Is there any way to define more than one update on the same share variable when calling theano.function()?
Example:
updates = []
updates.append([(self.W, self.W - 1)])
updates.append([(self.W, self.W - 2)])
train = th.function(inputs=[index], outputs=[cost], updates=updates)
will throw the error this shared variable already has an update expression:
ignore_bug_before` to at least "0.7".
tolerate_inplace_aliasing=tolerate_inplace_aliasing)
Traceback (most recent call last):
File "ae.py", line 340, in <module>
main()
File "ae.py", line 315, in main
ae.train(n_epochs=n_epochs, mini_batch_size=100, learning_rate=0.002, train_data= train_sentence_embeddings, test_data= test_sentence_embeddings)
File "ae.py", line 97, in train
givens={x:self.X[index:index+mini_batch_size,:]})
File "/usr/local/lib/python2.7/dist-packages/theano/compile/function.py", line 266, in function
profile=profile)
File "/usr/local/lib/python2.7/dist-packages/theano/compile/pfunc.py", line 489, in pfunc
no_default_updates=no_default_updates)
File "/usr/local/lib/python2.7/dist-packages/theano/compile/pfunc.py", line 198, in rebuild_collect_shared
(store_into, update_d[store_into]))
ValueError: ('this shared variable already has an update expression', (W, DimShuffle{1,0}.0))
I was interested in doing so as I have weight matrix that I need to update in such a way that different parts of it has a different update (which I do it using set_subtensor).
As explained in this thread on the theano-users mailing list...
You need to chain the set_subtensors so that you only end up with a single entry for the whole tensor in the updates list.
For example (code not like original because original does not actually include set_subtensor anywhere):
new_w = self.W
new_W = theano.tensor.set_subtensor(new_W[0], -1)
new_W = theano.tensor.set_subtensor(new_W[1], -2)
updates = [(self.W, new_W)]
train = theano.function(inputs=[index], outputs=[cost], updates=updates)