Gradient of scan fails when it involves a shared variable

[jessegrabowski]

Before

Currently, this graph has valid gradients with respect to mu and sigma:

mu = pt.dscalar('mu')
sigma = pt.dscalar('sigma')

epsilon = pt.random.normal(0, 1)
z = mu + sigma * epsilon

pt.grad(z, sigma).eval({mu:1, sigma:1})
# Out: Random draw from a N(0, 1)

But this graph does not:

def step(x, mu, sigma, rng):
    epsilon = pt.random.normal(0, 1, rng=rng)
    next_x = x + mu + sigma * epsilon
    return next_x, {rng:new_rng}

traj, updates = pytensor.scan(step, outputs_info=[x0], non_sequences=[mu, sigma, rng], n_steps=10)
pt.grad(traj[-1], sigma).eval({mu:1, sigma:1, x0:0})
# Out: Error, graph depends on a shared variable

After

I imagine that in cases where the "reparameterization trick" is used, stochastic gradients can be computed for scan graphs.

Context for the issue:

The "reparameterization trick" is well known in the machine learning literature as a way to get stochastic gradients from graphs with sampling operations. It seems like we already support this, because this graph can be differentiated:

epsilon = pt.random.normal(0, 1)
z = mu + sigma * epsilon

pt.grad(z, sigma).eval({mu:1, sigma:1})

But this graph cannot:

z= pt.random.normal(mu, sigma)
pt.grad(z, sigma).eval({mu:1, sigma:1})

The fact that even the "good" version breaks down in scan is I suppose a bug? Or a missing feature? Or neither? In the equation:

$$x_{t+1} = x_t + \mu + \sigma \epsilon_t$$ with $x_0$ given, it seems like:

$$\frac{\partial x_2}{\partial \sigma} =\frac{\partial}{\partial \sigma} x_0 + \mu + \sigma \epsilon_1 + \mu + \sigma \epsilon_2 = \epsilon_1 + \epsilon_2$$

I should get back the sum of the random draws for the sequence.

Context: I'm trying to use pytensor to compute greeks for options, which involves taking the derivative of sampled trajectories.

[jessegrabowski]

This works:

def step(x, epsilon, mu, sigma):
    next_x = x + mu + sigma * epsilon
    return next_x

rng = pytensor.shared(np.random.default_rng())
new_rng, epsilons = pt.random.normal(size=10, rng=rng)
traj, updates = pytensor.scan(step, outputs_info=[x0], non_sequences=[mu, sigma], sequences=[epsilons])

df_ds = pt.grad(traj[-1], sigma)
f = pytensor.function([x0, mu, sigma], df_ds, updates={rng:new_rng})

So as long as I can write the sequence as conditionally independent it works? It seems like it should be possible to get the gradient without doing that, though.

[ricardoV94]

This is a case I am not sure we want to be too clever for the user sake. If you're taking gradients in stochastic graphs perhaps you should know exactly what you're doing and do the reparametrization yourself (you can create your own suite of rewrites to change the graph before calling grad).

Note we never do any random rewrites by default (other than when certain distributions are missing in a backend) because depending on how the random generator routine is used it can alter the results. This is a decision Theano devs took for reproducibility/ease of debug that we can revisit, but should do so consciously.

[jessegrabowski]

I guess the tags I chose for this issue are quite bad because I don't think I want any kind of special automatic handling here. More that it seems like when scan is constructing it's gradient, it is failing because it's asking the random generator for a gradient, which it (obviously) doesn't have. Shouldn't these have a pass-through? If there's another complication (because an actual random variable -- NOT a generator -- is on the backwards graph) it can and should still error, I agree.

[ricardoV94]

Can you provide a full example? Your original one has new_rng but but that's not defined anywhere. strict=True may help.

The only issue with scans with gradient stuff I know is that they must be passed explicitly: https://github.com/pymc-devs/pytensor/issues/6

[jessegrabowski]

mu = pt.dscalar('mu')
sigma = pt.dscalar('sigma')
x0 = pt.dscalar('x0')
rng = pytensor.shared(np.random.default_rng(), 'rng')

def step(x, mu, sigma, rng):
    new_rng, epsilon = pm.Normal.dist(0, 1, rng=rng).owner.outputs
    next_x = x + mu + sigma * epsilon
    return next_x, {rng:new_rng}

traj, updates = pytensor.scan(step, outputs_info=[x0], non_sequences=[mu, sigma, rng], n_steps=10)
pt.grad(traj[-1], sigma)

Gives:

Traceback

---

NullTypeGradError Traceback (most recent call last) Cell In[132], line 10 7 return next_x, {rng:new_rng} 9 traj, updates = pytensor.scan(step, outputs_info=[x0], non_sequences=[mu, sigma, rng], n_steps=10) ---> 10 pt.grad(traj[-1], sigma).eval({mu:1, sigma:1, x0:0})

File ~/mambaforge/envs/cge-dev/lib/python3.11/site-packages/pytensor/gradient.py:616, in grad(cost, wrt, consider_constant, disconnected_inputs, add_names, known_grads, return_disconnected, null_gradients) 614 if isinstance(_rval[i].type, NullType): 615 if null_gradients == "raise": --> 616 raise NullTypeGradError( 617 f"grad encountered a NaN. {_rval[i].type.why_null}" 618 ) 619 else: 620 assert null_gradients == "return"

NullTypeGradError: grad encountered a NaN. This variable is Null because the grad method for input 3 (mu) of the Scan{scan_fn, while_loop=False, inplace=none} op is not implemented. Depends on a shared variable

[ricardoV94]

This seems to be an old known bug/limitation of Scan: https://groups.google.com/g/theano-users/c/dAwr1j8-QOY/m/8fmDmQPkPJkJ

Maybe something we can also address better in the Scan refactor, since we don't treat shared variables as magical entities anymore.

[jessegrabowski]

First refactor all of pytensor (and pymc) to remove shared variables? :D

[ricardoV94]

Shared variables are fine-ish, it's the treating them differently in PyTensor internals that's a source of unnecessary complexity.

This special treatment is also a thorn in OpFromGraph: https://github.com/pymc-devs/pytensor/issues/473

Which led me to basically reimplement it completely to be usable in PyMC: https://github.com/pymc-devs/pymc/pull/6947

[jessegrabowski]

I just read https://github.com/pymc-devs/pytensor/issues/473, so your thoughts on OpFromGraph and shared variables are fresh in my mind.

[ricardoV94]

In sum, I think shared variables should be explicit inputs everywhere except in the outer PyTensor function where there is an ambiguity of whether the call signature would require them or not

[ricardoV94]

I reopened, until we have a solution I think it's good to track the issue