AttributeError: 'ArrayBox' object has no attribute 'exp' (very simple code example)

[Duncanswilson]

I'm having trouble getting anything to work the way I'm defining my functions, it's probably user error, but my example is terse.

I'm just trying to define a simple neural network and keep running into issues using np.tanh() or np.exp() and the only similar errors I could find are from having nested data structures, which I don't have.

Here's my full code snippet:

import autograd.numpy as np
from autograd import grad

action_dim = 2
feature_dim = 6
layer_dim = 4 
w1 = np.random.rand(feature_dim, layer_dim)
b1 = np.random.rand(layer_dim)
w2 = np.random.rand(layer_dim, action_dim)
b2 = np.random.rand(action_dim)

def continuous_policy_function(w1, b1, w2, b2):
    x = observation.dot(w1) + b1
    x = 1.0 / (1.0 + np.exp(-x))  # sigmoid "squashing" function to interval [0,1]
    out = x.dot(w2) + b2
    return out 

grad_function = grad(continuous_policy_function)
observation = np.random.rand(feature_dim)
action = continuous_policy_function(w1, b1, w2, b2)
print(grad_function(w1, b1, w2, b2))

and the full trace:

Traceback (most recent call last):
  File "annoying_autograd.py", line 55, in <module>
    print(grad_function(w1, b1, w2, b2))
  File "/usr/local/Cellar/python3/3.6.4/Frameworks/Python.framework/Versions/3.6/lib/python3.6/site-packages/autograd/wrap_util.py", line 20, in nary_f
    return unary_operator(unary_f, x, *nary_op_args, **nary_op_kwargs)
  File "/usr/local/Cellar/python3/3.6.4/Frameworks/Python.framework/Versions/3.6/lib/python3.6/site-packages/autograd/differential_operators.py", line 24, in grad
    vjp, ans = _make_vjp(fun, x)
  File "/usr/local/Cellar/python3/3.6.4/Frameworks/Python.framework/Versions/3.6/lib/python3.6/site-packages/autograd/core.py", line 10, in make_vjp
    end_value, end_node =  trace(start_node, fun, x)
  File "/usr/local/Cellar/python3/3.6.4/Frameworks/Python.framework/Versions/3.6/lib/python3.6/site-packages/autograd/tracer.py", line 10, in trace
    end_box = fun(start_box)
  File "/usr/local/Cellar/python3/3.6.4/Frameworks/Python.framework/Versions/3.6/lib/python3.6/site-packages/autograd/wrap_util.py", line 15, in unary_f
    return fun(*subargs, **kwargs)
  File "annoying_autograd.py", line 29, in continuous_policy_function
    x = 1.0 / (1.0 + np.exp(-x))  # sigmoid "squashing" function to interval [0,1]
  File "/usr/local/Cellar/python3/3.6.4/Frameworks/Python.framework/Versions/3.6/lib/python3.6/site-packages/autograd/tracer.py", line 48, in f_wrapped
    return f_raw(*args, **kwargs)
AttributeError: 'ArrayBox' object has no attribute 'exp'

I'm using Python 3.6.4 and and autograd 1.2 which I pulled from pip this morning.

Please let me know what I'm doing wrong here! :)

[Duncanswilson]

So I managed to fix this on my end but with little explanation of what I was doing wrong.

One thing that is clear is that I was using the wrong grad function, but I also had to replace all uses of dot with the @ operator. note: (this also works if you replace calls like observation.dot(w1) with np.dot(observation, w1 )

... ¯_(ツ)_/¯ I'd love to know if there is something deeper going on here!

import autograd.numpy as np
from autograd import elementwise_grad as egrad  # for functions that vectorize over inputs

action_dim = 2
feature_dim = 6
layer_dim = 4 
w1 = np.random.rand(feature_dim, layer_dim)
b1 = np.random.rand(layer_dim)
w2 = np.random.rand(layer_dim, action_dim)
b2 = np.random.rand(action_dim)

def continuous_policy_function(w1, b1, w2, b2):
    x = observation @ w1 + b1
    x = 1.0 / (1.0 + np.exp(-x))  # sigmoid "squashing" function to interval [0,1]
    out = x @ w2 + b2
    return out 

grad_function = egrad(continuous_policy_function)
observation = np.random.rand(feature_dim)
action = continuous_policy_function(w1, b1, w2, b2)
print(grad_function(w1, b1, w2, b2))

[j-towns]

Hey there, have a read of the tutorial, there's more detail in there, including

we don't support the syntax A.dot(B); use the equivalent np.dot(A, B) instead