Initialice trace of MCMC /NUTS with MAP estimate?

1

Hi!

First, thanks for your attention and time. Second, I would like to initialize the sampling from MCMC/NUTS to the MAP estimate (which is estimated using autoguide) of the same model.

Following the baseball.py example (pyro/baseball.py at dev · pyro-ppl/pyro · GitHub)
I was able to do that in the older Pyro versions (!= 0.5.1) using the exec_traces from the svi and setting them as the initial_trace for NUTS. However now it has been changed to the initialize_model function that I really do not know what is it doing.

How can I initialize the trace with the MAP estimate?

Thank you very much,
Best wishes

2

@artistworking Using autoguide, you can obtain MAP point with guide.median(). As stated in the docs of initialize_model, that utility prepares all required information for MCMC (from a Pyro model). You can specify initial params of MCMC with

map_points = guide.median()
init_params = {name: transforms[name](value) for name, value in map_points.items()}
3

Hi!!
Thanks for your reply :slight_smile @fehiepsi : am trying to implement your suggestion but something is not clear to me:

If transforms is obtained from the initialize_model function, then this part :
transforms[name][value]
is not working because it throws:
TypeError: 'ComposeTransform' object is not subscriptable
due to Transforms looking like this (with the invertible transformation only…):

{'Value1': ComposeTransform(
    
), 'Value2': ComposeTransform(
    _InverseTransform(),
    _InverseTransform()
), 'Value3': ComposeTransform(
    
), 'Value4': ComposeTransform(
    _InverseTransform(),
    _InverseTransform()
)}

so I changed to this:

init_params = {name: transforms[name] (value) for name, value in map_points.items()}

but that does not apply any transformation to the map_point items. Init_params looks like this:

{‘value1’: tensor([[ -3.4214, 13.0987, -13.1515],
[ -0.6091, 10.6803, -15.8382]],
grad_fn=), ‘value2’: tensor([-0.2674, -0.6896, 0.1402], grad_fn=), ‘value3’: tensor([0.9876, -0.789, 0.5231], grad_fn=),‘value4’: tensor([ 5.4123, -1.2537, 5.1462], grad_fn=)}

which, if I use it when I run mcmc like this:

init_params, potential_fn, transforms, _ = initialize_model(model,model_args=(data_obs,), num_chains=chains,jit_compile=True,skip_jit_warnings=True)
map_points = _get_initial_trace(data_obs,average) #This returns guide.median()
init_params, potential_fn, transforms, _ = initialize_model(model,model_args=(data_obs,), num_chains=chains,jit_compile=True,skip_jit_warnings=True)
init_params = {name: transforms[name] (value).no_grad() for name, value in map_points.items()}
nuts_kernel = NUTS(potential_fn=potential_fn, max_tree_depth=5, target_accept_prob=0.8)
mcmc = MCMC(nuts_kernel, num_samples=samples, warmup_steps=warmup, num_chains=chains,initial_params=init_params, transforms=transforms)
mcmc.run(data_obs)

it results into this error:

RuntimeError: you can only change requires_grad flags of leaf variables. If you want to use a computed variable in a subgraph that doesn’t require differentiation use var_no_grad = var.detach()

I tried .no_grad() but it got this:

AttributeError: ‘Tensor’ object has no attribute ‘no_grad’

I am not sure if I misunderstood something, but is not exactly working…

Thank you very much for your help :slight_smile:

4

@artistworking Yeah, thanks for correcting me! That should be transforms[name](value). :smiley:

Regarding the RuntimeError, I think your MAP point has the .grad property left after SVI? If so, you should detach it first:

init_params = {k: v.detach() for k, v in init_params.items()}

or

map_points = {k: v.detach() for k, v in map_points.items()}
5

Oh yeah! Thanks to you too :slight_smile: . I knew I had to get rid of the grad but not sure how to .detach() worked, thanks!!!