Is it possible to call pm.fit('advi') and specify that I want to use the Adam optimizer for a faster inference? I assume that stochastic gradient descent is used by default. Thanks.
I suspect that you are looking for something like this:
pm.fit(n=1000, obj_optimizer=adam())
The various update methods can be found here.
Thanks.
I would like to explore my options with pymc version 4. I searched for info on the variational API and found the following for pymc3: Variational API quickstart — PyMC3 3.11.5 documentation
I have not found an updated version for pymc3 version 4. I am wondering whether the examples demonstrating the use of callbacks will have changed. Thanks.
I tried your suggestion and got an error message stating that name 'adam' is not defined. Here are the details.
Code:
def run_and_plot(model, seed, nb_iter=10000):
# with model:
# vi_fit2 = pm.fit(method='svgd', n=nb_iter, random_seed=seed)
with model:
vi_fit2 = pm.fit(method='advi', n=nb_iter, random_seed=seed, obj_optimizer=adam())
trace5 = vi_fit2.sample(10000)
pm.plot_trace(trace5);
fig, ax = plt.subplots(figsize=(8, 6))
plot_w = np.arange(K) + 1
ax.bar(plot_w - 0.5, trace5.posterior['w'].squeeze().mean(axis=0), width=1., lw=1, ec='w');
ax.set_xlim(0.5, K);
ax.set_xlabel('Component');
ax.set_ylabel('Posterior expected mixture weight');
plt.savefig("figure.png")
mean_w = np.mean(trace5.posterior['w'].squeeze(), axis=0)
nonzero_component = np.where(mean_w > 0.05)[0]
mean_theta = np.mean(trace5.posterior['theta'].squeeze(), axis=0)
print("mean_theta[nonzero_component]:\n", mean_theta[nonzero_component, :])
print("theta_actual:\n", theta_actual)
run_and_plot(model, seed=3437, nb_iter=10000)
Error:
---------------------------------------------------------------------------
NameError Traceback (most recent call last)
Input In [215], in <cell line: 1>()
----> 1 run_and_plot(model, seed=3437, nb_iter=10000)
Input In [214], in run_and_plot(model, seed, nb_iter)
1 def run_and_plot(model, seed, nb_iter=10000):
2 # with model:
3 # vi_fit2 = pm.fit(method='svgd', n=nb_iter, random_seed=seed)
5 with model:
----> 6 vi_fit2 = pm.fit(method='advi', n=nb_iter, random_seed=seed, obj_optimizer=adam())
8 trace5 = vi_fit2.sample(10000)
9 pm.plot_trace(trace5);
NameError: name 'adam' is not defined
A working example here or in the documentation would be nice! Thank you!
I tried a callback in the simplest way possible and got an error related to the number of argument expected in __call__. I am at a loss!
Code:
class Callback:
def __call__(self):
raise NotImplementedError
class Writeout(Callback):
def __init__(self):
pass
def __call__(self):
print("gordon")
callback = Writeout()
def run_and_plot(model, seed, nb_iter=10000):
# with model:
# vi_fit2 = pm.fit(method='svgd', n=nb_iter, random_seed=seed)
with model:
# the `callbacks` argument must be iterable (not the case in pymc3)
vi_fit2 = pm.fit(method='advi', n=nb_iter, random_seed=seed, callbacks=[callback])
# obj_optimizer=adam())
trace5 = vi_fit2.sample(10000)
pm.plot_trace(trace5);
fig, ax = plt.subplots(figsize=(8, 6))
plot_w = np.arange(K) + 1
ax.bar(plot_w - 0.5, trace5.posterior['w'].squeeze().mean(axis=0), width=1., lw=1, ec='w');
ax.set_xlim(0.5, K);
ax.set_xlabel('Component');
ax.set_ylabel('Posterior expected mixture weight');
plt.savefig("figure.png")
mean_w = np.mean(trace5.posterior['w'].squeeze(), axis=0)
nonzero_component = np.where(mean_w > 0.05)[0]
mean_theta = np.mean(trace5.posterior['theta'].squeeze(), axis=0)
print("mean_theta[nonzero_component]:\n", mean_theta[nonzero_component, :])
print("theta_actual:\n", theta_actual)
run_and_plot(model, seed=3437, nb_iter=100)
And the error:
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
Input In [251], in <cell line: 1>()
----> 1 run_and_plot(model, seed=3437, nb_iter=10000)
Input In [250], in run_and_plot(model, seed, nb_iter)
1 def run_and_plot(model, seed, nb_iter=10000):
2 # with model:
3 # vi_fit2 = pm.fit(method='svgd', n=nb_iter, random_seed=seed)
5 with model:
----> 6 vi_fit2 = pm.fit(method='advi', n=nb_iter, random_seed=seed, callbacks=[callback])
7 # obj_optimizer=adam())
9 trace5 = vi_fit2.sample(10000)
File ~/opt/anaconda3/envs/pymc4/lib/python3.9/site-packages/pymc/variational/inference.py:765, in fit(n, local_rv, method, model, random_seed, start, inf_kwargs, **kwargs)
763 else:
764 raise TypeError(f"method should be one of {set(_select.keys())} or Inference instance")
--> 765 return inference.fit(n, **kwargs)
File ~/opt/anaconda3/envs/pymc4/lib/python3.9/site-packages/pymc/variational/inference.py:144, in Inference.fit(self, n, score, callbacks, progressbar, **kwargs)
142 progress = range(n)
143 if score:
--> 144 state = self._iterate_with_loss(0, n, step_func, progress, callbacks)
145 else:
146 state = self._iterate_without_loss(0, n, step_func, progress, callbacks)
File ~/opt/anaconda3/envs/pymc4/lib/python3.9/site-packages/pymc/variational/inference.py:240, in Inference._iterate_with_loss(self, s, n, step_func, progress, callbacks)
238 progress.comment = f"Average Loss = {avg_loss:,.5g}"
239 for callback in callbacks:
--> 240 callback(self.approx, scores[: i + 1], i + s + 1)
241 except (KeyboardInterrupt, StopIteration) as e: # pragma: no cover
242 # do not print log on the same line
243 scores = scores[:i]
TypeError: __call__() takes 1 positional argument but 4 were given
Apologies, my code snippet was written quickly and should have been this:
pm.fit(n=1000, obj_optimizer=pm.adam())
The signature for the callback functions is this:
callbacks: list[function: (Approximation, losses, i) -> None]
The source of the callbacks can be found here, but it’s likely you just want pm.callbacks.CheckParametersConvergence(). So something like this:
with pm.Model() as model:
x = pm.Normal('x', mu=0, sigma=1)
y = pm.Normal('y', mu=x, sigma=1)
vi_fit2 = pm.fit(method='advi',
n=1000,
callbacks=[pm.callbacks.CheckParametersConvergence()],
obj_optimizer=pm.adam()
)
trace = vi_fit2.sample(10000)
I got it to work. However, now I have to learn to use the output of the callbacks. Will report back later. 
Now that I have callbacks working, I would like to access its output. I cannot figure out the logic. Is there information out there explaining how it works (aside from looking at the source code?) Thanks.
I cannot figure out how to access the data stored in the callback CheckParametersConvergence() . Any help would be appreciated! Thanks.
I think the quickstart illustrates how to interrogate the convergence.
I went through the Quickstart notebook in detail and tried things out before writing on discourse. Consider the following line:
with model:
mean_field = pm.fit(method="advi", callbacks=[CheckParametersConvergence()])
plt.plot(mean_field.hist);
I tried this (with pymc 4). I would like to do the following: update an array with my own diagnostics at a specified frequency. So the question is how to retrieve my array from the callback? I guess I must create my own callback? I might experiment with that.
It would be nice to check results every n iterations? Is that done by running pm.sample at a lower frequency, and simply restarting the simulation ?I have not tried this yet. Thanks!
I created my own class, copying CheckParametersConvergence. I was able to get the results allowed. I will now experiment with Tracker.