Dimension trouble initializing NUTS

[drewjohnston13]

When I try to sample, I'm seeing odd indexing errors as NUTS initializes. Here is my code:

with pm.Model() as model:

    # Set prior distributions.
    k = pm.Lognormal("k", mu=np.log(2), sigma=4, shape=(1,)) # Collision frequency
    A = pm.Lognormal("A", mu=np.log(9800.), sigma=1000, shape=(1,)) # Pre-exponential factor (rate)
    B = pm.Lognormal("B", mu=np.log(320.), sigma=100, shape=(1,)) # Proportionality Coefficient/Activation Energy
    C = pm.Lognormal("C", mu=np.log(1.5/10000), sigma=1/10000, shape=(1,)) # Scaled Heat Transfer Coefficient
    Cs = pm.Lognormal("Cs", mu=np.log(3.5), sigma=3, shape=(1,)) # Fuel Relative Disappearance Rate

    # Run the model.
    T_preds = theano_fem_solver(k, A, B, C, Cs)

    # Transform temperature to probabilities.
    probs = tt.nnet.sigmoid((T_preds-500)/50)

    # Calculate the likelihood.
    like = pm.Bernoulli('likelihood', p=probs, observed=obs)

    # MCMC
    trace = pm.sample(tune=2, draws=2, chains=1, cores=1, return_inferencedata=False)

Here, the traceback ends with an index error in this package during NUTS initialization:

File ~/miniconda3/envs/fenics_new/lib/python3.8/site-packages/theano/graph/op.py:130, in compute_test_value(node)
127 thunk.inputs = [storage_map[v] for v in node.inputs]
128 thunk.outputs = [storage_map[v] for v in node.outputs]
--> 130 required = thunk()
131 assert not required # We provided all inputs
133 for output in node.outputs:
134 # Check that the output has been computed

File ~/miniconda3/envs/fenics_new/lib/python3.8/site-packages/theano/graph/op.py:491, in Op.make_py_thunk..rval(p, i, o, n, params)
484 def rval(
485 p=p,
486 i=node_input_storage,
(...)
489 params=params_val,
490 ):
--> 491 r = p(n, [x[0] for x in i], o, params)
492 for o in node.outputs:
493 compute_map[o][0] = True

File ~/.local/lib/python3.8/site-packages/fenics_pymc3/core.py:50, in FenicsVJPOp.perform(self, node, inputs, outputs, params)
41 numpy_grads = evaluate_pullback(
42 fenics_output, fenics_inputs, tape, Δfenics_output
43 )
45 theano_grads = (
46 theano.gradient.grad_undefined(self, i, inputs[i]) if ng is None else ng
47 for i, ng in enumerate(numpy_grads)
48 )
---> 50 for i, tg in enumerate(theano_grads):
51 outputs[i][0] = tg

File ~/.local/lib/python3.8/site-packages/fenics_pymc3/core.py:46, in (.0)
40 fenics_output, fenics_inputs, tape = params
41 numpy_grads = evaluate_pullback(
42 fenics_output, fenics_inputs, tape, Δfenics_output
43 )
45 theano_grads = (
---> 46 theano.gradient.grad_undefined(self, i, inputs[i]) if ng is None else ng
47 for i, ng in enumerate(numpy_grads)
48 )
50 for i, tg in enumerate(theano_grads):
51 outputs[i][0] = tg
IndexError: list index out of range

Am I missing something here? Or is there an issue with the latest version of pymc3? I'd appreciate any help in resolving or understanding this.

[IvanYashchuk]

PyMC has changed a lot. I'm sorry I wasn't careful enough to pin the exact PyMC3 and Theano versions. I should find the time to do that finally.

[IvanYashchuk]

It seems I have pinned all the dependencies already in #8.

I'm sorry a lot has changed in PyMC. Yet another fork of Theano (https://www.pymc.io/blog/pytensor_announcement.html) doesn't look promising for a curious outsider to the world of PyMC like me. I'm afraid I won't be able to find the time to revive and update this demo to modern PyMC.