back to sgd for lin vec.

Author: orm011

vsms/search_loop_models.py

@@ -301,7 +301,7 @@ def fit_rank2(*, mod, X, y, batch_size, max_examples, valX=None, valy=None, logg
 
 def adjust_vec(vec, Xt, yt, learning_rate, loss_margin, max_examples, minibatch_size):
     vec = torch.from_numpy(vec).type(torch.float32)
-    mod = LookupVec(Xt.shape[1], margin=loss_margin, optimizer=torch.optim.Adam, learning_rate=learning_rate, init_vec=vec)
+    mod = LookupVec(Xt.shape[1], margin=loss_margin, optimizer=torch.optim.SGD, learning_rate=learning_rate, init_vec=vec)
     fit_rank2(mod=mod, X=Xt.astype('float32'), y=yt.astype('float'), 
             max_examples=max_examples, batch_size=minibatch_size,max_epochs=1)
     newvec = mod.vec.detach().numpy().reshape(1,-1)