<\/strong>Using The Numba JIT (Just in time Compiler)<\/h2>\n\n\n\nPython has a reputation for slow performance because it\u2019s fundamentally a scripting language. One way to get around this problem is to use the Numba JIT. Here\u2019s what that code looks like:<\/p>\n\n\n\n
First, use a timing decorator to get a grasp on the runtime of your functions:<\/p>\n\n\n\n
def timing(f):\n @wraps(f)\n def wrap(*args, **kwargs):\n ts = time()\n result = f(*args, **kwargs)\n te = time()\n print(f'fun: {f.__name__}, args: [{args}, {kwargs}] took: {te-ts} sec')\n return result\n return wrap<\/code><\/pre>\n\n\n\nNext, add a numba.jit decorator with the \u201cnopython\u201d keyword argument, and set to true. This will ensure that the code will be run by the JIT instead of regular python.<\/p>\n\n\n\n
@timing\n@numba.jit(nopython=True)\ndef expmean_jit(rea):\n \"\"\"Perform multiple mean calculations\"\"\"\n\n val = rea.mean() ** 2\n return val<\/code><\/pre>\n\n\n\nWhen you run it, you can see both a \u201cjit\u201d as well as a regular version being run via the command-line tool:<\/p>\n\n\n\n
$ python nuclearcli.py jit-test<\/code><\/p>\n\n\n\nRunning NO JIT\nfunc:'expmean' args:[(array([[1.0000e+00, 4.2080e+05, 4.2350e+05, ..., 1.0543e+06, 1.0485e+06,\n 1.0444e+06],\n [2.0000e+00, 5.4240e+05, 5.4670e+05, ..., 1.5158e+06, 1.5199e+06,\n 1.5253e+06],\n [3.0000e+00, 7.0900e+04, 7.1200e+04, ..., 1.1380e+05, 1.1350e+05,\n 1.1330e+05],\n ...,\n [1.5277e+04, 9.8900e+04, 9.8100e+04, ..., 2.1980e+05, 2.2000e+05,\n 2.2040e+05],\n [1.5280e+04, 8.6700e+04, 8.7500e+04, ..., 1.9070e+05, 1.9230e+05,\n 1.9360e+05],\n [1.5281e+04, 2.5350e+05, 2.5400e+05, ..., 7.8360e+05, 7.7950e+05,\n 7.7420e+05]], dtype=float32),), {}] took: 0.0007 sec<\/code><\/pre>\n\n\n\n$ python nuclearcli.py jit-test \u2013jit<\/p>\n\n\n\n
Running with JIT\nfunc:'expmean_jit' args:[(array([[1.0000e+00, 4.2080e+05, 4.2350e+05, ..., 1.0543e+06, 1.0485e+06,\n 1.0444e+06],\n [2.0000e+00, 5.4240e+05, 5.4670e+05, ..., 1.5158e+06, 1.5199e+06,\n 1.5253e+06],\n [3.0000e+00, 7.0900e+04, 7.1200e+04, ..., 1.1380e+05, 1.1350e+05,\n 1.1330e+05],\n ...,\n [1.5277e+04, 9.8900e+04, 9.8100e+04, ..., 2.1980e+05, 2.2000e+05,\n 2.2040e+05],\n [1.5280e+04, 8.6700e+04, 8.7500e+04, ..., 1.9070e+05, 1.9230e+05,\n 1.9360e+05],\n [1.5281e+04, 2.5350e+05, 2.5400e+05, ..., 7.8360e+05, 7.7950e+05,\n@click.option('--jit\/--no-jit', default=False)\n 7.7420e+05]], dtype=float32),), {}] took: 0.2180 sec<\/code><\/pre>\n\n\n\nHow does that work? Just a few lines of code allow for this simple toggle:<\/p>\n\n\n\n
@cli.command()\ndef jit_test(jit):\n rea = real_estate_array()\n if jit:\n click.echo(click.style('Running with JIT', fg='green'))\n expmean_jit(rea)\n else:\n click.echo(click.style('Running NO JIT', fg='red'))\n expmean(rea)<\/code><\/pre>\n\n\n\nIn some cases a JIT version could make code run thousands of times faster, but benchmarking is key. Another item to point out is the line:<\/p>\n\n\n\n
click.echo(click.style('Running with JIT', fg='green'))<\/code><\/pre>\n\n\n\nThis script allows for colored terminal output, which can be very helpful it creating sophisticated tools.<\/p>\n\n\n\n
<\/strong>Using the GPU with CUDA Python<\/h2>\n\n\n\nAnother way to nuclear power your code is to run it straight on a GPU. This example requires you run it on a machine with a CUDA enabled. Here\u2019s what that code looks like:<\/p>\n\n\n\n
@cli.command()\ndef cuda_operation():\n \"\"\"Performs Vectorized Operations on GPU\"\"\"\n\n x = real_estate_array()\n y = real_estate_array()\n\n print('Moving calculations to GPU memory')\n x_device = cuda.to_device(x)\n y_device = cuda.to_device(y)\n out_device = cuda.device_array(\n shape=(x_device.shape[0],x_device.shape[1]), dtype=np.float32)\n print(x_device)\n print(x_device.shape)\n print(x_device.dtype)\n\n print('Calculating on GPU')\n add_ufunc(x_device,y_device, out=out_device)\n\n out_host = out_device.copy_to_host()\n print(f'Calculations from GPU {out_host}')<\/code><\/pre>\n\n\n\nIt\u2019s useful to point out is that if the numpy array is first moved to the GPU, then a vectorized function does the work on the GPU. After that work is completed, then the data is moved from the GPU. By using a GPU there could be a monumental improvement to the code, depending on what it\u2019s running. The output from the command-line tool is shown below:<\/p>\n\n\n\n
$ python nuclearcli.py cuda-operation\nMoving calculations to GPU memory\n\n(10015, 259)\nfloat32\nCalculating on GPU\nCalculcations from GPU [[2.0000e+00 8.4160e+05 8.4700e+05 ... 2.1086e+06 2.0970e+06 2.0888e+06]\n [4.0000e+00 1.0848e+06 1.0934e+06 ... 3.0316e+06 3.0398e+06 3.0506e+06]\n [6.0000e+00 1.4180e+05 1.4240e+05 ... 2.2760e+05 2.2700e+05 2.2660e+05]\n ...\n [3.0554e+04 1.9780e+05 1.9620e+05 ... 4.3960e+05 4.4000e+05 4.4080e+05]\n [3.0560e+04 1.7340e+05 1.7500e+05 ... 3.8140e+05 3.8460e+05 3.8720e+05]\n [3.0562e+04 5.0700e+05 5.0800e+05 ... 1.5672e+06 1.5590e+06 1.5484e+06]]<\/code><\/pre>\n\n\n\n<\/strong>Running True Multi-Core Multithreaded Python using Numba<\/h2>\n\n\n\nOne common performance problem with Python is the lack of true, multi-threaded performance. This also can be fixed with Numba. Here\u2019s an example of some basic operations:<\/p>\n\n\n\n
@timing\n@numba.jit(parallel=True)\ndef add_sum_threaded(rea):\n \"\"\"Use all the cores\"\"\"\n\n x,_ = rea.shape\n total = 0\n for _ in numba.prange(x):\n total += rea.sum()\n print(total)\n\n@timing\ndef add_sum(rea):\n \"\"\"traditional for loop\"\"\"\n\n x,_ = rea.shape\n total = 0\n for _ in numba.prange(x):\n total += rea.sum()\n print(total)\n\n@cli.command()\n@click.option('--threads\/--no-jit', default=False)\ndef thread_test(threads):\n rea = real_estate_array()\n if threads:\n click.echo(click.style('Running with multicore threads', fg='green'))\n add_sum_threaded(rea)\n else:\n click.echo(click.style('Running NO THREADS', fg='red'))\n add_sum(rea)<\/code><\/pre>\n\n\n\nNote that the key difference between the parallel version is that it uses @numba.jit(parallel=True)<\/code> and numba.prange to spawn threads for iteration. Looking at the picture below, all of the CPUs are maxed out on the machine, but when almost the exact same code is run without the parallelization, it only uses a core.<\/p>\n\n\n\n![\"\"](\"https:\/\/kite.com\/wp-content\/uploads\/2019\/03\/graph.c2ee5067-1024x507.png\")
<\/figure><\/div>\n\n\n\n$ python nuclearcli.py thread-test<\/code><\/p>\n\n\n\n$ python nuclearcli.py thread-test --threads<\/code><\/p>\n\n\n\n<\/strong>KMeans Clustering<\/h2>\n\n\n\nOne more powerful thing that can be accomplished in a command-line tool is machine learning. In the example below, a KMeans clustering function is created with just a few lines of code. This clusters a pandas DataFrame into a default of 3 clusters.<\/p>\n\n\n\n
def kmeans_cluster_housing(clusters=3):\n \"\"\"Kmeans cluster a dataframe\"\"\"\n url = 'https:\/\/raw.githubusercontent.com\/noahgift\/socialpowernba\/master\/data\/nba_2017_att_val_elo_win_housing.csv'\n val_housing_win_df =pd.read_csv(url)\n numerical_df =(\n val_housing_win_df.loc[:,['TOTAL_ATTENDANCE_MILLIONS', 'ELO',\n 'VALUE_MILLIONS', 'MEDIAN_HOME_PRICE_COUNTY_MILLIONS']]\n )\n #scale data\n scaler = MinMaxScaler()\n scaler.fit(numerical_df)\n scaler.transform(numerical_df)\n #cluster data\n k_means = KMeans(n_clusters=clusters)\n kmeans = k_means.fit(scaler.transform(numerical_df))\n val_housing_win_df['cluster'] = kmeans.labels_\n return val_housing_win_df<\/code><\/pre>\n\n\n\nThe cluster number can be changed by passing in another number (as shown below) using click:<\/p>\n\n\n\n
@cli.command()\n@click.option('--num', default=3, help='number of clusters')\ndef cluster(num):\n df = kmeans_cluster_housing(clusters=num)\n click.echo('Clustered DataFrame')\n click.echo(df.head())<\/code><\/pre>\n\n\n\nFinally, the output of the Pandas DataFrame with the cluster assignment is show below. Note, it has cluster assignment as a column now.<\/p>\n\n\n\n
$ python -W nuclearcli.py cluster<\/code><\/p>\n\n\n\n
def timing(f):\n @wraps(f)\n def wrap(*args, **kwargs):\n ts = time()\n result = f(*args, **kwargs)\n te = time()\n print(f'fun: {f.__name__}, args: [{args}, {kwargs}] took: {te-ts} sec')\n return result\n return wrap<\/code><\/pre>\n\n\n\nNext, add a numba.jit decorator with the \u201cnopython\u201d keyword argument, and set to true. This will ensure that the code will be run by the JIT instead of regular python.<\/p>\n\n\n\n
@timing\n@numba.jit(nopython=True)\ndef expmean_jit(rea):\n \"\"\"Perform multiple mean calculations\"\"\"\n\n val = rea.mean() ** 2\n return val<\/code><\/pre>\n\n\n\nWhen you run it, you can see both a \u201cjit\u201d as well as a regular version being run via the command-line tool:<\/p>\n\n\n\n
$ python nuclearcli.py jit-test<\/code><\/p>\n\n\n\nRunning NO JIT\nfunc:'expmean' args:[(array([[1.0000e+00, 4.2080e+05, 4.2350e+05, ..., 1.0543e+06, 1.0485e+06,\n 1.0444e+06],\n [2.0000e+00, 5.4240e+05, 5.4670e+05, ..., 1.5158e+06, 1.5199e+06,\n 1.5253e+06],\n [3.0000e+00, 7.0900e+04, 7.1200e+04, ..., 1.1380e+05, 1.1350e+05,\n 1.1330e+05],\n ...,\n [1.5277e+04, 9.8900e+04, 9.8100e+04, ..., 2.1980e+05, 2.2000e+05,\n 2.2040e+05],\n [1.5280e+04, 8.6700e+04, 8.7500e+04, ..., 1.9070e+05, 1.9230e+05,\n 1.9360e+05],\n [1.5281e+04, 2.5350e+05, 2.5400e+05, ..., 7.8360e+05, 7.7950e+05,\n 7.7420e+05]], dtype=float32),), {}] took: 0.0007 sec<\/code><\/pre>\n\n\n\n$ python nuclearcli.py jit-test \u2013jit<\/p>\n\n\n\n
Running with JIT\nfunc:'expmean_jit' args:[(array([[1.0000e+00, 4.2080e+05, 4.2350e+05, ..., 1.0543e+06, 1.0485e+06,\n 1.0444e+06],\n [2.0000e+00, 5.4240e+05, 5.4670e+05, ..., 1.5158e+06, 1.5199e+06,\n 1.5253e+06],\n [3.0000e+00, 7.0900e+04, 7.1200e+04, ..., 1.1380e+05, 1.1350e+05,\n 1.1330e+05],\n ...,\n [1.5277e+04, 9.8900e+04, 9.8100e+04, ..., 2.1980e+05, 2.2000e+05,\n 2.2040e+05],\n [1.5280e+04, 8.6700e+04, 8.7500e+04, ..., 1.9070e+05, 1.9230e+05,\n 1.9360e+05],\n [1.5281e+04, 2.5350e+05, 2.5400e+05, ..., 7.8360e+05, 7.7950e+05,\n@click.option('--jit\/--no-jit', default=False)\n 7.7420e+05]], dtype=float32),), {}] took: 0.2180 sec<\/code><\/pre>\n\n\n\nHow does that work? Just a few lines of code allow for this simple toggle:<\/p>\n\n\n\n
@cli.command()\ndef jit_test(jit):\n rea = real_estate_array()\n if jit:\n click.echo(click.style('Running with JIT', fg='green'))\n expmean_jit(rea)\n else:\n click.echo(click.style('Running NO JIT', fg='red'))\n expmean(rea)<\/code><\/pre>\n\n\n\nIn some cases a JIT version could make code run thousands of times faster, but benchmarking is key. Another item to point out is the line:<\/p>\n\n\n\n
click.echo(click.style('Running with JIT', fg='green'))<\/code><\/pre>\n\n\n\nThis script allows for colored terminal output, which can be very helpful it creating sophisticated tools.<\/p>\n\n\n\n
<\/strong>Using the GPU with CUDA Python<\/h2>\n\n\n\nAnother way to nuclear power your code is to run it straight on a GPU. This example requires you run it on a machine with a CUDA enabled. Here\u2019s what that code looks like:<\/p>\n\n\n\n
@cli.command()\ndef cuda_operation():\n \"\"\"Performs Vectorized Operations on GPU\"\"\"\n\n x = real_estate_array()\n y = real_estate_array()\n\n print('Moving calculations to GPU memory')\n x_device = cuda.to_device(x)\n y_device = cuda.to_device(y)\n out_device = cuda.device_array(\n shape=(x_device.shape[0],x_device.shape[1]), dtype=np.float32)\n print(x_device)\n print(x_device.shape)\n print(x_device.dtype)\n\n print('Calculating on GPU')\n add_ufunc(x_device,y_device, out=out_device)\n\n out_host = out_device.copy_to_host()\n print(f'Calculations from GPU {out_host}')<\/code><\/pre>\n\n\n\nIt\u2019s useful to point out is that if the numpy array is first moved to the GPU, then a vectorized function does the work on the GPU. After that work is completed, then the data is moved from the GPU. By using a GPU there could be a monumental improvement to the code, depending on what it\u2019s running. The output from the command-line tool is shown below:<\/p>\n\n\n\n
$ python nuclearcli.py cuda-operation\nMoving calculations to GPU memory\n\n(10015, 259)\nfloat32\nCalculating on GPU\nCalculcations from GPU [[2.0000e+00 8.4160e+05 8.4700e+05 ... 2.1086e+06 2.0970e+06 2.0888e+06]\n [4.0000e+00 1.0848e+06 1.0934e+06 ... 3.0316e+06 3.0398e+06 3.0506e+06]\n [6.0000e+00 1.4180e+05 1.4240e+05 ... 2.2760e+05 2.2700e+05 2.2660e+05]\n ...\n [3.0554e+04 1.9780e+05 1.9620e+05 ... 4.3960e+05 4.4000e+05 4.4080e+05]\n [3.0560e+04 1.7340e+05 1.7500e+05 ... 3.8140e+05 3.8460e+05 3.8720e+05]\n [3.0562e+04 5.0700e+05 5.0800e+05 ... 1.5672e+06 1.5590e+06 1.5484e+06]]<\/code><\/pre>\n\n\n\n<\/strong>Running True Multi-Core Multithreaded Python using Numba<\/h2>\n\n\n\nOne common performance problem with Python is the lack of true, multi-threaded performance. This also can be fixed with Numba. Here\u2019s an example of some basic operations:<\/p>\n\n\n\n
@timing\n@numba.jit(parallel=True)\ndef add_sum_threaded(rea):\n \"\"\"Use all the cores\"\"\"\n\n x,_ = rea.shape\n total = 0\n for _ in numba.prange(x):\n total += rea.sum()\n print(total)\n\n@timing\ndef add_sum(rea):\n \"\"\"traditional for loop\"\"\"\n\n x,_ = rea.shape\n total = 0\n for _ in numba.prange(x):\n total += rea.sum()\n print(total)\n\n@cli.command()\n@click.option('--threads\/--no-jit', default=False)\ndef thread_test(threads):\n rea = real_estate_array()\n if threads:\n click.echo(click.style('Running with multicore threads', fg='green'))\n add_sum_threaded(rea)\n else:\n click.echo(click.style('Running NO THREADS', fg='red'))\n add_sum(rea)<\/code><\/pre>\n\n\n\nNote that the key difference between the parallel version is that it uses @numba.jit(parallel=True)<\/code> and numba.prange to spawn threads for iteration. Looking at the picture below, all of the CPUs are maxed out on the machine, but when almost the exact same code is run without the parallelization, it only uses a core.<\/p>\n\n\n\n<\/figure><\/div>\n\n\n\n$ python nuclearcli.py thread-test<\/code><\/p>\n\n\n\n$ python nuclearcli.py thread-test --threads<\/code><\/p>\n\n\n\n<\/strong>KMeans Clustering<\/h2>\n\n\n\nOne more powerful thing that can be accomplished in a command-line tool is machine learning. In the example below, a KMeans clustering function is created with just a few lines of code. This clusters a pandas DataFrame into a default of 3 clusters.<\/p>\n\n\n\n
def kmeans_cluster_housing(clusters=3):\n \"\"\"Kmeans cluster a dataframe\"\"\"\n url = 'https:\/\/raw.githubusercontent.com\/noahgift\/socialpowernba\/master\/data\/nba_2017_att_val_elo_win_housing.csv'\n val_housing_win_df =pd.read_csv(url)\n numerical_df =(\n val_housing_win_df.loc[:,['TOTAL_ATTENDANCE_MILLIONS', 'ELO',\n 'VALUE_MILLIONS', 'MEDIAN_HOME_PRICE_COUNTY_MILLIONS']]\n )\n #scale data\n scaler = MinMaxScaler()\n scaler.fit(numerical_df)\n scaler.transform(numerical_df)\n #cluster data\n k_means = KMeans(n_clusters=clusters)\n kmeans = k_means.fit(scaler.transform(numerical_df))\n val_housing_win_df['cluster'] = kmeans.labels_\n return val_housing_win_df<\/code><\/pre>\n\n\n\nThe cluster number can be changed by passing in another number (as shown below) using click:<\/p>\n\n\n\n
@cli.command()\n@click.option('--num', default=3, help='number of clusters')\ndef cluster(num):\n df = kmeans_cluster_housing(clusters=num)\n click.echo('Clustered DataFrame')\n click.echo(df.head())<\/code><\/pre>\n\n\n\nFinally, the output of the Pandas DataFrame with the cluster assignment is show below. Note, it has cluster assignment as a column now.<\/p>\n\n\n\n
$ python -W nuclearcli.py cluster<\/code><\/p>\n\n\n\n