder_core/detect/filters/plane/plane_filter.py:19
   ├─ 21.426 TileWalker.walk_out_of_brain_only  cellfinder_core/detect/filters/plane/tile_walker.py:56
   │  └─ 21.413 OutOfBrainTileFilter.is_low_average  cellfinder_core/detect/filters/plane/base_tile_filter.py:37
   └─ 12.690 enhance_peaks  cellfinder_core/detect/filters/plane/classical_filter.py:6
      ├─ 8.898 medfilt2d  scipy/signal/_signaltools.py:1828
      │  └─ 8.898 _medfilt2d  None
      ├─ 2.169 gaussian_filter  scipy/ndimage/_filters.py:279
      │  └─ 2.169 gaussian_filter1d  scipy/ndimage/_filters.py:211
      │     └─ 2.169 correlate1d  scipy/ndimage/_filters.py:84
      │        └─ 2.169 correlate1d  None
      └─ 1.203 laplace  scipy/ndimage/_filters.py:492
         └─ 1.182 generic_laplace  scipy/ndimage/_filters.py:450
            └─ 1.064 derivative2  scipy/ndimage/_filters.py:517
               └─ 1.064 correlate1d  scipy/ndimage/_filters.py:84
                  └─ 1.064 correlate1d  None```

[alessandrofelder]

• Memory usage levels are comparable across the three implementations (and very similar between cython and numba)
• With a tweak of passing a tile to is_low_average numba is comparable in speed to cython (14.5 and 13.5 seconds) whereas pure python is much slower (~20times for the non-scipy part)

Executed on a plane containing np.random 10k x 10k 16bit integers on my SWC macbook.

[alessandrofelder]

This suggests numba is a good way forward, for the 2d filter at least.

[dstansby]

Fixed by #91