Type error in PolicySaver.save()

[anmol438]

I have a DQN agent with policy of type <tf_agents.policies.greedy_policy.GreedyPolicy> to train a gym environment (CartPole-v1). I am using tf_agents 0.16.0 and gym 0.23.0
During saving the policy tf_agents.policies.policy_saver.PolicySaver I am having the following error:
policy_saver = PolicySaver(agent.policy)
policy_saver.save('./policy')

TypeError Traceback (most recent call last)
Cell In[8], line 1
----> 1 policy_saver = PolicySaver(agent.policy)
2 policy_saver.save('./policy')

File c:\Users\iitka.conda\envs\temp\lib\site-packages\tf_agents\policies\policy_saver.py:333, in PolicySaver.init(self, policy, batch_size, use_nest_path_signatures, seed, train_step, input_fn_and_spec, metadata)
326 get_initial_state_fn.get_concrete_function(*get_initial_state_input_specs)
328 train_step_fn = common.function(
329 lambda: saved_policy.train_step
330 ).get_concrete_function()
331 get_metadata_fn = common.function(
332 lambda: saved_policy.metadata
--> 333 ).get_concrete_function()
335 batched_time_step_spec = tf.nest.map_structure(
336 lambda spec: add_batch_dim(spec, [batch_size]), policy.time_step_spec
337 )
338 batched_time_step_spec = cast(ts.TimeStep, batched_time_step_spec)

File c:\Users\iitka.conda\envs\temp\lib\site-packages\tensorflow\python\eager\polymorphic_function\polymorphic_function.py:1227, in Function.get_concrete_function(self, *args, **kwargs)
1225 def get_concrete_function(self, *args, **kwargs):
1226 # Implements PolymorphicFunction.get_concrete_function.
-> 1227 concrete = self._get_concrete_function_garbage_collected(*args, **kwargs)
1228 concrete._garbage_collector.release() # pylint: disable=protected-access
1229 return concrete

File c:\Users\iitka.conda\envs\temp\lib\site-packages\tensorflow\python\eager\polymorphic_function\polymorphic_function.py:1197, in Function._get_concrete_function_garbage_collected(self, *args, **kwargs)
1195 if self._variable_creation_config is None:
1196 initializers = []
-> 1197 self._initialize(args, kwargs, add_initializers_to=initializers)
1198 self._initialize_uninitialized_variables(initializers)
1200 if self._created_variables:
1201 # In this case we have created variables on the first call, so we run the
1202 # version which is guaranteed to never create variables.

File c:\Users\iitka.conda\envs\temp\lib\site-packages\tensorflow\python\eager\polymorphic_function\polymorphic_function.py:695, in Function._initialize(self, args, kwds, add_initializers_to)
690 self._variable_creation_config = self._generate_scoped_tracing_options(
691 variable_capturing_scope,
692 tracing_compilation.ScopeType.VARIABLE_CREATION,
693 )
694 # Force the definition of the function for these arguments
--> 695 self._concrete_variable_creation_fn = tracing_compilation.trace_function(
696 args, kwds, self._variable_creation_config
697 )
699 def invalid_creator_scope(*unused_args, **unused_kwds):
700 """Disables variable creation."""

File c:\Users\iitka.conda\envs\temp\lib\site-packages\tensorflow\python\eager\polymorphic_function\tracing_compilation.py:178, in trace_function(args, kwargs, tracing_options)
175 args = tracing_options.input_signature
176 kwargs = {}
--> 178 concrete_function = _maybe_define_function(
179 args, kwargs, tracing_options
180 )
182 if not tracing_options.bind_graph_to_function:
183 concrete_function._garbage_collector.release() # pylint: disable=protected-access

File c:\Users\iitka.conda\envs\temp\lib\site-packages\tensorflow\python\eager\polymorphic_function\tracing_compilation.py:283, in _maybe_define_function(args, kwargs, tracing_options)
281 else:
282 target_func_type = lookup_func_type
--> 283 concrete_function = _create_concrete_function(
284 target_func_type, lookup_func_context, func_graph, tracing_options
285 )
287 if tracing_options.function_cache is not None:
288 tracing_options.function_cache.add(
289 concrete_function, current_func_context
290 )

File c:\Users\iitka.conda\envs\temp\lib\site-packages\tensorflow\python\eager\polymorphic_function\tracing_compilation.py:331, in _create_concrete_function(function_type, type_context, func_graph, tracing_options)
328 tracing_options.function_captures.merge_by_ref_with(graph_capture_container)
330 # Create a new FunctionType including captures and outputs.
--> 331 output_type = trace_type.from_value(
332 traced_func_graph.structured_outputs, type_context
333 )
334 traced_func_type = function_type_lib.FunctionType(
335 function_type.parameters.values(),
336 traced_func_graph.function_captures.capture_types,
337 return_annotation=output_type,
338 )
340 concrete_function = concrete_function_lib.ConcreteFunction.from_func_graph(
341 traced_func_graph,
342 traced_func_type,
(...)
348 shared_func_graph=False,
349 )

File c:\Users\iitka.conda\envs\temp\lib\site-packages\tensorflow\core\function\trace_type\trace_type_builder.py:144, in from_value(value, context)
142 if context.is_legacy_signature and isinstance(value, trace.TraceType):
143 return value
--> 144 elif isinstance(value, trace.SupportsTracingProtocol):
145 generated_type = value.tf_tracing_type(context)
146 if not isinstance(generated_type, trace.TraceType):

File c:\Users\iitka.conda\envs\temp\lib\site-packages\typing_extensions.py:647, in _ProtocolMeta.instancecheck(cls, instance)
645 for attr in cls.protocol_attrs:
646 try:
--> 647 val = inspect.getattr_static(instance, attr)
648 except AttributeError:
649 break

File c:\Users\iitka.conda\envs\temp\lib\inspect.py:1743, in getattr_static(obj, attr, default)
1740 dict_attr = _shadowed_dict(klass)
1741 if (dict_attr is _sentinel or
1742 type(dict_attr) is types.MemberDescriptorType):
-> 1743 instance_result = _check_instance(obj, attr)
1744 else:
1745 klass = obj

File c:\Users\iitka.conda\envs\temp\lib\inspect.py:1690, in _check_instance(obj, attr)
1688 instance_dict = {}
1689 try:
-> 1690 instance_dict = object.getattribute(obj, "dict")
1691 except AttributeError:
1692 pass

TypeError: this dict descriptor does not support '_DictWrapper' objects

What are the possible reasons and how can I resolve it?

[BaLinuss]

Hi,
I encountered the same problem with older code that worked previously.
Rolling back to Version 0.16.0 made it work again.
Note that you also need [email protected] to make it work.

Hope this works for you!