Skip to content

Latest commit

 

History

History
893 lines (687 loc) · 27.6 KB

README.md

File metadata and controls

893 lines (687 loc) · 27.6 KB

Numba

{:.no_toc}

* TOC {:toc}

The goal

"Numba is an open source JIT compiler that translates a subset of Python and NumPy code into fast machine code."

Questions to David Rotermund

pip install numba

A ~5 minute guide to Numba

For the example that will show you the options of optimization we need to understand the numba naming schema.

Numbers

For the function signatures we need to be able to translate the usual np.dtype into numpy.types.

For doing so we just replace np. by numba.types. .

Type name(s) Shorthand Comments
numba.types.boolean b1 represented as a byte
numba.types.uint8, byte u1 8-bit unsigned byte
numba.types.uint16 u2 16-bit unsigned integer
numba.types.uint32 u4 32-bit unsigned integer
numba.types.uint64 u8 64-bit unsigned integer
numba.types.int8, char i1 8-bit signed byte
numba.types.int16 i2 16-bit signed integer
numba.types.int32 i4 32-bit signed integer
numba.types.int64 i8 64-bit signed integer
numba.types.intc C int-sized integer
numba.types.uintc C int-sized unsigned integer
numba.types.intp pointer-sized integer
numba.types.uintp pointer-sized unsigned integer
numba.types.float32 f4 single-precision floating-point number
numba.types.float64, double f8 double-precision floating-point number
numba.types.complex64 c8 single-precision complex number
numba.types.complex128 c16 double-precision complex number

Arrays

If we have arrays in the function signature, which is a very likely senario, we might want to give as much information to numpy as possible about the numpy.ndarray. In some cases it is very benificial to make a np.ndarray an array with C memory layout and tell numba about it.

We can use the numpy function numpy.ascontiguousarray for converting a numpy array into a C memory layout.

We can also check a numpy array, let's call it X, if it is already in the C memory layout. This is done by looking at X.flags['C_CONTIGUOUS'].

Some example for array signatures are:

numba.types.float32[:] 1d array of float32 with no particular memory layout
numba.types.float32[:,:] 2d array of float32 with no particular memory layout
numba.types.float32[:,:,:] 3d array of float32 with no particular memory layout
numba.types.float32[::1] 1d array of float32 with C memory layout
numba.types.float32[:,::1] 2d array of float32 with C memory layout
numba.types.float32[:,:,::1] 3d array of float32 with C memory layout
numba.types.float32[::1,:] 2d array of float32 with Fortran memory layout
numba.types.float32[::1,:,:] 3d array of float32 with Fortran memory layout

An example (up to 350x faster)

For measuring the time used by the program I ran everything twice and took the second time. I did this is because the just-in-time compilation takes a moment for the first call of a function.

Basis code (7.76 sec)

This is the basic code without any optimizations.

import time
import numpy as np


def get_spike(
    h: np.ndarray, number_of_neurons: np.uint64, random_number: np.float64
) -> np.uint64:

    summation: np.float64 = np.float64(0.0)

    output: np.uint64 = np.uint64(number_of_neurons - 1)

    for i in range(0, np.uint64(number_of_neurons - 1)):
        summation += h[i]

        if random_number <= summation:
            output = np.uint64(i)
            return output

    return output


def main(
    number_of_iterations: np.uint64,
    number_of_neurons: np.uint64,
    random_number_spikes: np.ndarray,
    random_number_h: np.ndarray,
) -> np.ndarray:

    results = np.zeros((number_of_iterations), dtype=np.uint64)

    for i in range(0, number_of_iterations):
        h = random_number_h[i, :]
        h /= h.sum()
        results[i] = get_spike(h, number_of_neurons, random_number_spikes[i])

    return results


if __name__ == "__main__":
    number_of_iterations: np.uint64 = np.uint64(10000)
    number_of_neurons: np.uint64 = np.uint64(10000)
    myrng = np.random.default_rng()

    random_number_spikes = myrng.random((number_of_iterations), dtype=np.float64)
    random_number_h = myrng.random(
        (number_of_iterations, number_of_neurons), dtype=np.float64
    )

    start_time = time.perf_counter()
    results = main(
        number_of_iterations=number_of_iterations,
        number_of_neurons=number_of_neurons,
        random_number_spikes=random_number_spikes,
        random_number_h=random_number_h,
    )
    end_time = time.perf_counter()

    check_for_errors = np.sum([results >= number_of_neurons])
    if check_for_errors > 0:
        print("Something went really wrong! Panic!")
    print(f"{end_time-start_time:.5f} sec")
    print(results[0:10])

Optimization 1 (0.482sec)

We add just-in-time compilation to the function get_spike with @njit(cache=True). "To avoid compilation times each time you invoke a Python program, you can instruct Numba to write the result of function compilation into a file-based cache."

import time
import numpy as np
from numba import njit


@njit(cache=True)
def get_spike(
    h: np.ndarray, number_of_neurons: np.uint64, random_number: np.float64
) -> np.uint64:

    summation: np.float64 = np.float64(0.0)

    output: np.uint64 = np.uint64(number_of_neurons - 1)

    for i in range(0, np.uint64(number_of_neurons - 1)):
        summation += h[i]

        if random_number <= summation:
            output = np.uint64(i)
            return output

    return output


def main(
    number_of_iterations: np.uint64,
    number_of_neurons: np.uint64,
    random_number_spikes: np.ndarray,
    random_number_h: np.ndarray,
) -> np.ndarray:

    results = np.zeros((number_of_iterations), dtype=np.uint64)

    for i in range(0, number_of_iterations):
        h = random_number_h[i, :]
        h /= h.sum()
        results[i] = get_spike(h, number_of_neurons, random_number_spikes[i])

    return results


if __name__ == "__main__":
    number_of_iterations: np.uint64 = np.uint64(10000)
    number_of_neurons: np.uint64 = np.uint64(10000)
    myrng = np.random.default_rng()

    random_number_spikes = myrng.random((number_of_iterations), dtype=np.float64)
    random_number_h = myrng.random(
        (number_of_iterations, number_of_neurons), dtype=np.float64
    )

    start_time = time.perf_counter()
    results = main(
        number_of_iterations=number_of_iterations,
        number_of_neurons=number_of_neurons,
        random_number_spikes=random_number_spikes,
        random_number_h=random_number_h,
    )
    end_time = time.perf_counter()

    check_for_errors = np.sum([results >= number_of_neurons])
    if check_for_errors > 0:
        print("Something went really wrong! Panic!")
    print(f"{end_time-start_time:.5f} sec")
    print(results[0:10])

Optimization 2 (0.627sec)

We also add just-in-time compilation to the function main with @njit(cache=True).

import time
import numpy as np
from numba import njit


@njit(cache=True)
def get_spike(
    h: np.ndarray, number_of_neurons: np.uint64, random_number: np.float64
) -> np.uint64:

    summation: np.float64 = np.float64(0.0)

    output: np.uint64 = np.uint64(number_of_neurons - 1)

    for i in range(0, np.uint64(number_of_neurons - 1)):
        summation += h[i]

        if random_number <= summation:
            output = np.uint64(i)
            return output

    return output


@njit(cache=True)
def main(
    number_of_iterations: np.uint64,
    number_of_neurons: np.uint64,
    random_number_spikes: np.ndarray,
    random_number_h: np.ndarray,
) -> np.ndarray:

    results = np.zeros((number_of_iterations), dtype=np.uint64)

    for i in range(0, number_of_iterations):
        h = random_number_h[i, :]
        h /= h.sum()
        results[i] = get_spike(h, number_of_neurons, random_number_spikes[i])

    return results


if __name__ == "__main__":
    number_of_iterations: np.uint64 = np.uint64(10000)
    number_of_neurons: np.uint64 = np.uint64(10000)
    myrng = np.random.default_rng()

    random_number_spikes = myrng.random((number_of_iterations), dtype=np.float64)
    random_number_h = myrng.random(
        (number_of_iterations, number_of_neurons), dtype=np.float64
    )

    start_time = time.perf_counter()
    results = main(
        number_of_iterations=number_of_iterations,
        number_of_neurons=number_of_neurons,
        random_number_spikes=random_number_spikes,
        random_number_h=random_number_h,
    )
    end_time = time.perf_counter()

    check_for_errors = np.sum([results >= number_of_neurons])
    if check_for_errors > 0:
        print("Something went really wrong! Panic!")
    print(f"{end_time-start_time:.5f} sec")
    print(results[0:10])

Optimization 3 (0.619sec)

We add function signatures to the code with:

@njit(
    numba.types.uint64(numba.types.float64[:], numba.types.uint64, numba.types.float64),
    cache=True,
)
def get_spike(
    h: np.ndarray, number_of_neurons: np.uint64, random_number: np.float64
) -> np.uint64:


[...]

@njit(
    numba.types.uint64[:](
        numba.types.uint64,
        numba.types.uint64,
        numba.types.float64[:],
        numba.types.float64[:, :],
    ),
    cache=True,
)
def main(
    number_of_iterations: np.uint64,
    number_of_neurons: np.uint64,
    random_number_spikes: np.ndarray,
    random_number_h: np.ndarray,
) -> np.ndarray:
import time
import numpy as np
from numba import njit
import numba


@njit(
    numba.types.uint64(numba.types.float64[:], numba.types.uint64, numba.types.float64),
    cache=True,
)
def get_spike(
    h: np.ndarray, number_of_neurons: np.uint64, random_number: np.float64
) -> np.uint64:

    summation: np.float64 = np.float64(0.0)

    output: np.uint64 = np.uint64(number_of_neurons - 1)

    for i in range(0, np.uint64(number_of_neurons - 1)):
        summation += h[i]

        if random_number <= summation:
            output = np.uint64(i)
            return output

    return output


@njit(
    numba.types.uint64[:](
        numba.types.uint64,
        numba.types.uint64,
        numba.types.float64[:],
        numba.types.float64[:, :],
    ),
    cache=True,
)
def main(
    number_of_iterations: np.uint64,
    number_of_neurons: np.uint64,
    random_number_spikes: np.ndarray,
    random_number_h: np.ndarray,
) -> np.ndarray:

    results = np.zeros((number_of_iterations), dtype=np.uint64)

    for i in range(0, number_of_iterations):
        h = random_number_h[i, :]
        h /= h.sum()
        results[i] = get_spike(h, number_of_neurons, random_number_spikes[i])

    return results


if __name__ == "__main__":
    number_of_iterations: np.uint64 = np.uint64(10000)
    number_of_neurons: np.uint64 = np.uint64(10000)
    myrng = np.random.default_rng()

    random_number_spikes = myrng.random((number_of_iterations), dtype=np.float64)
    random_number_h = myrng.random(
        (number_of_iterations, number_of_neurons), dtype=np.float64
    )

    start_time = time.perf_counter()
    results = main(
        number_of_iterations=number_of_iterations,
        number_of_neurons=number_of_neurons,
        random_number_spikes=random_number_spikes,
        random_number_h=random_number_h,
    )
    end_time = time.perf_counter()

    check_for_errors = np.sum([results >= number_of_neurons])
    if check_for_errors > 0:
        print("Something went really wrong! Panic!")
    print(f"{end_time-start_time:.5f} sec")
    print(results[0:10])

Optimization 4 (0.419sec)

We tell numba about the C memory layout of the arrays with refining the function signature:

@njit(
    numba.types.uint64[::1](
        numba.types.uint64,
        numba.types.uint64,
        numba.types.float64[::1],
        numba.types.float64[:, ::1],
    ),
    cache=True,
)
import time
import numpy as np
from numba import njit
import numba


@njit(
    numba.types.uint64(numba.types.float64[:], numba.types.uint64, numba.types.float64),
    cache=True,
)
def get_spike(
    h: np.ndarray, number_of_neurons: np.uint64, random_number: np.float64
) -> np.uint64:

    summation: np.float64 = np.float64(0.0)

    output: np.uint64 = np.uint64(number_of_neurons - 1)

    for i in range(0, np.uint64(number_of_neurons - 1)):
        summation += h[i]

        if random_number <= summation:
            output = np.uint64(i)
            return output

    return output


@njit(
    numba.types.uint64[::1](
        numba.types.uint64,
        numba.types.uint64,
        numba.types.float64[::1],
        numba.types.float64[:, ::1],
    ),
    cache=True,
)
def main(
    number_of_iterations: np.uint64,
    number_of_neurons: np.uint64,
    random_number_spikes: np.ndarray,
    random_number_h: np.ndarray,
) -> np.ndarray:

    results = np.zeros((number_of_iterations), dtype=np.uint64)

    for i in range(0, number_of_iterations):
        h = random_number_h[i, :]
        h /= h.sum()
        results[i] = get_spike(h, number_of_neurons, random_number_spikes[i])

    return results


if __name__ == "__main__":
    number_of_iterations: np.uint64 = np.uint64(10000)
    number_of_neurons: np.uint64 = np.uint64(10000)
    myrng = np.random.default_rng()

    random_number_spikes = myrng.random((number_of_iterations), dtype=np.float64)
    random_number_h = myrng.random(
        (number_of_iterations, number_of_neurons), dtype=np.float64
    )

    start_time = time.perf_counter()
    results = main(
        number_of_iterations=number_of_iterations,
        number_of_neurons=number_of_neurons,
        random_number_spikes=random_number_spikes,
        random_number_h=random_number_h,
    )
    end_time = time.perf_counter()

    check_for_errors = np.sum([results >= number_of_neurons])
    if check_for_errors > 0:
        print("Something went really wrong! Panic!")
    print(f"{end_time-start_time:.5f} sec")
    print(results[0:10])

Optimization 5 (0.235sec)

We don't really need float64. Let's us switch to float32:

import time
import numpy as np
from numba import njit
import numba


@njit(
    numba.types.uint64(numba.types.float32[:], numba.types.uint64, numba.types.float32),
    cache=True,
)
def get_spike(
    h: np.ndarray, number_of_neurons: np.uint64, random_number: np.float32
) -> np.uint64:

    summation: np.float32 = np.float32(0.0)

    output: np.uint64 = np.uint64(number_of_neurons - 1)

    for i in range(0, np.uint64(number_of_neurons - 1)):
        summation += h[i]

        if random_number <= summation:
            output = np.uint64(i)
            return output

    return output


@njit(
    numba.types.uint64[::1](
        numba.types.uint64,
        numba.types.uint64,
        numba.types.float32[::1],
        numba.types.float32[:, ::1],
    ),
    cache=True,
)
def main(
    number_of_iterations: np.uint64,
    number_of_neurons: np.uint64,
    random_number_spikes: np.ndarray,
    random_number_h: np.ndarray,
) -> np.ndarray:

    results = np.zeros((number_of_iterations), dtype=np.uint64)

    for i in range(0, number_of_iterations):
        h = random_number_h[i, :]
        h /= h.sum()
        results[i] = get_spike(h, number_of_neurons, random_number_spikes[i])

    return results


if __name__ == "__main__":
    number_of_iterations: np.uint64 = np.uint64(10000)
    number_of_neurons: np.uint64 = np.uint64(10000)
    myrng = np.random.default_rng()

    random_number_spikes = myrng.random((number_of_iterations), dtype=np.float32)
    random_number_h = myrng.random(
        (number_of_iterations, number_of_neurons), dtype=np.float32
    )

    start_time = time.perf_counter()
    results = main(
        number_of_iterations=number_of_iterations,
        number_of_neurons=number_of_neurons,
        random_number_spikes=random_number_spikes,
        random_number_h=random_number_h,
    )
    end_time = time.perf_counter()

    check_for_errors = np.sum([results >= number_of_neurons])
    if check_for_errors > 0:
        print("Something went really wrong! Panic!")
    print(f"{end_time-start_time:.5f} sec")
    print(results[0:10])

Optimization 6 (0.144sec)

Let us activate fastmath

@njit(
    numba.types.uint64(numba.types.float32[:], numba.types.uint64, numba.types.float32),
    cache=True,
    fastmath=True,
)
[...]

@njit(
    numba.types.uint64[::1](
        numba.types.uint64,
        numba.types.uint64,
        numba.types.float32[::1],
        numba.types.float32[:, ::1],
    ),
    cache=True,
    fastmath=True,
)
import time
import numpy as np
from numba import njit
import numba


@njit(
    numba.types.uint64(numba.types.float32[:], numba.types.uint64, numba.types.float32),
    cache=True,
    fastmath=True,
)
def get_spike(
    h: np.ndarray, number_of_neurons: np.uint64, random_number: np.float32
) -> np.uint64:
    summation: np.float32 = np.float32(0.0)

    output: np.uint64 = np.uint64(number_of_neurons - 1)

    for i in range(0, np.uint64(number_of_neurons - 1)):
        summation += h[i]

        if random_number <= summation:
            output = np.uint64(i)
            return output

    return output


@njit(
    numba.types.uint64[::1](
        numba.types.uint64,
        numba.types.uint64,
        numba.types.float32[::1],
        numba.types.float32[:, ::1],
    ),
    cache=True,
    fastmath=True,
)
def main(
    number_of_iterations: np.uint64,
    number_of_neurons: np.uint64,
    random_number_spikes: np.ndarray,
    random_number_h: np.ndarray,
) -> np.ndarray:
    results = np.zeros((number_of_iterations), dtype=np.uint64)

    for i in range(0, number_of_iterations):
        h = random_number_h[i, :]
        h /= h.sum()
        results[i] = get_spike(h, number_of_neurons, random_number_spikes[i])

    return results


if __name__ == "__main__":
    number_of_iterations: np.uint64 = np.uint64(10000)
    number_of_neurons: np.uint64 = np.uint64(10000)
    myrng = np.random.default_rng()

    random_number_spikes = myrng.random((number_of_iterations), dtype=np.float32)
    random_number_h = myrng.random(
        (number_of_iterations, number_of_neurons), dtype=np.float32
    )

    start_time = time.perf_counter()
    results = main(
        number_of_iterations=number_of_iterations,
        number_of_neurons=number_of_neurons,
        random_number_spikes=random_number_spikes,
        random_number_h=random_number_h,
    )
    end_time = time.perf_counter()

    check_for_errors = np.sum([results >= number_of_neurons])
    if check_for_errors > 0:
        print("Something went really wrong! Panic!")
    print(f"{end_time-start_time:.5f} sec")
    print(results[0:10])

Optimization 7 (0.022sec)

We can run the function main in parallel. This can be activated by:

@njit(
    numba.types.uint64[::1](
        numba.types.uint64,
        numba.types.uint64,
        numba.types.float32[::1],
        numba.types.float32[:, ::1],
    ),
    cache=True,
    fastmath=True,
    parallel=True,
)

and then we need to replace range by prange.

import time
import numpy as np
from numba import njit, prange
import numba


@njit(
    numba.types.uint64(numba.types.float32[:], numba.types.uint64, numba.types.float32),
    cache=True,
    fastmath=True,
)
def get_spike(
    h: np.ndarray, number_of_neurons: np.uint64, random_number: np.float32
) -> np.uint64:
    summation: np.float32 = np.float32(0.0)

    output: np.uint64 = np.uint64(number_of_neurons - 1)

    for i in range(0, np.uint64(number_of_neurons - 1)):
        summation += h[i]

        if random_number <= summation:
            output = np.uint64(i)
            return output

    return output


@njit(
    numba.types.uint64[::1](
        numba.types.uint64,
        numba.types.uint64,
        numba.types.float32[::1],
        numba.types.float32[:, ::1],
    ),
    cache=True,
    fastmath=True,
    parallel=True,
)
def main(
    number_of_iterations: np.uint64,
    number_of_neurons: np.uint64,
    random_number_spikes: np.ndarray,
    random_number_h: np.ndarray,
) -> np.ndarray:
    results = np.zeros((number_of_iterations), dtype=np.uint64)

    for i in prange(0, number_of_iterations):
        h = random_number_h[i, :]
        h /= h.sum()
        results[i] = get_spike(h, number_of_neurons, random_number_spikes[i])

    return results


if __name__ == "__main__":
    number_of_iterations: np.uint64 = np.uint64(10000)
    number_of_neurons: np.uint64 = np.uint64(10000)
    myrng = np.random.default_rng()

    random_number_spikes = myrng.random((number_of_iterations), dtype=np.float32)
    random_number_h = myrng.random(
        (number_of_iterations, number_of_neurons), dtype=np.float32
    )

    start_time = time.perf_counter()
    results = main(
        number_of_iterations=number_of_iterations,
        number_of_neurons=number_of_neurons,
        random_number_spikes=random_number_spikes,
        random_number_h=random_number_h,
    )
    end_time = time.perf_counter()

    check_for_errors = np.sum([results >= number_of_neurons])
    if check_for_errors > 0:
        print("Something went really wrong! Panic!")
    print(f"{end_time-start_time:.5f} sec")

    print(results[0:10])

Failure is an option: Debugging

If something is too good to be true then maybe it is not true! Not only in the case of an email from a rich Nigerian prince you might want to debug the situation, also you shouldn't totally trust numba as well. As always: Check if the results are in the right region.

If Numba njit has a problem it usually gives you an error message and stops. This is the reason why we use njit instead of jit. If jit sees a problem it fixes it with slow Python code. And it does this silently. You will only notice the absence of an improvement. njit stops with an error message.

Only once we found a problem (with an earlier version of numba) where the use of parallel loops failed so beautiful that we got wrong results but a speed improvement 10000x. Looking on the speed improvement this was clearly a case of too good to be true.

But you are not helpless!

In the case of our example with prange we can activate debugging information

main.parallel_diagnostics(level=4)

and see that it did:

================================================================================
 Parallel Accelerator Optimizing:  Function main, <ipython-input-1-dc4cb05d144d>
 (30)  
================================================================================


Parallel loop listing for  Function main, <ipython-input-1-dc4cb05d144d> (30) 
---------------------------------------------------------------------------------|loop #ID
@njit(                                                                           | 
    numba.types.uint64[::1](                                                     | 
        numba.types.uint64,                                                      | 
        numba.types.uint64,                                                      | 
        numba.types.float32[::1],                                                | 
        numba.types.float32[:, ::1],                                             | 
    ),                                                                           | 
    cache=True,                                                                  | 
    fastmath=True,                                                               | 
    parallel=True,                                                               | 
)                                                                                | 
def main(                                                                        | 
    number_of_iterations: np.uint64,                                             | 
    number_of_neurons: np.uint64,                                                | 
    random_number_spikes: np.ndarray,                                            | 
    random_number_h: np.ndarray,                                                 | 
) -> np.ndarray:                                                                 | 
    results = np.zeros((number_of_iterations), dtype=np.uint64)------------------| #0
                                                                                 | 
    for i in prange(0, number_of_iterations):------------------------------------| #2
        h = random_number_h[i, :]                                                | 
        h /= h.sum()-------------------------------------------------------------| #1
        results[i] = get_spike(h, number_of_neurons, random_number_spikes[i])    | 
                                                                                 | 
    return results                                                               | 
--------------------------------- Fusing loops ---------------------------------
Attempting fusion of parallel loops (combines loops with similar properties)...
  Trying to fuse loops #0 and #2:
    - fusion failed: cross iteration dependency found between loops #0 and #2
----------------------------- Before Optimisation ------------------------------
Parallel region 0:
+--2 (parallel)
   +--1 (parallel)


--------------------------------------------------------------------------------
------------------------------ After Optimisation ------------------------------
Parallel region 0:
+--2 (parallel)
   +--1 (serial)


 
Parallel region 0 (loop #2) had 0 loop(s) fused and 1 loop(s) serialized as part
 of the larger parallel loop (#2).
--------------------------------------------------------------------------------
--------------------------------------------------------------------------------
 
---------------------------Loop invariant code motion---------------------------
Allocation hoisting:
No allocation hoisting found

Instruction hoisting:
loop #0:
  Has the following hoisted:
    $expr_out_var.15 = const(uint64, 0)
loop #2:
  Has the following hoisted:
    $const120.4 = const(NoneType, None)
    $const122.5 = const(NoneType, None)
    $124build_slice.6 = global(slice: <class 'slice'>)
    $124build_slice.7 = call $124build_slice.6($const120.4, $const122.5, func=$124build_slice.6, args=(Var($const120.4, <ipython-input-1-dc4cb05d144d>:50), Var($const122.5, <ipython-input-1-dc4cb05d144d>:50)), kws=(), vararg=None, varkwarg=None, target=None)
    $186load_global.16 = global(get_spike: CPUDispatcher(<function get_spike at 0x7fedb9fc27a0>))
  Failed to hoist the following:
    dependency: $126build_tuple.8 = build_tuple(items=[Var($parfor__index_18.95, <string>:2), Var($124build_slice.7, <ipython-input-1-dc4cb05d144d>:50)])
    dependency: h = getitem(value=random__number__h, index=$126build_tuple.8, fn=<built-in function getitem>)
    dependency: $206binary_subscr.22 = getitem(value=random__number__spikes, index=$parfor__index_18.95, fn=<built-in function getitem>)
    dependency: $220call.23 = call $push_global_to_block.94($h.1.22, number__of__neurons, $206binary_subscr.22, func=$push_global_to_block.94, args=[Var($h.1.22, <ipython-input-1-dc4cb05d144d>:51), Var(number__of__neurons, <ipython-input-1-dc4cb05d144d>:30), Var($206binary_subscr.22, <ipython-input-1-dc4cb05d144d>:52)], kws=(), vararg=None, varkwarg=None, target=None)
--------------------------------------------------------------------------------