Hi I am new to pymc3. I want to do a simple matrix multiplication involving a tensor X and coefficients beta.
If all in numpy, I could do beta * X, or np.matmul(X, beta), but I don’t know the equivalent in pymc3, but I have beta declared in pm
alpha = pm.Normal('alpha', 0, sd=10)
beta = pm.Normal('beta', 0, sd=0.5, shape=(D))
X itself is declared in numpy
X = np.random.normal(size=(N, D)) * 10
How do I implement
Y = alpha + beta * X
You can do:
Y = alpha + pm.math.dot(X, beta)
Thank you. That’s affirmative!
in tensorflow, I did a tensordot between a X tensor with shape (N, T, D) and a beta with shape (S, D). The idea is to dot alone the D axis on both tensors. in tensorflow this is expressed as
tf.tensordot(X, beta, [, ])
what would be the equivalent thing in pymc3 / theano?
Thank, it looks like this should work just like tensorflow, but it doesn’t. I have
x_logit = np.tensordot(X, beta, [, ])
where X has shape (1000, 9), and beta has shape (9, ), this line crashes with
if as_[axes_a[k]] != bs[axes_b[k]]:
IndexError: tuple index out of range
Then I changed to x_logit = tt.tensordot(X, beta, [, ])
now I get x_logit shape TensorVariable: Shape.0
I am of course expecting (1000,)
why doesn’t this work?
You are displaying the symbolic shape. You can check the shape conditioned on the current input (under PyMC3 at least) by doing x_logit.tag.test_value.shape
Ok, got it. Another reason I was confused is that the data exceeded memory capacity of the laptop, and the program simply quit silently without any error. Finally, I moved it to the server and started to see execution messages.
It helps to specify cores=1 in
pm.sample when you are testing large model, it makes the error easier to read as well.