So I am trying to do some modeling using a Plackett-Luce distribution. I found project on github that implements it here:
I am getting everything to work in terms of building a model, doing inference with SVI using AutoDiagonalNorm to the point that I know it’s working because it outputs fairly accurate loc and scale parameters:
guide.requires_grad_(False)
for name, value in pyro.get_param_store().items():
print(name, pyro.param(name))
AutoDiagonalNormal.loc Parameter containing:
tensor([-0.6924, -0.6104, -0.3954, 1.6901, 2.7612, -0.0822, 0.0528, 0.0675,
-0.0353, 0.0311, -0.1520, 0.8171, -0.0276, -0.1319, 0.9564, -0.7052,
0.6423, -0.0239, -0.4948, 0.2907, 0.2828])
AutoDiagonalNormal.scale tensor([0.4231, 0.2462, 0.1886, 0.3141, 0.2906, 0.2132, 0.3147, 0.1858, 0.1911,
0.1777, 0.1913, 0.2880, 0.1956, 0.1833, 0.3105, 0.2190, 0.2974, 0.2267,
0.2141, 0.2525, 0.1954])
So far so good. The trouble in paradise is that when I try to sample using Predictive I get the following error regarding batch_shape:
site_stats = {}
for k, v in samples.items():
site_stats[k] = {
"mean": torch.mean(v, 0),
"std": torch.std(v, 0),
"5%": v.kthvalue(int(len(v) * 0.05)),
"95%": v.kthvalue(int(len(v) * 0.95)),
}
return site_stats
predictive = Predictive(model, guide=guide, num_samples=800, return_sites=("obs","_RETURN"))
samples = predictive(data)
pred_summary = summary(samples)
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
<ipython-input-242-4df7bb95db13> in <module>
12
13 predictive = Predictive(model, guide=guide, num_samples=800, return_sites=("obs","_RETURN"))
---> 14 samples = predictive(data)
15 pred_summary = summary(samples)
~/pyro/xpm/pyro/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
720 result = self._slow_forward(*input, **kwargs)
721 else:
--> 722 result = self.forward(*input, **kwargs)
723 for hook in itertools.chain(
724 _global_forward_hooks.values(),
~/pyro/xpm/pyro/lib/python3.7/site-packages/pyro/infer/predictive.py in forward(self, *args, **kwargs)
199 parallel=self.parallel, model_args=args, model_kwargs=kwargs)
200 return _predictive(self.model, posterior_samples, self.num_samples, return_sites=return_sites,
--> 201 parallel=self.parallel, model_args=args, model_kwargs=kwargs)
202
203 def get_samples(self, *args, **kwargs):
~/pyro/xpm/pyro/lib/python3.7/site-packages/pyro/infer/predictive.py in _predictive(model, posterior_samples, num_samples, return_sites, return_trace, parallel, model_args, model_kwargs)
68 return_site_shapes = {}
69 for site in model_trace.stochastic_nodes + model_trace.observation_nodes:
---> 70 append_ndim = max_plate_nesting - len(model_trace.nodes[site]["fn"].batch_shape)
71 site_shape = (num_samples,) + (1,) * append_ndim + model_trace.nodes[site]['value'].shape
72 # non-empty return-sites
AttributeError: 'PlackettLuce' object has no attribute 'batch_shape'
I’m not sure what to do here. I tried using TorchDistributionMixin as the base class but that raises an error in the inference step:
ValueError: Error while computing log_prob at site 'obs':
shape mismatch: objects cannot be broadcast to a single shape: (21,) vs torch.Size([21, 20])
Trace Shapes:
Param Sites:
Sample Sites:
sigma dist |
value |
log_prob |
theta dist 20 |
value 20 |
log_prob 20 |
obs dist 21 |
value 21 20 |
guide.requires_grad_(False)
for name, value in pyro.get_param_store().items():
print(name, pyro.param(name))
Any help would be greatly appreciated!