Hi,
I’m trying out Bayesian Regression. I’ve run into RuntimeError: bool value of Tensor with more than one value is ambiguous while performing svi.step. train_X and train_y are torch tensors of size (torch.Size([120, 2]) and torch.Size([120])).
guide = AutoDiagonalNormal(model)
optim = Adam({'lr' : 0.03})
svi = SVI(model, guide, optim, loss=Trace_ELBO)
def train():
pyro.clear_param_store()
for j in range(num_iterations):
# calculate loss; apply gradients
loss = svi.step(train_X, train_y) # <-- ERROR
if j%100 == 0: # for every 100 steps
print("iteration {j} : loss [{loss}]".format(
j=j+1, loss=loss/len(data))
)
train()
I’m using torch-1.0.0 and pyro-ppl-0.3.0.
Would really appreciate your help in resolving this.
Traceback
<ipython-input-28-69443a3a9594> in train()
3 for j in range(num_iterations):
4 # calculate loss; apply gradients
----> 5 loss = svi.step(train_X, train_y)
6 if j%100 == 0: # for every 100 steps
7 print("iteration {j} : loss [{loss}]".format(
/usr/local/lib/python3.6/dist-packages/pyro/infer/svi.py in step(self, *args, **kwargs)
97 # get loss and compute gradients
98 with poutine.trace(param_only=True) as param_capture:
---> 99 loss = self.loss_and_grads(self.model, self.guide, *args, **kwargs)
100
101 params = set(site["value"].unconstrained()
/usr/local/lib/python3.6/dist-packages/pyro/infer/svi.py in _loss_and_grads(*args, **kwargs)
56 if loss_and_grads is None:
57 def _loss_and_grads(*args, **kwargs):
---> 58 loss_val = loss(*args, **kwargs)
59 loss_val.backward(retain_graph=True)
60 return loss_val
/usr/local/lib/python3.6/dist-packages/pyro/infer/elbo.py in __init__(self, num_particles, max_plate_nesting, max_iarange_nesting, vectorize_particles, strict_enumeration_warning, ignore_jit_warnings, retain_graph)
71 self.vectorize_particles = vectorize_particles
72 self.retain_graph = retain_graph
---> 73 if self.vectorize_particles and self.num_particles > 1:
74 self.max_plate_nesting += 1
75 self.strict_enumeration_warning = strict_enumeration_warning
RuntimeError: bool value of Tensor with more than one value is ambiguous
Edit
- bayesian_regression.py works just fine