Just-in-Time With Numba

Event: Remote Python Pizza 1.0
Date: 25 April 2020
Location: Remote (@ your couch!)

Facing slow processing times in your numerical codes? In this lightning talk, we will explore JIT compilation in Numba.

Transcript

1. Just-in-Time with Numba
   Presented by:
   Ong Chin Hwee (@ongchinhwee)
   25 April 2020
   Remote Python Pizza
   

   View Slide

2. About me
   Ong Chin Hwee 王敬惠
   ● Data Engineer @ ST Engineering
   ● Background in aerospace
   engineering + computational
   modelling
   ● Contributor to pandas 1.0 release
   ● Mentor team at BigDataX
   @ongchinhwee
   

   View Slide

3. Bottlenecks in a data science project
   ● Lack of data / Poor quality data
   ● Data Preprocessing
   ○ The 80/20 data science dilemma
   ■ In reality, it’s closer to 90/10
   ○ Slow processing speeds in Python!
   ■ Python runs on the interpreter, not compiled
   @ongchinhwee
   

   View Slide

4. Compiled vs Interpreted Languages
   Written Code Compiler
   Compiled Code
   in Target
   Language
   Linker
   Machine Code
   (executable)
   Loader
   Execution
   @ongchinhwee
   @ongchinhwee
   

   View Slide

5. Compiled vs Interpreted Languages
   Written Code Compiler
   Lower-level
   bytecode
   Virtual
   Machine
   Execution
   @ongchinhwee
   

   View Slide

6. What is Just-in-Time?
   Just-In-Time (JIT) compilation
   ● Converts source code into native machine code at
   runtime
   ● Is the reason why Java runs on a Virtual Machine (JVM)
   yet has comparable performance to compiled languages
   (C/C++ etc., Go)
   @ongchinhwee
   

   View Slide

7. Just-in-Time with Numba
   numba module
   ● Just-in-Time (JIT) compiler for Python that converts
   Python functions into machine code
   ● Can be used by simply applying a decorator (a wrapper)
   around functions to instruct numba to compile them
   ● Two modes of execution:
   ○ njit (nopython compilation of Numba-compatible code)
   ○ jit (object mode compilation with “loop-lifting”)
   @ongchinhwee
   

   View Slide

8. Numba Compiler Architecture
   Lower-level
   bytecode
   Numba
   interpreter
   Numba IR
   Lowering
   (codegen)
   LLVM IR
   @ongchinhwee
   Type
   inference
   Typed
   Numba IR
   Machine
   Code
   (executable)
   LLVM JIT
   Compiler
   IR: Intermediate Representation
   

   View Slide

9. Numba Compiler Architecture
   Lower-level
   bytecode
   Numba
   interpreter
   Numba IR
   Lowering
   (codegen)
   LLVM IR
   @ongchinhwee
   Type
   inference
   Typed
   Numba IR
   Machine
   Code
   (executable)
   LLVM JIT
   Compiler
   Numba frontend
   Numba backend
   IR: Intermediate Representation
   

   View Slide

10. Practical Implementation
    @ongchinhwee
    

    View Slide

11. Initialize File List in Directory
    import numpy as np
    import os
    import sys
    import time
    DIR = './chest_xray/train/NORMAL/'
    train_normal = [DIR + name for name in os.listdir(DIR)
    if os.path.isfile(os.path.join(DIR, name))]
    No. of images in
    ‘train/NORMAL’: 1431
    @ongchinhwee
    

    View Slide

12. With numba
    from PIL import Image
    from numba import jit
    @jit
    def image_proc(index):
    '''Convert + resize image'''
    im = Image.open(define_imagepath(index))
    im = im.convert("RGB")
    im_resized = np.array(im.resize((64,64)))
    return im_resized
    @ongchinhwee
    

    View Slide

13. With numba
    from PIL import Image
    from numba import jit
    @jit
    def image_proc(index):
    '''Convert + resize image'''
    im = Image.open(define_imagepath(index))
    im = im.convert("RGB")
    im_resized = np.array(im.resize((64,64)))
    return im_resized
    Code runs in object mode (@jit)
    @ongchinhwee
    

    View Slide

14. With numba
    start_cpu_time = time.clock()
    listcomp_output = np.array([image_resize(x) for x in
    train_normal])
    end_cpu_time = time.clock()
    total_tpe_time = end_cpu_time - start_cpu_time
    sys.stdout.write('List comprehension completed in {}
    seconds.\n'.format(
    total_tpe_time))
    Python-only:
    218.1 seconds
    After compilation:
    169.6 seconds
    @ongchinhwee
    

    View Slide

15. With numba
    import numpy as np
    from numba import njit
    @njit
    def square(a_list):
    squared_list = []
    '''Calculate square of number in a_list'''
    for x in a_list:
    squared_list.append(np.square(x))
    return squared_list
    @ongchinhwee
    

    View Slide

16. With numba
    import numpy as np
    from numba import njit
    @njit
    def square(a_list):
    squared_list = []
    '''Calculate square of number in a_list'''
    for x in a_list:
    squared_list.append(np.square(x))
    return squared_list
    Code runs in no-Python/native
    machine mode (@njit or
    @jit(nopython=true))
    @ongchinhwee
    

    View Slide

17. With numba
    a_list = np.array([i for i in range(1,100000)])
    start_cpu_time = time.time()
    listcomp_array_output = square(a_list)
    end_cpu_time = time.time()
    total_tpe_time = end_cpu_time - start_cpu_time
    sys.stdout.write(
    'Elapsed (after compilation) {}
    seconds.\n'.format(total_tpe_time))
    Python-only:
    0.51544 seconds
    After compilation:
    0.00585 seconds
    @ongchinhwee
    

    View Slide

18. Key Takeaways
    @ongchinhwee
    

    View Slide

19. Just-in-Time with numba
    ● Just-in-Time (JIT) compilation with numba
    ○ converts source code from non-compiled languages
    into native machine code at runtime
    ○ may not work for some functions/modules - these are
    still run on the interpreter
    ○ significantly enhances speedups provided by
    optimized numerical codes
    @ongchinhwee
    

    View Slide

20. Reach out to
    me!
    : ongchinhwee
    : @ongchinhwee
    : hweecat
    : https://ongchinhwee.me
    And check out my slides on:
    hweecat/talk_jit-numba
    

    View Slide