Merge pull request #32 from microsoft/0.1.3.4

Update to version 0.1.3.4

Author: allenanie

opto/trace/bundle.py

@@ -596,9 +596,25 @@ def is_valid_output(output):
             isinstance(output, tuple) and all([isinstance(o, Node) for o in output])
         )
 
-    def __get__(self, obj, objtype):
+
+    # Define __set_name__ and __get__ for FunModule to act as a descriptor.
+    def __get__(self, obj, db_type):
+        if obj is None:  # class method
+            return self
         # Support instance methods.
-        return functools.partial(self.__call__, obj)
+        method_name = f'__TRACE_RESERVED_bundle_{self.name}'  # NOTE we assume these are secret names not taken
+        obj_node_name = f'__TRACE_RESERVED_self_node'
+        if not hasattr(obj, obj_node_name):
+            setattr(obj, obj_node_name, node(obj))
+        if not hasattr(obj, method_name):
+            funmodule = copy.deepcopy(self)  # instance specific version
+            funmodule.forward = functools.partial(funmodule.forward, getattr(obj, obj_node_name))
+            setattr(obj, method_name, funmodule)
+        fun = getattr(obj, method_name)
+        assert fun is not self  # self is defined in the class level
+        assert isinstance(fun, FunModule), f"Expected {method_name} to be a FunModule, but got {type(fun)}"
+        # fun = functools.partial(self.__call__, obj)
+        return fun
 
     def detach(self):
         return copy.deepcopy(self)

opto/trace/containers.py

@@ -42,11 +42,14 @@ def parameters_dict(self):
         """
         parameters = {}
         for name, attr in inspect.getmembers(self):
+            if name.startswith('__TRACE_RESERVED_'):
+                # These are reserved for internal use.
+                continue
             if isinstance(attr, functools.partial):  # this is a class method
                 method = attr.func.__self__
                 if trainable_method(method):
                     parameters[name] = method.parameter
-            elif trainable_method(attr):  # method attribute
+            if trainable_method(attr):  # method attribute
                 parameters[name] = attr.parameter
             elif isinstance(attr, ParameterNode):
                 parameters[name] = attr

opto/trace/nodes.py

@@ -455,6 +455,7 @@ def __deepcopy__(self, memo):
                 setattr(result, k, defaultdict(list))
             else:
                 setattr(result, k, copy.deepcopy(v, memo))
+        GRAPH.register(result)
         return result
 
     def lt(self, other):

opto/version.py

@@ -1 +1 @@
-__version__ = "0.1.3.3"
+__version__ = "0.1.3.4"

tests/unit_tests/test_class_method.py

@@ -0,0 +1,149 @@
+from opto import trace
+from copy import deepcopy, copy
+
+@trace.model
+class Model:
+
+    @trace.bundle(trainable=True)
+    def forward(self, x):
+        return x + 1
+
+
+def test_case_two_models():
+    m1 = Model()
+    m2 = Model()
+
+    # Make sure the parameters are different
+    try:
+        assert m1.__TRACE_RESERVED_self_node is not m2.__TRACE_RESERVED_self_node
+    except AttributeError:
+        # These secrets attributes are not defined yet. They will only be defined after the bundled method is accessed.
+        pass
+
+    # The hidden nodes are defined now
+    assert len(m1.parameters()) == 1
+    assert len(m2.parameters()) == 1
+
+    # Make sure the parameters are different
+    assert m1.__TRACE_RESERVED_self_node is not m2.__TRACE_RESERVED_self_node  # they are defined now
+    assert m1.parameters()[0] is not m2.parameters()[0]
+
+    # check that the reserved node is the returned parameter
+    assert getattr(m1, '__TRACE_RESERVED_bundle_Model.forward').parameter is m1.parameters()[0]
+    assert getattr(m2, '__TRACE_RESERVED_bundle_Model.forward').parameter is m2.parameters()[0]
+
+    # each instance has a version different from the class' version
+    assert m1.forward != m2.forward
+    assert m1.forward != Model.forward
+    assert m2.forward.parameter == Model.forward.parameter == m1.forward.parameter
+
+    y1 = m1.forward(1)
+    y2 = m1.forward(2)
+
+    from opto.trace.utils import contain
+    # self is not duplicated
+    assert contain(y1.parents, m1.__TRACE_RESERVED_self_node)
+    assert contain(y2.parents, m1.__TRACE_RESERVED_self_node)
+    # assert m1.__TRACE_RESERVED_self_node in y1.parents
+    # assert m1.__TRACE_RESERVED_self_node in y2.parents
+    assert contain(y1.parents, m1.forward.parameter)
+    assert contain(y2.parents, m1.forward.parameter)
+    # assert m1.forward.parameter in y1.parents
+    # assert m1.forward.parameter in y2.parents
+    assert len(y1.parents) == 3  # since it's trainable
+    assert len(y2.parents) == 3
+
+def test_case_model_copy():
+    m1 = Model()
+    m2 = deepcopy(m1)
+
+    # Make sure the parameters are different
+    try:
+        assert m1.__TRACE_RESERVED_self_node is not m2.__TRACE_RESERVED_self_node
+    except AttributeError:
+        # These secrets attributes are not defined yet. They will only be defined after the bundled method is accessed.
+        pass
+
+    # The hidden nodes are defined now
+    assert len(m1.parameters()) == 1
+    assert len(m2.parameters()) == 1
+
+    # Make sure the parameters are different
+    assert m1.__TRACE_RESERVED_self_node is not m2.__TRACE_RESERVED_self_node  # they are defined now
+    assert m1.parameters()[0] is not m2.parameters()[0]
+
+    # check that the reserved node is the returned parameter
+    assert getattr(m1, '__TRACE_RESERVED_bundle_Model.forward').parameter is m1.parameters()[0]
+    assert getattr(m2, '__TRACE_RESERVED_bundle_Model.forward').parameter is m2.parameters()[0]
+
+    # each instance has a version different from the class' version
+    assert m1.forward is not m2.forward
+    assert m1.forward is not Model.forward
+    assert m2.forward.parameter == Model.forward.parameter == m1.forward.parameter
+
+    y1 = m1.forward(1)
+    y2 = m2.forward(2)
+
+    from opto.trace.utils import contain
+    # self is not duplicated
+    assert contain(y1.parents, m1.__TRACE_RESERVED_self_node)
+    assert contain(y2.parents, m2.__TRACE_RESERVED_self_node)
+    # assert m1.__TRACE_RESERVED_self_node in y1.parents
+    # assert m1.__TRACE_RESERVED_self_node in y2.parents
+    assert contain(y1.parents, m1.forward.parameter)
+    assert contain(y2.parents, m2.forward.parameter)
+
+    # assert m1.forward.parameter in y1.parents
+    # assert m1.forward.parameter in y2.parents
+    assert len(y1.parents) == 3  # since it's trainable
+    assert len(y2.parents) == 3
+
+def test_case_model_nested_copy():
+    m1 = Model()
+    m3 = deepcopy(m1)
+    m2 = deepcopy(m3)
+
+    # Make sure the parameters are different
+    try:
+        assert m1.__TRACE_RESERVED_self_node is not m2.__TRACE_RESERVED_self_node
+    except AttributeError:
+        # These secrets attributes are not defined yet. They will only be defined after the bundled method is accessed.
+        pass
+
+    # The hidden nodes are defined now
+    assert len(m1.parameters()) == 1
+    assert len(m2.parameters()) == 1
+
+    # Make sure the parameters are different
+    assert m1.__TRACE_RESERVED_self_node is not m2.__TRACE_RESERVED_self_node  # they are defined now
+    assert m1.parameters()[0] is not m2.parameters()[0]
+
+    # check that the reserved node is the returned parameter
+    assert getattr(m1, '__TRACE_RESERVED_bundle_Model.forward').parameter is m1.parameters()[0]
+    assert getattr(m2, '__TRACE_RESERVED_bundle_Model.forward').parameter is m2.parameters()[0]
+
+    # each instance has a version different from the class' version
+    assert m1.forward is not m2.forward
+    assert m1.forward is not Model.forward
+    assert m2.forward.parameter == Model.forward.parameter == m1.forward.parameter
+
+    y1 = m1.forward(1)
+    y2 = m2.forward(2)
+
+    from opto.trace.utils import contain
+    # self is not duplicated
+    assert contain(y1.parents, m1.__TRACE_RESERVED_self_node)
+    assert contain(y2.parents, m2.__TRACE_RESERVED_self_node)
+    # assert m1.__TRACE_RESERVED_self_node in y1.parents
+    # assert m1.__TRACE_RESERVED_self_node in y2.parents
+    assert contain(y1.parents, m1.forward.parameter)
+    assert contain(y2.parents, m2.forward.parameter)
+
+    # assert m1.forward.parameter in y1.parents
+    # assert m1.forward.parameter in y2.parents
+    assert len(y1.parents) == 3  # since it's trainable
+    assert len(y2.parents) == 3
+
+test_case_two_models()
+test_case_model_copy()
+test_case_model_nested_copy()