33..
44 _href from docs/source/index.rst
55
6- ===============
7- PyMC and Aesara
8- ===============
6+ =================
7+ PyMC and PyTensor
8+ =================
99
10- What is Aesara
11- ==============
10+ What is PyTensor
11+ ================
1212
13- Aesara is a package that allows us to define functions involving array
13+ PyTensor is a package that allows us to define functions involving array
1414operations and linear algebra. When we define a PyMC model, we implicitly
15- build up an Aesara function from the space of our parameters to
15+ build up an PyTensor function from the space of our parameters to
1616their posterior probability density up to a constant factor. We then use
1717symbolic manipulations of this function to also get access to its gradient.
1818
19- For a thorough introduction to Aesara see the
20- :doc: `aesara docs aesara :index
19+ For a thorough introduction to PyTensor see the
20+ :doc: `pytensor docs pytensor :index
2121but for the most part you don't need detailed knowledge about it as long
2222as you are not trying to define new distributions or other extensions
2323of PyMC. But let's look at a simple example to get a rough
@@ -33,8 +33,8 @@ arbitrarily chosen) function
3333
3434is similar to eg SymPy's `Symbol `)::
3535
36- import aesara
37- import aesara .tensor as at
36+ import pytensor
37+ import pytensor .tensor as at
3838 # We don't specify the dtype of our input variables, so it
3939 # defaults to using float64 without any special config.
4040 a = at.scalar('a')
@@ -56,16 +56,16 @@ do to compute the output::
5656 of the exponential of `inner `. Somewhat surprisingly, it
5757 would also have worked if we used `np.exp `. This is because numpy
5858 gives objects it operates on a chance to define the results of
59- operations themselves. Aesara variables do this for a large number
60- of operations. We usually still prefer the Aesara
59+ operations themselves. PyTensor variables do this for a large number
60+ of operations. We usually still prefer the PyTensor
6161 functions instead of the numpy versions, as that makes it clear that
6262 we are working with symbolic input instead of plain arrays.
6363
64- Now we can tell Aesara to build a function that does this computation.
65- With a typical configuration, Aesara generates C code, compiles it,
64+ Now we can tell PyTensor to build a function that does this computation.
65+ With a typical configuration, PyTensor generates C code, compiles it,
6666and creates a python function which wraps the C function::
6767
68- func = aesara .function([a, x, y], [out])
68+ func = pytensor .function([a, x, y], [out])
6969
7070We can call this function with actual arrays as many times as we want::
7171
@@ -75,15 +75,15 @@ We can call this function with actual arrays as many times as we want::
7575
7676 out = func(a_val, x_vals, y_vals)
7777
78- For the most part the symbolic Aesara variables can be operated on
79- like NumPy arrays. Most NumPy functions are available in `aesara .tensor
78+ For the most part the symbolic PyTensor variables can be operated on
79+ like NumPy arrays. Most NumPy functions are available in `pytensor .tensor
8080(which is typically imported as `at `). A lot of linear algebra operations
8181can be found in `at.nlinalg ` and `at.slinalg ` (the NumPy and SciPy
8282operations respectively). Some support for sparse matrices is available
83- in `aesara .sparse
84- see :mod: `the aesara api docs <aesara .tensor> `.
83+ in `pytensor .sparse
84+ see :mod: `the pytensor api docs <pytensor .tensor> `.
8585
86- A notable exception where Aesara variables do *not * behave like
86+ A notable exception where PyTensor variables do *not * behave like
8787NumPy arrays are operations involving conditional execution.
8888
8989Code like this won't work as expected::
@@ -123,16 +123,16 @@ Changing elements of an array is possible using `at.set_subtensor`::
123123 a = at.vector('a')
124124 b = at.set_subtensor(a[:10], 1)
125125
126- # is roughly equivalent to this (although aesara avoids
126+ # is roughly equivalent to this (although pytensor avoids
127127 # the copy if `a` isn't used anymore)
128128 a = np.random.randn(10)
129129 b = a.copy()
130130 b[:10] = 1
131131
132- How PyMC uses Aesara
132+ How PyMC uses PyTensor
133133====================
134134
135- Now that we have a basic understanding of Aesara we can look at what
135+ Now that we have a basic understanding of PyTensor we can look at what
136136happens if we define a PyMC model. Let's look at a simple example::
137137
138138 true_mu = 0.1
@@ -159,7 +159,7 @@ where with the normal likelihood :math:`N(x|μ,σ^2)`
159159
160160
161161space (the *free variables *) and the logp function. For each free variable
162- we generate an Aesara variable. And for each variable (observed or otherwise)
162+ we generate an PyTensor variable. And for each variable (observed or otherwise)
163163we add a term to the global logp. In the background something similar to
164164this is happening::
165165
@@ -177,7 +177,7 @@ So calling `pm.Normal()` modifies the model: It changes the logp function
177177of the model. If the `observed ` keyword isn't set it also creates a new
178178free variable. In contrast, `pm.Normal.dist() ` doesn't care about the model,
179179it just creates an object that represents the normal distribution. Calling
180- `logp ` on this object creates an Aesara variable for the logp probability
180+ `logp ` on this object creates an PyTensor variable for the logp probability
181181or log probability density of the distribution, but again without changing
182182the model in any way.
183183
@@ -209,8 +209,8 @@ is roughly equivalent to this::
209209 model.add_logp_term(pm.Normal.dist(mu, sigma).logp(data))
210210
211211The return values of the variable constructors are subclasses
212- of Aesara variables, so when we define a variable we can use any
213- Aesara operation on them::
212+ of PyTensor variables, so when we define a variable we can use any
213+ PyTensor operation on them::
214214
215215 design_matrix = np.array([[...]])
216216 with pm.Model() as model:
0 commit comments