allow fixed params in hmc

t-vi · Nov 24, 2017 · 3b4c2f4 · 3b4c2f4
1 parent bd48929
commit 3b4c2f4
Showing 1 changed file with 4 additions and 4 deletions.
diff --git a/candlegp/training/hmc.py b/candlegp/training/hmc.py
@@ -56,10 +56,10 @@ def logprob_grads():
     def thinning(thin_iterations, epsilon, lmin, lmax):
         logprob, grads = logprob_grads()
         for i in range(thin):
-            xs_prev = [p.data.clone() for p in model.parameters()]
+            xs_prev = [p.data.clone() for p in model.parameters() if p.requires_grad]
             grads_prev = grads
             logprob_prev = logprob
-            ps_init = [Variable(xs_prev[0].new(*p.size()).normal_()) for p in model.parameters()]
+            ps_init = [Variable(xs_prev[0].new(*p.size()).normal_()) for p in model.parameters() if p.requires_grad]
             ps = [p + 0.5 * epsilon * grad for p,grad in zip(ps_init, grads_prev)]
 
             max_iterations = int((torch.rand(1)*(lmax+1-lmin)+lmin)[0])
@@ -68,7 +68,7 @@ def thinning(thin_iterations, epsilon, lmin, lmax):
             proceed = True
             i_ps = 0
             while proceed and i_ps < max_iterations:
-                for x, p in zip(model.parameters(), ps):
+                for x, p in zip([p for p in model.parameters() if p.requires_grad], ps):
                     x.data += epsilon*p.data
                 _, grads = logprob_grads()
                 proceed = torch.stack([is_finite(grad).prod() for grad in grads], dim=0).prod().data[0]
@@ -91,7 +91,7 @@ def thinning(thin_iterations, epsilon, lmin, lmax):
                     proceed = False
                 # otherwise keep new
             if not proceed:
-                for p,x_prev in zip(model.parameters(), xs_prev):
+                for p,x_prev in zip([p for p in model.parameters() if p.requires_grad], xs_prev):
                     p.data = x_prev
                 logprob = logprob_prev
                 grads = grads_prev