tisimst/ad

Fast, transparent calculations of first and second-order automatic differentiation

ad's Issues

Does ad support any sparse matrix package?

ad supports numpy well, but doesn't work well with scipy.sparse. Many calculation involves sparse matrix. Does ad support one or plan to support?

error in ad setup command: use_2to3 is invalid.

`Collecting ad (from bundle_adjust==0.1.0.dev0)
Using cached ad-1.3.2.zip (26 kB)
Preparing metadata (setup.py) ... error
error: subprocess-exited-with-error

_ python setup.py egg_info did not run successfully.
_ exit code: 1
__> [1 lines of output]
error in ad setup command: use_2to3 is invalid.
[end of output]

note: This error originates from a subprocess, and is likely not a problem with pip.
error: metadata-generation-failed

_ Encountered error while generating package metadata.
__> See above for output.

`
How to solve it?

use cpython to convert derivative function to c code

Hi,

Can the derivative functions, Jacobian and Hessian, produced by gh() be converted to c code by cpython for speed up? Or any other way?

Thanks.

Installation issue: setuptools>=58 breaks support for use_2to3

Hello!

I am doing fresh installation of GPkit, a package that has ad as a dependency and I encountered some issues during installation.

Collecting ad
  Downloading ad-1.3.2.zip (26 kB)
    ERROR: Command errored out with exit status 1:
     command: /Users/philippe.kirschen/opt/anaconda3/envs/hops/bin/python -c 'import io, os, sys, setuptools, tokenize; sys.argv[0] = '"'"'/private/var/folders/6q/tfsts5jx6tnd_747qkgznwfwtz9v7b/T/pip-install-viy9b_j2/ad_0486c429c0824b5da1d683d85e2776af/setup.py'"'"'; __file__='"'"'/private/var/folders/6q/tfsts5jx6tnd_747qkgznwfwtz9v7b/T/pip-install-viy9b_j2/ad_0486c429c0824b5da1d683d85e2776af/setup.py'"'"';f = getattr(tokenize, '"'"'open'"'"', open)(__file__) if os.path.exists(__file__) else io.StringIO('"'"'from setuptools import setup; setup()'"'"');code = f.read().replace('"'"'\r\n'"'"', '"'"'\n'"'"');f.close();exec(compile(code, __file__, '"'"'exec'"'"'))' egg_info --egg-base /private/var/folders/6q/tfsts5jx6tnd_747qkgznwfwtz9v7b/T/pip-pip-egg-info-er2yb91p
         cwd: /private/var/folders/6q/tfsts5jx6tnd_747qkgznwfwtz9v7b/T/pip-install-viy9b_j2/ad_0486c429c0824b5da1d683d85e2776af/
    Complete output (1 lines):
    error in ad setup command: use_2to3 is invalid.
    ----------------------------------------

pip then went down through the versions until it got to 1.1.2 which then gave me a NameError that is presumably no longer relevant in the current versions of ad.

Anyway as the error above suggests it seems like use of use_2to3 is the issue. A quick search suggests that setuptools>=58 breaks support for use_2to3. I checked and my freshly-installed setuptools is version 58.0.4.

I confirmed that using setuptools<58 (57.5.0 seems like the latest version) fixes the issue on my end. Just wanted to give you a heads up!

NameError for long in __init__.py

On Mac with Python 3 I get this error:

$ python3.2 setup.py install --user
Traceback (most recent call last):
  File "setup.py", line 3, in <module>
    import ad
  File "/Users/deil/code/ad/ad/__init__.py", line 25, in <module>
    CONSTANT_TYPES = (float, int, long, complex)
NameError: name 'long' is not defined

2to3 suggests to replace long with int:
https://gist.github.com/cdeil/6384302

Is that OK or do you need long for some reason or Python version?

Speed

I really like this package but unless I am misusing it I think it suffers terribly from lack of speed. I am currently trying to calculate the greeks of a portfolio of 4 call options by two methods, namely AD and finite differences. Even when using the gradient function from ad the finite differences method still outperforms the ad implementation in terms of speed by quite a bit. I am still relatively new to python, so I may just be implementing it in a way so that it is slower. I have attached my code below, any advice would be appreciated.

from ad import adnumber
from ad.admath import *
import numpy as np
import time

time_start = time.clock()
def AD_vega_singlesim(M,S,v,T,r,delta,K,S1,v1,T1,r1,K1,S2,v2,T2,r2,K2,S3,v3,T3,r3,K3):
sensi1 = np.zeros(4)
sensi = np.zeros(4)
sensi2 = np.zeros(4)
sensi3 = np.zeros(4)
for j in range(1, int(M)):
z = np.random.normal(0,1)
price = ((S_exp((r - delta - .5_(v2))T + ( v * z_sqrt(T))) - K)* exp(-r_T))/M
price1 = ((S1.exp((r1 - delta - .5*(v12))T1 + ( v1 * z_sqrt(T1))) - K1)* exp(-r1_T1))/M
price2 = ((S2_exp((r2 - delta - .5(v2__2))T2 + ( v2 * z_sqrt(T2))) - K2) exp(-r2_T2))/M
price3 = ((S3_exp((r3 - delta - .5_(v3__2))T3 + ( v3 * z_sqrt(T3))) - K3) exp(-r3*T3))/M

    if price>0:
        sensi += price.gradient((r,v,T,S))
    if price1>0:
        sensi1 +=  price1.gradient((r1,v1,T1,S1))
    if price2>0:
        sensi2 +=  price2.gradient((r2,v2,T2,S2))
    if price3>0:
        sensi3 +=  price3.gradient((r3,v3,T3,S3))
return sensi,sensi1,sensi2,sensi3

print AD_vega_singlesim(1000,adnumber(100),adnumber(.2),adnumber(1.0),adnumber(.05),.01,100.0,adnumber(110),adnumber(.25),adnumber(2),adnumber(.06),125,adnumber(125),adnumber(.31),adnumber(2.5),adnumber(.07),122,
adnumber(150),adnumber(.13),adnumber(1.5),adnumber(.05),143)
time_elapsed = (time.clock() - time_start)
print time_elapsed

Add tests

As far as I can see ad doesn't have any unit tests.

Can you add some?
Or maybe as a starting point the examples in README.rst and user_guide.rst could be run as doctests (or be re-written as unit tests)?
(with this I could help)

Add ad.math functions to ADF objects.

Consider the following code:

>>> import ad
>>> import numpy as np
>>> x = ad.adnumber(1.0)
>>> y = np.sin(x)
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-20-bb2b9cec7df6> in <module>()
----> 1 y = np.sin(x)

AttributeError: 'ADV' object has no attribute 'sin'

Typically in the past I have allowed np to be replaced with a module like ad.math when I needed derivatives, but I believe that the ad package could become much simpler if the ADF class simply grew methods like those in ad.math. For example, the following seems to work nicely.

class ADF(object):
    ...
    def sin(self):
        return ad.admath.sin(self)

>>> import ad
>>> import numpy as np
>>> x = ad.adnumber(1.0)
>>> y = np.sin(x)
ad(0.8414709848078965)

Are there any issues to be aware of? (I don't recall seeing this behaviour before so maybe the call to self.sin etc. is a new feature in numpy? I can't seem to find a reference to this behaviour though. I am using version 1.11.0.)

Derivatives of Arrays

While ad integrates with numpy arrays by becoming arrays of adnumbers, it doesn't allow easy ways to take derivatives of the entire array. For example, let's say I have a function, and I want to plot it and its derivative:

def f(x):
    return x**2 exp(-x**2)

x = adnumber(linspace(0,4,200))
y = f(x)
plot(x, y)

But now how do I plot its derivative? I can't do dy = y.d(x), because y and x are both arrays; I have to do something a list comprehension:

dy = [v.d(v2) for v,v2 in zip(y,x)]
ddy = [v.d2(v2) for v,v2 in zip(y,x)]

plot(x, y)
plot(x, dy, '--')
plot(x, ddy, ':')

That works, but it would be really nice to have a function in ad that takes care of that, preferably by a numpy broadcasting function (so that it doesn't have to create a list, and then an array from the list).

See this notebook for an example.

P.S. Thank you so much for making this! Its an awesome library!