From 62a40ad97edb42d9030a7ef271656747d26cece0 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jure=20=C5=A0orn?= Date: Wed, 20 Jul 2022 15:24:11 +0200 Subject: [PATCH] A lot of changes in NumPy --- README.md | 73 ++++++++++++++++++++++++------------------------- index.html | 79 +++++++++++++++++++++++++++++------------------------- 2 files changed, 79 insertions(+), 73 deletions(-) diff --git a/README.md b/README.md index 072e9ed..a38526e 100644 --- a/README.md +++ b/README.md @@ -2559,18 +2559,10 @@ Profiling --------- ### Stopwatch ```python -from time import time -start_time = time() # Seconds since the Epoch. -... -duration = time() - start_time -``` - -#### High performance: -```python from time import perf_counter -start_time = perf_counter() # Seconds since the restart. +start_time = perf_counter() ... -duration = perf_counter() - start_time +duration_in_seconds = perf_counter() - start_time ``` ### Timing a Snippet @@ -2614,8 +2606,9 @@ Line # Mem usage Increment Line Contents ### Call Graph #### Generates a PNG image of the call graph with highlighted bottlenecks: ```python -# $ pip3 install pycallgraph2; brew/apt install graphviz +# $ pip3 install pycallgraph2; apt/brew install graphviz import pycallgraph2 as cg, datetime + filename = f'profile-{datetime.datetime.now():%Y%m%d%H%M%S}.png' drawer = cg.output.GraphvizOutput(output_file=filename) with cg.PyCallGraph(drawer): @@ -2633,62 +2626,70 @@ import numpy as np ``` ```python - = np.array() - = np.arange(from_inclusive, to_exclusive, ±step_size) - = np.ones() - = np.random.randint(from_inclusive, to_exclusive, ) + = np.array() # Returns 1d/2d NumPy array. + = np.zeros/ones() # Also np.full(, ). + = np.arange(from_inc, to_exc, ±step) # Also np.linspace(start, stop, num). + = np.random.randint(from_inc, to_exc, ) # Also np.random.random(). ``` ```python -.shape = - = .reshape() - = np.broadcast_to(, ) + = .reshape() # Also `.shape = `. + = .flatten() # Collapses array into one dimension. + = .squeeze() # Removes dimensions of length one. ``` ```python - = .sum(axis) -indexes = .argmin(axis) + = .sum/min/mean/var/std(axis) # Passed dimension gets aggregated. + = .argmin(axis) # Returns indexes of smallest elements. + = np.apply_along_axis(, axis, ) # Func can return a scalar or array. ``` -* **Shape is a tuple of dimension sizes.** -* **Axis is an index of the dimension that gets aggregated. Leftmost dimension has index 0.** +* **Shape is a tuple of dimension sizes. A 100x50 RGB image has shape (50, 100, 3).** +* **Axis is an index of the dimension that gets aggregated. Leftmost/outermost dimension has index 0. Summing a 100x50 RGB image along the axis 2 will return a greyscale image with shape (50, 100).** +* **Passing a tuple of axes will chain the operations like this: `'.(axis_1, keepdims=True).(axis_2).squeeze()'`.** ### Indexing ```bash - = <2d_array>[row_index, column_index] -<1d_view> = <2d_array>[row_index] -<1d_view> = <2d_array>[:, column_index] + = <2d_array>[row_index, column_index] # <3d_a>[table_i, row_i, column_i] +<1d_view> = <2d_array>[row_index] # <3d_a>[table_i, row_i] +<1d_view> = <2d_array>[:, column_index] # <3d_a>[table_i, :, column_i] ``` ```bash -<1d_array> = <2d_array>[row_indexes, column_indexes] -<2d_array> = <2d_array>[row_indexes] -<2d_array> = <2d_array>[:, column_indexes] +<1d_array> = <2d_array>[row_indexes, column_indexes] # <3d_a>[table_is, row_is, column_is] +<2d_array> = <2d_array>[row_indexes] # <3d_a>[table_is, row_is] +<2d_array> = <2d_array>[:, column_indexes] # <3d_a>[table_is, :, column_is] ``` ```bash -<2d_bools> = <2d_array> ><== -<1d_array> = <2d_array>[<2d_bools>] +<2d_bools> = <2d_array> ><== # <3d_array> ><== <1d_array> +<1d_array> = <2d_array>[<2d_bools>] # <3d_array>[<2d_bools>] ``` +* **All examples also allow assignments.** ### Broadcasting **Broadcasting is a set of rules by which NumPy functions operate on arrays of different sizes and/or dimensions.** ```python -left = [[0.1], [0.6], [0.8]] # Shape: (3, 1) -right = [ 0.1 , 0.6 , 0.8 ] # Shape: (3) +left = [[0.1], [0.6], [0.8]] # Shape: (3, 1) +right = [ 0.1 , 0.6 , 0.8 ] # Shape: (3,) ``` #### 1. If array shapes differ in length, left-pad the shorter shape with ones: ```python -left = [[0.1], [0.6], [0.8]] # Shape: (3, 1) -right = [[0.1 , 0.6 , 0.8]] # Shape: (1, 3) <- ! +left = [[0.1], [0.6], [0.8]] # Shape: (3, 1) +right = [[0.1 , 0.6 , 0.8]] # Shape: (1, 3) <- ! ``` #### 2. If any dimensions differ in size, expand the ones that have size 1 by duplicating their elements: ```python -left = [[0.1, 0.1, 0.1], [0.6, 0.6, 0.6], [0.8, 0.8, 0.8]] # Shape: (3, 3) <- ! -right = [[0.1, 0.6, 0.8], [0.1, 0.6, 0.8], [0.1, 0.6, 0.8]] # Shape: (3, 3) <- ! +left = [[0.1, 0.1, 0.1], # Shape: (3, 3) <- ! + [0.6, 0.6, 0.6], + [0.8, 0.8, 0.8]] + +right = [[0.1, 0.6, 0.8], # Shape: (3, 3) <- ! + [0.1, 0.6, 0.8], + [0.1, 0.6, 0.8]] ``` #### 3. If neither non-matching dimension has size 1, raise an error. diff --git a/index.html b/index.html index dd30471..66bd42a 100644 --- a/index.html +++ b/index.html @@ -54,7 +54,7 @@
- +
@@ -2096,19 +2096,13 @@ run(host='0.0.0.0', port='team': 'arsenal f.c.', 'odds': [2.09, 3.74, 3.68]} -

#Profiling

Stopwatch

from time import time
-start_time = time()                     # Seconds since the Epoch.
+
#Profiling
Stopwatch
from time import perf_counter
+start_time = perf_counter()
 ...
-duration = time() - start_time
+duration_in_seconds = perf_counter() - start_time
 

 
 
-
High performance:
from time import perf_counter
-start_time = perf_counter()             # Seconds since the restart.
-...
-duration = perf_counter() - start_time
-

-
 
Timing a Snippet
>>> from timeit import timeit
 >>> timeit("''.join(str(i) for i in range(100))",
 ...        number=10000, globals=globals(), setup='pass')
@@ -2139,8 +2133,9 @@ Line #         Mem usage      Increment   Line Contents
      3        38.012 MiB      0.344 MiB       a = [*range(10000)]
      4        38.477 MiB      0.465 MiB       b = {*range(10000)}
 

-
Call Graph
Generates a PNG image of the call graph with highlighted bottlenecks:
# $ pip3 install pycallgraph2; brew/apt install graphviz
+
Call Graph
Generates a PNG image of the call graph with highlighted bottlenecks:
# $ pip3 install pycallgraph2; apt/brew install graphviz
 import pycallgraph2 as cg, datetime
+
 filename = f'profile-{datetime.datetime.now():%Y%m%d%H%M%S}.png'
 drawer = cg.output.GraphvizOutput(output_file=filename)
 with cg.PyCallGraph(drawer):
@@ -2153,45 +2148,55 @@ drawer = cg.output.GraphvizOutput(output_file=filename)
 

 
 
-
<array> = np.array(<list/list_of_lists>)
-<array> = np.arange(from_inclusive, to_exclusive, ±step_size)
-<array> = np.ones(<shape>)
-<array> = np.random.randint(from_inclusive, to_exclusive, <shape>)
+
<array> = np.array(<list/list_of_lists>)                # Returns 1d/2d NumPy array.
+<array> = np.zeros/ones(<shape>)                        # Also np.full(<shape>, <el>).
+<array> = np.arange(from_inc, to_exc, ±step)            # Also np.linspace(start, stop, num).
+<array> = np.random.randint(from_inc, to_exc, <shape>)  # Also np.random.random(<shape>).
 

-
<array>.shape = <shape>
-<view>  = <array>.reshape(<shape>)
-<view>  = np.broadcast_to(<array>, <shape>)
+
<view>  = <array>.reshape(<shape>)                      # Also `<array>.shape = <shape>`.
+<array> = <array>.flatten()                             # Collapses array into one dimension.
+<view>  = <array>.squeeze()                             # Removes dimensions of length one.
 

-
<array> = <array>.sum(axis)
-indexes = <array>.argmin(axis)
+
<array> = <array>.sum/min/mean/var/std(axis)            # Passed dimension gets aggregated.
+<array> = <array>.argmin(axis)                          # Returns indexes of smallest elements.
+<array> = np.apply_along_axis(<func>, axis, <array>)    # Func can return a scalar or array.
 

 
    
-
  • Shape is a tuple of dimension sizes.
    • 
-
  • Axis is an index of the dimension that gets aggregated. Leftmost dimension has index 0.
    • 
+
  • Shape is a tuple of dimension sizes. A 100x50 RGB image has shape (50, 100, 3).
    • 
+
  • Axis is an index of the dimension that gets aggregated. Leftmost/outermost dimension has index 0. Summing a 100x50 RGB image along the axis 2 will return a greyscale image with shape (50, 100).
    • 
+
  • Passing a tuple of axes will chain the operations like this: '<array>.<method>(axis_1, keepdims=True).<method>(axis_2).squeeze()'.
    • 
 

-
Indexing
<el>       = <2d_array>[row_index, column_index]
-<1d_view>  = <2d_array>[row_index]
-<1d_view>  = <2d_array>[:, column_index]
+
Indexing
<el>       = <2d_array>[row_index, column_index]        # <3d_a>[table_i, row_i, column_i]
+<1d_view>  = <2d_array>[row_index]                      # <3d_a>[table_i, row_i]
+<1d_view>  = <2d_array>[:, column_index]                # <3d_a>[table_i, :, column_i]
 

 
-
<1d_array> = <2d_array>[row_indexes, column_indexes]
-<2d_array> = <2d_array>[row_indexes]
-<2d_array> = <2d_array>[:, column_indexes]
+
<1d_array> = <2d_array>[row_indexes, column_indexes]    # <3d_a>[table_is, row_is, column_is]
+<2d_array> = <2d_array>[row_indexes]                    # <3d_a>[table_is, row_is]
+<2d_array> = <2d_array>[:, column_indexes]              # <3d_a>[table_is, :, column_is]
 

-
<2d_bools> = <2d_array> ><== <el>
-<1d_array> = <2d_array>[<2d_bools>]
+
<2d_bools> = <2d_array> ><== <el>                       # <3d_array> ><== <1d_array>
+<1d_array> = <2d_array>[<2d_bools>]                     # <3d_array>[<2d_bools>]
 

-
Broadcasting
Broadcasting is a set of rules by which NumPy functions operate on arrays of different sizes and/or dimensions.
left  = [[0.1], [0.6], [0.8]]        # Shape: (3, 1)
-right = [ 0.1 ,  0.6 ,  0.8 ]        # Shape: (3)
+
    
+
  • All examples also allow assignments.
    • 
+

+
Broadcasting
Broadcasting is a set of rules by which NumPy functions operate on arrays of different sizes and/or dimensions.
left  = [[0.1], [0.6], [0.8]]                           # Shape: (3, 1)
+right = [ 0.1 ,  0.6 ,  0.8 ]                           # Shape: (3,)
 

 
 
-
1. If array shapes differ in length, left-pad the shorter shape with ones:
left  = [[0.1], [0.6], [0.8]]        # Shape: (3, 1)
-right = [[0.1 ,  0.6 ,  0.8]]        # Shape: (1, 3) <- !
+
1. If array shapes differ in length, left-pad the shorter shape with ones:
left  = [[0.1], [0.6], [0.8]]                           # Shape: (3, 1)
+right = [[0.1 ,  0.6 ,  0.8]]                           # Shape: (1, 3) <- !
 

 
-
2. If any dimensions differ in size, expand the ones that have size 1 by duplicating their elements:
left  = [[0.1, 0.1, 0.1], [0.6, 0.6, 0.6], [0.8, 0.8, 0.8]]  # Shape: (3, 3) <- !
-right = [[0.1, 0.6, 0.8], [0.1, 0.6, 0.8], [0.1, 0.6, 0.8]]  # Shape: (3, 3) <- !
+
2. If any dimensions differ in size, expand the ones that have size 1 by duplicating their elements:
left  = [[0.1,  0.1,  0.1],                             # Shape: (3, 3) <- !
+         [0.6,  0.6,  0.6],
+         [0.8,  0.8,  0.8]]
+
+right = [[0.1,  0.6,  0.8],                             # Shape: (3, 3) <- !
+         [0.1,  0.6,  0.8],
+         [0.1,  0.6,  0.8]]
 

 
 
3. If neither non-matching dimension has size 1, raise an error.
Example
For each point returns index of its nearest point ([0.1, 0.6, 0.8] => [1, 2, 1]):
>>> points = np.array([0.1, 0.6, 0.8])
@@ -2900,7 +2905,7 @@ $ pyinstaller script.py --add-data '<path>:.'