Hi,
Assume that I have 45 subjects, each with 68 features. I want to assign each subjects to K =3 clusters by creating a categorical variable called assignments whose size equals to the number of subjects: 45.
In my code, xt_collection is a matrix with size [45, 68, 3] = [nsubs, nfeatures, class K]. It means for each of the 45 subjects with 68 features, there are three possible assignments. When there’s no enumeration, xt_collection[np.arange(45),:,assignments] can lead to an outcome matrix with size [45, 68], which means for subject i = 0,…,44, pick the corresponding 68 elements from a [68,3] matrix according to the [68,assignments[i]]. So in the toy example, I would expect that the estimation of assignments equal to torch.zeros(45) since my actual observations are the same as column 0 of the xt_collection matrix. (All the subjects in this case come from cluster 0)
Without enumeration, the sampling works fine. However, when I use enumeration, there’s an error in dimention. One possible solution I came out is to replace the nested pyro.plate with a nested for loop, however, this leads to a very slow inference.
Could you please give me any suggestions?
Thank you so much for your help!
import torch
import os
import pyro
import numpy as np
from pyro.optim import Adam
from pyro.infer.autoguide import AutoNormal,init_to_value
from pyro.ops.indexing import Vindex
import pyro.distributions as dist
from pyro import poutine
from pyro.infer import SVI, TraceEnum_ELBO,TraceGraph_ELBO, config_enumerate, infer_discrete
from pyro.ops.special import safe_log
from pyro.ops.tensor_utils import convolve
from pyro.util import warn_if_nan
smoke_test = ('CI' in os.environ)
assert pyro.__version__.startswith('1.8.1')
x0_obs = 0.01* torch.ones([45,68])
x1_obs = torch.ones([45,68])
x2_obs = 10*torch.ones([45,68])
@config_enumerate
def model(xobs):
K = 3 # number of class to be signed
sigma = pyro.sample("sigma", dist.HalfNormal(1))
weights = pyro.sample('weights', dist.Dirichlet(torch.ones(K)/K))
with pyro.plate('nsubs', 45):
assignments = pyro.sample("assignments", dist.Categorical(weights))
xt_collection = torch.stack([x0_obs,x1_obs,x2_obs],axis = 2) # torch.Size([45, 68, 3]) = [nsubs, nfeatures, class K]
xt_keep = xt_collection[torch.arange(45),:,assignments]
with pyro.plate('Nfeatures',68):
with pyro.plate('Nsubs',45):
pyro.sample('obs', dist.Normal(xt_keep, sigma), obs= torch.tensor(x0_obs))
global global_guide, svi
pyro.set_rng_seed(31)
pyro.clear_param_store()
optim = pyro.optim.Adam({'lr': 0.01})
global_guide = AutoNormal(poutine.block(model, hide=['assignments']))
elbo = TraceEnum_ELBO()
svi = SVI(model, global_guide, optim, loss=elbo)
losses = []
for i in range(1001 if not smoke_test else 2):
loss = svi.step(x0_obs)
losses.append(loss)
if i % 10 == 0:
print("ELBO at iter i = "+str(i),loss)
KeyError: -4
The above exception was the direct cause of the following exception:
ValueError Traceback (most recent call last)
ValueError: Invalid tensor shape.
Allowed dims: -3, -2, -1
Actual shape: (3, 1, 45, 68)
Try adding shape assertions for your model's sample values and distribution parameters.
The above exception was the direct cause of the following exception:
ValueError Traceback (most recent call last)
/usr/local/lib/python3.7/dist-packages/pyro/ops/packed.py in pack(value, dim_to_symbol)
42 ]
43 )
---> 44 ) from e
45 value = value.squeeze()
46 value._pyro_dims = dims
ValueError: Error while packing tensors at site 'obs':
Invalid tensor shape.
Allowed dims: -3, -2, -1
Actual shape: (3, 1, 45, 68)
Try adding shape assertions for your model's sample values and distribution parameters.
Trace Shapes:
Param Sites:
Sample Sites:
sigma dist |
value |
log_prob |
weights dist | 3
value | 3
log_prob |
assignments dist 45 |
value 3 1 1 |
log_prob 3 1 45 |
obs dist 3 1 45 68 |
value 45 68 |
log_prob 3 1 45 68 |
