Spooky bug in Theano and/or Blocks.

wtf.py

import io
from blocks.bricks import Rectifier, MLP, Sequence, Logistic
from blocks.select import Selector
from theano import tensor
from theano.printing import debugprint


def wtf(call_b):
    sub_mlp = MLP([Rectifier(), Rectifier()], [5, 4, 3])
    a = Sequence([sub_mlp.apply, Logistic().apply])
    b = Sequence([sub_mlp.apply, Logistic().apply])
    x = tensor.matrix()
    y = a.apply(x)
    dprint1 = io.StringIO()
    debugprint(y, file=dprint1)

    if call_b:
        b.apply(x)
        dprint2 = io.StringIO()
        debugprint(y, file=dprint2)
        # SAME debugprint of y's graph
        assert dprint1.getvalue() == dprint2.getvalue()
    params = Selector([a]).get_parameters().values()

    # SAME contents in the params list.
    assert len(params) == 4
    assert sub_mlp.linear_transformations[0].W in params
    assert sub_mlp.linear_transformations[0].b in params
    assert sub_mlp.linear_transformations[1].W in params
    assert sub_mlp.linear_transformations[1].b in params

    # This will error if b.apply was called.
    print(tensor.grad(y.sum(), list(params)))

if __name__ == "__main__":
    print("Without calling b.apply:")
    wtf(call_b=False)
    print("--------------")
    print("With calling b.apply (result never used!):")
    wtf(call_b=True)

The answer: sub_mlp gets re-allocated during the implicit call to b.allocate() in line 18.

That's nasty. I will try to figure out why do not have a defence against multiple allocation.